Update from pytorch-transformers to transformers library

convai_evaluation.py

@@ -17,7 +17,7 @@
 from projects.convai2.eval_f1 import eval_f1, setup_args as setup_args_f1
 from projects.convai2.eval_ppl import eval_ppl, setup_args as setup_args_ppl
 from projects.convai2.build_dict import build_dict
-from pytorch_transformers import (OpenAIGPTDoubleHeadsModel, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer,
+from transformers import (OpenAIGPTDoubleHeadsModel, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer,
                                   GPT2DoubleHeadsModel, GPT2LMHeadModel, GPT2Tokenizer)
 
 from train import build_input_from_segments, pad_dataset, SPECIAL_TOKENS, add_special_tokens_

interact.py

@@ -12,7 +12,7 @@
 import torch
 import torch.nn.functional as F
 
-from pytorch_transformers import OpenAIGPTLMHeadModel, OpenAIGPTTokenizer, GPT2LMHeadModel, GPT2Tokenizer
+from transformers import OpenAIGPTLMHeadModel, OpenAIGPTTokenizer, GPT2LMHeadModel, GPT2Tokenizer
 from train import SPECIAL_TOKENS, build_input_from_segments, add_special_tokens_
 from utils import get_dataset, download_pretrained_model
 

requirements.txt

@@ -1,5 +1,5 @@
 torch
 pytorch-ignite
-pytorch-transformers>=1.2
+transformers==2.5.1
 tensorboardX==1.8
 tensorflow  # for tensorboardX

test_special_tokens.py

@@ -2,7 +2,7 @@
 import shutil
 import unittest
 
-from pytorch_transformers import OpenAIGPTTokenizer, GPT2Tokenizer
+from transformers import OpenAIGPTTokenizer, GPT2Tokenizer
 from train import ATTR_TO_SPECIAL_TOKEN, SPECIAL_TOKENS
 
 class TestSpecialTokenTreatment(unittest.TestCase):

train.py

@@ -16,14 +16,14 @@
 from ignite.metrics import Accuracy, Loss, MetricsLambda, RunningAverage
 from ignite.contrib.handlers import ProgressBar, PiecewiseLinear
 from ignite.contrib.handlers.tensorboard_logger import TensorboardLogger, OutputHandler, OptimizerParamsHandler
-from pytorch_transformers import (AdamW, OpenAIGPTDoubleHeadsModel, OpenAIGPTTokenizer,
+from transformers import (AdamW, OpenAIGPTDoubleHeadsModel, OpenAIGPTTokenizer,
                                   GPT2DoubleHeadsModel, GPT2Tokenizer, WEIGHTS_NAME, CONFIG_NAME)
 
 from utils import get_dataset, make_logdir
 
 SPECIAL_TOKENS = ["<bos>", "<eos>", "<speaker1>", "<speaker2>", "<pad>"]
 ATTR_TO_SPECIAL_TOKEN = {'bos_token': '<bos>', 'eos_token': '<eos>', 'pad_token': '<pad>',
-                         'additional_special_tokens': ('<speaker1>', '<speaker2>')}
+                         'additional_special_tokens': ['<speaker1>', '<speaker2>']}
 MODEL_INPUTS = ["input_ids", "mc_token_ids", "lm_labels", "mc_labels", "token_type_ids"]
 PADDED_INPUTS = ["input_ids", "lm_labels", "token_type_ids"]
 
@@ -42,7 +42,7 @@ def pad_dataset(dataset, padding=0):
     """ Pad the dataset. This could be optimized by defining a Dataset class and padding at the batch level, but this is simpler. """
     max_l = max(len(x) for x in dataset["input_ids"])
     for name in PADDED_INPUTS:
-        dataset[name] = [x + [padding if name != "lm_labels" else -1] * (max_l - len(x)) for x in dataset[name]]
+        dataset[name] = [x + [padding if name != "lm_labels" else -100] * (max_l - len(x)) for x in dataset[name]]
     return dataset
 
 
@@ -62,9 +62,9 @@ def build_input_from_segments(persona, history, reply, tokenizer, lm_labels=Fals
     instance["input_ids"] = list(chain(*sequence))
     instance["token_type_ids"] = [speaker2 if i % 2 else speaker1 for i, s in enumerate(sequence) for _ in s]
     instance["mc_token_ids"] = len(instance["input_ids"]) - 1
-    instance["lm_labels"] = [-1] * len(instance["input_ids"])
+    instance["lm_labels"] = [-100] * len(instance["input_ids"])
     if lm_labels:
-        instance["lm_labels"] = ([-1] * sum(len(s) for s in sequence[:-1])) + [-1] + sequence[-1][1:]
+        instance["lm_labels"] = ([-100] * sum(len(s) for s in sequence[:-1])) + [-100] + sequence[-1][1:]
     return instance
 
 
@@ -227,7 +227,7 @@ def inference(engine, batch):
 
     # Prepare metrics - note how we compute distributed metrics
     RunningAverage(output_transform=lambda x: x).attach(trainer, "loss")
-    metrics = {"nll": Loss(torch.nn.CrossEntropyLoss(ignore_index=-1), output_transform=lambda x: (x[0][0], x[1][0])),
+    metrics = {"nll": Loss(torch.nn.CrossEntropyLoss(ignore_index=-100), output_transform=lambda x: (x[0][0], x[1][0])),
                "accuracy": Accuracy(output_transform=lambda x: (x[0][1], x[1][1]))}
     metrics.update({"average_nll": MetricsLambda(average_distributed_scalar, metrics["nll"], args),
                     "average_accuracy": MetricsLambda(average_distributed_scalar, metrics["accuracy"], args)})
@@ -260,7 +260,7 @@ def inference(engine, batch):
 
     # On the main process: close tensorboard logger and rename the last checkpoint (for easy re-loading with OpenAIGPTModel.from_pretrained method)
     if args.local_rank in [-1, 0] and args.n_epochs > 0:
-        os.rename(checkpoint_handler._saved[-1][1][-1], os.path.join(log_dir, WEIGHTS_NAME))  # TODO: PR in ignite to have better access to saved file paths (cleaner)
+        os.rename(os.path.join(log_dir, checkpoint_handler._saved[-1][1]), os.path.join(log_dir, WEIGHTS_NAME))  # TODO: PR in ignite to have better access to saved file paths (cleaner)
         tb_logger.close()
 
 if __name__ == "__main__":

utils.py

@@ -11,7 +11,7 @@
 
 import torch
 
-from pytorch_transformers import cached_path
+from transformers import cached_path
 
 PERSONACHAT_URL = "https://s3.amazonaws.com/datasets.huggingface.co/personachat/personachat_self_original.json"
 HF_FINETUNED_MODEL = "https://s3.amazonaws.com/models.huggingface.co/transfer-learning-chatbot/gpt_personachat_cache.tar.gz"