Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[ "What task did you fine-tuned it on?\r\nYou can convert it by running the `convert_xlnet_checkpoint_to_pytorch.py` script with a `--finetuning_task` argument (see [here](https://github.com/huggingface/pytorch-transformers/blob/master/pytorch_transformers/convert_xlnet_checkpoint_to_pytorch.py#L94-L97))\r\n", "binary classification (sentiment) on my dataset\r\nI tried the following\r\n`import torch\r\nfrom pytorch_transformers import XLNetConfig\r\nfrom pytorch_transformers import XLNetTokenizer\r\nfrom pytorch_transformers import XLNetModel\r\nconfig = XLNetConfig.from_pretrained('./')\r\ntokenizer = XLNetTokenizer.from_pretrained('./')\r\nmodel = XLNetModel(config)\r\n\r\ninput_ids = torch.tensor(tokenizer.encode(\"Apple stocks increase\")).unsqueeze(0) # Batch size 1\r\noutputs = model(input_ids)\r\nlast_hidden_states = outputs[0]`\r\nhowever, the output is not binary, the output tensor is :\r\ntensor([[[-1.0645, -1.1170, -0.5549, ..., 0.7089, 0.2862, -0.2995],\r\n [ 0.0198, -1.6598, -0.1760, ..., 0.1167, -0.1545, 0.1777],\r\n [ 0.6514, -0.5614, -1.2180, ..., 1.0796, 1.1217, 1.0362]]],\r\n grad_fn=<PermuteBackward>)\r\n\r\n", "the converting was done using:\r\n`pytorch_transformers xlnet $TRANSFO_XL_CHECKPOINT_PATH $TRANSFO_XL_CONFIG_PAH $PYTORCH_DUMP_OUTPUT . `\r\n\r\nI did not specified the task, because I though that it is not important.", "It is important to match the last (classification) layer otherwise the conversion will fail.\r\nIf you give me more information on your TF training (like which script you usde for instance) I may be able to help.", "I was using the:\r\n`train_command = \"python xlnet/run_classifier.py \\\r\n --do_train=True \\\r\n --do_eval=True \\\r\n --eval_all_ckpt=True \\\r\n --task_name=spam \\\r\n --data_dir=\"+DATA_DIR+\" \\\r\n --output_dir=\"+OUTPUT_DIR+\" \\\r\n --model_dir=\"+CHECKPOINT_DIR+\" \\\r\n --uncased=False \\\r\n --spiece_model_file=\"+PRETRAINED_MODEL_DIR+\"/spiece.model \\\r\n --model_config_path=\"+PRETRAINED_MODEL_DIR+\"/xlnet_config.json \\\r\n --init_checkpoint=\"+PRETRAINED_MODEL_DIR+\"/xlnet_model.ckpt \\\r\n --max_seq_length=128 \\\r\n --train_batch_size=8 \\\r\n --eval_batch_size=8 \\\r\n --num_hosts=1 \\\r\n --num_core_per_host=1 \\\r\n --learning_rate=2e-5 \\\r\n --train_steps=4000 \\\r\n --warmup_steps=500 \\\r\n --save_steps=500 \\\r\n --iterations=500\"\r\n\r\n! {train_command}`\r\nIt is similar to the imdb sentiment task\r\nThe colab is here\r\nhttps://colab.research.google.com/drive/1nfWEEDxPOE8myb-hwGdoXs9nVXM3AVcz", "I have modified the:\r\nclass ImdbProcessor(DataProcessor):", "Here is the model after model.eval():\r\n`XLNetModel(\r\n (word_embedding): Embedding(32000, 1024)\r\n (layer): ModuleList(\r\n (0): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (1): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (2): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (3): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (4): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (5): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (6): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (7): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (8): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (9): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (10): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (11): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (12): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (13): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (14): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (15): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (16): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (17): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (18): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (19): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (20): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (21): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (22): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (23): XLNetLayer(\r\n (rel_attn): XLNetRelativeAttention(\r\n (layer_norm): XLNetLayerNorm()\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (ff): XLNetFeedForward(\r\n (layer_norm): XLNetLayerNorm()\r\n (layer_1): Linear(in_features=1024, out_features=4096, bias=True)\r\n (layer_2): Linear(in_features=4096, out_features=1024, bias=True)\r\n (dropout): Dropout(p=0.1)\r\n )\r\n (dropout): Dropout(p=0.1)\r\n )\r\n )\r\n (dropout): Dropout(p=0.1)\r\n)\r\n`", "You'll have to modify the conversion script because this is not a standard task and the number of labels won't be found in the list (see top of the conversion script) you should add `'spam': 2` in the list (if you have two labels indeed).\r\n\r\nThe other option is to directly load the TF model in PyTorch and save the pytorch model afterwards with something like this:\r\n```\r\nconfig = XLNetConfig.from_pretrained('xlnet-large-cased', num_labels=2, finetuning_task='spam')\r\nmodel = XLNetForSequenceClassification.from_pretrained('path/to/your/tf/model.ckpt.index', config=config, from_tf=True)\r\nmodel.save_pretrained('pytorch_model_saving_directory')\r\n```", "Perfect, tnx.\r\nJust a minor question if I have classification model with 4 labels, I do the same changes spam:4?", "Yes, change my `2` to `4`", "Hi Thom,\r\nI did:\r\n`import torch\r\nfrom pytorch_transformers import XLNetConfig\r\nfrom pytorch_transformers import XLNetForSequenceClassification\r\nconfig = XLNetConfig.from_pretrained('xlnet-large-cased', num_labels=2, finetuning_task='sentiment')\r\nmodel = XLNetForSequenceClassification.from_pretrained('../model.ckpt-2500', config=config, from_tf=True)\r\nmodel.save_pretrained('./test')`\r\n\r\nAfterwards I get in the test folder the config.json and pytorch_model.bin.\r\n\r\nI tried to run:\r\n'import torch\r\nfrom pytorch_transformers import XLNetConfig\r\nfrom pytorch_transformers import XLNetTokenizer\r\nfrom pytorch_transformers import XLNetModel\r\nconfig = XLNetConfig.from_pretrained('./')\r\ntokenizer = XLNetTokenizer.from_pretrained('./')\r\nmodel = XLNetModel(config)\r\n\r\ninput_ids = torch.tensor(tokenizer.encode(\"Apple stocks increase\")).unsqueeze(0) # Batch size 1\r\noutputs = model(input_ids)\r\nlast_hidden_states = outputs[0]\r\nprint(last_hidden_states)\r\n\r\n\r\nThere was an error:\r\nModel name './' was not found in model name list (xlnet-base-cased, xlnet-large-cased). We assumed './' was a path or url but couldn't find tokenizer filesat this path or url.\r\n\r\nSo I included the spiece.model into the same test folder, but still the results is:\r\ntensor([[[ 0.0156, -0.9046, 0.9789, ..., -0.7764, 1.1309, 0.2862],\r\n [ 0.4416, 0.0665, 2.0020, ..., 0.4117, 0.9779, -1.0588],\r\n [ 0.6908, -0.0000, 0.1479, ..., 0.0032, 0.9871, -0.9482]]],\r\n grad_fn=<PermuteBackward>)\r\n\r\nDo u know what I am doing wrong? ", "While converting this is some of the warnings:\r\nWeights not copied to PyTorch model: beta1_power, beta2_power, global_step, model/classification_finsent/logit/bias, model/classification_finsent/logit/bias/Adam, model/classification_finsent/logit/bias/Adam_1, model/classification_finsent/logit/kernel, model/classification_finsent/logit/kernel/Adam, model/classification_finsent/logit/kernel/Adam_1", "In addition, I have tried the other approach, to modify the script, but the results are the same. I think that the problem is that the weights from the classification_finsent are not copied to pytorch model. Any suggestions on this", "I have tried also in this way:\r\n`import torch\r\nfrom pytorch_transformers import XLNetConfig\r\nfrom pytorch_transformers import XLNetTokenizer\r\nfrom pytorch_transformers import XLNetForSequenceClassification\r\n\r\nconfig = XLNetConfig.from_pretrained('/home/igor/Falcon/XLNet/sentiment_model/PyTorch/pytorch_transformer_script/test')\r\ntokenizer = XLNetTokenizer.from_pretrained('/home/igor/Falcon/XLNet/sentiment_model/PyTorch/pytorch_transformer_script/test')\r\nconfig.output_hidden_states=True\r\nmodel = XLNetForSequenceClassification(config)\r\n\r\ninput_ids = torch.tensor(tokenizer.encode(\"Apple stocks increase rapidly\")).unsqueeze(0) # Batch size 1\r\nlabels = torch.tensor([1]).unsqueeze(0) # Batch size 1\r\noutputs = model(input_ids, labels=labels)\r\nprint(outputs[0])`\r\n\r\nand the output now is:\r\ntensor(0.3944, grad_fn=<NllLossBackward>)\r\n", "In the last attempt I did:\r\n```python\r\nimport torch\r\nfrom pytorch_transformers import XLNetConfig\r\nfrom pytorch_transformers import XLNetTokenizer\r\nfrom pytorch_transformers import XLNetForSequenceClassification\r\n\r\nconfig = XLNetConfig.from_pretrained('xlnet-large-cased', num_labels=2, finetuning_task='finsent')\r\ntokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased') #I am using the default tokenizer\r\n\r\nmodel = XLNetForSequenceClassification.from_pretrained('/home/igor/Falcon/XLNet/sentiment_model/model.ckpt-2500',config=config, from_tf=True) #this is my finetuned model\r\nmodel.eval() #in evaluation mode\r\ninput_ids = torch.tensor(tokenizer.encode(\"Apple stock increase and they are overwhelmed with the success\")).unsqueeze(0) # Batch size 1\r\nlabels = torch.tensor([1]).unsqueeze(0) # Batch size 1\r\noutputs = model(input_ids, labels=labels)\r\nloss, logits = outputs[:2]\r\nprint(torch.nn.functional.softmax(logits.data))\r\n```\r\n\r\nHowever I am getting different prediction for the same input data, I did model.eval() to stop the dropout, but still the inference looks random.\r\n\r\nIn addition here is my config file printed:\r\n```json\r\n{\r\n \"attn_type\": \"bi\",\r\n \"bi_data\": false,\r\n \"clamp_len\": -1,\r\n \"d_head\": 64,\r\n \"d_inner\": 4096,\r\n \"d_model\": 1024,\r\n \"dropatt\": 0.1,\r\n \"dropout\": 0.1,\r\n \"end_n_top\": 5,\r\n \"ff_activation\": \"gelu\",\r\n \"finetuning_task\": \"finsent\",\r\n \"init\": \"normal\",\r\n \"init_range\": 0.1,\r\n \"init_std\": 0.02,\r\n \"initializer_range\": 0.02,\r\n \"layer_norm_eps\": 1e-12,\r\n \"max_position_embeddings\": 512,\r\n \"mem_len\": null,\r\n \"n_head\": 16,\r\n \"n_layer\": 24,\r\n \"n_token\": 32000,\r\n \"num_labels\": 2,\r\n \"output_attentions\": false,\r\n \"output_hidden_states\": false,\r\n \"reuse_len\": null,\r\n \"same_length\": false,\r\n \"start_n_top\": 5,\r\n \"summary_activation\": \"tanh\",\r\n \"summary_last_dropout\": 0.1,\r\n \"summary_type\": \"last\",\r\n \"summary_use_proj\": true,\r\n \"torchscript\": false,\r\n \"untie_r\": true\r\n}\r\n```\r\nIt looks that I getting closer, but still the inference is strange", "@thomwolf I finally succeed to import the checkpoint model and infer. Still I am not sure if this a valid approach:\r\n```python\r\nimport torch\r\nfrom pytorch_transformers import XLNetConfig\r\nfrom pytorch_transformers import XLNetTokenizer\r\nfrom pytorch_transformers import XLNetForSequenceClassification\r\n\r\n\r\nseed = 0\r\ntorch.manual_seed(seed)\r\nif torch.cuda.is_available():\r\n\ttorch.cuda.manual_seed_all(seed)\r\n#load the initial config file from the XLNet model\r\nconfig = XLNetConfig.from_pretrained('xlnet_config.json', num_labels=2, finetuning_task='finsent')\r\n#the tokenizer I am using is the initial one (spiece.model)\r\ntokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')\r\n\r\nmodel = XLNetForSequenceClassification.from_pretrained('/sentiment_over_imdb/model.ckpt-4000',config=config, from_tf=True)\r\nmodel.eval()\r\n\r\ndef sentiment(data):\r\n\tinput_ids = torch.tensor(tokenizer.encode(data)).unsqueeze(0) # Batch size 1\r\n\tlabels = torch.tensor([1]).unsqueeze(0) # Batch size 1\r\n\r\n\toutputs = model(input_ids, labels=labels)\r\n\tloss, logits = outputs[:2]\r\n\toutput = torch.nn.functional.softmax(logits.data, dim=1)\r\n\tsent = output.tolist()[0][1]\r\n return sent\r\n```" ]

[]

[]

[ "Indeed, I will remove this doctring, there is no attention_mask on GPT-2.", "> Indeed, I will remove this doctring, there is no attention_mask on GPT-2.\r\n\r\nBut what to do if I do want to avoid computing attention on the paddings in the input sequences.", "@Saner3 @thomwolf I have same question? don't we need that for paddings?", "GPT-2 is a model with absolute position embeddings (like Bert) so you should always pad on the right to get best performances for this model (will add this information to the doc_string).\r\n\r\nAs it's a causal model (only attend to the left context), also means that the model will not attend to the padding tokens (which are on the right) for any real token anyway.\r\n\r\nSo in conclusion, no need to take special care of avoiding attention on padding.\r\n\r\nJust don't use the output of the padded tokens for anything as they don't contain any reliable information (which is obvious I hope).", "@thomwolf thanks much, and great job!" ]

[]

[]

[ "HI,\r\nI have the same problem!\r\nWhat was the solution here?", "Me too.", "Need more information like version of python/pytorch/transformers (all the information requested in the issue templates actually)", "> Need more information like version of python/pytorch/transformers (all the information requested in the issue templates actually)\r\n\r\nI am experiencing this issue as well with the BertForNextSentencePrediction model and not having much luck with a solution. I'm using macOS Mojave 10.14.6, python 3.7, pytorch 1.3.1 and transformers 2.2.1. \r\n\r\nPlease let me know if there is any more details I can provide. Thanks!", "You should open a new issue with a clean code example we can test and the associate full error message." ]

[]

[]

[ "Ok for that, thanks @wschin" ]

[]

[]

[ "It's now `run_glue.py`", "Hi @thomwolf \r\nDoc needs changes from run_bert_classifier to run_glue\r\nhttps://huggingface.co/pytorch-transformers/examples.html", "Hey, just another headsup @thomwolf \r\nThis Doc also needs changing for the run_bert_classifier.py:\r\nhttps://huggingface.co/transformers/v1.1.0/examples.html#introduction", "That's an old version of the doc @mtwright (notice the `v1.1.0`), you should check out the up to date one:\r\n\r\nhttps://huggingface.co/transformers/examples.html#introduction", "documentation is still broken, by the way", "like, here, points to a bunch of files that do not exist. also not quite sure if these instructions work anymore:\r\n\r\nhttps://huggingface.co/transformers/converting_tensorflow_models.html" ]

[ { "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 because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "What is the full line of the assertion you are testing?\r\nAnd what is your input text?", "I think the mistake was with me, sorry" ]

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

[]

[ "XLNet makes heavy use of `torch.einsum()` but I'm not sure this method is fp16 compatible.\r\nIt's also quite slow currently so maybe in the mid/long-term it would be good to change these einsum to standard matmul. I won't have time to do that very soon though.", "As as a suggestion, you can add ```apex.amp.register_half_function(torch, 'einsum')``` somewhere near the top of your driver script (examples/run_squad.py for instance).\r\n\r\nThis forces `amp` to cast the inputs to einsum to `torch.half` before executing, allowing you to get the perf benefits of fp16 + TensorCores when appropriate.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Hi, @slayton58, could `torch.einsum` be automatically processed by `apex.amp` now?\r\nYour response will be appreciated!" ]

[]

[]

[ "Thanks!" ]

[]

[]

[ "I have same issue. ", "Thanks @seanie12!" ]

[]

[]

[ "Added:\r\n\r\nI also found that `_convert_token_to_id()` in tokenization_gpt2.py (line 182) uses `unk_token`, which is initially `None` in GPT2 tokenizer. This line of code can also lead to bugs.\r\n~~~~\r\ndef _convert_token_to_id(self, token):\r\n \"\"\" Converts a token (str/unicode) in an id using the vocab. \"\"\"\r\n if token in self.encoder:\r\n return self.encoder.get(token)\r\n return self.encoder.get(self.unk_token)\r\n~~~~", "Indeed, GPT-2 doesn't have a `unk_token` since it's supposed to be able to encode any string but this does have some unintended consequences since we also use the fact that a tokenizer returns the `unk_token` to check whether a token is in the current vocabulary or not.\r\n\r\nI'll see how we can update this in the most coherent way. Probably mapping the special `<|endoftext|>` token as `unk_token` and using the same logic as the other models (returning the `unk_token` when the token is not in the vocabulary) is the simplest way to fix it.", "What will be the quick fix for this? When running `run_generation.py` in examples, I could resolve the `None' error by adding special tokens to the tokenizer like below:\r\n\r\n` special_tokens = {\"cls_token\":\"[CLS]\", \"unk_token\":\"[UNK]\"} `\r\n` tokenier = tokenizer_class.from_pretrained(\"gpt2\", cls_token=\"[CLS]\", unk_token=\"[UNK]\")`\r\n` tokenizer.add_special_tokens(special_tokens) `\r\n\r\nBut then, I got a CUDA error (probably) due to the different embedding size of the model and tokenizer. So, I resized the model's token embedding size from 50257 to 50259 like below:\r\n\r\n` model.resize_token_embeddings(len(tokenizer)) `\r\n\r\nThen, it tokenizes the tokens correctly with the additional token encoders that have `tokenizer.added_tokens_encoder.keys()` with [CLS] and [UNK]. But, regardless of input, the gpt2 output seems to be wrong: a sequence [CLS] [CLS] ...\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n", "@dykang \r\nA quick fix would be cleaning the generated sentence by replacing unwanted tokens as \"\".\r\nHowever, it is reasonable that the GPT2 outputs weird sentences given inputs including \"[CLS]\" as in your example because the new word embeddings are not trained. ", "Thanks, @ZHAOTING. However, GH-910 seems to only add unk token though. @thomwolf, would the PR be generalizable to add any special tokens as my earlier comment above? When special tokens are added, how do existing pre-trained gpt2 models work properly? ", "Adding some context. Is it possible to add [CLS] and [SEP] tokens to gpt2-medium in a non destructive way. After finetuning a bit following the structure indicated here: \r\n```\r\n # (Default, BERT/XLM pattern): [CLS] + A + [SEP] + B + [SEP]\r\n # (XLNet/GPT pattern): A + [SEP] + B + [SEP] + [CLS]\r\n```\r\nit is clear that GPT2 will attempt to predict the BPE tokens for [CLS] as ` [C LS]`, but adding the [SEP] and [CLS] tokens produce only an output of [CLS] / [SEP] tokens despite any top_k, top_p, or temperature settings.\r\n\r\n\r\n", "I had a similar problem.\r\n\r\nI tried to finetune a gpt2 model using the simpletransformers library. \r\n\r\nThe error seems to originate from\r\n\"/usr/local/lib/python3.6/site-packages/transformers/optimization.py\"\r\nThis is the file where the optimizer AdamW is implemented.\r\n\r\nIn my case, I made a backup of the \r\n\"scheduler.pt\" and \"optimizer.pt\" \r\nfiles in my saved checkpoint.\r\n\r\nIn file \r\n\"/usr/local/lib/python3.6/site-packages/simpletransformers/language_modeling/language_modeling_model.py\"\r\nthe optimizer and scheduler were also loaded from the checkpoint but this broke the code.\r\nIf these two files cannot be found in the path then your code will proceed with the rest of it.\r\n\r\nThis however remains a bug as your new optimizer and scheduler will start from scratch ignoring any \"knowledge\" they already contained about your previous optimization.\r\n\r\n " ]

[]

[]

[ "Changing that line still causes an error in line 568, BertAdam has to be changed to AdamW as well and the warmup kwarg has to be removed.", "#797 (specifically d6522e28732fd14a926440ef5f315e6a8e13792c) ", "have fix the error! I tested it on toy dataset.", "@shibing624 this bug can be closed" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/pytorch-transformers/pull/797?src=pr&el=h1) Report\n> Merging [#797](https://codecov.io/gh/huggingface/pytorch-transformers/pull/797?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/pytorch-transformers/commit/5fe0b378d899f81eb0a7f2db0c4eb0234748e915?src=pr&el=desc) will **not change** coverage.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/pytorch-transformers/pull/797?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #797 +/- ##\n=======================================\n Coverage 78.91% 78.91% \n=======================================\n Files 34 34 \n Lines 6193 6193 \n=======================================\n Hits 4887 4887 \n Misses 1306 1306\n```\n\n\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-transformers/pull/797?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-transformers/pull/797?src=pr&el=footer). Last update [5fe0b37...a7ba27b](https://codecov.io/gh/huggingface/pytorch-transformers/pull/797?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Hi, do you want to update these examples to the new `pytorch-transformers` API at the same time?\r\nModels now return `tuple` so we should take the first element of the model output as the loss and we should also update `BertAdam` to `AdamW`.", "ok， I have a try, I‘m stilll use the old API. ", "Hi, I have updated the opt to the new API , please check. ", "Thanks @yzy5630!" ]

[]

[]

[ "Thanks Stefan!" ]

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

[]

[ "I also tried to finetune xlnet base on squad 2.0 but the numbers on dev are pretty bad\r\n`Results: {'exact': 3.0405120862461046, 'f1': 6.947601433150003, 'total': 11873, 'HasAns_exact': 6.056005398110662, 'HasAns_f1': 13.881388632893048, 'HasAns_total': 5928, 'NoAns_exact': 0.0336417157275021, 'NoAns_f1': 0.0336417157275021, 'NoAns_total': 5945, 'best_exact': 50.07159100480081, 'best_exact_thresh': 0.0, 'best_f1': 50.07159100480081, 'best_f1_thresh': 0.0}`", "I suspect something is wrong with the evaluation code. Looking into it now.", "@tbright17 Nothing wrong with evaluation. Accuracy and evaluation loss aren't changed during training. I used my own evaluation script, I used old BertAdam or OpenAIAdam optimizers without success.\r\n@thomwolf Can you help?", "I'll give a look, I've only tested XLNet on STS-B for the moment. You should check the hyper-parameters as well, they probably won't be the same as the ones of STS-B (some are mentioned in the XLNet paper).", "First thing that comes to mind is that SST-2 is ~10 times bigger than STS-B (see the [GLUE paper](https://arxiv.org/abs/1804.07461)) so you need to increase the number of training step a lot if you want to do at least one full epoch on SST-2 training dataset (here you use the value for STS-B). And you should probably do several epochs, e.g. we do 6-7 epochs on STS-B). Check some examples of recommended hyper-parameters table 8 of the [xlnet paper](http://arxiv.org/abs/1906.08237).\r\n\r\nYou can also directly specify the number of epochs instead of the maximum number of steps in the script. You can see all the hyper-parameters of the script with `python ./run_glue.py --help`.", "> First thing that comes to mind is that SST-2 is ~10 times bigger than STS-B (see the [GLUE paper](https://arxiv.org/abs/1804.07461)) so you need to increase the number of training step a lot if you want to do at least one full epoch on SST-2 training dataset (here you use the value for STS-B). And you should probably do several epochs, e.g. we do 6-7 epochs on STS-B). Check some examples of recommended hyper-parameters table 8 of the [xlnet paper](http://arxiv.org/abs/1906.08237).\r\n> \r\n> You can also directly specify the number of epochs instead of the maximum number of steps in the script. You can see all the hyper-parameters of the script with `python ./run_glue.py --help`.\r\n\r\nI trained STS-B task with the same problem. You can see the following output with evaluation of every 100 steps (I added train and evaluation loss in output):\r\n\r\n```\r\n07/17/2019 13:09:55 - INFO - __main__ - ***** Running evaluation *****\r\n07/17/2019 13:09:55 - INFO - __main__ - Num examples = 1500\r\n07/17/2019 13:09:55 - INFO - __main__ - Batch size = 8\r\n07/17/2019 13:10:09 - INFO - __main__ - ***** Eval results *****\r\n07/17/2019 13:10:09 - INFO - __main__ - corr = -0.05367882385720809\r\n07/17/2019 13:10:09 - INFO - __main__ - eval_loss = 2.8412214481133096##################################################################################################################| 188/188 [00:14<00:00, 13.41it/s]\r\n07/17/2019 13:10:09 - INFO - __main__ - pearson = -0.041275192\r\n07/17/2019 13:10:09 - INFO - __main__ - spearmanr = -0.06608245566229025\r\n07/17/2019 13:10:09 - INFO - __main__ - Training loss: 307.258519500494\r\n 07/17/2019 13:10:41 - INFO - __main__ - Loading features from cached file ...glue_data/STS-B/cached_dev_xlnet-large-cased_128_sts-b | 199/719 [01:18<03:25, 2.53it/s]\r\n07/17/2019 13:10:41 - INFO - __main__ - ***** Running evaluation *****\r\n07/17/2019 13:10:41 - INFO - __main__ - Num examples = 1500\r\n07/17/2019 13:10:41 - INFO - __main__ - Batch size = 8\r\n07/17/2019 13:10:56 - INFO - __main__ - ***** Eval results *****\r\n07/17/2019 13:10:56 - INFO - __main__ - corr = 0.13943037650184956\r\n07/17/2019 13:10:56 - INFO - __main__ - eval_loss = 2.3762524007482733##################################################################################################################| 188/188 [00:14<00:00, 13.29it/s]\r\n07/17/2019 13:10:56 - INFO - __main__ - pearson = 0.13502572\r\n07/17/2019 13:10:56 - INFO - __main__ - spearmanr = 0.1438350282350605\r\n07/17/2019 13:10:56 - INFO - __main__ - Training loss: 533.9101385176182\r\n 07/17/2019 13:11:28 - INFO - __main__ - Loading features from cached file .../glue_data/STS-B/cached_dev_xlnet-large-cased_128_sts-b | 299/719 [02:05<02:56, 2.39it/s]\r\n07/17/2019 13:11:28 - INFO - __main__ - ***** Running evaluation *****\r\n07/17/2019 13:11:28 - INFO - __main__ - Num examples = 1500\r\n07/17/2019 13:11:28 - INFO - __main__ - Batch size = 8\r\n07/17/2019 13:11:42 - INFO - __main__ - ***** Eval results *****\r\n07/17/2019 13:11:42 - INFO - __main__ - corr = -0.0830871973267994\r\n07/17/2019 13:11:42 - INFO - __main__ - eval_loss = 2.5565993221516305##################################################################################################################| 188/188 [00:14<00:00, 13.20it/s]\r\n07/17/2019 13:11:42 - INFO - __main__ - pearson = -0.08915693\r\n07/17/2019 13:11:42 - INFO - __main__ - spearmanr = -0.077017461524765\r\n07/17/2019 13:11:42 - INFO - __main__ - Training loss: 761.6802722513676\r\n 07/17/2019 13:12:15 - INFO - __main__ - Loading features from cached file .../glue_data/STS-B/cached_dev_xlnet-large-cased_128_sts-b | 399/719 [02:52<02:18, 2.32it/s]\r\n07/17/2019 13:12:15 - INFO - __main__ - ***** Running evaluation *****\r\n07/17/2019 13:12:15 - INFO - __main__ - Num examples = 1500\r\n07/17/2019 13:12:15 - INFO - __main__ - Batch size = 8\r\n07/17/2019 13:12:29 - INFO - __main__ - ***** Eval results *****\r\n07/17/2019 13:12:29 - INFO - __main__ - corr = -0.08715267932681456\r\n07/17/2019 13:12:29 - INFO - __main__ - eval_loss = 2.398741365113157###################################################################################################################| 188/188 [00:14<00:00, 13.12it/s]\r\n07/17/2019 13:12:29 - INFO - __main__ - pearson = -0.08428703\r\n07/17/2019 13:12:29 - INFO - __main__ - spearmanr = -0.09001832616862088\r\n07/17/2019 13:12:29 - INFO - __main__ - Training loss: 974.8287971913815\r\n```\r\n\r\nHow you can see training loss is increasing, eval loss is almost the same, other metrics fluctuate around 0.", "@thomwolf So, it looks like training is happening but in opposite direction for some reason", "Maybe you haven't fully read the [explanation](https://github.com/huggingface/pytorch-transformers#fine-tuning-xlnet-model-on-the-sts-b-regression-task) accompanying the STS-B example in the readme?\r\n\r\nIt says \"On this machine we thus have a batch size of 32, please increase `gradient_accumulation_steps` to reach the same batch size if you have a smaller machine.\"", "@avostryakov Did you try to reduce the learning rate? I had a similar issue training with the TensorFlow version XLNet on only one GPU. I tried reducing the learning rate from 5e-5 to 1e-5, and it worked. Wish this can help you.", "@thomwolf @tbright17 I got similar numbers like you Squad 2.0. Seems that the model probably isn't learning much. I'll print out the losses to explore. Also should we change the LR as well? \r\n : the best I got with fine-tuning on Squad 2.0 with a `train_batch_size=8` and `gas=1` all others are default on a single v100 gpu was:\r\n `07/16/2019 16:21:43 - INFO - __main__ - Results: {'exact': 26.438136949380947, 'f1': 28.470459931964722, 'total': 11873, 'HasAns_exact': 0.08434547908232119, 'HasAns_f1': 4.154819630940996, 'HasAns_total': 5928, 'NoAns_exact': 52.716568544995795, 'NoAns_f1': 52.716568544995795, 'NoAns_total': 5945, 'best_exact': 50.07159100480081, 'best_exact_thresh': 0.0, 'best_f1': 50.07159100480081, 'best_f1_thresh': 0.0}`", "May also be a problem of batch size, the authors use a batch size between 32 and 128 in the paper. \r\n\r\nWhat effective batch size do you have (printed during training)?\r\n\r\nWhile we reproduce the official XLNet number on STS-B, I still have to work a bit on the SQuAD example for XLNet, the XLNet authors used a complex pre- and post-processing of the data (smarter than Bert's) that I haven't fully integrated into our `run_squad` example yet.", "> Maybe you haven't fully read the [explanation accompanying the STS-B example in the readme](https://github.com/huggingface/pytorch-transformers#fine-tuning-xlnet-model-on-the-sts-b-regression-task)?\r\n> \r\n> It says \"On this machine we thus have a batch size of 32, please increase `gradient_accumulation_steps` to reach the same batch size if you have a smaller machine.\"\r\n\r\n@thomwolf You are right, STS-B started to train with batch size 32 and gradient_accumulation_steps = 2. Now I'm wondering why it so heavily depends on batch size. But it doesn't help for STS-2, I set max_steps=5000 (it's 5 epochs) and training and evaluation loss didn't change at all during training. I'm trying to train with learning rate 1e-5 how it was recommended by @alexpython1988 ", "@thomwolf maybe. Also my sequence length is `384`: the authors did mention they prolly did 512. Here's my batch size related printout: I think the number of examples seem a lil low. No? I think Squad has about 150K examples (ha and na questions) and with the `doc_stride` I think it should be more than 150k examples (I think).\r\n\r\n`07/15/2019 13:23:32 - INFO - __main__ - ***** Running training *****`\r\n`07/15/2019 13:23:32 - INFO - __main__ - Num examples = 133947`\r\n`07/15/2019 13:23:32 - INFO - __main__ - Num Epochs = 3`\r\n`07/15/2019 13:23:32 - INFO - __main__ - Instantaneous batch size per GPU = 4`\r\n`07/15/2019 13:23:32 - INFO - __main__ - Total train batch size (w. parallel, distributed & accumulation) = 4`\r\n`07/15/2019 13:23:32 - INFO - __main__ - Gradient Accumulation steps = 1`\r\n`07/15/2019 13:23:32 - INFO - __main__ - Total optimization steps = 100461`\r\n\r\n\r\nI saw in the [renatoviolin's repo](https://github.com/renatoviolin/xlnet/blob/master/run_squad_GPU.py) that they have the following which gives them `86F1` on a RTX2080: \r\n`flags.DEFINE_integer(\"max_seq_length\",\r\n default=512, help=\"Max sequence length\")\r\nflags.DEFINE_integer(\"max_query_length\",\r\n default=64, help=\"Max query length\")\r\nflags.DEFINE_integer(\"doc_stride\",\r\n default=128, help=\"Doc stride\")\r\nflags.DEFINE_integer(\"max_answer_length\",\r\n default=64, help=\"Max answer length\")`\r\n\r\nAlso, lr is different than ours (`5e-5` in this repo):\r\n`flags.DEFINE_float(\"learning_rate\", default=3e-5, help=\"initial learning rate\")`\r\n", "Learning rate = 1e-5 helps to train STS-2 together with batch size 32 and accumulation steps = 2. I need more experiments but it works. Thanks, @thomwolf, and @alexpython1988!", "Great to hear, good job and good luck @avostryakov! Feel free to share good hyper-parameters if you find a nice set and I can add them to the documentation (with credits).", "> May also be a problem of batch size, the authors use a batch size between 32 and 128 in the paper.\r\n> \r\n> What effective batch size do you have (printed during training)?\r\n> \r\n> While we reproduce the official XLNet number on STS-B, I still have to work a bit on the SQuAD example for XLNet, the XLNet authors used a complex pre- and post-processing of the data (smarter than Bert's) that I haven't fully integrated into our `run_squad` example yet.\r\n\r\nI was using per_gpu_train_batch 8 for squad 2.0. Powerful model is hard to tune maybe", "> Great to hear, good job and good luck @avostryakov! Feel free to share good hyper-parameters if you find a nice set and I can add them to the documentation (with credits).\r\n\r\n@thomwolf My the best result for SST-2 so far is 94.15 of accuracy (in xlnet's article 95.6). It's better than BERT-large. I trained with the following parameters:\r\n\r\n```\r\npython ./examples/run_glue.py \\\r\n --model_type xlnet \\\r\n --model_name_or_path xlnet-large-cased \\\r\n --do_train \\\r\n --evaluate_during_training \\\r\n --do_eval \\\r\n --logging_steps 500 \\\r\n --save_steps 3000 \\\r\n --task_name=sst-2 \\\r\n --data_dir=${GLUE_DIR}/SST-2 \\\r\n --output_dir=./proc_data/sst-2 \\\r\n --max_seq_length=128 \\\r\n --learning_rate 1e-5 \\\r\n --per_gpu_eval_batch_size=8 \\\r\n --per_gpu_train_batch_size=8 \\\r\n --gradient_accumulation_steps=1 \\\r\n --max_steps=16000 \\\r\n --model_name=xlnet-large-cased \\\r\n --overwrite_output_dir \\\r\n --overwrite_cache \\\r\n --warmup_steps=120 \\\r\n --fp16\r\n```", "@thomwolf Ok, the last result for SST-2 almost matched with XLNet article: Accuracy 95.4:\r\n\r\n```\r\npython ./examples/run_glue.py \\\r\n --model_type xlnet \\\r\n --model_name_or_path xlnet-large-cased \\\r\n --do_train \\\r\n --evaluate_during_training \\\r\n --do_eval \\\r\n --logging_steps 400 \\\r\n --save_steps 3000 \\\r\n --task_name=sst-2 \\\r\n --data_dir=${GLUE_DIR}/SST-2 \\\r\n --output_dir=./proc_data/sst-2 \\\r\n --max_seq_length=128 \\\r\n --learning_rate 1e-5 \\\r\n --per_gpu_eval_batch_size=16 \\\r\n --per_gpu_train_batch_size=16 \\\r\n --gradient_accumulation_steps=1 \\\r\n --max_steps=8000 \\\r\n --model_name=xlnet-large-cased \\\r\n --overwrite_output_dir \\\r\n --overwrite_cache \\\r\n --warmup_steps=120 \\\r\n --fp16\r\n```\r\n\r\nThank you for your work!", "This is great @avostryakov! Thanks for sharing the results!\r\nI'm editing the issue title until I've time to add the hyperparameters to the doc.", "Hi, how could I finetune the model for text generation? Is it possible just having raw text for the finetuning?", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "Hi Jason,\r\n\r\nCan you give me a little more information on the model-loading workflow you are using so I can understand the whys and wherefores of these proposed modifications?", "Hey Thomas,\r\n\r\nSorry for the delay. My thinking is this: the `from_pretrained` method current does two things: resolve the path/archive for loading a pretrained model, and the specialized model loading logic (e.g. handing the fact that the current model may have different heads from those in the loaded weights). My proposed change is to separate the two and allow the user to just do the second. \r\nThis would be useful in cases where the user already has access to the `state_dict` in memory (e.g. if they have a different model saving workflow/format).", "Hi Jason,\r\nThere is a `state_dict` option in `from_pretrained` that, I think, let you do just that!\r\nSee here for instance: https://huggingface.co/pytorch-transformers/main_classes/model.html#pytorch_transformers.PreTrainedModel.from_pretrained", "Closing this for now. Feel free to re-open if the provided solution doesn't solve your problem, Jason." ]

[]

[]

[ "Damned, missed that one, you are right.\r\nAdding the missing doc-string:\r\n```\r\n**pooler_output**: ``torch.FloatTensor`` of shape ``(batch_size, hidden_size)``\r\n Last layer hidden-state of the first token of the sequence (classification token)\r\n further processed by a Linear layer and a Tanh activation function. The Linear\r\n layer weights are trained from the next sentence prediction (classification)\r\n objective during Bert pretraining. This output is usually *not* a good summary\r\n of the semantic content of the input, you're often better with averaging or pooling\r\n the sequence of hidden-states for the whole input sequence.\r\n```\r\nWe'll probably do a small release in a few days once we have gathered all the feedbacks from the main release. In the meantime, I'll set up PyTorch-Hub so people can get the models from master.", "A minor edit with the final two optional outputs `hidden_states` and `attentions` are tuples, not lists.", "cc @LysandreJik :)", "The documentation is outdated regarding that issue. It should probably be re-compiled :-)" ]

[]

[]

[]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/pytorch-transformers/pull/791?src=pr&el=h1) Report\n> Merging [#791](https://codecov.io/gh/huggingface/pytorch-transformers/pull/791?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/pytorch-transformers/commit/b33a385091de604afb566155ec03329b84c96926?src=pr&el=desc) will **not change** coverage.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/pytorch-transformers/pull/791?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #791 +/- ##\n=======================================\n Coverage 78.91% 78.91% \n=======================================\n Files 34 34 \n Lines 6193 6193 \n=======================================\n Hits 4887 4887 \n Misses 1306 1306\n```\n\n\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-transformers/pull/791?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-transformers/pull/791?src=pr&el=footer). Last update [b33a385...9d381e7](https://codecov.io/gh/huggingface/pytorch-transformers/pull/791?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[]

[]

[ "I'm currently finishing to add the documentation but just use `XLNetModel` instead of `BertModel` in the usage example with `BertModel`", "Thanks a lot, @thomwolf for the quick reply. I'll try it out.", "Here is an example now: https://huggingface.co/pytorch-transformers/model_doc/xlnet.html#pytorch_transformers.XLNetModel", "@thomwolf, I tried the following snippet. The similarity score changes every time I run the cell. That is, the embeddings or the weights are changing every time. Is this related to dropout?\r\n\r\n\r\n```\r\nconfig = XLNetConfig.from_pretrained('xlnet-large-cased')\r\n\r\ntokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')\r\nmodel = XLNetModel(config)\r\ninput_ids = torch.tensor(tokenizer.encode(\"The apple juice is sour.\")).unsqueeze(0) \r\ninput_ids_2 = torch.tensor(tokenizer.encode(\"The orange juice is sweet.\")).unsqueeze(0) \r\n\r\noutputs = model(input_ids)\r\noutputs_2 = model(input_ids_2)\r\nlast_hidden_states = outputs[0] \r\nlast_hidden_states_2 = outputs_2[0]\r\n\r\napple = last_hidden_states[0][1]\r\norange = last_hidden_states_2[0][1]\r\n\r\nx = apple\r\ny = orange\r\ncos_sim = dot(x.detach().numpy(),y.detach().numpy())/(norm(x.detach().numpy())*norm(y.detach().numpy()))\r\nprint(cos_sim)\r\n\r\n```\r\n\r\n", "For me logits values changes as well ... using exactly the same settings as mentioned in the example.\r\n\r\nHave you found a way to fix that?", "@Oxi84 put `model.eval()` before you make the predictions. This fixed the problem of changing weights for me.", "Thanks. For me it works when call like that:\r\n\r\n tokenizer = XLNetTokenizer.from_pretrained(\"xlnet-large-cased\")\r\n model = XLNetLMHeadModel.from_pretrained(\"xlnet-large-cased\")\r\n model.eval()\r\n\r\nHowever accuracy seems to be much lower that for Bert - with the code i wrote here: https://github.com/huggingface/pytorch-transformers/issues/846\r\n\r\nDid you find that the accuracy is good or bad? I compared with Bert on few examples for masked word prediction and most of XLNet predicted word with the highest probability do not fit at all. \r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n", "@kushalj001 hi, how can I get the sentence vector", "Hi, so it seems that creating a model with a configuration is primarily the problem here:\r\n`model = XLNetLMHeadModel.from_pretrained(\"xlnet-large-cased\")`\r\nyields consistent outputs, but \r\n`config = XLNetConfig.from_pretrained(\"xlnet-large-uncased\")`\r\n`model = XLNetModel(config)` \r\ndoes not at all. \r\nMy question is, how is it possible to set configuration states (like getting hidden states of the model). I have run the glue STS-B fine tuning code to customize the model which is stored at `./proc_data/sts-b-100`, but when I load the model using code like this to get hidden states:\r\n\r\n`config = XLNetConfig.from_pretrained('./proc_data/sts-b-110/')`\r\n`config.output_hidden_states=True`\r\n`tokenizer = XLNetTokenizer.from_pretrained('././proc_data/sts-b-110/')`\r\n`model = XLNetForSequenceClassification(config)`\r\n\r\nI get results that vary wildly across runs. \r\n\r\nSpecifically, I would like to get the hidden states of each layer from the fine tuned model and correlate it to the actual text similarity. I was thinking I'd load the model with XLNetForSequenceClassification, get all the hidden states setting the configuration to output hidden states and do such a correlation. Is my approach incorrect?", "Looking at run_glue, it seems that actually outputs[1] is used for prediction? This is confusing because all the examples use [0] and the documentation is not very clear.\r\n `outputs = model(**inputs)`\r\n `tmp_eval_loss, logits = outputs[:2]`\r\nFrom run_glue.py\r\n", "Ok, I figured the logits and loss issue out - the issue is that for XLNetForSequenceClassification, the second index does in fact have logits while the first has loss.", "@thomwolf @Oxi84 while calculating word-embeddings of a document, i.e multiple sentences, is it necessary to pass the document sentence-wise? For my dataset, I removed punctuation as a part of the pre-processing step. So now, my whole document goes into the model. Does this hurt the model's performance? Does it make a lot of difference in terms of capturing the context of words?\r\nThanks", "It should improve acuracy if the text is longer, but still for me Bert is way better ... on 20-40 words long text.", "> It should improve acuracy if the text is longer, but still for me Bert is way better ... on 20-40 words long text.\r\n\r\nYeah, even for my experiments, BERT simply outperfoms XLNet. Still don't know why though.\r\nWhen you say \"it should improve accuracy\", you mean that feeding sentences to calculate word-vec would be better, right?", "Did you managed to try tensorflow version of XLNet, there is a chance it might be different from the pytorch version?", "Maybe there is some bug, but its unlikely since the bechmark results with the XLnet pytorch are the same. But I gues this would the first thing to try to recheck.", "> Did you managed to try tensorflow version of XLNet, there is a chance it might be different from the pytorch version?\r\n\r\nAny simple way of doing this?", "any updates regarding this issue? ", "@kushalj001 why remove the punctuation ? Is it domain specific or to improve accuracy?", "> @kushalj001 why remove the punctuation ? Is it domain specific or to improve accuracy?\r\n\r\nMy dataset had a lot of random punctuation, ie misplaced single and double-quotes.\r\nBut also, do punctuations add any valuable information to the text? Apart from the period (which can be used to break a large para into sentences), does keeping other punctuation symbols make sense? ", "I will close this issue which dates back before we had the clean documentation up here: https://huggingface.co/pytorch-transformers/\r\n\r\nPlease open a new issue with a clear explanation of your specific problem if you have related issues." ]

[ { "id": 1834059054, "node_id": "MDU6TGFiZWwxODM0MDU5MDU0", "url": "https://api.github.com/repos/huggingface/transformers/labels/Ex:%20Generation", "name": "Ex: Generation", "color": "06EFF8", "default": false, "description": "Natural Language Generation" } ]

[ { "login": "patrickvonplaten", "id": 23423619, "node_id": "MDQ6VXNlcjIzNDIzNjE5", "avatar_url": "https://avatars.githubusercontent.com/u/23423619?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patrickvonplaten", "html_url": "https://github.com/patrickvonplaten", "followers_url": "https://api.github.com/users/patrickvonplaten/followers", "following_url": "https://api.github.com/users/patrickvonplaten/following{/other_user}", "gists_url": "https://api.github.com/users/patrickvonplaten/gists{/gist_id}", "starred_url": "https://api.github.com/users/patrickvonplaten/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patrickvonplaten/subscriptions", "organizations_url": "https://api.github.com/users/patrickvonplaten/orgs", "repos_url": "https://api.github.com/users/patrickvonplaten/repos", "events_url": "https://api.github.com/users/patrickvonplaten/events{/privacy}", "received_events_url": "https://api.github.com/users/patrickvonplaten/received_events", "type": "User", "site_admin": false } ]

[ "I tried it, but text quality is lowered a lot and inference time does not change at all.\r\n\r\nI simply changed `perm_mask` to be 0 over initial context and 1 over generated tokens.\r\n\r\n---\r\n\r\nInput :\r\n\r\n> In Seoul, you can do a lot of things ! For example you can\r\n\r\nGenerated text with full bidirectionality :\r\n\r\n> buy grocery stores and restaurants, or even buy liquor, tobacco, etc. Then you can go to the mall. Then you can visit shopping mall. Then you can go to the university, then you can visit an outdoor pool. You can visit the cinema. You can visit art galleries. Then you can visit a garden.<eop> Etc. etc. etc. After all, if you can buy items and enjoy them, then yes, you can enjoy them in Seoul. It is that simple.\r\n\r\nGenerated text with bidirectionality over context tokens, and unidirectionality over generated tokens :\r\n\r\n> buy tons Free do hotel on you whichT Seoul, list and do you coffee non can many of you sit- shopping People you river boatou. and Koreans in long you into graduate train/ by teacher college c people there ho sister formst to in city plain daughtera kayak cat.: years World home. still home later N will plan yearses street his looks a marriage different by tell it too stunning out to what ice by person a, people a bag.\r\n\r\n**Why is it that bad ?**", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Hi @Colanim,\r\n\r\nThanks for the issue - sorry that we overlooked it! \r\n\r\nI will take a closer look into this. GPT2 uses key value state caching whet doing generation. Not sure whether XLNet does something similar. Will see if it'd be easy to add or not!", "Sorry to answer that late. `XLNet` is known to be rather slow for text generation due to the needed padding to get it started. \r\n\r\n`XLNet` uses `mems` which is similar to `past` to have a longer memory span.\r\nSince the quality seems to degrade much when applying your suggestion, I don't think trying to add a `XLNet` enhancement for generation is of high priority at the moment...Sorry! But feel free to open a PR if you have a good solution :-) " ]

[ { "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 because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "You can write your own code like the prediction phase [here](https://github.com/huggingface/pytorch-pretrained-BERT/blob/78462aad6113d50063d8251e27dbaadb7f44fbf0/examples/run_squad.py#L345) ", "@Swathygsb have you figured it out? I have the same use case as you and I'm struggling to understand the source code. ", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

[ { "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 because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "Yeah this is solved in the coming release" ]

[]

[]

[ "This will be possible in the new release out soon.", "I find a method that can get the words embeddings.Thank you all the same!\r\nself.model = BertModel.from_pretrained(config.bert_path)\r\nself.word_emb = self.model.embeddings" ]

[]

[]

[ "Because that's what Bert's authors do in the official TF code:\r\nhttps://github.com/google-research/bert/blob/bee6030e31e42a9394ac567da170a89a98d2062f/modeling.py#L231", "Just wanted to point out for future reference the motivation has been answered by the original BERT authors in [[this GitHub issue]](https://github.com/google-research/bert/issues/43)." ]

[]

[]

[ "I have added a test suite that tests both the `tie_weights` function as well as the `resize_token_embeddings`", "# [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781?src=pr&el=h1) Report\n> Merging [#781](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781?src=pr&el=desc) into [xlnet](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/commit/50e62a4cb4d503e3559b88838b8cf9f745fef516?src=pr&el=desc) will **decrease** coverage by `0.23%`.\n> The diff coverage is `93.05%`.\n\n[](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## xlnet #781 +/- ##\n=========================================\n- Coverage 78.84% 78.6% -0.24% \n=========================================\n Files 35 34 -1 \n Lines 6092 6122 +30 \n=========================================\n+ Hits 4803 4812 +9 \n- Misses 1289 1310 +21\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [pytorch\\_transformers/modeling\\_xlnet.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781/diff?src=pr&el=tree#diff-cHl0b3JjaF90cmFuc2Zvcm1lcnMvbW9kZWxpbmdfeGxuZXQucHk=) | `78.79% <100%> (+0.44%)` | :arrow_up: |\n| [pytorch\\_transformers/tests/modeling\\_xlnet\\_test.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781/diff?src=pr&el=tree#diff-cHl0b3JjaF90cmFuc2Zvcm1lcnMvdGVzdHMvbW9kZWxpbmdfeGxuZXRfdGVzdC5weQ==) | `95.86% <100%> (+0.08%)` | :arrow_up: |\n| [...rch\\_transformers/tests/modeling\\_transfo\\_xl\\_test.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781/diff?src=pr&el=tree#diff-cHl0b3JjaF90cmFuc2Zvcm1lcnMvdGVzdHMvbW9kZWxpbmdfdHJhbnNmb194bF90ZXN0LnB5) | `94.33% <100%> (+0.1%)` | :arrow_up: |\n| [pytorch\\_transformers/modeling\\_xlm.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781/diff?src=pr&el=tree#diff-cHl0b3JjaF90cmFuc2Zvcm1lcnMvbW9kZWxpbmdfeGxtLnB5) | `86.48% <100%> (+0.3%)` | :arrow_up: |\n| [pytorch\\_transformers/modeling\\_bert.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781/diff?src=pr&el=tree#diff-cHl0b3JjaF90cmFuc2Zvcm1lcnMvbW9kZWxpbmdfYmVydC5weQ==) | `87.69% <100%> (+0.17%)` | :arrow_up: |\n| [pytorch\\_transformers/tests/modeling\\_gpt2\\_test.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781/diff?src=pr&el=tree#diff-cHl0b3JjaF90cmFuc2Zvcm1lcnMvdGVzdHMvbW9kZWxpbmdfZ3B0Ml90ZXN0LnB5) | `84.21% <66.66%> (-3.29%)` | :arrow_down: |\n| [pytorch\\_transformers/tests/modeling\\_openai\\_test.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781/diff?src=pr&el=tree#diff-cHl0b3JjaF90cmFuc2Zvcm1lcnMvdGVzdHMvbW9kZWxpbmdfb3BlbmFpX3Rlc3QucHk=) | `84.21% <66.66%> (-0.79%)` | :arrow_down: |\n| [pytorch\\_transformers/tests/modeling\\_xlm\\_test.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781/diff?src=pr&el=tree#diff-cHl0b3JjaF90cmFuc2Zvcm1lcnMvdGVzdHMvbW9kZWxpbmdfeGxtX3Rlc3QucHk=) | `72.13% <75%> (ø)` | :arrow_up: |\n| [pytorch\\_transformers/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781/diff?src=pr&el=tree#diff-cHl0b3JjaF90cmFuc2Zvcm1lcnMvbW9kZWxpbmdfZ3B0Mi5weQ==) | `75.62% <75%> (-5.39%)` | :arrow_down: |\n| [pytorch\\_transformers/modeling\\_transfo\\_xl.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781/diff?src=pr&el=tree#diff-cHl0b3JjaF90cmFuc2Zvcm1lcnMvbW9kZWxpbmdfdHJhbnNmb194bC5weQ==) | `57.32% <76.92%> (+0.32%)` | :arrow_up: |\n| ... and [14 more](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781?src=pr&el=footer). Last update [50e62a4...2918b7d](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/781?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[ { "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 because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "Oh yes you are right, thanks it's fixed in the coming release." ]

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

[]

[ "@tholor and @timoeller may have some insights on these", "Hey Sebastian, thanks for using the German Bert and digging into its details. The mysterious [unused3001] token was actually a special comma symbol to get rid of [UNK] tokens in some of our training texts. But we covered it up later on in the process + didn't anticipate it would be coming back to us : ) \r\n\r\nSo agreed, it is unwanted behaviour. \r\n\r\nThough TL;DR, we don't believe it is impacting either pretraining or downstream task training. \r\n\r\n\r\n\r\nApparently the token at index 0 (= [unused3001]) is used as padding token in TF Bert and pytorch Bert and the implementations do not really care if it is called [unused3001] [PAD] or [something].\r\n\r\nTo be a bit more intuitive we now swapped [unused3001] and [PAD] in the vocab files (pytorch and TF) only. \r\nMight be that future code somehow substitutes \"[PAD]\" input strings, which could cause problems.\r\n\r\nThe only thing that seems worrisome to us is that the embedding values for this padding token are non-zero (and change over the course of training) for our German Bert but also for Googles open sourced models. \r\n\r\nI tried to check how the padding embedding is handled in TF but am not familiar with debugging there...\r\n\r\nMaybe you want to dig more into it and raise an issue in the original TF Bert repro? Maybe this closed issue could be related to a rather unwanted padding embedding handling: https://github.com/google-research/bert/issues/113\r\n\r\nHope that helps, good luck!", "Thanks a lot for the input @Timoeller (not quite sure who Christian is, though ;) ). I also got the feeling that it doesn't really impact downstream applications (NER in this case). At least not heavily.\r\n\r\nI'll do some more experiments and raise an issue with the original repo if it feel it is warranted. \r\n\r\nThanks again and all the best,\r\nSebastian\r\n\r\n", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/775?src=pr&el=h1) Report\n> Merging [#775](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/775?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/commit/78462aad6113d50063d8251e27dbaadb7f44fbf0?src=pr&el=desc) will **decrease** coverage by `0.1%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/775?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #775 +/- ##\n==========================================\n- Coverage 61.5% 61.39% -0.11% \n==========================================\n Files 19 19 \n Lines 4026 4025 -1 \n==========================================\n- Hits 2476 2471 -5 \n- Misses 1550 1554 +4\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/775?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [pytorch\\_pretrained\\_bert/file\\_utils.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/775/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvZmlsZV91dGlscy5weQ==) | `66.44% <0%> (-1.35%)` | :arrow_down: |\n| [pytorch\\_pretrained\\_bert/optimization.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/775/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvb3B0aW1pemF0aW9uLnB5) | `73.52% <0%> (-0.74%)` | :arrow_down: |\n| [pytorch\\_pretrained\\_bert/tokenization\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/775/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uX2dwdDIucHk=) | `81.91% <0%> (-0.54%)` | :arrow_down: |\n| [pytorch\\_pretrained\\_bert/tokenization\\_transfo\\_xl.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/775/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uX3RyYW5zZm9feGwucHk=) | `31.85% <0%> (-0.19%)` | :arrow_down: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/775?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/775?src=pr&el=footer). Last update [78462aa...b72f755](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/775?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[]

[]

[ "Indeed, I've now added the text padding trick of Aman (add some padding text to have longer inputs) and the quality is really a lot higher.\r\n\r\nWill merge the xlnet branch in master and release on Monday." ]

[]

[]

[]

[]

[]

[ "Hi @laifi, \r\n\r\nI cannot reproduce this issue. Are you sure that you run with the latest code from master branch? It looks suspicious to me that `tokenizer = BertTokenizer.from_pretrained('bert-base-german-cased')` doesn't find the model. \r\nCan you please check if you have [the according line](https://github.com/huggingface/pytorch-pretrained-BERT/blob/78462aad6113d50063d8251e27dbaadb7f44fbf0/pytorch_pretrained_bert/tokenization.py#L37) in your PRETRAINED_VOCAB_ARCHIVE_MAP? \r\n\r\nFor your second approach with downloaded files: \r\n- be aware that model packaging changed lately from archives to individual files for vocab, model and config (see [here](https://github.com/huggingface/pytorch-pretrained-BERT/pull/688#issuecomment-502991015)). If you really want to download manually you should download the [.bin](https://int-deepset-models-bert.s3.eu-central-1.amazonaws.com/pytorch/bert-base-german-cased-pytorch_model.bin), [bert_config.json](https://int-deepset-models-bert.s3.eu-central-1.amazonaws.com/pytorch/bert-base-german-cased-config.json) and the [vocab file](https://int-deepset-models-bert.s3.eu-central-1.amazonaws.com/pytorch/bert-base-german-cased-vocab.txt) to a folder called \"bert-base-german-cased\"\r\n- `from_pretrained` expects a model name or path not a .bin . You should try: BertTokenizer.from_pretrained('YOUR_PATH_TO/bert-base-german-cased') \r\n\r\nHope that helps!", "Thank you @tholor , i installed the package with pip and i cannot find 'bert-german-cased' in PRETRAINED_VOCAB_ARCHIVE_MAP\r\nNow , i tried to reinstall the package from source and it's working .\r\n", "@laifi I am keep getting the same error as the one that you got:\r\n\r\n> UnicodeDecodeError: 'utf-8' codec can't decode byte 0x80 in position 0: invalid start byte\r\n\r\nI also tried to reinstall it, how did you fix it?\r\n", "> @laifi I am keep getting the same error as the one that you got:\r\n> \r\n> > UnicodeDecodeError: 'utf-8' codec can't decode byte 0x80 in position 0: invalid start byte\r\n> \r\n> I also tried to reinstall it, how did you fix it?\r\n\r\n@shaked571 , i have just uninstalled the pip package and installed it again from source (try to not keep any cache for the package).\r\n**PS: the issue is fixed in the last migration from pytorch-pretrained-bert to pytorch-transformers .**", "Hi,\r\nI also run into the same issue when I try this piece of code in google colab. \r\ntokenizer = BertTokenizer.from_pretrained('bert-base-german-cased')", "Hi,\r\nI also have the same issue. Using \r\n```\r\nfrom transformers import AutoTokenizer\r\ntokenizer = AutoTokenizer.from_pretrained(\"bert-base-german-cased\")\r\n```\r\nsolves the problem for me" ]

[]

[]

[ "If you want to evaluate only, you have to set `--output_dir` to the path of your previously trained model. Otherwise, the script will use the original model." ]

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

[]

[ "you can use the path to the folder containing your fine-tuned model as `--bert_model`.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "This model was WIP. Fixed now." ]

[]

[]

[ "`run_gpt2` has top-K which is better than beam-search for high-entropy tasks like open-domain generation. The coming release example (currently on the xlnet branch to be merged with master on Monday) will have top-K and Nucleus sampling (see Holtzman et al. http://arxiv.org/abs/1904.09751)", "Hi,\r\nIs it possible to include beam search decoding in ```run_generation.py``` ?", "hope that beam search appears in run_generation.py", "We'll add it.\r\ncc @rlouf ", "@thomwolf I see that run_generation.py has disappeared and beam_search does not exist anymore, nor in transformers/generate. Where could we find the implementation of batch beam_search in this repo ?", "You can’t... so far. We are reworking the API for greedy decoding and sampling, and will work on beam search afterwards." ]

[]

[]

[]

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

[]

[ "As far as I know, the pytorch code of XLNet is not completely ready now. But you could find it in the branch `xlnet` and the classifier code is nearly ready in the file `example/run_xlnet_classifier.py`. I have successfully fine-tuned it on the SST-2 task (which belongs to GLUE) with following args:\r\n\r\n```shell\r\npython run_xlnet_classifier.py \\\r\n--data_dir ..\\glue_data\\SST-2 \\\r\n--task_name sst-2 \\\r\n--output_dir sst_model \\\r\n--do_train \\\r\n--do_eval \\\r\n--max_seq_length 128 \\\r\n--train_batch_size 64 \\\r\n--learning_rate 5e-6\r\n```", "@SivilTaram Cant be fine tuned on external data ? is the tensorflow version ready ?", "@AhmedBahaaElDinMohammed Sure you could fine-tune it on external data, which means you should process your data and construct train/validate `examples` as SST-2 does. You could see `example/utlis_glue.py` for more details to handle your external data :)\r\n\r\nThe tensorflow is ready, you could refer to the original repo for help. This repo is only for pytorch version, thanks.", "> I have successfully fine-tuned it on the SST-2 task (which belongs to GLUE) \r\n\r\n@SivilTaram Would it be possible to fine-tune it on SQuAD 2.0? or alternatively, convert a fine-tuned model from the original repo/tensorflow?", "@edanweis Not ready now. Please wait the author to complete the awesome work :) Or you could watch the updates of PR [here](https://github.com/huggingface/pytorch-pretrained-BERT/pull/711).", "Has anyone tried fp16 for xlnet?\r\nI tried it and found that the memory was half, but it was slower than fp32(even when I used the same GPU memory). \r\nEnvironment: v100, cuda 10.0, torch 1.1\r\nThe environment is ok, because I tried bert + fp16 and it was much faster than fp32.\r\nI thought it is the problem of torch.einsum, but I am not that sure. \r\nGuys, do you have the same problem ?", "@SivilTaram Following the latest release 0.6.2, I am trying to convert my tf checkpoints:\r\n\r\n```\r\nexport TRANSFO_XL_CHECKPOINT_PATH=home/edanweis/xlnet/model/squad\r\nexport TRANSFO_XL_CONFIG_PATH=home/edanweis/xlnet/model/squad\r\nexport FINETUNING_TASK=squad\r\n\r\n\r\npytorch_transformers xlnet \\\r\n $TRANSFO_XL_CHECKPOINT_PATH \\\r\n $TRANSFO_XL_CONFIG_PATH \\\r\n $PYTORCH_DUMP_OUTPUT \\\r\n $FINETUNING_TASK \\\r\n```\r\nBut getting `pytorch_transformers/__main__.py\", line 111, in main FINETUNING_TASK) UnboundLocalError: local variable 'FINETUNING_TASK' referenced before assignment`", "@SivilTaram Did you try to finetune XLNet with the last code (release 1.0) using examples/run_glue.py? Everything works but accuracy didn't change and every time is around 0.50? It looks like it didn't train at all. I used the following script:\r\n\r\n```\r\nexport GLUE_DIR=/path/to/glue\r\n\r\npython ./examples/run_glue.py \\\r\n --model_type xlnet \\\r\n --model_name_or_path xlnet-large-cased \\\r\n --do_train \\\r\n --do_eval \\\r\n --evaluate_during_training \\\r\n --logging_steps 500 \\\r\n --save_steps 1000 \\\r\n --task_name=sst-2 \\\r\n --data_dir=${GLUE_DIR}/SST-2 \\\r\n --output_dir=./proc_data/sst-2 \\\r\n --max_seq_length=128 \\\r\n --per_gpu_eval_batch_size=8 \\\r\n --per_gpu_train_batch_size=8 \\\r\n --gradient_accumulation_steps=1 \\\r\n --max_steps=8000 \\\r\n --model_name=xlnet-large-cased \\\r\n --overwrite_output_dir \\\r\n --overwrite_cache \\\r\n --warmup_steps=120\r\n```", "@avostryakov I do not yet. I guess you could explore if the loss decrease as expected? There should be loss logs, along with tensorboard logs.", "@SivilTaram Evaluation loss isn't changed, training loss is increased. It looks like something wrong with the optimization process during training.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "Yes we could add this. You mean tasks like GLUE or SQuAD?", "> Yes we could add this. You mean tasks like GLUE or SQuAD?\r\n\r\nYes! exactly!\r\nPlease add this, thanks!", "@thomwolf Are you still working on the code to finetune the GPT2 language model (not classification task)? Thanks.", "@experiencor @thomwolf also curious about GPT2 LM finetuning issue, thanks!", "We'll add an example for fine-tuning the models (probably refactor the Bert's one at the same time) this month.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "> We'll add an example for fine-tuning the models (probably refactor the Bert's one at the same time) this month.\r\n\r\nDid you add the example ? looking for an example of fine tuning gpt-2 for downstream tasks.", "+1. Also interested." ]

[ { "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 try stacking a linear layer over-top of BERT that takes as input the BERT sequence representation + your features. You would have to fine-tune through all of BERT.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "> I am trying to fine-tune BERT for a sequence classification task where in addition to the sequences, I have extra features such as the writer age, tags, etc. I want to use those extra features, and I was thinking about concatenating them to the input of the final linear layer.\r\n> Is there a way of doing such a thing? If not, what is the best way for integrating extra features in the fine-tuning process?\r\n\r\nDid you find a way to add extra features then fine-tuning BERT?" ]

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

[]

[ "also ran into this. I think they forgot to upload the file/make it public. You can find the vocab file on the original google repo\r\n\r\nhttps://github.com/google-research/bert", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[ { "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 because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "Not really providing a solution here, but have you considered https://github.com/allenai/scibert instead?\r\nAllenAI provides PyTorch weights, and through tests they claim their model is superior https://arxiv.org/pdf/1903.10676.pdf on their suite of tasks. For that and for ease of use, it may be a valid alternative.", "Hello, I am trying to figure out how to load the SciBert weights. I see that you can use\r\n\r\n```\r\n# Simple serialization for models and tokenizers\r\nmodel.save_pretrained('./directory/to/save/') # save\r\nmodel = model_class.from_pretrained('./directory/to/save/') # re-load\r\ntokenizer.save_pretrained('./directory/to/save/') # save\r\n```\r\n\r\nSo my guess is to download them from here\r\nhttps://github.com/allenai/scibert#pytorch-models\r\n\r\nUntar them, then point to that directory\r\n\r\n```\r\nmodel = model_class.from_pretrained('DIRECTORY/TO/DOWNLOADED/UNZIPPED/SCIBERT/Pytorch.bin') \r\n```\r\n\r\nUnless there is more to that, the part I am confused about is that SciBert also has it's own vocab and 'vocab.txt' file. I am wondering how to point to that file, and not the default one. \r\n\r\nEdit\r\n\r\nFound the answer\r\n\r\nhttps://github.com/huggingface/pytorch-transformers/issues/69#issuecomment-443215315\r\n\r\nyou can just do a direct path to it", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "How much memory does your GPU have. You can check this by running `nvidia-smi`.", "I also has same phenomena. Also, the learning time become slower and much GPU consumption occur, both of which I think is natural, regarding parameters BERT has.\r\n\r\nThe substitutional way is that, no fine-tuning and dump. \r\nI mean, feed your sequence to Bert and dump your layers. In the training process of your task, simply load sequence with your dumped result of Bert.", "Try `--reduce_memory `, it worked on mine with base multilingual uncased BERT with batch size= 4 on my single 2080Ti.\r\n\r\n\r\n```\r\npython3 finetune_on_pregenerated.py \r\n--pregenerated_data training/ \r\n--bert_model bert-base-multilingual-uncased \r\n--do_lower_case \r\n--output_dir finetuned_lm/ \r\n--epochs 3 \r\n--reduce_memory \r\n--train_batch_size 4\r\n```", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

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

[]

[ "Your snippet is too short to see what type of error there is, can you extract a larger one?", "Hi \r\nyes\r\n\r\n", "Pull the latest changes from master and report if that helps.", "it has been disappeared in pregenerate script, and appeared in finetuning on pregenerated data script\r\n\r\n", "You are getting these errors because you are using Python 3.5 and the code is making use of f-strings which are introduced in Python 3.6. You could try using Python 3.6 or change the source code replacing f-string with str.format syntax.\r\n\r\n@thomwolf The readme says repository supports Python 3.5+. Does that mean Python 3.5 is supported as well? If yes, I think we should change f-strings for the format syntax. \r\n", "Well, only the library code supports Python 3.5+, I don't check the examples which are mostly contributed by the community.\r\nIf you want to fix Python 3.5 support for the examples I'm happy to welcome a PR.\r\n", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

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

[]

[ "You need to install the master version (not with pip or conda) : \r\n```\r\ngit clone https://github.com/huggingface/pytorch-pretrained-BERT.git\r\ncd pytorch-pretrained-BERT\r\npython setup.py install\r\n```\r\n\r\nThen you can use it like this : \r\n```\r\nmodel = BertModel.from_pretrained('bert-base-uncased',\r\n output_attentions=True,\r\n keep_multihead_output=True)\r\nmodel.eval() # turn off dropout layers\r\nattn = model(tokens)[0]\r\n```\r\n\r\nTell me if I'm misinterpreting your problem", "Thank you a lot for the help, I didn't expect this to only work on the current release. \r\n\r\nHowever, I think with this I found a problem in the BERT encoder module:\r\n\r\n```\r\ndef forward(self, hidden_states, attention_mask, output_all_encoded_layers=True, head_mask=None):\r\n all_encoder_layers = []\r\n all_attentions = []\r\n for i, layer_module in enumerate(self.layer):\r\n hidden_states = layer_module(hidden_states, attention_mask, head_mask[i])\r\n```\r\nThe forward function by default gets `None` for the `head_mask` parameter. Then, however, it indexes it, which causes an error. I think it would be nice to handle this case.", "Hi. I want to do something similar but with the **BertForQuestionAnswering** model.\r\n\r\nThe BertModel is the general BERT model that is used to classify whether a sentence is the next sentence or not. I want to get the attention values for QuestionAnswering while I pass a new paragraph and a question as inputs. I want to use the **BertForQuestionAnswering** model (which is pretrained on SQuAD if I am not wrong) and get the self-attention values on the question words. Is it possible to achieve this in a similar way as mentioned above?\r\n\r\n**NOTE:** I know the above method gives attention values of the pre-trained model. I want to get attention values of the model when I feed a new input question to the model. Something similar to what can be done using [BertViz](https://github.com/jessevig/bertviz) (although I do not want to visualize attention, just want to get the values).\r\n\r\nThanks.", "Hi, this will be in the next release (release date sometime next week).\r\nThere will be attention/hidden-state output options for all the models.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[ { "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 because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[ { "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 because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "Yes, this library is not made for training a model from scratch.\r\n\r\nYou should use one of the libraries I referred to here: https://github.com/huggingface/pytorch-pretrained-BERT/issues/543#issuecomment-491207121\r\n\r\nI might give it a look one day but not in the short-term.", "@thomwolf Thank you so much for the info! :)\r\n\r\nJust to share, I quickly did a benchmark of XLM (this one fits my needs the most out of your three recommendations). \r\n\r\n**Sentences/s (for the specs I mentioned above):**\r\n\r\nBatch size | Official TF BERT | HuggingFace PyTorch BERT | XLM PyTorch BERT\r\n-- | -- | -- | --\r\n128 over 1 GPU | 610 | 288 | 575\r\n250 over 1 GPU | 647 | OOM | 625\r\n500 over 1 GPU | 665 | OOM | 650\r\n700 over 1 GPU | N/A | OOM | OOM\r\n900 over 1 GPU | 667 | OOM | OOM\r\n1000 over 1 GPU | OOM | OOM | OOM\r\n128 over 4 GPUs | 889 (1.5x) | 779 (2.7x) | N/A\r\n512 over 4 GPUs | 1522 (2.3x) | 1018 (3.?x) | N/A\r\n1000 over 4 GPUs | 1798 (2.?x) | OOM | N/A\r\n2000 over 4 GPUs | 1946 (2.?x) | OOM | N/A\r\n3600 over 4 GPUs | 1991 (3.0x) | OOM | N/A\r\n4000 over 4 GPUs | OOM | OOM | N/A\r\n\r\nNote: Only spent 2 hours on XLM, not sure if I set the vocab to be exactly the same size as the others, but they should be in the same ballpark.\r\n\r\nI haven't got a chance to benchmark the multi-GPU XLM. But in general, it looks like:\r\n1. The TensorFlow implementation uses memory more efficiently.\r\n2. PyTorch's multi-GPU scaling seems better.\r\n3. PyTorch itself is not slower than TF.\r\n\r\nn", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Hi @thomwolf , \r\nI was trying to fine-tune pytorch-transformers's gpt2 (124M) on a V100 16GB GPU. But I am not able to accommodate more than the batch_size of 2. I am using seq-length of 1024 tokens. \r\n\r\nThis might be evident from above comments but I am new to training NNs so wanted to confirm if fine tuning would also cause OOM as in training from scratch? If so, then is only option available to finetune gpt2 is to use original tensorfolow implementation?\r\n\r\nThanks", "Hi @SKRohit, with the GPT-2 model you can either fine-tune it with a batch size of 4 and a sequence of 512 tokens, or a batch size of 2 and a sequence of 1024 tokens, like what you've tried. We have had good results with a batch size of 4 and a sequence of 512 in our experiments.\r\n\r\nIf you want a bigger batch size, you can set up gradient accumulation, which would allow you to put larger to much larger batch sizes. You can find an example of gradient accumulation applied to fine-tuning in our [language model fine-tuning example](https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_lm_finetuning.py).", "Yes, @LysandreJik I am using gradient accumulation. I found max possible batch_size = 2 to be too small given this [comment](https://github.com/openai/gpt-2/issues/150#issuecomment-529153176) so asked to make sure there is no error in my code or any issue with my gcloud gpu. \r\nAlso, have you finetuned gpt2 architectures using mixed_precision (mp) training? Did you find any difference in performance of mp trained gpt2 in comparison to without mp?\r\nAnd I am referring to fine-tuning script provided in `pytorch_transformers` repo 👍 .\r\n\r\nThanks. ", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "mark", "@thomwolf What is the bottleneck in HuggingFace transformers pretraining comparing to Tensorflow and other PyTorch implementations?", "I also find the Transformers library to be more memory hungry. It seems to be even slower with Pytorch than TF, too.\r\n\r\nOn the flip side, it is really easy to use. I guess if you have big datasets and 2x slower is critically insufficient, it's not a good option. But if the difference is just half a day or less, it may not be that bad." ]

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

[]

[ "You are right Ethan.\r\nI'm refactoring the examples which were a bit rotten, let's include this fix as well.", "Great! Either way the examples are a great starting point :)\r\n\r\nI'm also wondering if tensorboard is [only logging](https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/examples/run_classifier.py#L345) the training loss for the last forward pass for a batch (if several are required / when using gradient accumulation)? A fix would be to maintain a variable ```tr_batch_loss``` (similar to ```tr_loss```) for each full training batch (reset after each parameter update) and log that instead.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "You can only load from a tensorflow checkpoint in a `BertForPretraing` model.\r\nI will add a check.\r\nAlternatively, you should use the conversion script to make a pytorch model and then you can import the resulting pytorch model in any type of Bert model.", "I use the BertForPretraining.from_pretrained().bert to get the BertModel from the tensorflow checkpoint, I think it is useful", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "You are probably using the \"wrong bert\" . Install bert-tensorflow. There are two packages with the same name\r\n ", "same for distilbert. Using base_model.distilbert will solve the problem" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/748?src=pr&el=h1) Report\n> Merging [#748](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/748?src=pr&el=desc) into [xlnet](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/commit/708877958a308a0f0e8fd199f8f327e4797f1583?src=pr&el=desc) will **increase** coverage by `0.21%`.\n> The diff coverage is `96.06%`.\n\n[](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/748?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## xlnet #748 +/- ##\n=========================================\n+ Coverage 71.5% 71.72% +0.21% \n=========================================\n Files 35 35 \n Lines 5587 5633 +46 \n=========================================\n+ Hits 3995 4040 +45 \n- Misses 1592 1593 +1\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/748?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [...\\_pretrained\\_bert/tests/modeling\\_transfo\\_xl\\_test.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/748/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdGVzdHMvbW9kZWxpbmdfdHJhbnNmb194bF90ZXN0LnB5) | `94.23% <100%> (ø)` | :arrow_up: |\n| [pytorch\\_pretrained\\_bert/model\\_utils.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/748/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxfdXRpbHMucHk=) | `92.61% <100%> (+0.04%)` | :arrow_up: |\n| [pytorch\\_pretrained\\_bert/modeling.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/748/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmcucHk=) | `87.55% <100%> (+0.47%)` | :arrow_up: |\n| [pytorch\\_pretrained\\_bert/modeling\\_openai.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/748/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfb3BlbmFpLnB5) | `79.5% <100%> (+0.11%)` | :arrow_up: |\n| [pytorch\\_pretrained\\_bert/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/748/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfZ3B0Mi5weQ==) | `81.57% <100%> (+0.15%)` | :arrow_up: |\n| [pytorch\\_pretrained\\_bert/modeling\\_xlnet.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/748/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfeGxuZXQucHk=) | `74.17% <83.33%> (+0.14%)` | :arrow_up: |\n| [...torch\\_pretrained\\_bert/tests/model\\_tests\\_commons.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/748/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdGVzdHMvbW9kZWxfdGVzdHNfY29tbW9ucy5weQ==) | `97.08% <97.22%> (-0.01%)` | :arrow_down: |\n| [pytorch\\_pretrained\\_bert/tokenization\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/748/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uX2dwdDIucHk=) | `82.44% <0%> (-1.07%)` | :arrow_down: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/748?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/748?src=pr&el=footer). Last update [7088779...b43b130](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/748?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[]

[]

[ "I would not advise to use them for training from scratch. See #751 for discussion and links." ]

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

[]

[ "I might be wrong, but I think GPT2Tokenizer uses byte pair encoding, a form of subword-level encoding. On an intuitive level, this is a between character-level and word level, and akin to breaking the word apart by syllable (in reality it's breaking the word apart by the highest frequency patterns). I know some people use Sentence-Piece tokenization for working with Chinese in BERT, so it might be worthwhile to see if there's a similar effort for a Sentence-Piece GPT-2", "@DEBADRIBASAK \r\nCan you share the steps for fine-tuning on hindi dataset? Thanks.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/745?src=pr&el=h1) Report\n> Merging [#745](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/745?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/commit/dad3c7a485b7ffc6fd2766f349e6ee845ecc2eee?src=pr&el=desc) will **decrease** coverage by `0.05%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/745?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #745 +/- ##\n==========================================\n- Coverage 62.27% 62.22% -0.06% \n==========================================\n Files 18 18 \n Lines 3979 3979 \n==========================================\n- Hits 2478 2476 -2 \n- Misses 1501 1503 +2\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/745?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [pytorch\\_pretrained\\_bert/tokenization\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/745/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uX2dwdDIucHk=) | `82.44% <0%> (-1.07%)` | :arrow_down: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/745?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/745?src=pr&el=footer). Last update [dad3c7a...64b2a82](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/745?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

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

[]

[ "I second this issue #735 ", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]