add training script

codeparrot_training.py

@@ -0,0 +1,202 @@
+from transformers import GPT2LMHeadModel, AutoTokenizer
+from transformers import AdamW, get_scheduler, set_seed
+from datasets import load_dataset
+from accelerate import Accelerator
+import datasets, transformers
+from huggingface_hub import Repository
+
+from torch.utils.data import IterableDataset
+from torch.utils.data.dataloader import DataLoader
+from torch.utils.tensorboard import SummaryWriter
+from argparse import Namespace
+import torch
+import logging
+import wandb
+
+
+class ConstantLengthDataset(IterableDataset):
+
+    def __init__(self, tokenizer, dataset, seq_length=1024, batch_size=3,
+                 num_of_sequences=1024, chars_per_token=3.6):
+        self.tokenizer = tokenizer
+        self.concatenation_token = tokenizer.bos_token
+        self.dataset = dataset
+        self.seq_length = seq_length
+        self.batch_size = batch_size
+        self.input_characters = seq_length * chars_per_token * num_of_sequences
+
+    def __iter__(self):
+        iterator = iter(self.dataset)
+        more_examples = True
+        batch = []
+        while more_examples:
+            buffer = ''
+            while True:
+                if len(buffer) >= self.input_characters:
+                    break
+                try:
+                    next_example = next(iterator)['content']
+                    buffer = buffer + self.concatenation_token + next_example
+                except StopIteration:
+                    more_examples = False
+                    break
+
+            tokenized_input = tokenizer(buffer, truncation=True,
+                                        max_length=self.seq_length,
+                                        return_overflowing_tokens=True)
+
+            for input_ids in tokenized_input['input_ids']:
+                if len(input_ids) == self.seq_length:
+                    batch.append(input_ids)
+                if len(batch) == self.batch_size:
+                    yield torch.tensor(batch)
+                    batch = []
+
+def setup_logging(project_name):
+    logger = logging.getLogger(__name__)
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO,)
+    if accelerator.is_main_process: # we only want to setup logging once
+        wandb.init(project=project_name, config=args)
+        tb_writer = SummaryWriter()
+        tb_writer.add_hparams(vars(args), {'0': 0})
+        logger.setLevel(logging.INFO)
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_info()
+    else:
+        logger.setLevel(logging.ERROR)
+        datasets.utils.logging.set_verbosity_error()
+        transformers.utils.logging.set_verbosity_error()
+    return logger, tb_writer
+
+def create_dataloaders(dataset_name):
+    train_data = load_dataset(dataset_name+'-train', split="train",
+                              streaming=True)
+    train_data = train_data.shuffle(buffer_size=args.shuffle_buffer)
+    valid_data = load_dataset(dataset_name+'-valid', split="train",
+                              streaming=True)
+
+    train_dataset = ConstantLengthDataset(tokenizer, train_data,
+                                        seq_length=args.seq_length,
+                                        batch_size=args.train_batch_size)
+    valid_dataset = ConstantLengthDataset(tokenizer, valid_data,
+                                        seq_length=args.seq_length,
+                                        batch_size=args.valid_batch_size)
+
+    train_dataloader=DataLoader(train_dataset, batch_size=args.train_batch_size)
+    eval_dataloader=DataLoader(valid_dataset, batch_size=args.valid_batch_size)
+    return train_dataloader, eval_dataloader
+
+def get_grouped_params(model, no_decay=["bias", "LayerNorm.weight"]):
+    params_with_wd, params_without_wd = [], []
+    for n, p in model.named_parameters():
+        if any(nd in n for nd in no_decay): params_without_wd.append(p)
+        else: params_with_wd.append(p)
+    return [{'params': params_with_wd, 'weight_decay': args.weight_decay},
+            {'params': params_without_wd, 'weight_decay': 0.0}]
+
+def log_metrics(step, metrics):
+    logger.info(f"Step {step}: {metrics}")
+    if accelerator.is_main_process:
+        wandb.log(metrics)
+        [tb_writer.add_scalar(k, v, step) for k, v in metrics.items()]
+
+def evaluate():
+    model.eval()
+    losses = []
+    for step, batch in enumerate(eval_dataloader):
+        with torch.no_grad():
+            outputs = model(batch[0], labels=batch[0])
+        loss = outputs.loss.repeat(args.valid_batch_size)
+        losses.append(accelerator.gather(loss))
+        if args.max_eval_steps > 0 and step >= args.max_eval_steps: break
+    loss = torch.mean(torch.cat(losses))
+    try: perplexity = torch.exp(loss)
+    except OverflowError: perplexity = float("inf")
+    return loss.item(), perplexity.item()
+
+# Hyperparameters
+project_name = 'transformersbook/codeparrot-small'
+dataset_name = 'transformersbook/codeparrot'
+config = {"train_batch_size": 4,
+          "valid_batch_size": 4,
+          "weight_decay": 0.1,
+          "shuffle_buffer": 1000,
+          "learning_rate": 5e-4,
+          "lr_scheduler_type": "cosine",
+          "num_warmup_steps": 1000,
+          "gradient_accumulation_steps": 8,
+          "max_train_steps": 4096,
+          "max_eval_steps": 1024,
+          "seq_length": 1024,
+          "seed": 1,
+          "save_checkpoint_steps":4096,}
+args = Namespace(**config)
+set_seed(args.seed)
+
+# Accelerator
+accelerator = Accelerator()
+samples_per_step = accelerator.state.num_processes * args.train_batch_size
+
+# Logging
+logger, tb_writer = setup_logging(project_name.split("/")[1])
+logger.info(accelerator.state)
+run_name = wandb.run.name
+
+# Load model and tokenizer
+hf_repo = Repository("./", clone_from=project_name, revision=run_name)
+model = GPT2LMHeadModel.from_pretrained("./")
+tokenizer = AutoTokenizer.from_pretrained("./")
+
+# Load dataset and dataloader
+train_dataloader, eval_dataloader = create_dataloaders(dataset_name)
+
+# Prepare the optimizer and learning rate scheduler
+optimizer = AdamW(get_grouped_params(model), lr=args.learning_rate)
+lr_scheduler = get_scheduler(name=args.lr_scheduler_type, optimizer=optimizer,
+                             num_warmup_steps=args.num_warmup_steps,
+                             num_training_steps=args.max_train_steps,)
+def get_lr(): return optimizer.param_groups[0]['lr']
+
+# Prepare everything with our `accelerator`.
+model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
+    model, optimizer, train_dataloader, eval_dataloader)
+
+# Train model
+model.train()
+completed_steps = 0
+for step, batch in enumerate(train_dataloader, start=1):
+    loss = model(batch[0], labels=batch[0]).loss
+    log_metrics(step, {'lr': get_lr(), 'samples': step*samples_per_step,
+                       'steps': completed_steps, 'loss/train': loss.item()})
+    loss = loss / args.gradient_accumulation_steps
+    accelerator.backward(loss)
+    if step % args.gradient_accumulation_steps == 0:     
+        optimizer.step()
+        lr_scheduler.step()
+        optimizer.zero_grad()
+        completed_steps += 1
+    if step % args.save_checkpoint_steps == 0:
+        logger.info('Evaluating and saving model checkpoint')
+        eval_loss, perplexity = evaluate()
+        log_metrics(step, {'loss/eval': eval_loss, 'perplexity': perplexity})
+        accelerator.wait_for_everyone()
+        unwrapped_model = accelerator.unwrap_model(model)
+        unwrapped_model.save_pretrained("./")
+        if accelerator.is_main_process:
+            hf_repo.push_to_hub(commit_message=f'step {step}')
+        model.train()
+    if completed_steps >= args.max_train_steps:
+        break
+
+# Evaluate and save the last checkpoint
+logger.info('Evaluating and saving model after training')
+eval_loss, perplexity = evaluate()
+log_metrics(step, {'loss/eval': eval_loss, 'perplexity': perplexity})
+accelerator.wait_for_everyone()
+unwrapped_model = accelerator.unwrap_model(model)
+unwrapped_model.save_pretrained("./")
+if accelerator.is_main_process:
+    try: hf_repo.push_to_hub(commit_message=f'final model')
+    except: logger.info('No changes to previously saved model.')