Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[ "Hey @marcoabrate it looks like you're seeing the same error as here https://github.com/huggingface/transformers/issues/9146. This should be fixed on transformers master and the latest Ray nightly wheels. Can you try with those and see if that fixes this? You can install the latest Ray nightly wheels by following the instructions here: https://docs.ray.io/en/master/installation.html#daily-releases-nightlies.", "Hi @amogkam, thank you for your quick reply.\r\nI am now on HF transformers master and I am installing raytune for Python 3.6 with\r\n\r\n`pip install -U \"https://s3-us-west-2.amazonaws.com/ray-wheels/latest/ray-2.0.0.dev0-cp36-cp36m-manylinux2014_x86_64.whl\"`\r\n\r\nHowever, on Google Colab I still get the same error:\r\n\r\n```\r\n[INFO|trainer.py:358] 2021-02-02 11:04:25,269 >> Using amp fp16 backend\r\n02/02/2021 11:04:25 - INFO - __main__ - *** Hyperparameters Search ***\r\n02/02/2021 11:04:25 - INFO - ray.tune.ray_trial_executor - Initializing Ray automatically.For cluster usage or custom Ray initialization, call `ray.init(...)` before `tune.run`.\r\n2021-02-02 11:04:26,687\tINFO services.py:1182 -- View the Ray dashboard at http://127.0.0.1:8265\r\ntcmalloc: large alloc 1236656128 bytes == 0x7fc271b46000 @ 0x7fc54e359615 0x591e47 0x4cc179 0x4cc2db 0x566a71 0x5a4cd1 0x5a4fb8 0x7fc4a06ddc7c 0x7fc4a06e4bfa 0x7fc4a06e4fe2 0x7fc4a06e634f 0x7fc4a06e3a39 0x7fc4a06e4afc 0x7fc4a06e634f 0x7fc4a06e3a39 0x7fc4a06e4fe2 0x7fc4a06e634f 0x7fc4a06e3a39 0x7fc4a06e5d13 0x7fc4a06e643c 0x7fc4a06e3a39 0x7fc4a06e4f45 0x7fc4a06e67fa 0x7fc4a06e3a39 0x7fc4a06e5d5e 0x7fc4a06e643c 0x7fc4a06e3a39 0x7fc4a06e4f45 0x7fc4a06e67fa 0x7fc4a06e3a39 0x7fc4a06e5d5e\r\ntcmalloc: large alloc 1545822208 bytes == 0x7fc215910000 @ 0x7fc54e359615 0x591e47 0x4cc179 0x4cc2db 0x566a71 0x5a4cd1 0x5a4fb8 0x7fc4a06ddc7c 0x7fc4a06e4bfa 0x7fc4a06e4fe2 0x7fc4a06e634f 0x7fc4a06e3a39 0x7fc4a06e4afc 0x7fc4a06e634f 0x7fc4a06e3a39 0x7fc4a06e4fe2 0x7fc4a06e634f 0x7fc4a06e3a39 0x7fc4a06e5d13 0x7fc4a06e643c 0x7fc4a06e3a39 0x7fc4a06e4f45 0x7fc4a06e67fa 0x7fc4a06e3a39 0x7fc4a06e5d5e 0x7fc4a06e643c 0x7fc4a06e3a39 0x7fc4a06e4f45 0x7fc4a06e67fa 0x7fc4a06e3a39 0x7fc4a06e5d13\r\nTraceback (most recent call last):\r\n File \"/usr/local/lib/python3.6/dist-packages/redis/connection.py\", line 706, in send_packed_command\r\n sendall(self._sock, item)\r\n File \"/usr/local/lib/python3.6/dist-packages/redis/_compat.py\", line 9, in sendall\r\n return sock.sendall(*args, **kwargs)\r\nConnectionResetError: [Errno 104] Connection reset by peer\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nTraceback (most recent call last):\r\n File \"/content/drive/My Drive/MAGMA: Summarization/transformers_last/transformers/examples/seq2seq/finetune_trainer.py\", line 432, in <module>\r\n main()\r\n File \"/content/drive/My Drive/MAGMA: Summarization/transformers_last/transformers/examples/seq2seq/finetune_trainer.py\", line 346, in main\r\n resources_per_trial = {'gpu': 1})\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/trainer.py\", line 1188, in hyperparameter_search\r\n best_run = run_hp_search(self, n_trials, direction, **kwargs)\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/integrations.py\", line 220, in run_hp_search_ray\r\n analysis = ray.tune.run(_objective, config=trainer.hp_space(None), num_samples=n_trials, **kwargs)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/tune.py\", line 338, in run\r\n restore=restore)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/experiment.py\", line 149, in __init__\r\n self._run_identifier = Experiment.register_if_needed(run)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/experiment.py\", line 294, in register_if_needed\r\n register_trainable(name, run_object)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/registry.py\", line 71, in register_trainable\r\n _global_registry.register(TRAINABLE_CLASS, name, trainable)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/registry.py\", line 124, in register\r\n self.flush_values()\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/registry.py\", line 146, in flush_values\r\n _internal_kv_put(_make_key(category, key), value, overwrite=True)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/_private/client_mode_hook.py\", line 47, in wrapper\r\n return func(*args, **kwargs)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/experimental/internal_kv.py\", line 35, in _internal_kv_put\r\n updated = worker.redis_client.hset(key, \"value\", value)\r\n File \"/usr/local/lib/python3.6/dist-packages/redis/client.py\", line 3050, in hset\r\n return self.execute_command('HSET', name, *items)\r\n File \"/usr/local/lib/python3.6/dist-packages/redis/client.py\", line 900, in execute_command\r\n conn.send_command(*args)\r\n File \"/usr/local/lib/python3.6/dist-packages/redis/connection.py\", line 726, in send_command\r\n check_health=kwargs.get('check_health', True))\r\n File \"/usr/local/lib/python3.6/dist-packages/redis/connection.py\", line 718, in send_packed_command\r\n (errno, errmsg))\r\nredis.exceptions.ConnectionError: Error 104 while writing to socket. Connection reset by peer.\r\n02/02/2021 11:04:38 - INFO - wandb.sdk.internal.internal - Internal process exited\r\n```", "I confirm with HF Transformers master and the latest ray[tune] version available using pip, the Trainer function works as expected.\r\n\r\nThank you for your help." ]

[]

[]

[ "I don't think optuna properly optimizes GPU memory. We don't have support from them so you may be better using ray-tune, where the maintainers happily reply to question on our GitHub in case of problems. ", "Thank you. I was using Optuna because with RayTune I get an error even before the first trial starts. I will open an issue about the RayTune error." ]

[]

[]

[]

[]

[]

[ "HI @Aunsiels you are right, thanks! Feel free to open PR to fix it :)\r\n\r\nIf `train_file` is `None` it should use `validation_file` to get the extension" ]

[]

[]

[ "Hmmm I would say the issue here comes from `tqdm`. Do you know how come there is no distribution for `tqdm`? Managing to install it as a standalone would be a good first step imo.", "pinging @philschmid or @n1t0 who might know about Google's AI Platform (and other ways to deploy in the cloud)", "> Hmmm I would say the issue here comes from `tqdm`. Do you know how come there is no distribution for `tqdm`? Managing to install it as a standalone would be a good first step imo.\r\n\r\nI tried to precise the link of `tqdm` package but without success\r\n```python\r\nfrom setuptools import setup\r\n\r\nsetup(\r\n name='customerPredictionCustomerReview',\r\n version='0.1',\r\n scripts=['predictor.py'],\r\n install_requires=[\"tqdm\", \"transformers==4.2.2\"],\r\n dependency_links=[\r\n \"https://files.pythonhosted.org/packages/80/02/8f8880a4fd6625461833abcf679d4c12a44c76f9925f92bf212bb6cefaad/tqdm-4.56.0-py2.py3-none-any.whl\"]\r\n)\r\n```", "@iElsha I am going to take a look later at why your deployment into Google Cloud AI-platform with a customer prediction routine might not work. \r\n\r\nIn addition, Google offers different other services to deploy `transformers` in the cloud. The easiest way I think is to use [managed Cloud Run](https://cloud.google.com/run). With Cloud Run you can deploy highly scalable containerized applications on a fully managed serverless platform it supports currently up to 8GB of memory and 4 CPUs. You just have to build a `flask` or `fastAPI` container and deploy it. \r\n\r\nAnother possible solution could be `GKE`, Google's managed Kubernetes service when you want to scale your application or want to be more flexible in terms of configuration. `GKE` supports `Cloud Run` too. So it is possible to use your `Cloud Run` container out-of-the-box on `GKE`. \r\n\r\nAnd last but not least there is [App Engine](https://cloud.google.com/appengine/docs/standard/python3/quickstart) a highly scalable fully managed platform. \r\n\r\n\r\n", "> > Hmmm I would say the issue here comes from `tqdm`. Do you know how come there is no distribution for `tqdm`? Managing to install it as a standalone would be a good first step imo.\r\n> \r\n> I tried to precise the link of `tqdm` package but without success\r\n> \r\n> ```python\r\n> from setuptools import setup\r\n> \r\n> setup(\r\n> name='customerPredictionCustomerReview',\r\n> version='0.1',\r\n> scripts=['predictor.py'],\r\n> install_requires=[\"tqdm\", \"transformers==4.2.2\"],\r\n> dependency_links=[\r\n> \"https://files.pythonhosted.org/packages/80/02/8f8880a4fd6625461833abcf679d4c12a44c76f9925f92bf212bb6cefaad/tqdm-4.56.0-py2.py3-none-any.whl\"]\r\n> )\r\n> ```\r\n\r\n@iElsha do you have the complete code somewhere available? like in a Github Repository? I would like to try to recreate the error.", "Thanks for the quick reply @philschmid \r\n\r\nHere's the github link: https://github.com/iElsha/ICC-Customer-system-AI\r\n\r\nAnd there the deployment commands:\r\n\r\n```shell\r\npython setup.py sdist --formats=gztar\r\ngsutil cp cloudDeploy/dist/customerPredictionCustomerReview-0.1.tar.gz gs://customer_system/src/\r\n\r\n# Create the model project once and update the gcloud tool\r\ngcloud ai-platform models create customerReviewModel --regions europe-west1 --project <YourProjectId>\r\ngcloud components install beta\r\n\r\n# create & delete command to manage the version\r\ngcloud beta ai-platform versions create v01 --model customerReviewModel --runtime-version 2.2 --python-version 3.7 --origin gs://customer_system/model --package-uris gs://customer_system/src/customerPredictionCustomerReview-0.1.tar.gz --prediction-class predictor.MyPredictor --project <YourProjectId>\r\ngcloud beta ai-platform versions delete v01 --model customerReviewModel --project <YourProjectId>\r\n```\r\n\r\n---\r\n\r\n**Edit 01/02 - 16:30:**\r\n\r\nI also tried the solution with `App Engine` (F4 - memory 1024MB) but it seems that it can not load TensorFlow properly:\r\n\r\n```json\r\n{\r\n \"textPayload\": \"2021-02-01 15:00:03.883430: W tensorflow/stream_executor/platform/default/dso_loader.cc:60] Could not load dynamic library 'libcudart.so.11.0'; dlerror: libcudart.so.11.0: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /layers/google.python.pip/pip/lib\",\r\n}\r\n```\r\n\r\nI am going to try with pytorch", "I also tried to see where does the error comes from by only installing `tqdm` and not `transformer`. it worked and I succeed to deploy, meaning that the issue might come from `transformer` or somewhere else but `tqdm` seems fine?\r\n", "@iElsha I could reproduce the error. After that, I researched and found that the Google Cloud AI-Platform `Custom prediction routines` is in BETA and not official GA and others have the same problem with installing packages. [Issue 1](https://stackoverflow.com/questions/62816129/how-do-you-override-google-ai-platforms-standard-librarys-i-e-upgrade-scikit) [Issue 2](https://stackoverflow.com/questions/64781326/getting-create-version-failed-bad-model-detected-with-error-on-ai-platform-wh) \r\nI think the issue is not from `transformers` side. You can create an Issue [at Google official Issue tracker](https://issuetracker.google.com/issues/new?component=187220&template=1161235) or try to create [a custom container for online prediction with AI-Platform](https://cloud.google.com/ai-platform/prediction/docs/custom-container-requirements) or use Cloud Run. I found this [blog post](https://chatbotslife.com/deploying-transformer-models-1350876016f) where a GPT-2 model is used. \r\n", "> Another possible solution could be `GKE`, Google's managed Kubernetes service when you want to scale your application or want to be more flexible in terms of configuration. `GKE` supports `Cloud Run` too. So it is possible to use your `Cloud Run` container out-of-the-box on `GKE`.\r\n\r\nAs you suggested it works with cloud Run, just with a docker container.\r\n\r\nI before tried on AppEngine, where I was with TensorFlow (2G memory), but TensorFlow couldn't load there due to a missing dependency in the system. I switched to PyTorch and it worked for a few requests but exceed the memory and makes the service unavailable.\r\n\r\nCloud Run with a docker container and flask is, for now, the correct solution to deploy the transformers pipeline. I used a 4G & 1VCPU as settings with PyTorch, which seems lighter & faster to load on a cold boot than TensorFlow.\r\n\r\nThanks for the help\r\n", "@iElsha Would be very interesting if you can at some point share about the operational aspects of Cloud Run (request latency distribution, scalability from simulated traffic, cost)! We could even write a blogpost about it.", "> > > Hmmm I would say the issue here comes from `tqdm`. Do you know how come there is no distribution for `tqdm`? Managing to install it as a standalone would be a good first step imo.\r\n> > \r\n> > \r\n> > I tried to precise the link of `tqdm` package but without success\r\n> > ```python\r\n> > from setuptools import setup\r\n> > \r\n> > setup(\r\n> > name='customerPredictionCustomerReview',\r\n> > version='0.1',\r\n> > scripts=['predictor.py'],\r\n> > install_requires=[\"tqdm\", \"transformers==4.2.2\"],\r\n> > dependency_links=[\r\n> > \"https://files.pythonhosted.org/packages/80/02/8f8880a4fd6625461833abcf679d4c12a44c76f9925f92bf212bb6cefaad/tqdm-4.56.0-py2.py3-none-any.whl\"]\r\n> > )\r\n> > ```\r\n> \r\n> @iElsha do you have the complete code somewhere available? like in a Github Repository? I would like to try to recreate the error.\r\n\r\nThis kind of makes me feel that the issue is not with GCP Custom Prediction routines but some way `tqdm` and `transformers` are interacting when installing this way. I am able to install several other packages, including `tqdm` in a custom prediction routine build - but I cannot install `transformers`.", "> > Hmmm I would say the issue here comes from `tqdm`. Do you know how come there is no distribution for `tqdm`? Managing to install it as a standalone would be a good first step imo.\r\n> \r\n> I tried to precise the link of `tqdm` package but without success\r\n> \r\n> ```python\r\n> from setuptools import setup\r\n> \r\n> setup(\r\n> name='customerPredictionCustomerReview',\r\n> version='0.1',\r\n> scripts=['predictor.py'],\r\n> install_requires=[\"tqdm\", \"transformers==4.2.2\"],\r\n> dependency_links=[\r\n> \"https://files.pythonhosted.org/packages/80/02/8f8880a4fd6625461833abcf679d4c12a44c76f9925f92bf212bb6cefaad/tqdm-4.56.0-py2.py3-none-any.whl\"]\r\n> )\r\n> ```\r\n\r\n install_requires=[\"tqdm-wheel\"] will help you installed the library I guess so because I had also similar kind of problems with libraries and I installed it this way.\r\nI think it will help you too." ]

[ { "id": 2648621985, "node_id": "MDU6TGFiZWwyNjQ4NjIxOTg1", "url": "https://api.github.com/repos/huggingface/transformers/labels/Feature%20request", "name": "Feature request", "color": "FBCA04", "default": false, "description": "Request for a new feature" } ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "Pinging @abhishekkrthakur ", "@abhishekkrthakur I have try to convert a mlm convbert to transformers one, this is the convert code\r\n\r\n```\r\nimport torch\r\nimport os\r\n\r\nimport tensorflow as tf\r\n\r\nfrom transformers import ConvBertConfig, ConvBertForMaskedLM, ConvBertPreTrainedModel\r\nfrom transformers.utils import logging\r\nfrom operator import attrgetter\r\n\r\nlogger = logging.get_logger(__name__)\r\n\r\nconfig_file = \"weights/convbert_base_mlm/config.json\"\r\ntf_path = \"tf_weights/ft_local/model.ckpt-490000\"\r\npytorch_dump_path = \"weights/convbert_base_mlm\"\r\nconfig = ConvBertConfig.from_json_file(config_file)\r\n\r\n#model = ConvBertPreTrainedModel(config)\r\nmodel = ConvBertForMaskedLM(config)\r\n\r\ndef load_tf_weights_in_convbert(model, config, tf_checkpoint_path):\r\n \"\"\"Load tf checkpoints in a pytorch model.\"\"\"\r\n try:\r\n import tensorflow as tf\r\n except ImportError:\r\n logger.error(\r\n \"Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see \"\r\n \"https://www.tensorflow.org/install/ for installation instructions.\"\r\n )\r\n raise\r\n tf_path = os.path.abspath(tf_checkpoint_path)\r\n logger.info(\"Converting TensorFlow checkpoint from {}\".format(tf_path))\r\n # Load weights from TF model\r\n init_vars = tf.train.list_variables(tf_path)\r\n tf_data = {}\r\n for name, shape in init_vars:\r\n logger.info(\"Loading TF weight {} with shape {}\".format(name, shape))\r\n array = tf.train.load_variable(tf_path, name)\r\n tf_data[name] = array\r\n\r\n param_mapping = {\r\n \"convbert.embeddings.word_embeddings.weight\": \"electra/embeddings/word_embeddings\",\r\n \"convbert.embeddings.position_embeddings.weight\": \"electra/embeddings/position_embeddings\",\r\n \"convbert.embeddings.token_type_embeddings.weight\": \"electra/embeddings/token_type_embeddings\",\r\n \"convbert.embeddings.LayerNorm.weight\": \"electra/embeddings/LayerNorm/gamma\",\r\n \"convbert.embeddings.LayerNorm.bias\": \"electra/embeddings/LayerNorm/beta\",\r\n \"convbert.embeddings_project.weight\": \"electra/embeddings_project/kernel\",\r\n \"convbert.embeddings_project.bias\": \"electra/embeddings_project/bias\",\r\n \"generator_predictions.LayerNorm.weight\": \"generator_predictions/LayerNorm/gamma\",\r\n \"generator_predictions.LayerNorm.bias\": \"generator_predictions/LayerNorm/beta\",\r\n \"generator_predictions.dense.weight\": \"generator_predictions/dense/kernel\",\r\n \"generator_predictions.dense.bias\": \"generator_predictions/dense/bias\",\r\n \"generator_lm_head.bias\": \"generator_predictions/output_bias\"\r\n }\r\n if config.num_groups > 1:\r\n group_dense_name = \"g_dense\"\r\n else:\r\n group_dense_name = \"dense\"\r\n\r\n for j in range(config.num_hidden_layers):\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.self.query.weight\"\r\n ] = f\"electra/encoder/layer_{j}/attention/self/query/kernel\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.self.query.bias\"\r\n ] = f\"electra/encoder/layer_{j}/attention/self/query/bias\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.self.key.weight\"\r\n ] = f\"electra/encoder/layer_{j}/attention/self/key/kernel\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.self.key.bias\"\r\n ] = f\"electra/encoder/layer_{j}/attention/self/key/bias\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.self.value.weight\"\r\n ] = f\"electra/encoder/layer_{j}/attention/self/value/kernel\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.self.value.bias\"\r\n ] = f\"electra/encoder/layer_{j}/attention/self/value/bias\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.self.key_conv_attn_layer.depthwise.weight\"\r\n ] = f\"electra/encoder/layer_{j}/attention/self/conv_attn_key/depthwise_kernel\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.self.key_conv_attn_layer.pointwise.weight\"\r\n ] = f\"electra/encoder/layer_{j}/attention/self/conv_attn_key/pointwise_kernel\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.self.key_conv_attn_layer.bias\"\r\n ] = f\"electra/encoder/layer_{j}/attention/self/conv_attn_key/bias\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.self.conv_kernel_layer.weight\"\r\n ] = f\"electra/encoder/layer_{j}/attention/self/conv_attn_kernel/kernel\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.self.conv_kernel_layer.bias\"\r\n ] = f\"electra/encoder/layer_{j}/attention/self/conv_attn_kernel/bias\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.self.conv_out_layer.weight\"\r\n ] = f\"electra/encoder/layer_{j}/attention/self/conv_attn_point/kernel\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.self.conv_out_layer.bias\"\r\n ] = f\"electra/encoder/layer_{j}/attention/self/conv_attn_point/bias\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.output.dense.weight\"\r\n ] = f\"electra/encoder/layer_{j}/attention/output/dense/kernel\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.output.LayerNorm.weight\"\r\n ] = f\"electra/encoder/layer_{j}/attention/output/LayerNorm/gamma\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.output.dense.bias\"\r\n ] = f\"electra/encoder/layer_{j}/attention/output/dense/bias\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.attention.output.LayerNorm.bias\"\r\n ] = f\"electra/encoder/layer_{j}/attention/output/LayerNorm/beta\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.intermediate.dense.weight\"\r\n ] = f\"electra/encoder/layer_{j}/intermediate/{group_dense_name}/kernel\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.intermediate.dense.bias\"\r\n ] = f\"electra/encoder/layer_{j}/intermediate/{group_dense_name}/bias\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.output.dense.weight\"\r\n ] = f\"electra/encoder/layer_{j}/output/{group_dense_name}/kernel\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.output.dense.bias\"\r\n ] = f\"electra/encoder/layer_{j}/output/{group_dense_name}/bias\"\r\n param_mapping[\r\n f\"convbert.encoder.layer.{j}.output.LayerNorm.weight\"\r\n ] = f\"electra/encoder/layer_{j}/output/LayerNorm/gamma\"\r\n param_mapping[f\"convbert.encoder.layer.{j}.output.LayerNorm.bias\"] = f\"electra/encoder/layer_{j}/output/LayerNorm/beta\"\r\n\r\n for param in model.named_parameters():\r\n param_name = param[0]\r\n retriever = attrgetter(param_name)\r\n result = retriever(model)\r\n tf_name = param_mapping[param_name]\r\n value = torch.from_numpy(tf_data[tf_name])\r\n logger.info(f\"TF: {tf_name}, PT: {param_name} \")\r\n if tf_name.endswith(\"/kernel\"):\r\n if not tf_name.endswith(\"/intermediate/g_dense/kernel\"):\r\n if not tf_name.endswith(\"/output/g_dense/kernel\"):\r\n value = value.T\r\n if tf_name.endswith(\"/depthwise_kernel\"):\r\n value = value.permute(1, 2, 0) # 2, 0, 1\r\n if tf_name.endswith(\"/pointwise_kernel\"):\r\n value = value.permute(2, 1, 0) # 2, 1, 0\r\n if tf_name.endswith(\"/conv_attn_key/bias\"):\r\n value = value.unsqueeze(-1)\r\n result.data = value\r\n return model\r\n\r\nmodel = load_tf_weights_in_convbert(model, config, tf_path)\r\nmodel.save_pretrained(pytorch_dump_path)\r\n```", "@RyanHuangNLP good idea! do you want to make a PR? Or should I fix it? ", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "@RyanHuangNLP I have a question regarding your script. I tried to extract the generator from the tf checkpoint, but it seems the size mismatch. I reduce the hidden_size by 4 (25%), as in the electra config file, and num_attention_head to 4. Does your script is currently converting the discriminator instead?", "@Shiro-LK my script is not for the electra one, that is for mlm one, may be you should first convert the generator parameters names to the discriminator. It is important to check the parameter name", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Hi @mmoya01 \r\n\r\nThis error happens if `torch.utils.checkpoint` is not imported. This is fixed on master now, see #9626", "@patil-suraj thank you, that work", "I still get the same error (when training DeBERTa-V3-base) on a colab GPU with Trainsformers==4.12\r\n\r\nI using\r\n`model.gradient_checkpointing_enable() # to decrease memory usage\r\n`\r\nBefore doing normal training via the HF trainer.\r\n\r\n(It's fixed if I run this:) \r\n`from torch.utils.checkpoint import checkpoint\r\n`\r\n", "I get the same error when running training on `DebertaForSequenceClassification` using the Trainer API with `gradient_checkpointing` set to True.\r\n\r\n@MoritzLaurer 's solution works for this also", "> \r\n\r\nthanks, it worked! " ]

[]

[]

[ "You're right, but changing the class won't help us as `PreTrainedTokenizerBase` is not documented either. A quick fix would be to just add the `encode` method to the doc in `PreTrainedTokenizer` so that the reference gets resolved.\r\n\r\n> Also it seems that all modules have this\r\n> What are input IDs? <../glossary.html#input-ids>__\r\n> but not where input_ids are documented, instead after the mask - shouldn't that line be 2 records up?\r\n\r\nIt seems to be only in the T5 model from a quick look. This is missing in the `input_ids` arg (but it should also be in the `decoder_input_ids` args). Did you find other models where it's missing?\r\n\r\nCan do a quick PR to fix those tomorrow morning.", "> You're right, but changing the class won't help us as PreTrainedTokenizerBase is not documented either.\r\n\r\nIsn't this the documentation?\r\n\r\nhttps://huggingface.co/transformers/internal/tokenization_utils.html#transformers.tokenization_utils_base.PreTrainedTokenizerBase.encode\r\n\r\n> It seems to be only in the T5 model from a quick look. This is missing in the input_ids arg (but it should also be in the decoder_input_ids args). Did you find other models where it's missing?\r\n\r\nIt's hard to devise a full-proof detector, because the input varies, but here is a quick attempt that catches a few of missing ones:\r\n```\r\n# from the root of the repo\r\ngrep -Inr -A30 'input_ids (:obj:' src/transformers/models/ | \\\r\nperl -ne '$x .= $_; END { for (split /--/, $x) { s/attention_mask.*//msg; print if !/What are input IDs/ } }'\r\n```\r\nBasically I'm trying to match every instance of the `*_input_ids` doc entries (assuming they all have the same pattern), I dump the subsequent text and then I check whether there is a matching \"What are input IDs\" in the next few lines. I also snip out any text after `attention_mask` to avoid overlap with entries like `decoder_input_ids`, which may have this pointer.\r\n\r\nIt dumps output where it's most likely missing, like in this entry:\r\n```\r\nsrc/transformers/models/t5/modeling_tf_t5.py:942: decoder_input_ids (:obj:`tf.Tensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):\r\nsrc/transformers/models/t5/modeling_tf_t5.py-943- Provide for sequence to sequence training. T5 uses the :obj:`pad_token_id` as the starting token for\r\nsrc/transformers/models/t5/modeling_tf_t5.py-944- :obj:`decoder_input_ids` generation. If :obj:`past_key_values` is used, optionally only the last\r\nsrc/transformers/models/t5/modeling_tf_t5.py-945- :obj:`decoder_input_ids` have to be input (see :obj:`past_key_values`).\r\nsrc/transformers/models/t5/modeling_tf_t5.py-946-\r\nsrc/transformers/models/t5/modeling_tf_t5.py-947- To know more on how to prepare :obj:`decoder_input_ids` for pretraining take a look at `T5 Training\r\nsrc/transformers/models/t5/modeling_tf_t5.py-948- <./t5.html#training>`__. If :obj:`decoder_input_ids` and :obj:`decoder_inputs_embeds` are both unset,\r\nsrc/transformers/models/t5/modeling_tf_t5.py-949- :obj:`decoder_input_ids` takes the value of :obj:`input_ids`.\r\n```\r\n\r\nThe detected chunks are just double newline separated. There are probably a few false positives, but most seem to be true positives. You have the file and the line for the context.\r\n\r\nAnd in many places where \"What are input IDs\" are, in the same place the corresponding entry for attention is missing.\r\n\r\nAlso note I only scanned under `/models/`, there is more in non-model files, but I think it's by design.\r\n\r\n", "> Isn't this the documentation?\r\n>https://huggingface.co/transformers/internal/tokenization_utils.html#transformers.tokenization_utils_base.PreTrainedTokenizerBase.encode\r\n\r\nAh yes, but this is under the \"internal\" tools, so let's have the subclasses show the documentation since I doubt users will go that far down.\r\n\r\nWill try your magic perl, thanks!", "Well, I meant that the xref link could link to that page. So it's not about users browsing to it, but sphinx resolving to that doc. \r\n\r\nUnless I'm missing something and you are talking about something else." ]

[]

[]

[]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "Mmmm probably. This is a bit tricky to make sure it doesn't break anything but makes more sense. I'll try to look at this on Monday.", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "I have the same question.", "Hello! You're using TensorFlow models (see the `TF` prefix) but you're asking the tokenizer to return PyTorch tensors. You should either stick to full PyTorch (remove the `TF` prefix) or full TF (ask the tokenizer to return `tf` values)", "I met the same issue, I did not know how to fix it\r\n```\r\ntensor([[ 0, 24948, 5357, 88, 14, 397, 1176, 6724, 7, 35297,\r\n 18109, 5814, 16, 43, 167, 4446, 37361, 381, 2, 1,\r\n 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]],\r\n device='cuda:0')\r\ntensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],\r\n device='cuda:0')\r\n---------------------------------------------------------------------------\r\nAttributeError Traceback (most recent call last)\r\n[<ipython-input-13-1309f9063eea>](https://localhost:8080/#) in <module>\r\n 1 _, tokenizer = load_pho_bert()\r\n----> 2 infer('Cảm ơn bạn đã chạy thử model của mình. Chúc một ngày tốt lành nha!', tokenizer)\r\n\r\n2 frames\r\n[/usr/local/lib/python3.7/dist-packages/keras/engine/input_spec.py](https://localhost:8080/#) in display_shape(shape)\r\n 269 \r\n 270 def display_shape(shape):\r\n--> 271 return str(tuple(shape.as_list()))\r\n 272 \r\n 273 \r\n\r\nAttributeError: 'torch.Size' object has no attribute 'as_list'\r\n```", "> Hello! You're using TensorFlow models (see the `TF` prefix) but you're asking the tokenizer to return PyTorch tensors. You should either stick to full PyTorch (remove the `TF` prefix) or full TF (ask the tokenizer to return `tf` values)\r\n\r\nPlease help me how to fix this problem? How can I change my code?\r\ndef infer(text, tokenizer, max_len=120):\r\n device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')\r\n print(device)\r\n class_names = ['thế giới', 'thể thao', 'văn hóa', 'vi tính']\r\n\r\n model = tf.keras.models.load_model('./models/cnn_nlp_text_classification_4_classer.h5')\r\n\r\n encoded_review = tokenizer.encode_plus(\r\n text,\r\n max_length=max_len,\r\n truncation=True,\r\n add_special_tokens=True,\r\n padding='max_length',\r\n return_attention_mask=True,\r\n return_token_type_ids=False,\r\n return_tensors='pt',\r\n )\r\n\r\n input_ids = encoded_review['input_ids'].to(device)\r\n print(input_ids.shape)\r\n attention_mask = encoded_review['attention_mask'].to(device)\r\n print(attention_mask.shape)\r\n\r\n output = model(input_ids, attention_mask)\r\n ==> error happen here\r\n" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "The following is part of the reference code after my deletion:\r\n\r\n```\r\ndef get_binary_data(args, macro, tokenizer, read_data=False):\r\n if macro.local_rank not in [-1, 0]:\r\n torch.distributed.barrier()\r\n\r\n train_file = [\"a.txt\", \"b.txt\", \"c.txt\"]\r\n valid_file = [\"a_valid.txt\", \"b_valid.txt\", \"c_valid.txt\"]\r\n train_data, vaild_data = [],[]\r\n cached_features_file = args.train_path+\"utt_generator\"\r\n if read_data == False:\r\n for t_f in train_file:\r\n train_utterance = file_reader(args.train_path+t_f)\r\n datasets = get_data_loaders(train_utterance, tokenizer)\r\n train_data.extend(datasets)\r\n for t_f in valid_file:\r\n vaild_utterance = file_reader(args.train_path+t_f)\r\n datasets = get_data_loaders(vaild_utterance, tokenizer)\r\n vaild_data.extend(datasets)\r\n else:\r\n read_data = torch.load(cached_features_file)\r\n train_data = read_data[\"train_data\"]\r\n vaild_data = read_data[\"dev_data\"]\r\n\r\n train_len = len(train_data)\r\n logger.info(\"train len:%d, valid len:%d.\"%(len(train_data), len(vaild_data)))\r\n\r\n train_batch_size = args.batch_size * max(1, macro.n_gpu)\r\n train_sampler = RandomSampler(train_data) if macro.local_rank == -1 else DistributedSampler(train_data)\r\n train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=train_batch_size, collate_fn=collate_fn)\r\n\r\n eval_batch_size = args.batch_size * max(1, macro.n_gpu)\r\n eval_sampler = SequentialSampler(vaild_data)\r\n eval_dataloader = DataLoader(vaild_data, sampler=eval_sampler, batch_size=eval_batch_size, collate_fn=collate_fn_test)\r\n\r\n if macro.local_rank == 0:\r\n torch.distributed.barrier()\r\n return train_dataloader, eval_dataloader, train_len\r\n\r\ndef train(model, training_data, optimizer, device, scheduler, args, macro):\r\n model.train()\r\n batch_idx = 0\r\n epoch_loss = 0\r\n logging_loss = 0\r\n global global_step\r\n global tb_writer\r\n for batch in tqdm(\r\n training_data,\r\n mininterval=2,\r\n desc=\" - (Traning) \",\r\n leave=False,\r\n disable=macro.local_rank not in [-1, 0]\r\n ):\r\n batch_idx += 1\r\n input_ids, lm_labels, token_type_ids, attention_mask = list(map(lambda x: x.to(device), batch))\r\n (lm_loss), *_ = model(input_ids, token_type_ids=token_type_ids, lm_labels=lm_labels, attention_mask=attention_mask)\r\n\r\n if macro.n_gpu > 1:\r\n lm_loss = lm_loss.mean() # mean() to average on multi-gpu parallel (not distributed) training\r\n loss = lm_loss / args.gradient_accumulation_steps\r\n loss.backward()\r\n epoch_loss += loss.item()\r\n if (batch_idx) % args.gradient_accumulation_steps == 0:\r\n torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)\r\n optimizer.step()\r\n optimizer.zero_grad()\r\n scheduler.step()\r\n return epoch_loss/batch_idx\r\n\r\n@func_time\r\ndef train_epoch(model, train_data, dev_data, optimizer, device, scheduler, best_result, args, tokenizer, epoch, macro):\r\n if macro.local_rank in [-1, 0]:\r\n logger.info(\"----\" * 5)\r\n logger.info('Epoch: {}'.format(epoch))\r\n train_loss = train(model, train_data, optimizer, device, scheduler, args, macro)\r\n if epoch>=12:\r\n if macro.local_rank in [-1, 0]:\r\n model_to_evaluate = model.module if hasattr(model, \"module\") else model\r\n vaild_bleu = evaluate(model_to_evaluate, dev_data, device, tokenizer, args)\r\n if best_result < vaild_bleu:\r\n best_result = vaild_bleu\r\n torch.save(model_to_evaluate.state_dict(), args.output_model_path + \"_valid\")\r\n logger.info('save:{}'.format(args.output_model_path + \"_valid\"))\r\n logger.info(\"Val. Bleu: %4f\" % (vaild_bleu))\r\n if epoch%5==0:\r\n if macro.local_rank == -1 or torch.distributed.get_rank() == 0:\r\n model_to_save = model.module if hasattr(model, \"module\") else model\r\n torch.save(model_to_save.state_dict(), args.output_model_path+\"_\"+str(epoch))\r\n logger.info('save:{}'.format(args.output_model_path+\"_\"+str(epoch)))\r\n if macro.local_rank in [-1, 0]:\r\n logger.info(\"Train Loss:%.5f\"%(train_loss))\r\n logger.info(\"----\"*5)\r\n return best_result\r\n\r\ndef main():\r\n parser = argparse.ArgumentParser()\r\n parser.add_argument('--no_cuda', action=\"store_true\")\r\n parser.add_argument(\"--local_rank\", type=int, default=-1, help=\"local_rank for distributed training on gpus\")\r\n macro = parser.parse_args()\r\n args = OptionSet()\r\n\r\n args.model_name = \"test\"\r\n args.model_name = args.model_name+\"_distributed_nogrid\"\r\n\r\n global logger\r\n logger = create_logger(args)\r\n\r\n # Setup CUDA, GPU & distributed training\r\n if macro.local_rank == -1 or macro.no_cuda:\r\n device = torch.device(\"cuda\" if torch.cuda.is_available() and not macro.no_cuda else \"cpu\")\r\n macro.n_gpu = 0 if macro.no_cuda else torch.cuda.device_count()\r\n else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs\r\n torch.cuda.set_device(macro.local_rank)\r\n device = torch.device(\"cuda\", macro.local_rank)\r\n torch.distributed.init_process_group(backend=\"nccl\")\r\n macro.n_gpu = 1\r\n # args.device = device\r\n set_seed(macro, args.seed)\r\n # Setup logging\r\n logging.basicConfig(\r\n format=\"%(asctime)s - %(levelname)s - %(name)s - %(message)s\",\r\n datefmt=\"%m/%d/%Y %H:%M:%S\",\r\n level=logging.INFO if macro.local_rank in [-1, 0] else logging.WARN,\r\n )\r\n logger.warning(\r\n \"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s\",\r\n macro.local_rank,\r\n device,\r\n macro.n_gpu,\r\n bool(macro.local_rank != -1),\r\n # args.fp16,\r\n )\r\n # Load pretrained model and tokenizer\r\n if macro.local_rank not in [-1, 0]:\r\n # Make sure only the first process in distributed training will download model & vocab\r\n torch.distributed.barrier()\r\n\r\n logger.info('using device:{}'.format(device))\r\n model, _, tokenizer = creat_model()\r\n if macro.local_rank == 0:\r\n # Make sure only the first process in distributed training will download model & vocab\r\n torch.distributed.barrier()\r\n model = model.to(device)\r\n global PAD_idx\r\n PAD_idx = tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[-3])\r\n train_data, dev_data, train_len = get_binary_data(args, macro, tokenizer, read_data=False)\r\n # Training phase.\r\n logger.info(\"Start training.\")\r\n instances_num = train_len\r\n train_steps = int(instances_num * args.epochs_num / args.batch_size) + 1\r\n\r\n logger.info('Batch size: {}'.format(args.batch_size))\r\n logger.info('The number of training instances:{}'.format(instances_num))\r\n\r\n num_parameters = 0\r\n parameters = model.parameters()\r\n for parameter in parameters:\r\n num_parameters += parameter.numel()\r\n logger.info('number of model parameters: {}'.format(num_parameters))\r\n\r\n decoder_layer = list(map(id, model.decoder.layers.parameters()))\r\n encoder_para = filter(lambda p: id(p) not in (decoder_layer), model.parameters())\r\n optimizer_grouped_parameters = [\r\n {'params': encoder_para, 'lr': args.learning_rate, 'weight_decay_rate': 0.01},\r\n {'params': model.decoder.layers.parameters(), 'lr': args.learning_rate * 5, 'weight_decay_rate': 0.01}\r\n ]\r\n optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, correct_bias=False)\r\n scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=train_steps*args.warmup, num_training_steps=train_steps)\r\n\r\n # multi-gpu training (should be after apex fp16 initialization)\r\n if macro.n_gpu > 1:\r\n model = torch.nn.DataParallel(model)\r\n\r\n # Distributed training (should be after apex fp16 initialization)\r\n if macro.local_rank != -1:\r\n model = torch.nn.parallel.DistributedDataParallel(\r\n model, device_ids=[macro.local_rank], output_device=macro.local_rank, find_unused_parameters=True\r\n )\r\n\r\n best_result = 0.0\r\n for epoch in range(1, args.epochs_num+1):\r\n best_result = train_epoch(model, train_data, dev_data, optimizer, device, scheduler, best_result, args, tokenizer, epoch, macro)\r\n\r\n```", "This is the result of my experiment:\r\n\r\n", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "You could subclass the `TrainingArguments` class and add more fields to it. You could refer to `https://github.com/huggingface/transformers/blob/master/src/transformers/training_args_seq2seq.py` for an example https://github.com/huggingface/transformers/blob/master/src/transformers/training_args_seq2seq.py", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "@LysandreJik knows better for the centralized logging system so I'll defer to him." ]

[]

[]

[]

[]

[]

[ "Please use the [forums](https://discuss.huggingface.co/) for questions like this. We keep issues for bugs or feature requests only." ]

[ { "id": 2796628563, "node_id": "MDU6TGFiZWwyNzk2NjI4NTYz", "url": "https://api.github.com/repos/huggingface/transformers/labels/WIP", "name": "WIP", "color": "234C99", "default": false, "description": "Label your PR/Issue with WIP for some long outstanding Issues/PRs that are work in progress" } ]

[ { "login": "stas00", "id": 10676103, "node_id": "MDQ6VXNlcjEwNjc2MTAz", "avatar_url": "https://avatars.githubusercontent.com/u/10676103?v=4", "gravatar_id": "", "url": "https://api.github.com/users/stas00", "html_url": "https://github.com/stas00", "followers_url": "https://api.github.com/users/stas00/followers", "following_url": "https://api.github.com/users/stas00/following{/other_user}", "gists_url": "https://api.github.com/users/stas00/gists{/gist_id}", "starred_url": "https://api.github.com/users/stas00/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/stas00/subscriptions", "organizations_url": "https://api.github.com/users/stas00/orgs", "repos_url": "https://api.github.com/users/stas00/repos", "events_url": "https://api.github.com/users/stas00/events{/privacy}", "received_events_url": "https://api.github.com/users/stas00/received_events", "type": "User", "site_admin": false } ]

[ "I'm definitely not an expert on logging so I'll leave it to @LysandreJik and @sgugger here. The idea of adding new `multi-process` logging functionality sounds very reasonable to me though!", "I would advocate for using the `logger.warning` in cases where you want to display something from all processes. While in the library we should be strict about what we want to display as info/warn/error, I think we can be a bit more flexible and use the logger verbosity differently, as it's actively defined in the scripts. Of course, this is only if you're defining the logs in the script and not in the library, otherwise we'll need to reconsider.\r\n\r\nIf we don't want to go down this road, we could also have an approach defined on the log levels. We could potentially define additional log levels, between INFO and WARN that could get the job done.\r\n\r\n", "Oh, I like an intermediate level between INFO and WARN, like `IMPORTANT_INFO`? (`ALL_PROCESS_INFO` wouldn't make sense as it would not technically control the number of processes).", "I'm talking about the library here, Trainer that is. But, of course, the same should apply to custom scripts.\r\n\r\nI think a custom logger is a better idea. In particular since it needs to log the rank of the process - currently I have to add it manually.\r\n\r\nI don't think we should mess with levels. These should be sacrosanct. This is because it could make things very confusing for the user.\r\n\r\nBut being able to retrieve a 2nd logger object that logs for all processes with the format that includes a process rank and using the same log level would be useful. \r\n\r\nThough need to decide whether:\r\na. Such logger would be not logging anything unless invoked in a distributed environment.\r\nb. Or perhaps it's actually better for it to be identical to the normal logger under non-distributed env, so logs aren't missed - it's just it'll not include process rank.", "> I'm talking about the library here, Trainer that is.\r\n\r\nThe problem is that only Trainer knows when it's executed in a distributed training but logs are in all parts of the library. Though maybe this new logger will only be used inside the `Trainer`?\r\n\r\n(Sorry my 3yo made a wrong click.) ", "> The problem is that only Trainer knows when it's executed in a distributed training but logs are in all parts of the library. \r\n\r\nNot really. We have `torch.distributed.get_rank()` for most things to know whether we are under distributed, though the logger shouldn't be initialized until first use since the dist env comes a bit later in the game. Down the road if we have other methods that defined multiproc we will just try those too or provide one for a user to run if it's non-standard and it'll set a multi-proc flag in the logging library.\r\n\r\nIt probably could/should copy the same format from normal logger, but embed process rank into it.\r\n\r\ni.e. perhaps it can be created on the fly and require no special handling on the user side (or trainer side).\r\n\r\n> Though maybe this new logger will only be used inside the Trainer?\r\n\r\nNo, user scripts will need it too. It's not trainer-specific. Think `model.parallelize()` `model.enable_pipeline` (new)\r\n\r\nAnd just to clarify - this is for logging inside the model's code - perhaps I will find a way to abstract it out, but it still would be outside of Trainer and in the core library.\r\n\r\nAn example of this is building a custom device_map for PP or MP specific to the process and logging that. This would be the same whether it was called from Trainer or user's code.\r\n\r\n> (Sorry my 3yo made a wrong click.)\r\n\r\nYou have competition growing ;)", "I have no issues with having a second logger used in multi-process environments. It would be nice if it could be handled by the `logging` class of `transformers`, so as to have a single front-end for the logging, otherwise we'll end up confusing the users as much as if we add intermediate logging levels.\r\n\r\nDo we actually need a second logger though, wouldn't it be simpler to adapt the formatter for those particular logs?", "> I have no issues with having a second logger used in multi-process environments. It would be nice if it could be handled by the `logging` class of `transformers`, so as to have a single front-end for the logging, [...]\r\n\r\nYes, that is what I had in mind.\r\n\r\n> Do we actually need a second logger though, wouldn't it be simpler to adapt the formatter for those particular logs?\r\n\r\nCould you please give a example of what you have in mind? \r\n\r\nI have no attachment whatsoever to how this is done. So if you already have an idea on how to make this work I'm all ears.\r\n\r\nThank you.", "Ok I gave it more thought and my proposal of using formatters was a mistake, it won't be possible this way. I looked for a solution using filters, but, alas, the logs are already filtered by the levels before they're handled by the filters.\r\n\r\nThinking about it further, I'm not 100% sure how I can see several loggers here as we already have one logger per module. If you know how to do it cleanly, by all means, please do!\r\n\r\n---\r\n\r\nAs an aside, I don't think having intermediate levels is a bad thing. The `logging` utility has an `addLevelName` method, and this specific use-case seems perfect. I understand why adding many different level names will get harder to understand, but this is the addition of one level for a situation that would benefit from it.\r\n\r\nIt would only require adding a level, which has a name. Here's how it could look like:\r\n\r\n```py\r\nimport logging\r\nimport sys\r\n\r\n# Add level > 30\r\nlogging.addLevelName(35, \"MODEL_PARALLEL_INFO\")\r\n\r\n# Setup logging like we do in our scripts\r\nlogger = logging.getLogger(__name__)\r\nlogging.basicConfig(\r\n format=\"%(asctime)s - %(levelname)s - %(name)s - %(message)s\",\r\n datefmt=\"%m/%d/%Y %H:%M:%S\",\r\n handlers=[logging.StreamHandler(sys.stdout)],\r\n)\r\n\r\nfor i in range(3):\r\n # Simulate our scripts' logging\r\n logger.setLevel(logging.INFO if i == 0 else logging.WARN)\r\n\r\n logger.warning(f\"Initiating {i}\")\r\n logger.info(\"Random information\")\r\n logger.log(logging.getLevelName(\"MODEL_PARALLEL_INFO\"), f\"Device Map of {i}: []\")\r\n```\r\n\r\nThis logs the following:\r\n\r\n```\r\n02/03/2021 15:40:28 - WARNING - __main__ - Initiating 0\r\n02/03/2021 15:40:28 - INFO - __main__ - Random information\r\n02/03/2021 15:40:28 - MODEL_PARALLEL_INFO - __main__ - Device Map of 0: []\r\n02/03/2021 15:40:28 - WARNING - __main__ - Initiating 1\r\n02/03/2021 15:40:28 - MODEL_PARALLEL_INFO - __main__ - Device Map of 1: []\r\n02/03/2021 15:40:28 - WARNING - __main__ - Initiating 2\r\n02/03/2021 15:40:28 - MODEL_PARALLEL_INFO - __main__ - Device Map of 2: []\r\n```\r\n\r\nHappy to drop the idea if you still think this would make things confusing for the user. I do agree that it could make things confusing for users not making use of the scripts and wondering why an INFO statement slipped when they set their verbosity to `warn`.", "I think we can have the intermediate level be intermediate -> so not shown when the verbosity is set to `warn`. I expect the scripts to switch the statement from\r\n```\r\nlogger.setLevel(logging.INFO if i == 0 else logging.WARN)\r\nlogger.parallel_info(f\"Initiating {i}\")\r\n```\r\nto\r\n```\r\nlogger.setLevel(logging.INFO if i == 0 else logging.MODEL_PARALLEL_INFO)\r\nlogger.parallel_info(f\"Initiating {i}\")\r\n```\r\n\r\nI disagree with the name though, as this is a bit too specific ;-) PARALLEL_INFO is enough IMO\r\n", "To the naming:\r\n\r\nWell, PP is too specific - not very generic either.\r\n\r\nWhat we have here is 2 specific events, which may happen under any distributed training. So the common is that it's distributed, the separate is:\r\n1. log only once for multiple processes (avoid duplicated logging)\r\n2. log for every process (only for unique per-process logging)\r\n\r\nand single process training is a special case of distributed with n_procs=1, so for a single proc both should be logged.\r\n\r\nSo I think this is what the name should reflect and not the specific circumstance it's used in.\r\n\r\n------------------\r\n\r\nTo the implementation, thank you for your specific code suggestions @LysandreJik and @sgugger - please let me experiment with your proposals and try other things out and I will come back to you.\r\n\r\n", "Apologies for the delay, here is how I see a simple solution that doesn't break any conventions.\r\n\r\nWe create a second logger. Just need to think how to make it appear if the user didn't explicitly configure one and make it globally available from other modules.\r\n\r\nHere is a possible implementation:\r\n```\r\n# logger.py\r\n\r\nimport logging\r\nimport sys\r\nimport os\r\n\r\nlocal_rank = int(os.environ.get(\"LOCAL_RANK\", -1))\r\n\r\n# normal logger\r\nlogger = logging.getLogger(__name__)\r\nhandler_shared = logging.StreamHandler(sys.stdout)\r\nformatter_shared = logging.Formatter('%(asctime)s - %(levelname)s - %(name)s - %(message)s')\r\nhandler_shared.setFormatter(formatter_shared)\r\nlogger.addHandler(handler_shared)\r\n\r\n# rank-specific logger\r\nif local_rank != -1:\r\n logger_rank_specific = logging.getLogger(__name__ + \"rank_specific\")\r\n handler_rank_specific = logging.StreamHandler(sys.stdout)\r\n formatter_rank_specific = logging.Formatter(f'%(asctime)s - %(levelname)s - p{local_rank} - %(name)s - %(message)s')\r\n handler_rank_specific.setFormatter(formatter_rank_specific)\r\n logger_rank_specific.addHandler(handler_rank_specific)\r\nelse:\r\n logger_rank_specific = logger\r\n\r\n# the 2nd logger is just for special info that each process should print\r\nlogger_rank_specific.setLevel(logging.INFO)\r\n# set normal logger to just the main process INFO\r\nlogger.setLevel(logging.INFO if local_rank < 1 else logging.WARN)\r\n\r\n# test\r\nlogger.warning(f\"Initiating\")\r\nlogger.info(\"Random information\")\r\n\r\nlogger_rank_specific.info(f\"Device Map: {[1] * local_rank}\")\r\n\r\n```\r\n\r\nDist test:\r\n```\r\n$ python -m torch.distributed.launch --nproc_per_node 4 ./logger.py\r\n2021-02-11 19:57:45,889 - WARNING - __main__ - Initiating\r\n2021-02-11 19:57:45,889 - INFO - __main__ - Random information\r\n2021-02-11 19:57:45,889 - INFO - p0 - __main__rank_specific - Device Map: []\r\n2021-02-11 19:57:45,897 - WARNING - __main__ - Initiating\r\n2021-02-11 19:57:45,898 - INFO - p1 - __main__rank_specific - Device Map: [1]\r\n2021-02-11 19:57:45,905 - WARNING - __main__ - Initiating\r\n2021-02-11 19:57:45,905 - INFO - p2 - __main__rank_specific - Device Map: [1, 1]\r\n2021-02-11 19:57:45,914 - WARNING - __main__ - Initiating\r\n2021-02-11 19:57:45,914 - INFO - p3 - __main__rank_specific - Device Map: [1, 1, 1]\r\n```\r\n\r\nNon-dist test:\r\n```\r\n$ python ./logger.py\r\n2021-02-11 20:21:16,716 - WARNING - __main__ - Initiating\r\n2021-02-11 20:21:16,717 - INFO - __main__ - Random information\r\n2021-02-11 20:21:16,717 - INFO - __main__ - Device Map: []\r\n```\r\n\r\nAll works.\r\n\r\nNot sure what to call the second logger, open to suggestions.\r\n\r\nWhat do you think?\r\n\r\nThank you.", "I am fine with using a second logger like this, I guess it could be called `multiprocess_logger` and that its name could be `+ \"rank_specific\"` like you said. Should we add a method `get_multiprocess_logger` in the logging module so that people don't have to remember the `__name__ + \"rank_specific\"` part? This would give an API like:\r\n```\r\nfrom .util import logging\r\n\r\nlogger = logging.get_logger(__name__)\r\nmultiprocess_logger = logging.get_multiprocess_logger(__name__)\r\n```\r\nin the modules where we need the `multiprocess_logger`. And the `set_verbosity_xxx` methods would affect both transformers logger.\r\n\r\nFor the scripts we would still need to do it manually though.", "Yes, of course, we will have it all nicely wrapped up. If @LysandreJik is in agreement, I will work on a PR.\r\n\r\nIt'd be nice to have a somewhat shorter name for `multiprocess_logger`, but the one you proposed works too. Perhaps, in reversed to aid the completion? `logger_multiproc` or `logger_multiprocess` or `logger_mp`?\r\n\r\nAlso I'm not sure with `__name__ + \"rank_specific\"` - should it be in sync with the variable name - whichever we choose?\r\n\r\nHmm, what if instead of changing the format for that logger to be\r\n```\r\nlogging.Formatter(f'%(asctime)s - %(levelname)s - p{local_rank} - %(name)s - %(message)s')\r\n```\r\nWe keep the exact same format, but we simply append the actual rank to the name?\r\n```\r\nlogger_rank_specific = logging.getLogger(__name__ + f\"local_rank_{local_rank}\")\r\nlogging.Formatter(f'%(asctime)s - %(levelname)s - %(name)s - %(message)s')\r\n```\r\nBut again, either way works. Just one less thing to modify in this case.\r\n\r\n", "Thanks for writing everything out, I'm ok with your proposal!" ]

[]

[]

[]

[]

[]

[ "I'm guessing you're running the trainer code block from the sequence classification example verbatim. You want `DistilBertForTokenClassification` not `DistilBertForSequenceClassification`, so comment out:\r\n\r\n> model = DistilBertForSequenceClassification.from_pretrained(\"distilbert-base-uncased\") ", "> I'm guessing you're running the trainer code block from the sequence classification example verbatim. You want `DistilBertForTokenClassification` not `DistilBertForSequenceClassification`, so comment out:\r\n> \r\n> > model = DistilBertForSequenceClassification.from_pretrained(\"distilbert-base-uncased\")\r\n\r\nDoh. Indeed, thanks for the catch.", "Had to use DistilBertForTokenClassification to reproduce the example. " ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "All fast tokenizers have this feature, just pass along `return_offsets_mapping=True` in your call to the tokenizer. Also note that fast tokenizers are used by default when `AutoTokenizer` is called.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "> Thanks for flagging the doc was incorrect! You changes are not entirely correct either so I made some suggestions.\r\n\r\nThanks for fixing my tweaks - I like the changes so committed them 😃 ", "Sorry for missing that last suggestion of yours - should be ready to go now!", "Yes, thanks a lot!" ]

[]

[]

[ "Hi! Could you install an earlier version of `transformers` to see if it works? I believe it was tested with `transformers==3.0.1`", "It works fine up to and including version v4.2.2 but is broken in versions above that", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Pinging @julien-c ", "> I wanted to check if your conditions set a default license under which models are uploaded when not specified?\r\n\r\nNo, that's really the model author's call. But we will try to make it easier/more straightforward for a user to pick one in the future.\r\n\r\n> In order to have a missing license added when it is missing, could you please advise on the standard way to proceed?\r\n\r\nA GH issue is fine I think, otherwise a thread on [discuss.huggingface.co](https://discuss.huggingface.co) would work well too.", "Thank you for the clarification!" ]

[]

[]

[]

[]

[]

[ "Currently `SortishSampler` will only work with the `finetune_trainer.py` scripts. It will be supported in `run_seq2seq.py` in the soon. And to answer your question, yes it works with multiple GPUs, and you won't need to enable distributer parameter if the training is launched on multi GPUs using `torch.distributed.launch`, it'll be enabled automatically. ", "thanks a lot." ]

[]

[]

[]

[]

[]

[ "I don't think this is something that should be done in `save_vocabulary`. You have the option in `save_pretrained` to set `legacy_format` to `False` to generate that `tokenizer.json` file. I'm not an expert in the tokenization side with all the stuff that was added for backward compatibility so I don't know if there is a better option.\r\n\r\nI wasn't aware havin this file was mandatory for some models to use the fast tokenizer. Are you sure you have sentencepiece installed? It might be due to this that the conversion slow to fast does not work automatically\r\n\r\nAnyhow, once we have found the right way to generate that `tokenizer.json` file, it should be added on the model sharing doc page, next to the section on how to generate TF/PyTorch checkpoints, so that people know what to do to have the most complete model on the hub.", "I don't have a problem to add it anywhere else, who do we tag on this?\r\n\r\n1. Let the code speak for itself:\r\n```\r\npython -c \"from transformers import T5Tokenizer, T5TokenizerFast; mname_from='sshleifer/t5-tinier-random'; tokenizer = T5Tokenizer.from_pretrained(mname_from); tokenizer_fast = T5TokenizerFast.from_pretrained(mname_from)\"\r\nTraceback (most recent call last):\r\n File \"<string>\", line 1, in <module>\r\n File \"/mnt/disc1/data/trash/src/transformers/src/transformers/tokenization_utils_base.py\", line 1762, in from_pretrained\r\n return cls._from_pretrained(\r\n File \"/mnt/disc1/data/trash/src/transformers/src/transformers/tokenization_utils_base.py\", line 1835, in _from_pretrained\r\n tokenizer = cls(*init_inputs, **init_kwargs)\r\n File \"/mnt/disc1/data/trash/src/transformers/src/transformers/models/t5/tokenization_t5_fast.py\", line 139, in __init__\r\n super().__init__(\r\n File \"/mnt/disc1/data/trash/src/transformers/src/transformers/tokenization_utils_fast.py\", line 86, in __init__\r\n fast_tokenizer = TokenizerFast.from_file(fast_tokenizer_file)\r\nException: No such file or directory (os error 2)\r\n```\r\n\r\n2. If `footokenizer.from_pretrained()` fetches `tokenizer.json` then `footokenizer.save_pretrained()` must save it too.\r\n \r\n> I wasn't aware havin gthis file was mandatory for some models to use the fast tokenizer. Are you sure you have sentencepiece installed? It might be due to this that the conversion slow to fast does not work automatically\r\n\r\n```\r\npip install sentencepiece\r\nRequirement already satisfied: sentencepiece in /mnt/nvme1/anaconda3/envs/main-38/lib/python3.8/site-packages (0.1.91)\r\n```\r\n\r\nIf you look at the trace it is hunting for that file and can't find it.\r\n\r\n> Anyhow, once we have found the right way to generate that tokenizer.json file, it should be added on the model sharing doc page, next to the section on how to generate TF/PyTorch checkpoints, so that people know what to do to have the most complete model on the hub.\r\n\r\nAgreed!\r\n\r\n@LysandreJik, @n1t0 ", "ok, so as @sgugger suggested on slack, the fast tokenizer saving will be handled on the core-level some time in the future, so I removed that part from this PR, leaving just the logger part." ]

[]

[]

[ "If I understood correctly the user's issue, the problem was that any value was accepted. We can add \"True\" in `ENV_VAR_TRUE_VALUES` which seems to be missing, but if I set `WAND_DISABLED=False` for instance, I would expect wandb to not be disabled.\r\n\r\nIn any case those env variables are now deprecated (have to make a PR to issue a proper warning) since we have the `report_to` training argument that allows the user to set the reporting platform they want to use.", "> If I understood correctly the user's issue, the problem was that any value was accepted. We can add \"True\" in `ENV_VAR_TRUE_VALUES` which seems to be missing, but if I set `WAND_DISABLED=False` for instance, I would expect wandb to not be disabled.\r\n\r\nThat is how I implemented it originally for this PR, but then read user's issue that triggered the PR that broke the original setting, and the issue writer requested a plain - any `WANDB_DISABLED` value. I'm fine with either. Do you want me to recode it to add `True`?\r\n\r\nAlso it needs to be documented, so that this disabling is solid and doesn't get changed again and again. If it's documented with just `Yes` that is already supported that is good enough for me.\r\n\r\n> In any case those env variables are now deprecated (have to make a PR to issue a proper warning) since we have the `report_to` training argument that allows the user to set the reporting platform they want to use.\r\n\r\nWell, except this new feature doesn't help in this particular case. As you can see from https://github.com/huggingface/transformers/issues/9623 and problems as recent as yesterday wandb is still a problem, even if you don't purposefully activate it or even have it installed. I won't be trying to fix this if it worked. \r\n\r\nPerhaps the default `report_to` should be `None` and have an option for `All` to ease up for those who want them all?\r\n\r\nWhatever the outcome, please let's fix so that if one doesn't have wandb installed it shouldn't break things.\r\n\r\nThank you.\r\n", "> Do you want me to recode it to add True?\r\n\r\nYes, just as I said, adding `True` to the `ENV_VAR_TRUE_VALUES` should be enough to have this work (it's an oversight that `True` is not in that constant).\r\n\r\n> Also it needs to be documented, so that this disabling is solid and doesn't get changed again and again. If it's documented with just Yes that is already supported that is good enough for me.\r\n\r\nBy all means, please add documentation in this PR. For now it's documented with the [callback](https://huggingface.co/transformers/main_classes/callback.html#transformers.integrations.WandbCallback) but I'm open to any suggestion to make this better.\r\n\r\n> Whatever the outcome, please let's fix so that if one doesn't have wandb installed it shouldn't break things\r\n\r\nThe bug in #9623 with wandb not installed is linked to something weird in your env as I haven't been able to reproduce it by following your steps. I can add stronger checks that wandb is a proper module by checking its version/authors (like is done for [datasets](https://github.com/huggingface/transformers/blob/22121e813e2d043feb4484865ab5871870cb9dc3/src/transformers/file_utils.py#L130) but I have no idea if it will solve your bug or not (since I have no reproducer on my side).\r\n\r\nIf wandb is installed and you pass along `--report_to []`, you should not see either\r\n```\r\nwandb.errors.error.Error: You must call wandb.init() before wandb.log()\r\n```\r\nnor\r\n```\r\nAttributeError: module 'wandb' has no attribute 'ensure_configured'\r\n``` \r\nas the callback is not passed to the Trainer.\r\n\r\n> Perhaps the default report_to should be None and have an option for All to ease up for those who want them all?\r\n\r\nAs I explained before, that switch will be done in v5, as it is a breaking change.", "Thank you for the feedback, @sgugger - PR updated as requested, plus synced trainer_tf with the same solution." ]

[]

[]

[ "Solved it by installing tensorflow-gpu=2.3.0 & cuda 10.1\r\n\r\nFollowing this guide:\r\nhttps://medium.com/analytics-vidhya/tensorflow-2-3-0-with-gpu-support-on-windows-10-f975a552ea7c\r\n\r\nUse this command to install gpu2.3.0\r\npython -m pip install https://storage.googleapis.com/tensorflow/windows/gpu/tensorflow_gpu-2.3.0-cp37-cp37m-win_amd64.whl" ]

[]

[]

[ "I'm not sure where you have seen those objects: they are nowhere in the transformers library. The library provides `EncoderDecoderModel`, see the [encoder/decoder doc page](https://huggingface.co/transformers/model_doc/encoderdecoder.html).", "Hi, I was reading this (https://medium.com/huggingface/encoder-decoders-in-transformers-a-hybrid-pre-trained-architecture-for-seq2seq-af4d7bf14bb8). I also found someone reporting some issue while using the same object here\r\nhttps://github.com/huggingface/transformers/issues/2206.\r\nPerhaps I am looking at some older version ?\r\n\r\n", "This is indeed from an older version (I guess 2 something or even 1 something). ", "Thanks. I will look at the EncoderDecoderModel" ]

[ { "id": 2604155188, "node_id": "MDU6TGFiZWwyNjA0MTU1MTg4", "url": "https://api.github.com/repos/huggingface/transformers/labels/Benchmarks", "name": "Benchmarks", "color": "2DF372", "default": false, "description": "Issues related to Memory regressions in tests and scripts" }, { "id": 2796628563, "node_id": "MDU6TGFiZWwyNzk2NjI4NTYz", "url": "https://api.github.com/repos/huggingface/transformers/labels/WIP", "name": "WIP", "color": "234C99", "default": false, "description": "Label your PR/Issue with WIP for some long outstanding Issues/PRs that are work in progress" } ]

[ { "login": "stas00", "id": 10676103, "node_id": "MDQ6VXNlcjEwNjc2MTAz", "avatar_url": "https://avatars.githubusercontent.com/u/10676103?v=4", "gravatar_id": "", "url": "https://api.github.com/users/stas00", "html_url": "https://github.com/stas00", "followers_url": "https://api.github.com/users/stas00/followers", "following_url": "https://api.github.com/users/stas00/following{/other_user}", "gists_url": "https://api.github.com/users/stas00/gists{/gist_id}", "starred_url": "https://api.github.com/users/stas00/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/stas00/subscriptions", "organizations_url": "https://api.github.com/users/stas00/orgs", "repos_url": "https://api.github.com/users/stas00/repos", "events_url": "https://api.github.com/users/stas00/events{/privacy}", "received_events_url": "https://api.github.com/users/stas00/received_events", "type": "User", "site_admin": false } ]

[ "I was thinking about one feature if possible,\r\n\r\nHow about when we run an example script a benchmarking script is automatically run and store the results in one file if the user passes an optional argument.\r\n\r\nWhen the user uploads the model on the model hub we can directly sort the model based on benchmarking results file.", "All data files on the model hub for the same model arch will give the same speed performance results, since they are just data points. \r\n\r\nTherefore it's the model code that needs to be benchmarked (and the trainer if there is more than one).\r\n\r\nAnd given that currently we have only one model implementation of each there is nothing to compare it to.\r\n\r\nThe main idea of this issue is to do regression testing, to ensure that we don't accidentally make models slower while changing the code. For an example of this happening, please see: https://github.com/huggingface/transformers/pull/11218" ]

[]

[]

[ "Hi @ari9dam \r\nPlease use the [forum](https://discuss.huggingface.co/) for such questions, and there's a discussion about this in the post \r\nhttps://discuss.huggingface.co/t/t5-finetuning-tips/684" ]

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Pinging @mfuntowicz ", "> Morgan was mentioning that the transpose_for_score method was called right after the Q/K/V projection, but that there was no need to split this dimension if we're not doing head masking.\r\nWhat do you think? Maybe that's some work for another PR, though.\r\n\r\nI think it seems doable but not sure, I prefer to keep things like this to be sure we revert properly as it was before and we get at least a proper version, and we can take care of this change in another PR.", "Sounds good to me!", "LGTM! @patrickvonplaten feel free to merge if you approve the changes ^^", "I won't have time to do a proper review today (can do it tomorrow), but feel free to merge without me if @LysandreJik and @sgugger are ok with it", "@patrickvonplaten if you can take a look at it today and merge it if it's fine with you, that would be great" ]

[ { "id": 1843244711, "node_id": "MDU6TGFiZWwxODQzMjQ0NzEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/New%20model", "name": "New model", "color": "fbca04", "default": false, "description": "" } ]

[]

[]

[ { "id": 1260952223, "node_id": "MDU6TGFiZWwxMjYwOTUyMjIz", "url": "https://api.github.com/repos/huggingface/transformers/labels/Discussion", "name": "Discussion", "color": "22870e", "default": false, "description": "Discussion on a topic (keep it focused or open a new issue though)" } ]

[]

[ "Just did, thanks a lot @LysandreJik, the form is super quick to fill-in and interesting!\r\nEveryone, we're waiting for you", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "@sgugger @LysandreJik \r\nCould you help me review these code ?", "> Hi there! Thanks for updating your example. We have now created a `research_projects` project for the examples not directly maintained by the core team, and I think the `run_mlm_wwm` script and the chine_ref file could all go there in a new folder. Would you mind adjusting your PR in that direction?\r\n\r\nSure, maybe move `run_chinese_ref.py` to `research_projects` folder and leave `run_mlm_wwm.py` in where it was would be better ? And I don't know which folder is better ?\r\nThe two files are independent, we could move it to anywhere.", "The `run_mlm_wwm` file is not maintained by us directly and it only works for BERT-models, compared to the other examples, so I think it can all go together there. You can create a new folder named `mlm_wwm` (since it's not just Chinese) for instance and have the specific requirements in the `requirements.txt` file there?", "> The `run_mlm_wwm` file is not maintained by us directly and it only works for BERT-models, compared to the other examples, so I think it can all go together there. You can create a new folder named `mlm_wwm` (since it's not just Chinese) for instance and have the specific requirements in the `requirements.txt` file there?\r\n\r\ndone!", "Last thing is to run `make style` to make sure the files are properly formatted, let me know if you have any issue doing this!", "> Last thing is to run `make style` to make sure the files are properly formatted, let me know if you have any issue doing this!\r\n\r\nyeah, seem my previous PR also failed in format :(\r\nI got error as follow:\r\n```\r\n#!/bin/bash -eo pipefail\r\nblack --check examples tests src utils\r\nwould reformat /home/circleci/transformers/examples/research_projects/mlm_wwm/run_chinese_ref.py\r\nwould reformat /home/circleci/transformers/src/transformers/trainer.py\r\nOh no! 💥 💔 💥\r\n2 files would be reformatted, 706 files would be left unchanged.\r\n\r\nExited with code exit status 1\r\n```\r\nBut I formate my code.\r\n\r\n\r\nMaybe you could help me do this part ?", "@sgugger My pleasure. Maybe you could help me fix the formate error :(\r\nMy python version `3.9.1` black `20.8b1`, why I got diff result in CI." ]

[ { "id": 1920687293, "node_id": "MDU6TGFiZWwxOTIwNjg3Mjkz", "url": "https://api.github.com/repos/huggingface/transformers/labels/Fast%20Tokenizers", "name": "Fast Tokenizers", "color": "b60205", "default": false, "description": "" } ]

[]

[ "cc'ing @n1t0 on this in case he didn't see it", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Hello, thanks for opening an issue! We try to keep the github issues for bugs/feature requests.\r\nCould you ask your question on the [forum](https://discusss.huggingface.co) instead? You'll get more help there.\r\n\r\nThe docs regarding the maintained trainer are available [here](https://github.com/huggingface/transformers/tree/master/examples/seq2seq#new-script) and may be useful to you.\r\nThanks!", "Oh sorry about that, Thank you very much" ]

[]

[]

[ "I wouldn't be surprised that ONNX serializes each of the layers as independant layers when they're all repeated. I don't know enough about the ONNX export to know it that's the issue or what to do to fix it though.\r\n\r\nDo you get similar increases in size with other models? With BERT for example?", "Hi @LysandreJik , I faced the issue only on albert models so far, exporting to onnx for other BERT models worked fine and use them for prediction as well.", " I solved the issue using [this code](https://github.com/thehetpandya/onnx-shared-weights-remove/blob/main/onnx_remove_shared_weights.ipynb) that removes shared weights from the ONNX model." ]

[]

[]

[ "I am trying to understand the use of sortish sampler. Right now it is not used in the run_seq2seq.py script. ", "I found it. It is in the t**rainer_seq2seq.p**y script." ]

[]

[]

[ "Hi! Thanks for opening an issue. I see two issues with your setup here:\r\n- Why are you using `from_pretrained` to load the `RobertaModel` inside your pre-trained model? You should just initialize a `RobertaModel` from the configuration imo.\r\n- Instead of `PreTrainedModel`, I would instead use `RobertaPreTrainedModel`.\r\n\r\nSee the below script for an example of what I would recommend. I'm saving & reloading the model to make sure that all the weights get saved/loaded:\r\n\r\n```py\r\nfrom transformers import RobertaModel, RobertaConfig, logging\r\nfrom transformers.models.roberta.modeling_roberta import RobertaPreTrainedModel\r\nimport torch\r\n\r\nlogging.set_verbosity_info()\r\n\r\nclass MaskClassifier(RobertaPreTrainedModel):\r\n def __init__(self, config):\r\n super().__init__(config=config)\r\n self.roberta = RobertaModel(config)\r\n self.max_mask = 10\r\n self.hidden_size = config.hidden_size\r\n self.linear1 = torch.nn.Linear(2 * self.hidden_size, self.hidden_size)\r\n self.linear2 = torch.nn.Linear(self.hidden_size, self.max_mask + 1)\r\n self.softmax = torch.nn.Softmax(dim=1)\r\n\r\n self.init_weights()\r\n\r\nmodel = MaskClassifier.from_pretrained(\"roberta-base\")\r\n```\r\n\r\nLet's see the logs now, for the first load using the `roberta-base` checkpoint:\r\n\r\n```\r\nSome weights of the model checkpoint at roberta-base were not used when initializing MaskClassifier: ['lm_head.bias', 'lm_head.dense.weight', 'lm_head.dense.bias', 'lm_head.layer_norm.weight', 'lm_head.layer_norm.bias', 'lm_head.decoder.weight']\r\n- This IS expected if you are initializing MaskClassifier from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\r\n- This IS NOT expected if you are initializing MaskClassifier from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\r\nSome weights of MaskClassifier were not initialized from the model checkpoint at roberta-base and are newly initialized: ['roberta.embeddings.position_ids', 'linear1.weight', 'linear1.bias', 'linear2.weight', 'linear2.bias']\r\n```\r\n\r\nThe warning tells you: you're not using the `lm_head` weights, and the following layers are initialized: `linear1` and `linear2`.\r\nSince you're not using the LM head, and the two layers are the ones you just added, then there's nothing to worry about.\r\n\r\nLet's try saving the model and reloading it again:\r\n\r\n```py\r\nmodel.save_pretrained(\"here\")\r\nMaskClassifier.from_pretrained(\"here\")\r\n```\r\n\r\nThe logs show:\r\n```\r\nAll model checkpoint weights were used when initializing MaskClassifier.\r\nAll the weights of MaskClassifier were initialized from the model checkpoint at here.\r\n```\r\n\r\nSuccess :tada: ", "Thanks a lot!!! It works ", "@LysandreJik \r\n\r\nI really appreciate your help! You saved me from nightmares... \r\nActually I have one more custom model, and I tried the same structure you showed me, but it fails to load the weights. The only difference is that I'm using RobertaForMaskedLM, not RobertaModel here. \r\n\r\nModel Structure\r\n```\r\nclass MaskedLM(RobertaPreTrainedModel):\r\n def __init__(self, config):\r\n super().__init__(config=config)\r\n self.roberta = RobertaForMaskedLM(config)\r\n # self.tokenizer = RobertaTokenizer.from_pretrained('roberta-base')\r\n self.refinement_num = 3\r\n # self.mask_id = self.tokenizer.convert_tokens_to_ids([tokenizer.mask_token])[0] # 50264\r\n self.init_weights()\r\n def forward( ... )\r\n```\r\nInitialize Model\r\n```\r\nmodel = MaskedLM.from_pretrained('roberta-base')\r\n```\r\n\r\nError Message \r\n```\r\nSome weights of the model checkpoint at roberta-base were not used when initializing MaskedLM: ['lm_head.bias', 'lm_head.dense.weight', 'lm_head.dense.bias', 'lm_head.layer_norm.weight', 'lm_head.layer_norm.bias', 'lm_head.decoder.weight', 'roberta.embeddings.word_embeddings.weight', 'roberta.embeddings.position_embeddings.weight', 'roberta.embeddings.token_type_embeddings.weight', 'roberta.embeddings.LayerNorm.weight', 'roberta.embeddings.LayerNorm.bias', 'roberta.encoder.layer.0.attention.self.query.weight', ... ]\r\n- This IS expected if you are initializing MaskedLM from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\r\n- This IS NOT expected if you are initializing MaskedLM from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\r\nSome weights of MaskedLM were not initialized from the model checkpoint at roberta-base and are newly initialized: ['roberta.roberta.embeddings.position_ids', 'roberta.roberta.embeddings.word_embeddings.weight', 'roberta.roberta.embeddings.position_embeddings.weight', 'roberta.roberta.embeddings.token_type_embeddings.weight', 'roberta.roberta.embeddings.LayerNorm.weight', 'roberta.roberta.embeddings.LayerNorm.bias', 'roberta.roberta.encoder.layer.0.attention.self.query.weight', ... ]\r\n```\r\n\r\nI don't know why this model has '**roberta.roberta**.embeddings.position_ids', not '**roberta**.embeddings.position_ids'", "Hmmm, the issue here is that there is a difference between `RobertaModel`, which has the following weights:\r\n```\r\nembeddings.position_ids\r\nembeddings.xxx\r\n[...]\r\n```\r\nand `RobertaForMaskedLM`, which contains `RobertaModel` under the `roberta` prefix:\r\n```\r\nroberta.embeddings.position_ids\r\nroberta.embeddings.xxx\r\n[...]\r\nlm_head.dense\r\nlm_head.bias\r\n[...]\r\n```\r\n\r\nI'm not entirely sure of what you're trying to achieve as I don't see your forward function, but I think you could prevent a lot of pain by redefining your model somewhat like `RobertaForMaskedLM` is setup:\r\n\r\n\r\n```py\r\n# Import the RobertaLMHead\r\nfrom transformers.models.roberta.modeling_roberta import RobertaPreTrainedModel, RobertaLMHead\r\n\r\n\r\nclass MaskedLM(RobertaPreTrainedModel):\r\n def __init__(self, config):\r\n super().__init__(config=config)\r\n\r\n # Create the RoBERTa model and its head like in the MaskedLM layer\r\n self.roberta = RobertaModel(config)\r\n self.lm_head = RobertaLMHead(config)\r\n\r\n self.refinement_num = 3\r\n self.init_weights()\r\n\r\n def forward( ... )\r\n outputs = self.roberta(xxx)\r\n sequence_output = outputs[0]\r\n prediction_scores = self.lm_head(sequence_output)\r\n\r\n # Do your stuff!\r\n```\r\n\r\nThis way you can load the checkpoint seamlessly in your model, as the naming with the prefixes will be correct.", "Thanks!! I tried to build a MaskedLM with some refinements. After predicting multiple <mask> tokens, mask two random predicted tokens and predict them again. Anyway thanks a lot 🤗 🤗 🤗 ", "Glad I could help!", "> (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\r\n\r\nHi~ \r\nI just have a question about why ” (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).“ IS NOT expected? the knowledge in pretrained model is same as your task that you are going to finetune?\r\nThanks~\r\nBest,\r\nPengbo" ]

[]

[]

[ "hi @mansimane \r\n\r\nIn your code in `TrainingArguments`, `adam_epsilon` is set to `1e06`, which quite a large value, I believe it's a typo, should be 1e-6 as mentioned in the comment. This could be the reason for `nan` gradients. ", "Thanks @patil-suraj for the catch. I fixed the Adam epsilon, but still some gradients are becoming infinity and nan after first backward pass. Following is the config I tried \r\n\r\n```python\r\ntraining_args = TrainingArguments(\r\n output_dir=\"./deberta\",\r\n overwrite_output_dir=True,\r\n\r\n num_train_epochs=1000,\r\n per_gpu_train_batch_size=32,\r\n learning_rate=1e-10,\r\n\r\n warmup_steps=10000,\r\n weight_decay=0.01,\r\n adam_beta1=0.9,\r\n adam_beta2=0.999,\r\n adam_epsilon=1e-6,\r\n max_grad_norm=1.0,\r\n save_steps=10_000,\r\n save_total_limit=2,\r\n logging_first_step=False,\r\n logging_steps=1,\r\n max_steps=10000,\r\n gradient_accumulation_steps=1,\r\n\r\n)\r\n```", "Hi, \r\n\r\nsorry for the late reply. I tested MLM with `DeBertaForMaskedLM` using the `run_mlm.py` script, and everything seems to be working fine. So it seems like a hyperparameter issue (I would suggest using the same hyperparameter values as this script). Your learning rate for example seems way too low.\r\n\r\nMy Google colab to reproduce: https://colab.research.google.com/drive/1Rk5JoBTzK0I8J3FjG2R4J9HCeOrUpRTt?usp=sharing", "I am having the same issue but with MobileBert after loading a pre-trained model. I trained from scratch a LM 23000 steps. Now loading the model mobilebert.from_pretrained() to reload the model and keep training. Now when I try to keep training the loss i NaN. I have removed all related to learning rate in the training args and the nans keep appearing.\r\n\r\n```\r\nfrom transformers import Trainer, TrainingArguments\r\n\r\ntraining_args = TrainingArguments(\r\n output_dir=\"./mobile_linear_att_4Heads_8L_128_512_03layerdrop_shared_all_dataset_1\",\r\n overwrite_output_dir=True,\r\n num_train_epochs=1,\r\n per_device_train_batch_size=95,\r\n save_steps=50,\r\n save_total_limit=2,\r\n logging_first_step=True,\r\n logging_steps=50,\r\n gradient_accumulation_steps=8,\r\n fp16=True,\r\n dataloader_num_workers=19,\r\n)\r\n\r\ntrainer = Trainer(\r\n model=model,\r\n args=training_args,\r\n data_collator=data_collator,\r\n train_dataset=big_dataset,\r\n tokenizer=tokenizer)\r\n\r\n\r\ntrainer.train()\r\n```\r\n\r\nEDIT: After some debugging I looked into the \"trainer_state.json\" and I have seen that before finishing the last training I got NaNs into the model so, it is nothing related to learning rate o something at this moment.\r\n\r\n```\r\n{\r\n \"cuda max_memory_reserved\": 23460839424,\r\n \"cuda memory cached\": 23460839424,\r\n \"cuda memory consumption\": 111139328,\r\n \"epoch\": 0.99,\r\n \"learning_rate\": 0.0004937288135593219,\r\n \"loss\": 4.5816,\r\n \"num_parameters\": 5920442,\r\n \"step\": 22900\r\n },\r\n {\r\n \"cuda max_memory_reserved\": 23460839424,\r\n \"cuda memory cached\": 23460839424,\r\n \"cuda memory consumption\": 111139328,\r\n \"epoch\": 0.99,\r\n \"learning_rate\": 0.0004934745762711864,\r\n \"loss\": NaN,\r\n \"num_parameters\": 5920442,\r\n \"step\": 22950\r\n },\r\n```\r\n\r\nEDIT2: I think that my issue is related to the scheduler in the learning rate. I am trying to train in batches of 20% of the dataset, so the learning rate scheduler I think, it calculate the learning rate based on the epoch and not on the current step, so I hardcoded in: \r\n\r\n```\r\nself.lr_scheduler = get_scheduler(\r\n self.args.lr_scheduler_type,\r\n self.optimizer,\r\n num_warmup_steps=self.args.warmup_steps,\r\n num_training_steps=num_training_steps, # <- here I hardcoded the calculated final (20%+20%+20%...) training steps\r\n )\r\n```\r\n\r\nSo when I was approximating the final of the training in the first 20% it got something weird.", "it's a pain to train on shards of (bookcorpus + wikipedia + openwebtext) I am processing the 20% of each one because I dont have more than 1 TB of disk. But I am figthing with the learning rate scheduler, because I have to do engineering to train on all the dataset. ", "Thank you @NielsRogge . I was able to train DeBERTa with run_mlm.py script. Not sure what was the issue in my code, it gave nan after trying learning rate that you used as well. ", "@mansimane are you using fp16 or fp32 ?" ]

[]

[]

[]

[]

[]

[ "> when you port this to the new run_seq2seq, it would be great to try to find a way to make this not use any special code for a given model\r\n\r\nI'm working on it in #9844, it's not finished though. We might need to add `get_input_embeddings` and `get_pos_embeddings` methods to every s2s model, to avoid special cases.", "If we need to add some methods to deal with the special cases, I would prefer it (otherwise the script might fail with new seq2seq models)." ]

[]

[]

[ "the doc link you shared if for V2.2.0, we have updated all the examples in a recent version. You can find the new ner examples here https://github.com/huggingface/transformers/tree/master/examples/token-classification", "Thank you @patil-suraj " ]

[]

[]

[ "@stancld can you please rebase on master in order to solve the conflicts?", "Thanks! The PR should be merged once @LysandreJik and @patrickvonplaten will have reviewed it.", "Thanks for fixing it!" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Updated with review comments. Please let me know if/when it's okay to merge :) ", "Good for me, thanks a lot!", "this is much better, thank you for the adjustment, @abhishekkrthakur " ]

[]

[]

[ "Hey @alexvaca0 thanks for bringing this up. Indeed our example is outdated. I updated it here https://github.com/amogkam/ray/blob/hf-pbt/python/ray/tune/examples/pbt_transformers/pbt_transformers.py and when running on the latest Ray wheels and the latest transformer release (4.2.2), I am seeing perturbations happening as expected\r\n\r\n\r\nI'll also try this out on transformers master branch and see if it's working on that as well, and will update here again. Thanks!", "I just tried on transformers master and am seeing it work as well. Please let me know if this updated example works for you @alexvaca0! ", "Thank you so much for your quick and very helpful answer @amogkam I'm using also the latest ray wheels and master version of transformers, therefore that example should work for me too! I'm going to try it as soon as I can so that I can tell you if it works for me or not! Thank you :) ", "I just checked your code and I don't find any change with respect to the official example except for the evaluation strategy, which was steps and now it's epochs, but maybe I'm missing something. I'll try first that example and then I'll try to apply it to my dataset. I have one more question regarding PBT with transformers: I'm observing that from time to time the models \"re-start\" from the beggining (that is, a model that had trained for 1.53 epochs suddenly returns to step 0 and starts from there). In my configuration I set number of epochs to 10, expecting that each of the 4 models in the population trains for 10 epochs, but mutating their configurations in the process. However, they'd never reach that number of epochs if they continue restarting from the beginning... Is there something I'm missing here? Do you think this issue will be also solved with the new training script? @amogkam Thank you !! :) ", "@amogkam \r\n\r\n2021-01-29 15:01:01,078\tWARNING trial_runner.py:370 -- Trial Runner checkpointing failed: Checkpoint must not be in-memory.\r\n\r\nIt still throws this error, and I've checked that another anomaly still persists: models not always re-start training from the point where they left, but start from the beginning again... Any clues why this may be happening? It's strange that this doesn't occur always, as sometimes models do re-start training from the point where they left...", "Hey @alexvaca0, could you share what your stdout looks like please? Also is this with any modifications to the example, and can you share the full code that you are using? Thanks!", "Could you give me your email so that I can share it with you that way? :) @amogkam ", "Hey yes you can send it to [email protected]", "Great! Already sent :) @amogkam ", "I am having the same problem ```Trial Runner checkpointing failed: Checkpoint must not be in-memory.```, but it does some times manage to create checkpoints, as I have ```PopulationBasedTraining: 4 checkpoints, 2 perturbs```\r\nModels seem to start training from the beginning again most of the time. I'm guessing it happens when the checkpoiting fails.\r\n\r\nI also notice that I am getting some errors and warnings:\r\n\r\n>WARNING function_runner.py:541 -- Function checkpointing is disabled. This may result in unexpected behavior when using checkpointing features or certain schedulers. To enable, set the train function arguments to be `func(config, checkpoint_dir=None)\r\n\r\n>ERROR syncer.py:72 -- Log sync requires rsync to be installed.\r\n\r\nDid you find a solution? \r\n\r\nAlso, I don't really understand how the ```perturbation_interval``` and ```time_attr``` arguments work together. It seems to consider the models for perturbations at the number of ```logging_steps``` I set in ```TrainingArguments```, but as I understand it, it is supposed to do so after every training step (so after every minibatch?) when ```time_attr=training_iteration``` and ```perturbation_interval=1.``` Since that's how it seems to work, I set ```checkpoint_freq``` to the same value as ```logging_steps``` in ```TrainingArguments```\r\n\r\n\r\nHere's what I think is the relevant part of my code:\r\n\r\n``` python\r\nclass UCCTrainer(Trainer):\r\n def compute_loss(self, model, inputs, return_outputs=False):\r\n outputs = model(\r\n input_ids=inputs['input_ids'],\r\n attention_mask=inputs['attention_mask'],\r\n token_type_ids=inputs['token_type_ids']\r\n )\r\n loss = th.nn.BCEWithLogitsLoss()(outputs['logits'], inputs['labels'])\r\n return (loss, outputs) if return_outputs else loss\r\n\r\n\r\ndef model_init():\r\n return BertForSequenceClassification.from_pretrained(\r\n config.MODEL_NAME, return_dict=True\r\n )\r\n\r\n\r\ndef objective(metrics):\r\n try:\r\n return metrics[config.COMPUTE_OBJECTIVE]\r\n except KeyError:\r\n return metrics[f'eval_{config.COMPUTE_OBJECTIVE}']\r\n\r\n\r\ndef hp_space(trial):\r\n return {\r\n 'learning_rate': tune.uniform(1e-5, 5e-5),\r\n 'num_train_epochs': tune.choice([2, 3, 4, 5]),\r\n 'seed': tune.choice(range(1, 50)),\r\n 'weight_decay': tune.uniform(0.0, 0.3),\r\n 'per_device_train_batch_size': tune.choice([10, 15, 20])\r\n }\r\n\r\n\r\ndef compute_metrics(eval_pred: EvalPrediction):\r\n scores = eval_pred.predictions # np.array 4427x2\r\n labels = eval_pred.label_ids # np.array 4427x2\r\n pred = np.argmax(scores, axis=1)\r\n labels_flat = np.argmax(labels, axis=1)\r\n return get_binary_metrics(pred, labels_flat)\r\n\r\nif __name__ == '__main__':\r\n os.environ['WANDB_WATCH'] = 'all'\r\n tokenizer = BertTokenizer.from_pretrained(\r\n config.MODEL_NAME,\r\n do_lower_case=config.DO_LOWER_CASE\r\n )\r\n train_df = dataframe_from_json('data/train_balanced.json')\r\n train_binary = make_binary_df(train_df)\r\n train_data = UCCDataset(train_binary, tokenizer, config.MAX_LEN)\r\n total_steps = len(train_data)/config.TRAIN_BATCH_SIZE\r\n warmup_steps = round(0.1*total_steps)\r\n training_args = TrainingArguments(\r\n output_dir=config.OUTPUT_DIR,\r\n do_train=True,\r\n do_eval=True,\r\n evaluation_strategy='steps',\r\n learning_rate=config.LEARNING_RATE,\r\n weight_decay=0.1,\r\n logging_steps=config.LOG_INTERVAL,\r\n seed=1,\r\n disable_tqdm=True,\r\n report_to=['wandb'],\r\n run_name=config.RUN_NAME,\r\n load_best_model_at_end=config.LOAD_BEST_LAST,\r\n metric_for_best_model=config.COMPUTE_OBJECTIVE,\r\n logging_first_step=True,\r\n lr_scheduler_type='linear',\r\n warmup_steps=warmup_steps\r\n )\r\n val_df = get_clean_df(pd.read_csv('data/val.csv'))\r\n val_binary = make_binary_df(val_df)\r\n val_data = UCCDataset(val_df, tokenizer, config.MAX_LEN)\r\n model_config = BertConfig(\r\n vocab_size=tokenizer.vocab_size,\r\n pretrained_model_name_or_path=config.MODEL_NAME,\r\n num_labels=config.N_LABELS,\r\n return_dict=True\r\n )\r\n\r\n trainer = UCCTrainer(\r\n args=training_args,\r\n train_dataset=train_data,\r\n eval_dataset=val_data,\r\n tokenizer=tokenizer,\r\n model_init=model_init,\r\n compute_metrics=compute_metrics\r\n )\r\n\r\n ray_scheduler = PopulationBasedTraining(\r\n time_attr='training_iteration',\r\n metric=f'eval_{config.COMPUTE_OBJECTIVE}',\r\n mode='max',\r\n perturbation_interval=1,\r\n hyperparam_mutations={\r\n 'learning_rate': tune.uniform(1e-5, 5e-5),\r\n 'num_train_epochs': tune.choice([2, 3, 4, 5]),\r\n 'seed': tune.choice(range(1, 50)),\r\n 'weight_decay': tune.uniform(0.0, 0.3),\r\n 'per_device_train_batch_size': tune.choice([10, 15, 20])\r\n }\r\n )\r\n best_model = trainer.hyperparameter_search(\r\n hp_space=hp_space,\r\n compute_objective=objective,\r\n n_trials=3,\r\n direction='maximize',\r\n backend='ray',\r\n # the following arguments are kwargs for tune.run\r\n scheduler=ray_scheduler,\r\n name='testmars5',\r\n resources_per_trial={'cpu': 1, 'gpu': 1},\r\n keep_checkpoints_num=3,\r\n checkpoint_score_attr=\"training_iteration\",\r\n checkpoint_freq=config.LOG_INTERVAL\r\n )\r\n```", "Hey folks, is this still an issue?\r\n\r\ncc @jwa018 @khrystynaFaryna" ]

[]

[]

[]

[]

[]

[ "Use \"AlbertTokenizer\" rather than \"AutoTokenizer\", this should solve your issue.\r\nPlease, check the updated notebook version.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Prot_albert tokenizer is returning none type, what changed?" ]

[]

[]

[ "Hmmmm the -100 id should be linked to the ignored values. It shouldn't try to decode this. Pinging @sgugger ", "Not sure why there would be `-100` in the labels with the old script. Note that we are not maintaining that one anymore and will replace it with `run_seq2seq` which is almost ready for use (misses a few features you don't seem to be using in your command anyway).\r\n\r\nIf you really need the old one, you should add the line\r\n```\r\nlabels = np.where(labels != -100, labels, tokenizer.pad_token_id)\r\n```\r\nin the `compute_metric` function to replace the -100s by the pad token id.", "We also fixed this very recently (yesterday) in the model, see: https://github.com/huggingface/transformers/commit/74f16b82765a05eccee45e80d79370202a958873 => so you should also be able to run: \r\n```\r\npython finetune_trainer.py --learning_rate 3e-5 --fp16 --evaluation_strategy steps --predict_with_generate --model_name_or_path moussaKam/barthez --data_dir xsum --do_train --do_eval --output_dir welcome_back --per_device_train_batch_size 4 --task summarization --max_target_length 50 --overwrite_output_dir --eval_steps 50 --n_val 20\r\n```\r\non master now.\r\n\r\nHowever, as @sgugger points out, we strongly recommend using the `run_seq2seq.py` script from now on as we won't continue maintaining `finetune_trainer.py` anymore." ]

[]

[]

[ "Hi there, thanks for your PR!\r\nThe examples scripts are kept simple and without too much functionality so users can easily understand and tweak them for their needs (they are just examples, they do not mean to cover **everything**). As you saw, it's super easy to add things like this option, if needed. The PR will stay this to demonstrate how, but I don't think we will merge it.", "As far as I know this is already done in the Tokenizer logic. You can define this lower casing option in the `tokenizer_config.json` - and this is done for quite a lot models. I dont't see any reason to have this as an extra cli option 🤔 ", "I just tried to be useful :)", "No problem at all. If you want to make sure something you are working on will be accepted, don't hesitate to open an issue about it first, that way we can tell you if it's desirable or not :-)\r\nBe sure to check the [good first issues](https://github.com/huggingface/transformers/issues?q=is%3Aissue+is%3Aopen+label%3A%22Good+First+Issue%22) if you want to try something else!" ]

[]

[]

[]

[]

[]

[ "It is used quite a lot! Here for example:\r\n\r\nhttps://github.com/huggingface/transformers/blob/4c3ae89ad3215c3252ebf8ce964795ba8813d810/src/transformers/models/electra/modeling_electra.py#L194-L196\r\n\r\nActually just Ctrl+F \"position_embedding_type\" in this file and you should be able to find out where it's used :) (11 occurrences)", "thanks " ]

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "I noticed this myself, `DataCollatorForWholeWordMask` seems to be very specific to the BERT tokenizer. It seems that it should be using the special tokens mask, and the word_ids() from the tokenizer rather than rely on [CLS],[SEP] tokens and subwords starting with ## (so it fails with metaspace for example).\r\n\r\nEdit: It also calls `self._tensorize_batch` which as far as I can see isn't implemented, so I assume this class isn't maintained?", "It's a bit old but I wanted to share my quick fix for Roberta-like tokenizers (I think this can be more general propose, but I just needed this for Herbert tokenizer):\r\n```\r\ndef _whole_word_mask(self, input_tokens: List[str], max_predictions=512):\r\n \"\"\"\r\n Get 0/1 labels for masked tokens with whole word mask proxy\r\n \"\"\"\r\n if not isinstance(self.tokenizer, (BertTokenizer, BertTokenizerFast,\r\n RobertaTokenizer, RobertaTokenizerFast,\r\n XLMRobertaTokenizer, XLMRobertaTokenizerFast,\r\n HerbertTokenizer, HerbertTokenizerFast,\r\n XLMTokenizer)):\r\n warnings.warn(\r\n \"DataCollatorForWholeWordMask is only suitable for BertTokenizer or RobertaTokenizer-like tokenizers. \"\r\n \"Please refer to the documentation for more information.\"\r\n )\r\n\r\n cand_indexes = []\r\n special_tokens = [val for key, val in self.tokenizer.special_tokens_map.items()\r\n if key not in ['unk_token', 'mask_token']]\r\n is_bert_tokenizer = isinstance(self.tokenizer, (BertTokenizer, BertTokenizerFast))\r\n for (i, token) in enumerate(input_tokens):\r\n if token in special_tokens:\r\n continue\r\n\r\n if is_bert_tokenizer:\r\n if len(cand_indexes) >= 1 and token.startswith(\"##\"):\r\n cand_indexes[-1].append(i)\r\n else:\r\n cand_indexes.append([i])\r\n else: # Roberta-like tokenizers have </w> token at the end to indicate end of word\r\n # edge case for chinese (##) are added in DataCollatorForWholeWordMask\r\n if token.startswith(\"##\"):\r\n token = token[2:]\r\n if token.endswith(\"</w>\"):\r\n token = token[:-4]\r\n if len(cand_indexes) == 0:\r\n cand_indexes.append([i])\r\n else:\r\n cand_indexes[-1].append(i)\r\n\r\n if token.endswith(\"</w>\"):\r\n cand_indexes.append([])\r\n\r\n if len(cand_indexes[-1]) == 0:\r\n cand_indexes = cand_indexes[:-1]\r\n\r\n random.shuffle(cand_indexes)\r\n num_to_predict = min(max_predictions, max(1, int(round(len(input_tokens) * self.mlm_probability))))\r\n masked_lms = []\r\n covered_indexes = set()\r\n for index_set in cand_indexes:\r\n if len(masked_lms) >= num_to_predict:\r\n break\r\n # If adding a whole-word mask would exceed the maximum number of\r\n # predictions, then just skip this candidate.\r\n if len(masked_lms) + len(index_set) > num_to_predict:\r\n continue\r\n is_any_index_covered = False\r\n for index in index_set:\r\n if index in covered_indexes:\r\n is_any_index_covered = True\r\n break\r\n if is_any_index_covered:\r\n continue\r\n for index in index_set:\r\n covered_indexes.add(index)\r\n masked_lms.append(index)\r\n\r\n if len(covered_indexes) != len(masked_lms):\r\n raise ValueError(\"Length of covered_indexes is not equal to length of masked_lms.\")\r\n mask_labels = [1 if i in covered_indexes else 0 for i in range(len(input_tokens))]\r\n return mask_labels\r\n```", "If anybody else has this issue. I fixed it for RoBERTa by adding a few lines that deal with the way RoBERTa tokenizes. Note that it's not a general purpose solution for other LMs. The previous comment did not work for me. See here:\r\n\r\nhttps://github.com/RikVN/transformers/blob/main/src/transformers/data/data_collator.py#L948" ]

[]

[]

[ "OK, I can reproduce the problem with just google/mt5-small and 2 gpus:\r\n```\r\nexport BS=1; PYTHONPATH=../../src USE_TF=0 deepspeed --num_gpus=2 ./finetune_trainer.py --model_name_or_path google/mt5-small --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_eval --do_predict --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size 1 --per_device_train_batch_size 1 --predict_with_generate --eval_steps 25000 --sortish_sampler --task translation_en_to_ro --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 5 --n_train 60 --n_val 10 --n_test 10 --deepspeed ds_config.json --fp16\r\n```\r\n\r\nWe will get it sorted out today.", "ok, the problem had nothing to do with DeepSpeed, it's just a seq2seq neglect. \r\n\r\nThe fix is:\r\n\r\n```\r\ndiff --git a/examples/seq2seq/utils.py b/examples/seq2seq/utils.py\r\nindex 8b24bfda..303b89f7 100644\r\n--- a/examples/seq2seq/utils.py\r\n+++ b/examples/seq2seq/utils.py\r\n@@ -563,7 +563,7 @@ def freeze_embeds(model):\r\n \"\"\"Freeze token embeddings and positional embeddings for bart, just token embeddings for t5.\"\"\"\r\n model_type = model.config.model_type\r\n\r\n- if model_type == \"t5\":\r\n+ if model_type in [\"t5\", \"mt5\"]:\r\n freeze_params(model.shared)\r\n for d in [model.encoder, model.decoder]:\r\n freeze_params(d.embed_tokens)\r\n```\r\n\r\nPlease let me know if you can manage to apply this fix. I will make a proper PR later, but it'll take some work, since I need to make a tiny mt5 model and add a test.\r\n\r\nYou can just edit the file if you don't know how to apply a patch. ", "The fix should be merged shortly https://github.com/huggingface/transformers/pull/9879\r\n", "I can solve the `--freeze_embeds` bug now, thanks for your help! @stas00 \r\n\r\nAs for questions 3 and 4, I noticed that the title of the issue has been edited. I don't know if these questions are caused by the model or the seq2seq trainer. Maybe I should raise them in a new issue?", "Oh, you wrote those items as steps to reproduce the problem, so I didn't know that those were issues that needed to/could be fixed. \r\n\r\nOnce I discovered that the issue you posted was unrelated to DeepSpeed I took the liberty to adjust the subject.\r\n\r\nIn general, yes, let's try to keep each issue separate, so that it makes it much easier to track things and not let things fall between the cracks.\r\n\r\nBack to your follow up question:\r\n\r\nLooking just at the params:\r\n\r\n- t5-3b ~10GB\r\n- mt5-xl ~15GB\r\n\r\nSo the 2nd model is substantially larger, and if t5-3b fit tightly onto a 24GB card it's not surprising that the larger model didn't. \r\n\r\nand in addition to model params you also need to allocate memory for:\r\n- inputs\r\n- gradients\r\n- optimizer states\r\n\r\n\r\nI tried mt5-xl on 4x 40gb gpu setup and it worked, but took ~29GB on each GPU, so there is the problem - you're 5GB short.\r\n\r\nThe command I run was:\r\n```\r\nexport BS=1; PYTHONPATH=../../src USE_TF=0 /usr/bin/time -v deepspeed --num_gpus=4 ./finetune_trainer.py --model_name_or_path google/mt5-xl --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_eval --do_predict --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size 1 --per_device_train_batch_size 1 --predict_with_generate --eval_steps 25000 --sortish_sampler --task translation_en_to_ro --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 5 --n_train 60 --n_val 10 --n_test 10 --deepspeed ds_config.json --fp16\r\n```\r\n\r\nYou may try to tweak the buffer sizes in `ds_config.json` but I think the gap is too big. \r\n\r\nI'm working on a 2D Parallelism solution that will combine pipe|model-parallelism w/ ZeRO-DP (DeepSpeed), which should enable such feats with huge models, but it might take some time. The docs aren't quite there so it takes a lot of trial and error to move forward. You may want to track this PR https://github.com/huggingface/transformers/pull/9765 for updates.\r\n\r\nAlternatively when fairscale or DeepSpeed releases ZeRO phase 3, you shouldn't have a problem loading this model onto 4x 24GB gpus. Currently the problem is that the model params are too big w/o phase 3. In phase 3 params are partitioned too - problem solved.\r\n", "> I tried mt5-xl on 4x 40gb gpu setup and it worked, but took ~29GB on each GPU, so there is the problem - you're 5GB short.\r\n\r\nThat's help a lot! Thank you!\r\n\r\nI am also looking forward to ZeRO stage 3 and your pipe|model-parallelism. Hope one day we can working on it. Thank you again!", "> And I got the overflow problem. This is not surprising me because MT5-large seems not fixed FP16 yet.\r\n\r\nDid you get `nan` loss or gradient overflow warning ? And yes, fp16 is still not working for mT5-large\r\n\r\n> I assume that T5-3b and MT5-xl should be in the same order of magnitude\r\n\r\nmT5-xl is actually quite bigger than T5-3b for two reasons\r\n1. It's vocab_size is huge (250112), which results in bigger token_embedding layer and final linear layer.\r\n2. It's based on t51.1 which uses `gated-gelu` activation instead of `relu`. `gated-gelu` adds one extra linear layer in every feed-forward layer.", "@patil-suraj That's very helpful! Thank you a lot!\r\n\r\nNow I understand that there are many differences between mT5-xl and T5-3b, and I will set up separate experiments for them in the future. By the way, do you have any plans to repair the FP16 in mt5-large/xl ?", "Dear @patil-suraj, here you have mentioned for mt5-small you have made it work with fp16? since you did not mention this model, do you mind telling me how you made it work? I am having a hard time with mt5-small with fp16 thanks a lot for your advice ", "I have a similar error here\r\n\r\n```python\r\nfrom transformers import T5TokenizerFast, MT5ForConditionalGeneration\r\n\r\ntokenizer = T5TokenizerFast.from_pretrained('google/mt5-base') # \"google/mt5-base\" \"google/mt5-large\" \"google/mt5-xl\"\r\n\r\nmodel = MT5ForConditionalGeneration.from_pretrained('google/mt5-base', return_dict=True)\r\n\r\ncondition = \"translate English to German: \"\r\ninput = \"My name is Azeem and I live in India\"\r\n\r\n# You can also use \"translate English to French\" and \"translate English to Romanian\"\r\ninput_ids = tokenizer(condition+input, return_tensors=\"pt\").input_ids # Batch size 1\r\n\r\noutputs = model.generate(input_ids)\r\n\r\ndecoded = tokenizer.decode(outputs[0], skip_special_tokens=True)\r\n\r\nprint(decoded)\r\n```\r\n\r\n\r\nStacktrace:\r\n```\r\n---------------------------------------------------------------------------\r\nAttributeError Traceback (most recent call last)\r\n[<ipython-input-8-f9822d331a70>](https://localhost:8080/#) in <module>()\r\n 3 tokenizer = T5TokenizerFast.from_pretrained('google/mt5-base') # \"google/mt5-base\" \"google/mt5-large\" \"google/mt5-xl\"\r\n 4 \r\n----> 5 model = AutoModelForSeq2SeqLM.from_pretrained('google/mt5-base', return_dict=True)\r\n 6 \r\n 7 condition = \"translate English to German: \"\r\n\r\n8 frames\r\n[/usr/local/lib/python3.7/dist-packages/transformers/configuration_utils.py](https://localhost:8080/#) in __getattribute__(self, key)\r\n 250 if key != \"attribute_map\" and key in super().__getattribute__(\"attribute_map\"):\r\n 251 key = super().__getattribute__(\"attribute_map\")[key]\r\n--> 252 return super().__getattribute__(key)\r\n 253 \r\n 254 def __init__(self, **kwargs):\r\n\r\nAttributeError: 'MT5Config' object has no attribute 'relative_attention_max_distance'\r\n```\r\n\r\n@stas00 any idea? I'm using HF master:\r\n```\r\n!pip install git+https://github.com/huggingface/transformers.git\r\n```", "@loretoparisi \r\n\r\nThis is because T5Config now has `relative_attention_max_distance` attribute introduced in the #16155 which was missing from `MT5Config`. Fix is here #16170\r\n" ]