Merge pull request #3 from erenup/run_multiple_choice_merge

Run multiple choice merge

README.md

@@ -288,7 +288,7 @@ This is the model provided as `bert-large-uncased-whole-word-masking-finetuned-s
 ### `run_generation.py`: Text generation with GPT, GPT-2, Transformer-XL and XLNet
 
 A conditional generation script is also included to generate text from a prompt.
-The generation script includes the [tricks](https://github.com/rusiaaman/XLNet-gen#methodology) proposed by by Aman Rusia to get high quality generation with memory models like Transformer-XL and XLNet (include a predefined text to make short inputs longer).
+The generation script includes the [tricks](https://github.com/rusiaaman/XLNet-gen#methodology) proposed by Aman Rusia to get high quality generation with memory models like Transformer-XL and XLNet (include a predefined text to make short inputs longer).
 
 Here is how to run the script with the small version of OpenAI GPT-2 model:
 

docs/source/pretrained_models.rst

@@ -79,10 +79,10 @@ Here is the full list of the currently provided pretrained models together with
 |                   |                                                            | | XLM English model                                                                                                                   |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``xlm-mlm-ende-1024``                                      | | 6-layer, 1024-hidden, 8-heads                                                                                                       |
-|                   |                                                            | | XLM English-German Multi-language model                                                                                             |
+|                   |                                                            | | XLM English-German model trained on the concatenation of English and German wikipedia                                               |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``xlm-mlm-enfr-1024``                                      | | 6-layer, 1024-hidden, 8-heads                                                                                                       |
-|                   |                                                            | | XLM English-French Multi-language model                                                                                             |
+|                   |                                                            | | XLM English-French model trained on the concatenation of English and French wikipedia                                               |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``xlm-mlm-enro-1024``                                      | | 6-layer, 1024-hidden, 8-heads                                                                                                       |
 |                   |                                                            | | XLM English-Romanian Multi-language model                                                                                           |
@@ -93,11 +93,11 @@ Here is the full list of the currently provided pretrained models together with
 |                   | ``xlm-mlm-tlm-xnli15-1024``                                | | 12-layer, 1024-hidden, 8-heads                                                                                                      |
 |                   |                                                            | | XLM Model pre-trained with MLM + TLM on the `15 XNLI languages <https://github.com/facebookresearch/XNLI>`__.                       |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``xlm-clm-enfr-1024``                                      | | 12-layer, 1024-hidden, 8-heads                                                                                                      |
-|                   |                                                            | | XLM English model trained with CLM (Causal Language Modeling)                                                                       |
+|                   | ``xlm-clm-enfr-1024``                                      | | 6-layer, 1024-hidden, 8-heads                                                                                                       |
+|                   |                                                            | | XLM English-French model trained with CLM (Causal Language Modeling) on the concatenation of English and French wikipedia           |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``xlm-clm-ende-1024``                                      | | 6-layer, 1024-hidden, 8-heads                                                                                                       |
-|                   |                                                            | | XLM English-German Multi-language model trained with CLM (Causal Language Modeling)                                                 |
+|                   |                                                            | | XLM English-German model trained with CLM (Causal Language Modeling) on the concatenation of English and German wikipedia           |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | RoBERTa           | ``roberta-base``                                           | | 12-layer, 768-hidden, 12-heads, 125M parameters                                                                                     |
 |                   |                                                            | | RoBERTa using the BERT-base architecture                                                                                            |

examples/contrib/README.md

+# Community contributed examples
+
+This folder contains examples which are not actively maintained (mostly contributed by the community).
+
+Using these examples together with a recent version of the library usually requires to make small (sometimes big) adaptations to get the scripts working.

examples/single_model_scripts/run_openai_gpt.py → examples/contrib/run_openai_gpt.py

@@ -153,9 +153,11 @@ def main():
     # This loading functions also add new tokens and embeddings called `special tokens`
     # These new embeddings will be fine-tuned on the RocStories dataset
     special_tokens = ['_start_', '_delimiter_', '_classify_']
-    tokenizer = OpenAIGPTTokenizer.from_pretrained(args.model_name, special_tokens=special_tokens)
-    special_tokens_ids = list(tokenizer.convert_tokens_to_ids(token) for token in special_tokens)
-    model = OpenAIGPTDoubleHeadsModel.from_pretrained(args.model_name, num_special_tokens=len(special_tokens))
+    tokenizer = OpenAIGPTTokenizer.from_pretrained(args.model_name)
+    tokenizer.add_tokens(special_tokens)
+    special_tokens_ids = tokenizer.convert_tokens_to_ids(special_tokens)
+    model = OpenAIGPTDoubleHeadsModel.from_pretrained(args.model_name)
+    model.resize_token_embeddings(len(tokenizer))
     model.to(device)
 
     # Load and encode the datasets
@@ -221,7 +223,7 @@ def main():
             for step, batch in enumerate(tqdm_bar):
                 batch = tuple(t.to(device) for t in batch)
                 input_ids, mc_token_ids, lm_labels, mc_labels = batch
-                losses = model(input_ids, mc_token_ids, lm_labels, mc_labels)
+                losses = model(input_ids, mc_token_ids=mc_token_ids, lm_labels=lm_labels, mc_labels=mc_labels)
                 loss = args.lm_coef * losses[0] + losses[1]
                 loss.backward()
                 scheduler.step()
@@ -258,7 +260,7 @@ def main():
             batch = tuple(t.to(device) for t in batch)
             input_ids, mc_token_ids, lm_labels, mc_labels = batch
             with torch.no_grad():
-               _, mc_loss, _, mc_logits = model(input_ids, mc_token_ids, lm_labels, mc_labels)
+               _, mc_loss, _, mc_logits = model(input_ids, mc_token_ids=mc_token_ids, lm_labels=lm_labels, mc_labels=mc_labels)
 
             mc_logits = mc_logits.detach().cpu().numpy()
             mc_labels = mc_labels.to('cpu').numpy()

examples/single_model_scripts/run_swag.py → examples/contrib/run_swag.py

@@ -670,4 +670,4 @@ def main():
 
 
 if __name__ == "__main__":
-    main()
+    main()
\ No newline at end of file

examples/single_model_scripts/run_transfo_xl.py → examples/contrib/run_transfo_xl.py

@@ -113,7 +113,7 @@ def main():
         with torch.no_grad():
             mems = None
             for idx, (data, target, seq_len) in enumerate(eval_iter):
-                ret = model(data, target, mems)
+                ret = model(data, lm_labels=target, mems=mems)
                 loss, _, mems = ret
                 loss = loss.mean()
                 total_loss += seq_len * loss.item()

examples/lm_finetuning/README.md

-# BERT Model Finetuning using Masked Language Modeling objective
-
-## Introduction
-
-The three example scripts in this folder can be used to **fine-tune** a pre-trained BERT model using the pretraining objective (combination of masked language modeling and next sentence prediction loss). In general, pretrained models like BERT are first trained with a pretraining objective (masked language modeling and next sentence prediction for BERT) on a large and general natural language corpus. A classifier head is then added on top of the pre-trained architecture and the model is quickly fine-tuned on a target task, while still (hopefully) retaining its general language understanding. This greatly reduces overfitting and yields state-of-the-art results, especially when training data for the target task are limited.
-
-The [ULMFiT paper](https://arxiv.org/abs/1801.06146) took a slightly different approach, however, and added an intermediate step in which the model is fine-tuned on text **from the same domain as the target task and using the pretraining objective** before the final stage in which the classifier head is added and the model is trained on the target task itself. This paper reported significantly improved results from this step, and found that they could get high-quality classifications even with only tiny numbers (<1000) of labelled training examples, as long as they had a lot of unlabelled data from the target domain.
-
-Although this wasn't covered in the original BERT paper, domain-specific fine-tuning of Transformer models has [recently been reported by other authors](https://arxiv.org/pdf/1905.05583.pdf), and they report performance improvements as well.
-
-## Input format
-
-The scripts in this folder expect a single file as input, consisting of untokenized text, with one **sentence** per line, and one blank line between documents. The reason for the sentence splitting is that part of BERT's training involves a _next sentence_ objective in which the model must predict whether two sequences of text are contiguous text from the same document or not, and to avoid making the task _too easy_, the split point between the sequences is always at the end of a sentence. The linebreaks in the file are therefore necessary to mark the points where the text can be split.
-
-## Usage
-
-There are two ways to fine-tune a language model using these scripts. The first _quick_ approach is to use [`simple_lm_finetuning.py`](./simple_lm_finetuning.py). This script does everything in a single script, but generates training instances that consist of just two sentences. This is quite different from the BERT paper, where (confusingly) the NextSentence task concatenated sentences together from each document to form two long multi-sentences, which the paper just referred to as _sentences_. The difference between this simple approach and the original paper approach can have a significant effect for long sequences since two sentences will be much shorter than the max sequence length. In this case, most of each training example will just consist of blank padding characters, which wastes a lot of computation and results in a model that isn't really training on long sequences.
-
-As such, the preferred approach (assuming you have documents containing multiple contiguous sentences from your target domain) is to use [`pregenerate_training_data.py`](./pregenerate_training_data.py) to pre-process your data into training examples following the methodology used for LM training in the original BERT paper and repository. Since there is a significant random component to training data generation for BERT, this script includes an option to generate multiple _epochs_ of pre-processed data, to avoid training on the same random splits each epoch. Generating an epoch of data for each training epoch should result a better final model, and so we recommend doing so.
-
-You can then train on the pregenerated data using [`finetune_on_pregenerated.py`](./finetune_on_pregenerated.py), and pointing it to the folder created by [`pregenerate_training_data.py`](./pregenerate_training_data.py). Note that you should use the same `bert_model` and case options for both! Also note that `max_seq_len` does not need to be specified for the [`finetune_on_pregenerated.py`](./finetune_on_pregenerated.py) script, as it is inferred from the training examples.
-
-There are various options that can be tweaked, but they are mostly set to the values from the BERT paper/repository and default values should make sense. The most relevant ones are:
-
-- `--max_seq_len`: Controls the length of training examples (in wordpiece tokens) seen by the model. Defaults to 128 but can be set as high as 512. Higher values may yield stronger language models at the cost of slower and more memory-intensive training.
-- `--fp16`: Enables fast half-precision training on recent GPUs.
-
-In addition, if memory usage is an issue, especially when training on a single GPU, reducing `--train_batch_size` from the default 32 to a lower number (4-16) can be helpful, or leaving `--train_batch_size` at the default and increasing `--gradient_accumulation_steps` to 2-8. Changing `--gradient_accumulation_steps` may be preferable as alterations to the batch size may require corresponding changes in the learning rate to compensate. There is also a `--reduce_memory` option for both the `pregenerate_training_data.py` and `finetune_on_pregenerated.py` scripts that spills data to disc in shelf objects or numpy memmaps rather than retaining it in memory, which significantly reduces memory usage with little performance impact.
-
-## Examples
-
-### Simple fine-tuning
-
-```
-python3 simple_lm_finetuning.py 
---train_corpus my_corpus.txt 
---bert_model bert-base-uncased 
---do_lower_case 
---output_dir finetuned_lm/
---do_train
-```
-
-### Pregenerating training data
-
-```
-python3 pregenerate_training_data.py
---train_corpus my_corpus.txt
---bert_model bert-base-uncased
---do_lower_case
---output_dir training/
---epochs_to_generate 3
---max_seq_len 256
-```
-
-### Training on pregenerated data
-
-```
-python3 finetune_on_pregenerated.py
---pregenerated_data training/
---bert_model bert-base-uncased
---do_lower_case
---output_dir finetuned_lm/
---epochs 3
-```

examples/lm_finetuning/finetune_on_pregenerated.py

-from argparse import ArgumentParser
-from pathlib import Path
-import os
-import torch
-import logging
-import json
-import random
-import numpy as np
-from collections import namedtuple
-from tempfile import TemporaryDirectory
-
-from torch.utils.data import DataLoader, Dataset, RandomSampler
-from torch.utils.data.distributed import DistributedSampler
-from tqdm import tqdm
-
-from pytorch_transformers import WEIGHTS_NAME, CONFIG_NAME
-from pytorch_transformers.modeling_bert import BertForPreTraining
-from pytorch_transformers.tokenization_bert import BertTokenizer
-from pytorch_transformers.optimization import AdamW, WarmupLinearSchedule
-
-InputFeatures = namedtuple("InputFeatures", "input_ids input_mask segment_ids lm_label_ids is_next")
-
-log_format = '%(asctime)-10s: %(message)s'
-logging.basicConfig(level=logging.INFO, format=log_format)
-
-
-def convert_example_to_features(example, tokenizer, max_seq_length):
-    tokens = example["tokens"]
-    segment_ids = example["segment_ids"]
-    is_random_next = example["is_random_next"]
-    masked_lm_positions = example["masked_lm_positions"]
-    masked_lm_labels = example["masked_lm_labels"]
-
-    assert len(tokens) == len(segment_ids) <= max_seq_length  # The preprocessed data should be already truncated
-    input_ids = tokenizer.convert_tokens_to_ids(tokens)
-    masked_label_ids = tokenizer.convert_tokens_to_ids(masked_lm_labels)
-
-    input_array = np.zeros(max_seq_length, dtype=np.int)
-    input_array[:len(input_ids)] = input_ids
-
-    mask_array = np.zeros(max_seq_length, dtype=np.bool)
-    mask_array[:len(input_ids)] = 1
-
-    segment_array = np.zeros(max_seq_length, dtype=np.bool)
-    segment_array[:len(segment_ids)] = segment_ids
-
-    lm_label_array = np.full(max_seq_length, dtype=np.int, fill_value=-1)
-    lm_label_array[masked_lm_positions] = masked_label_ids
-
-    features = InputFeatures(input_ids=input_array,
-                             input_mask=mask_array,
-                             segment_ids=segment_array,
-                             lm_label_ids=lm_label_array,
-                             is_next=is_random_next)
-    return features
-
-
-class PregeneratedDataset(Dataset):
-    def __init__(self, training_path, epoch, tokenizer, num_data_epochs, reduce_memory=False):
-        self.vocab = tokenizer.vocab
-        self.tokenizer = tokenizer
-        self.epoch = epoch
-        self.data_epoch = epoch % num_data_epochs
-        data_file = training_path / f"epoch_{self.data_epoch}.json"
-        metrics_file = training_path / f"epoch_{self.data_epoch}_metrics.json"
-        assert data_file.is_file() and metrics_file.is_file()
-        metrics = json.loads(metrics_file.read_text())
-        num_samples = metrics['num_training_examples']
-        seq_len = metrics['max_seq_len']
-        self.temp_dir = None
-        self.working_dir = None
-        if reduce_memory:
-            self.temp_dir = TemporaryDirectory()
-            self.working_dir = Path(self.temp_dir.name)
-            input_ids = np.memmap(filename=self.working_dir/'input_ids.memmap',
-                                  mode='w+', dtype=np.int32, shape=(num_samples, seq_len))
-            input_masks = np.memmap(filename=self.working_dir/'input_masks.memmap',
-                                    shape=(num_samples, seq_len), mode='w+', dtype=np.bool)
-            segment_ids = np.memmap(filename=self.working_dir/'segment_ids.memmap',
-                                    shape=(num_samples, seq_len), mode='w+', dtype=np.bool)
-            lm_label_ids = np.memmap(filename=self.working_dir/'lm_label_ids.memmap',
-                                     shape=(num_samples, seq_len), mode='w+', dtype=np.int32)
-            lm_label_ids[:] = -1
-            is_nexts = np.memmap(filename=self.working_dir/'is_nexts.memmap',
-                                 shape=(num_samples,), mode='w+', dtype=np.bool)
-        else:
-            input_ids = np.zeros(shape=(num_samples, seq_len), dtype=np.int32)
-            input_masks = np.zeros(shape=(num_samples, seq_len), dtype=np.bool)
-            segment_ids = np.zeros(shape=(num_samples, seq_len), dtype=np.bool)
-            lm_label_ids = np.full(shape=(num_samples, seq_len), dtype=np.int32, fill_value=-1)
-            is_nexts = np.zeros(shape=(num_samples,), dtype=np.bool)
-        logging.info(f"Loading training examples for epoch {epoch}")
-        with data_file.open() as f:
-            for i, line in enumerate(tqdm(f, total=num_samples, desc="Training examples")):
-                line = line.strip()
-                example = json.loads(line)
-                features = convert_example_to_features(example, tokenizer, seq_len)
-                input_ids[i] = features.input_ids
-                segment_ids[i] = features.segment_ids
-                input_masks[i] = features.input_mask
-                lm_label_ids[i] = features.lm_label_ids
-                is_nexts[i] = features.is_next
-        assert i == num_samples - 1  # Assert that the sample count metric was true
-        logging.info("Loading complete!")
-        self.num_samples = num_samples
-        self.seq_len = seq_len
-        self.input_ids = input_ids
-        self.input_masks = input_masks
-        self.segment_ids = segment_ids
-        self.lm_label_ids = lm_label_ids
-        self.is_nexts = is_nexts
-
-    def __len__(self):
-        return self.num_samples
-
-    def __getitem__(self, item):
-        return (torch.tensor(self.input_ids[item].astype(np.int64)),
-                torch.tensor(self.input_masks[item].astype(np.int64)),
-                torch.tensor(self.segment_ids[item].astype(np.int64)),
-                torch.tensor(self.lm_label_ids[item].astype(np.int64)),
-                torch.tensor(self.is_nexts[item].astype(np.int64)))
-
-
-def main():
-    parser = ArgumentParser()
-    parser.add_argument('--pregenerated_data', type=Path, required=True)
-    parser.add_argument('--output_dir', type=Path, required=True)
-    parser.add_argument("--bert_model", type=str, required=True, help="Bert pre-trained model selected in the list: bert-base-uncased, "
-                             "bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.")
-    parser.add_argument("--do_lower_case", action="store_true")
-    parser.add_argument("--reduce_memory", action="store_true",
-                        help="Store training data as on-disc memmaps to massively reduce memory usage")
-
-    parser.add_argument("--epochs", type=int, default=3, help="Number of epochs to train for")
-    parser.add_argument("--local_rank",
-                        type=int,
-                        default=-1,
-                        help="local_rank for distributed training on gpus")
-    parser.add_argument("--no_cuda",
-                        action='store_true',
-                        help="Whether not to use CUDA when available")
-    parser.add_argument('--gradient_accumulation_steps',
-                        type=int,
-                        default=1,
-                        help="Number of updates steps to accumulate before performing a backward/update pass.")
-    parser.add_argument("--train_batch_size",
-                        default=32,
-                        type=int,
-                        help="Total batch size for training.")
-    parser.add_argument('--fp16',
-                        action='store_true',
-                        help="Whether to use 16-bit float precision instead of 32-bit")
-    parser.add_argument('--loss_scale',
-                        type=float, default=0,
-                        help="Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
-                        "0 (default value): dynamic loss scaling.\n"
-                        "Positive power of 2: static loss scaling value.\n")
-    parser.add_argument("--warmup_steps", 
-                        default=0, 
-                        type=int,
-                        help="Linear warmup over warmup_steps.")
-    parser.add_argument("--adam_epsilon", 
-                        default=1e-8, 
-                        type=float,
-                        help="Epsilon for Adam optimizer.")
-    parser.add_argument("--learning_rate",
-                        default=3e-5,
-                        type=float,
-                        help="The initial learning rate for Adam.")
-    parser.add_argument('--seed',
-                        type=int,
-                        default=42,
-                        help="random seed for initialization")
-    args = parser.parse_args()
-
-    assert args.pregenerated_data.is_dir(), \
-        "--pregenerated_data should point to the folder of files made by pregenerate_training_data.py!"
-
-    samples_per_epoch = []
-    for i in range(args.epochs):
-        epoch_file = args.pregenerated_data / f"epoch_{i}.json"
-        metrics_file = args.pregenerated_data / f"epoch_{i}_metrics.json"
-        if epoch_file.is_file() and metrics_file.is_file():
-            metrics = json.loads(metrics_file.read_text())
-            samples_per_epoch.append(metrics['num_training_examples'])
-        else:
-            if i == 0:
-                exit("No training data was found!")
-            print(f"Warning! There are fewer epochs of pregenerated data ({i}) than training epochs ({args.epochs}).")
-            print("This script will loop over the available data, but training diversity may be negatively impacted.")
-            num_data_epochs = i
-            break
-    else:
-        num_data_epochs = args.epochs
-
-    if args.local_rank == -1 or args.no_cuda:
-        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        n_gpu = torch.cuda.device_count()
-    else:
-        torch.cuda.set_device(args.local_rank)
-        device = torch.device("cuda", args.local_rank)
-        n_gpu = 1
-        # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
-        torch.distributed.init_process_group(backend='nccl')
-    logging.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
-        device, n_gpu, bool(args.local_rank != -1), args.fp16))
-
-    if args.gradient_accumulation_steps < 1:
-        raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
-                            args.gradient_accumulation_steps))
-
-    args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
-
-    random.seed(args.seed)
-    np.random.seed(args.seed)
-    torch.manual_seed(args.seed)
-    if n_gpu > 0:
-        torch.cuda.manual_seed_all(args.seed)
-
-    if args.output_dir.is_dir() and list(args.output_dir.iterdir()):
-        logging.warning(f"Output directory ({args.output_dir}) already exists and is not empty!")
-    args.output_dir.mkdir(parents=True, exist_ok=True)
-
-    tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
-
-    total_train_examples = 0
-    for i in range(args.epochs):
-        # The modulo takes into account the fact that we may loop over limited epochs of data
-        total_train_examples += samples_per_epoch[i % len(samples_per_epoch)]
-
-    num_train_optimization_steps = int(
-        total_train_examples / args.train_batch_size / args.gradient_accumulation_steps)
-    if args.local_rank != -1:
-        num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
-
-    # Prepare model
-    model = BertForPreTraining.from_pretrained(args.bert_model)
-    # We don't need to manually call model.half() following Apex's recommend
-    # if args.fp16:
-    #     model.half()
-    model.to(device)
-    if args.local_rank != -1:
-        try:
-            from apex.parallel import DistributedDataParallel as DDP
-        except ImportError:
-            raise ImportError(
-                "Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
-        model = DDP(model)
-    elif n_gpu > 1:
-        model = torch.nn.DataParallel(model)
-
-    # Prepare optimizer
-    param_optimizer = list(model.named_parameters())
-    no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
-    optimizer_grouped_parameters = [
-        {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
-         'weight_decay': 0.01},
-        {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
-    ]
-
-    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps,
-                                     t_total=num_train_optimization_steps)
-
-    if args.fp16:
-        try:
-            # from apex.optimizers import FP16_Optimizer
-            # from apex.optimizers import FusedAdam
-            from apex import amp
-        except ImportError:
-            raise ImportError(
-                "Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
-
-        # This below line of code is the main upgrade of Apex Fp16 implementation. I chose opt_leve="01"
-        # because it's recommended for typical use by Apex. We can make it configured
-        model, optimizer = amp.initialize(model, optimizer, opt_level="O1")
-
-    # We don't need to use FP16_Optimizer wrapping over FusedAdam as well. Now Apex supports all Pytorch Optimizer
-
-    #     optimizer = FusedAdam(optimizer_grouped_parameters,
-    #                           lr=args.learning_rate,
-    #                           bias_correction=False,
-    #                           max_grad_norm=1.0)
-    #     if args.loss_scale == 0:
-    #         optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
-    #     else:
-    #         optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
-    # else:
-    #     optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    # scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=num_train_optimization_steps)
-
-    global_step = 0
-    logging.info("***** Running training *****")
-    logging.info(f"  Num examples = {total_train_examples}")
-    logging.info("  Batch size = %d", args.train_batch_size)
-    logging.info("  Num steps = %d", num_train_optimization_steps)
-    model.train()
-    for epoch in range(args.epochs):
-        epoch_dataset = PregeneratedDataset(epoch=epoch, training_path=args.pregenerated_data, tokenizer=tokenizer,
-                                            num_data_epochs=num_data_epochs, reduce_memory=args.reduce_memory)
-        if args.local_rank == -1:
-            train_sampler = RandomSampler(epoch_dataset)
-        else:
-            train_sampler = DistributedSampler(epoch_dataset)
-        train_dataloader = DataLoader(epoch_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
-        tr_loss = 0
-        nb_tr_examples, nb_tr_steps = 0, 0
-        with tqdm(total=len(train_dataloader), desc=f"Epoch {epoch}") as pbar:
-            for step, batch in enumerate(train_dataloader):
-                batch = tuple(t.to(device) for t in batch)
-                input_ids, input_mask, segment_ids, lm_label_ids, is_next = batch
-                outputs = model(input_ids, segment_ids, input_mask, lm_label_ids, is_next)
-                loss = outputs[0]
-                if n_gpu > 1:
-                    loss = loss.mean() # mean() to average on multi-gpu.
-                if args.gradient_accumulation_steps > 1:
-                    loss = loss / args.gradient_accumulation_steps
-                if args.fp16:
-                    # I depricate FP16_Optimizer's backward func and replace as Apex document
-                    # optimizer.backward(loss)
-                    with amp.scale_loss(loss, optimizer) as scaled_loss:
-                        scaled_loss.backward()
-                else:
-                    loss.backward()
-                tr_loss += loss.item()
-                nb_tr_examples += input_ids.size(0)
-                nb_tr_steps += 1
-                pbar.update(1)
-                mean_loss = tr_loss * args.gradient_accumulation_steps / nb_tr_steps
-                pbar.set_postfix_str(f"Loss: {mean_loss:.5f}")
-                if (step + 1) % args.gradient_accumulation_steps == 0:
-                    optimizer.step()
-                    scheduler.step()  # Update learning rate schedule
-                    optimizer.zero_grad()
-                    global_step += 1
-
-    # Save a trained model
-    if args.local_rank == -1 or torch.distributed.get_rank() == 0:
-        logging.info("** ** * Saving fine-tuned model ** ** * ")
-        model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
-        model_to_save.save_pretrained(args.output_dir)
-        tokenizer.save_pretrained(args.output_dir)
-
-
-if __name__ == '__main__':
-    main()

examples/lm_finetuning/pregenerate_training_data.py

-from argparse import ArgumentParser
-from pathlib import Path
-from tqdm import tqdm, trange
-from tempfile import TemporaryDirectory
-import shelve
-from multiprocessing import Pool
-
-from random import random, randrange, randint, shuffle, choice
-from pytorch_transformers.tokenization_bert import BertTokenizer
-import numpy as np
-import json
-import collections
-
-class DocumentDatabase:
-    def __init__(self, reduce_memory=False):
-        if reduce_memory:
-            self.temp_dir = TemporaryDirectory()
-            self.working_dir = Path(self.temp_dir.name)
-            self.document_shelf_filepath = self.working_dir / 'shelf.db'
-            self.document_shelf = shelve.open(str(self.document_shelf_filepath),
-                                              flag='n', protocol=-1)
-            self.documents = None
-        else:
-            self.documents = []
-            self.document_shelf = None
-            self.document_shelf_filepath = None
-            self.temp_dir = None
-        self.doc_lengths = []
-        self.doc_cumsum = None
-        self.cumsum_max = None
-        self.reduce_memory = reduce_memory
-
-    def add_document(self, document):
-        if not document:
-            return
-        if self.reduce_memory:
-            current_idx = len(self.doc_lengths)
-            self.document_shelf[str(current_idx)] = document
-        else:
-            self.documents.append(document)
-        self.doc_lengths.append(len(document))
-
-    def _precalculate_doc_weights(self):
-        self.doc_cumsum = np.cumsum(self.doc_lengths)
-        self.cumsum_max = self.doc_cumsum[-1]
-
-    def sample_doc(self, current_idx, sentence_weighted=True):
-        # Uses the current iteration counter to ensure we don't sample the same doc twice
-        if sentence_weighted:
-            # With sentence weighting, we sample docs proportionally to their sentence length
-            if self.doc_cumsum is None or len(self.doc_cumsum) != len(self.doc_lengths):
-                self._precalculate_doc_weights()
-            rand_start = self.doc_cumsum[current_idx]
-            rand_end = rand_start + self.cumsum_max - self.doc_lengths[current_idx]
-            sentence_index = randrange(rand_start, rand_end) % self.cumsum_max
-            sampled_doc_index = np.searchsorted(self.doc_cumsum, sentence_index, side='right')
-        else:
-            # If we don't use sentence weighting, then every doc has an equal chance to be chosen
-            sampled_doc_index = (current_idx + randrange(1, len(self.doc_lengths))) % len(self.doc_lengths)
-        assert sampled_doc_index != current_idx
-        if self.reduce_memory:
-            return self.document_shelf[str(sampled_doc_index)]
-        else:
-            return self.documents[sampled_doc_index]
-
-    def __len__(self):
-        return len(self.doc_lengths)
-
-    def __getitem__(self, item):
-        if self.reduce_memory:
-            return self.document_shelf[str(item)]
-        else:
-            return self.documents[item]
-
-    def __enter__(self):
-        return self
-
-    def __exit__(self, exc_type, exc_val, traceback):
-        if self.document_shelf is not None:
-            self.document_shelf.close()
-        if self.temp_dir is not None:
-            self.temp_dir.cleanup()
-
-
-def truncate_seq_pair(tokens_a, tokens_b, max_num_tokens):
-    """Truncates a pair of sequences to a maximum sequence length. Lifted from Google's BERT repo."""
-    while True:
-        total_length = len(tokens_a) + len(tokens_b)
-        if total_length <= max_num_tokens:
-            break
-
-        trunc_tokens = tokens_a if len(tokens_a) > len(tokens_b) else tokens_b
-        assert len(trunc_tokens) >= 1
-
-        # We want to sometimes truncate from the front and sometimes from the
-        # back to add more randomness and avoid biases.
-        if random() < 0.5:
-            del trunc_tokens[0]
-        else:
-            trunc_tokens.pop()
-
-MaskedLmInstance = collections.namedtuple("MaskedLmInstance",
-                                          ["index", "label"])
-
-def create_masked_lm_predictions(tokens, masked_lm_prob, max_predictions_per_seq, whole_word_mask, vocab_list):
-    """Creates the predictions for the masked LM objective. This is mostly copied from the Google BERT repo, but
-    with several refactors to clean it up and remove a lot of unnecessary variables."""
-    cand_indices = []
-    for (i, token) in enumerate(tokens):
-        if token == "[CLS]" or token == "[SEP]":
-            continue
-        # Whole Word Masking means that if we mask all of the wordpieces
-        # corresponding to an original word. When a word has been split into
-        # WordPieces, the first token does not have any marker and any subsequence
-        # tokens are prefixed with ##. So whenever we see the ## token, we
-        # append it to the previous set of word indexes.
-        #
-        # Note that Whole Word Masking does *not* change the training code
-        # at all -- we still predict each WordPiece independently, softmaxed
-        # over the entire vocabulary.
-        if (whole_word_mask and len(cand_indices) >= 1 and token.startswith("##")):
-            cand_indices[-1].append(i)
-        else:
-            cand_indices.append([i])
-
-    num_to_mask = min(max_predictions_per_seq,
-                      max(1, int(round(len(tokens) * masked_lm_prob))))
-    shuffle(cand_indices)
-    masked_lms = []
-    covered_indexes = set()
-    for index_set in cand_indices:
-        if len(masked_lms) >= num_to_mask:
-            break
-        # If adding a whole-word mask would exceed the maximum number of
-        # predictions, then just skip this candidate.
-        if len(masked_lms) + len(index_set) > num_to_mask:
-            continue
-        is_any_index_covered = False
-        for index in index_set:
-            if index in covered_indexes:
-                is_any_index_covered = True
-                break
-        if is_any_index_covered:
-            continue
-        for index in index_set:
-            covered_indexes.add(index)
-
-            masked_token = None
-            # 80% of the time, replace with [MASK]
-            if random() < 0.8:
-                masked_token = "[MASK]"
-            else:
-                # 10% of the time, keep original
-                if random() < 0.5:
-                    masked_token = tokens[index]
-                # 10% of the time, replace with random word
-                else:
-                    masked_token = choice(vocab_list)
-            masked_lms.append(MaskedLmInstance(index=index, label=tokens[index]))
-            tokens[index] = masked_token
-
-    assert len(masked_lms) <= num_to_mask
-    masked_lms = sorted(masked_lms, key=lambda x: x.index)
-    mask_indices = [p.index for p in masked_lms]
-    masked_token_labels = [p.label for p in masked_lms]
-
-    return tokens, mask_indices, masked_token_labels
-
-
-def create_instances_from_document(
-        doc_database, doc_idx, max_seq_length, short_seq_prob,
-        masked_lm_prob, max_predictions_per_seq, whole_word_mask, vocab_list):
-    """This code is mostly a duplicate of the equivalent function from Google BERT's repo.
-    However, we make some changes and improvements. Sampling is improved and no longer requires a loop in this function.
-    Also, documents are sampled proportionally to the number of sentences they contain, which means each sentence
-    (rather than each document) has an equal chance of being sampled as a false example for the NextSentence task."""
-    document = doc_database[doc_idx]
-    # Account for [CLS], [SEP], [SEP]
-    max_num_tokens = max_seq_length - 3
-
-    # We *usually* want to fill up the entire sequence since we are padding
-    # to `max_seq_length` anyways, so short sequences are generally wasted
-    # computation. However, we *sometimes*
-    # (i.e., short_seq_prob == 0.1 == 10% of the time) want to use shorter
-    # sequences to minimize the mismatch between pre-training and fine-tuning.
-    # The `target_seq_length` is just a rough target however, whereas
-    # `max_seq_length` is a hard limit.
-    target_seq_length = max_num_tokens
-    if random() < short_seq_prob:
-        target_seq_length = randint(2, max_num_tokens)
-
-    # We DON'T just concatenate all of the tokens from a document into a long
-    # sequence and choose an arbitrary split point because this would make the
-    # next sentence prediction task too easy. Instead, we split the input into
-    # segments "A" and "B" based on the actual "sentences" provided by the user
-    # input.
-    instances = []
-    current_chunk = []
-    current_length = 0
-    i = 0
-    while i < len(document):
-        segment = document[i]
-        current_chunk.append(segment)
-        current_length += len(segment)
-        if i == len(document) - 1 or current_length >= target_seq_length:
-            if current_chunk:
-                # `a_end` is how many segments from `current_chunk` go into the `A`
-                # (first) sentence.
-                a_end = 1
-                if len(current_chunk) >= 2:
-                    a_end = randrange(1, len(current_chunk))
-
-                tokens_a = []
-                for j in range(a_end):
-                    tokens_a.extend(current_chunk[j])
-
-                tokens_b = []
-
-                # Random next
-                if len(current_chunk) == 1 or random() < 0.5:
-                    is_random_next = True
-                    target_b_length = target_seq_length - len(tokens_a)
-
-                    # Sample a random document, with longer docs being sampled more frequently
-                    random_document = doc_database.sample_doc(current_idx=doc_idx, sentence_weighted=True)
-
-                    random_start = randrange(0, len(random_document))
-                    for j in range(random_start, len(random_document)):
-                        tokens_b.extend(random_document[j])
-                        if len(tokens_b) >= target_b_length:
-                            break
-                    # We didn't actually use these segments so we "put them back" so
-                    # they don't go to waste.
-                    num_unused_segments = len(current_chunk) - a_end
-                    i -= num_unused_segments
-                # Actual next
-                else:
-                    is_random_next = False
-                    for j in range(a_end, len(current_chunk)):
-                        tokens_b.extend(current_chunk[j])
-                truncate_seq_pair(tokens_a, tokens_b, max_num_tokens)
-
-                assert len(tokens_a) >= 1
-                assert len(tokens_b) >= 1
-
-                tokens = ["[CLS]"] + tokens_a + ["[SEP]"] + tokens_b + ["[SEP]"]
-                # The segment IDs are 0 for the [CLS] token, the A tokens and the first [SEP]
-                # They are 1 for the B tokens and the final [SEP]
-                segment_ids = [0 for _ in range(len(tokens_a) + 2)] + [1 for _ in range(len(tokens_b) + 1)]
-
-                tokens, masked_lm_positions, masked_lm_labels = create_masked_lm_predictions(
-                    tokens, masked_lm_prob, max_predictions_per_seq, whole_word_mask, vocab_list)
-
-                instance = {
-                    "tokens": tokens,
-                    "segment_ids": segment_ids,
-                    "is_random_next": is_random_next,
-                    "masked_lm_positions": masked_lm_positions,
-                    "masked_lm_labels": masked_lm_labels}
-                instances.append(instance)
-            current_chunk = []
-            current_length = 0
-        i += 1
-
-    return instances
-
-
-def create_training_file(docs, vocab_list, args, epoch_num):
-    epoch_filename = args.output_dir / "epoch_{}.json".format(epoch_num)
-    num_instances = 0
-    with epoch_filename.open('w') as epoch_file:
-        for doc_idx in trange(len(docs), desc="Document"):
-            doc_instances = create_instances_from_document(
-                docs, doc_idx, max_seq_length=args.max_seq_len, short_seq_prob=args.short_seq_prob,
-                masked_lm_prob=args.masked_lm_prob, max_predictions_per_seq=args.max_predictions_per_seq,
-                whole_word_mask=args.do_whole_word_mask, vocab_list=vocab_list)
-            doc_instances = [json.dumps(instance) for instance in doc_instances]
-            for instance in doc_instances:
-                epoch_file.write(instance + '\n')
-                num_instances += 1
-    metrics_file = args.output_dir / "epoch_{}_metrics.json".format(epoch_num)
-    with metrics_file.open('w') as metrics_file:
-        metrics = {
-            "num_training_examples": num_instances,
-            "max_seq_len": args.max_seq_len
-        }
-        metrics_file.write(json.dumps(metrics))
-
-
-def main():
-    parser = ArgumentParser()
-    parser.add_argument('--train_corpus', type=Path, required=True)
-    parser.add_argument("--output_dir", type=Path, required=True)
-    parser.add_argument("--bert_model", type=str, required=True,
-                        choices=["bert-base-uncased", "bert-large-uncased", "bert-base-cased",
-                                 "bert-base-multilingual-uncased", "bert-base-chinese", "bert-base-multilingual-cased"])
-    parser.add_argument("--do_lower_case", action="store_true")
-    parser.add_argument("--do_whole_word_mask", action="store_true",
-                        help="Whether to use whole word masking rather than per-WordPiece masking.")
-    parser.add_argument("--reduce_memory", action="store_true",
-                        help="Reduce memory usage for large datasets by keeping data on disc rather than in memory")
-
-    parser.add_argument("--num_workers", type=int, default=1,
-                        help="The number of workers to use to write the files")
-    parser.add_argument("--epochs_to_generate", type=int, default=3,
-                        help="Number of epochs of data to pregenerate")
-    parser.add_argument("--max_seq_len", type=int, default=128)
-    parser.add_argument("--short_seq_prob", type=float, default=0.1,
-                        help="Probability of making a short sentence as a training example")
-    parser.add_argument("--masked_lm_prob", type=float, default=0.15,
-                        help="Probability of masking each token for the LM task")
-    parser.add_argument("--max_predictions_per_seq", type=int, default=20,
-                        help="Maximum number of tokens to mask in each sequence")
-
-    args = parser.parse_args()
-
-    if args.num_workers > 1 and args.reduce_memory:
-        raise ValueError("Cannot use multiple workers while reducing memory")
-
-    tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
-    vocab_list = list(tokenizer.vocab.keys())
-    with DocumentDatabase(reduce_memory=args.reduce_memory) as docs:
-        with args.train_corpus.open() as f:
-            doc = []
-            for line in tqdm(f, desc="Loading Dataset", unit=" lines"):
-                line = line.strip()
-                if line == "":
-                    docs.add_document(doc)
-                    doc = []
-                else:
-                    tokens = tokenizer.tokenize(line)
-                    if tokens:
-                        doc.append(tokens)
-            if doc:
-                docs.add_document(doc)  # If the last doc didn't end on a newline, make sure it still gets added
-        if len(docs) <= 1:
-            exit("ERROR: No document breaks were found in the input file! These are necessary to allow the script to "
-                 "ensure that random NextSentences are not sampled from the same document. Please add blank lines to "
-                 "indicate breaks between documents in your input file. If your dataset does not contain multiple "
-                 "documents, blank lines can be inserted at any natural boundary, such as the ends of chapters, "
-                 "sections or paragraphs.")
-
-        args.output_dir.mkdir(exist_ok=True)
-
-        if args.num_workers > 1:
-            writer_workers = Pool(min(args.num_workers, args.epochs_to_generate))
-            arguments = [(docs, vocab_list, args, idx) for idx in range(args.epochs_to_generate)]
-            writer_workers.starmap(create_training_file, arguments)
-        else:
-            for epoch in trange(args.epochs_to_generate, desc="Epoch"):
-                create_training_file(docs, vocab_list, args, epoch)
-
-
-if __name__ == '__main__':
-    main()

examples/lm_finetuning/simple_lm_finetuning.py

-# coding=utf-8
-# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
-# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""BERT finetuning runner."""
-
-from __future__ import absolute_import, division, print_function, unicode_literals
-
-import argparse
-import logging
-import os
-import random
-from io import open
-
-import numpy as np
-import torch
-from torch.utils.data import DataLoader, Dataset, RandomSampler
-from torch.utils.data.distributed import DistributedSampler
-from tqdm import tqdm, trange
-
-from pytorch_transformers import WEIGHTS_NAME, CONFIG_NAME
-from pytorch_transformers.modeling_bert import BertForPreTraining
-from pytorch_transformers.tokenization_bert import BertTokenizer
-from pytorch_transformers.optimization import AdamW, WarmupLinearSchedule
-
-logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
-                    datefmt='%m/%d/%Y %H:%M:%S',
-                    level=logging.INFO)
-logger = logging.getLogger(__name__)
-
-
-class BERTDataset(Dataset):
-    def __init__(self, corpus_path, tokenizer, seq_len, encoding="utf-8", corpus_lines=None, on_memory=True):
-        self.vocab = tokenizer.vocab
-        self.tokenizer = tokenizer
-        self.seq_len = seq_len
-        self.on_memory = on_memory
-        self.corpus_lines = corpus_lines  # number of non-empty lines in input corpus
-        self.corpus_path = corpus_path
-        self.encoding = encoding
-        self.current_doc = 0  # to avoid random sentence from same doc
-
-        # for loading samples directly from file
-        self.sample_counter = 0  # used to keep track of full epochs on file
-        self.line_buffer = None  # keep second sentence of a pair in memory and use as first sentence in next pair
-
-        # for loading samples in memory
-        self.current_random_doc = 0
-        self.num_docs = 0
-        self.sample_to_doc = [] # map sample index to doc and line
-
-        # load samples into memory
-        if on_memory:
-            self.all_docs = []
-            doc = []
-            self.corpus_lines = 0
-            with open(corpus_path, "r", encoding=encoding) as f:
-                for line in tqdm(f, desc="Loading Dataset", total=corpus_lines):
-                    line = line.strip()
-                    if line == "":
-                        self.all_docs.append(doc)
-                        doc = []
-                        #remove last added sample because there won't be a subsequent line anymore in the doc
-                        self.sample_to_doc.pop()
-                    else:
-                        #store as one sample
-                        sample = {"doc_id": len(self.all_docs),
-                                  "line": len(doc)}
-                        self.sample_to_doc.append(sample)
-                        doc.append(line)
-                        self.corpus_lines = self.corpus_lines + 1
-
-            # if last row in file is not empty
-            if self.all_docs[-1] != doc:
-                self.all_docs.append(doc)
-                self.sample_to_doc.pop()
-
-            self.num_docs = len(self.all_docs)
-
-        # load samples later lazily from disk
-        else:
-            if self.corpus_lines is None:
-                with open(corpus_path, "r", encoding=encoding) as f:
-                    self.corpus_lines = 0
-                    for line in tqdm(f, desc="Loading Dataset", total=corpus_lines):
-                        if line.strip() == "":
-                            self.num_docs += 1
-                        else:
-                            self.corpus_lines += 1
-
-                    # if doc does not end with empty line
-                    if line.strip() != "":
-                        self.num_docs += 1
-
-            self.file = open(corpus_path, "r", encoding=encoding)
-            self.random_file = open(corpus_path, "r", encoding=encoding)
-
-    def __len__(self):
-        # last line of doc won't be used, because there's no "nextSentence". Additionally, we start counting at 0.
-        return self.corpus_lines - self.num_docs - 1
-
-    def __getitem__(self, item):
-        cur_id = self.sample_counter
-        self.sample_counter += 1
-        if not self.on_memory:
-            # after one epoch we start again from beginning of file
-            if cur_id != 0 and (cur_id % len(self) == 0):
-                self.file.close()
-                self.file = open(self.corpus_path, "r", encoding=self.encoding)
-
-        t1, t2, is_next_label = self.random_sent(item)
-
-        # tokenize
-        tokens_a = self.tokenizer.tokenize(t1)
-        tokens_b = self.tokenizer.tokenize(t2)
-
-        # combine to one sample
-        cur_example = InputExample(guid=cur_id, tokens_a=tokens_a, tokens_b=tokens_b, is_next=is_next_label)
-
-        # transform sample to features
-        cur_features = convert_example_to_features(cur_example, self.seq_len, self.tokenizer)
-
-        cur_tensors = (torch.tensor(cur_features.input_ids),
-                       torch.tensor(cur_features.input_mask),
-                       torch.tensor(cur_features.segment_ids),
-                       torch.tensor(cur_features.lm_label_ids),
-                       torch.tensor(cur_features.is_next))
-
-        return cur_tensors
-
-    def random_sent(self, index):
-        """
-        Get one sample from corpus consisting of two sentences. With prob. 50% these are two subsequent sentences
-        from one doc. With 50% the second sentence will be a random one from another doc.
-        :param index: int, index of sample.
-        :return: (str, str, int), sentence 1, sentence 2, isNextSentence Label
-        """
-        t1, t2 = self.get_corpus_line(index)
-        if random.random() > 0.5:
-            label = 0
-        else:
-            t2 = self.get_random_line()
-            label = 1
-
-        assert len(t1) > 0
-        assert len(t2) > 0
-        return t1, t2, label
-
-    def get_corpus_line(self, item):
-        """
-        Get one sample from corpus consisting of a pair of two subsequent lines from the same doc.
-        :param item: int, index of sample.
-        :return: (str, str), two subsequent sentences from corpus
-        """
-        t1 = ""
-        t2 = ""
-        assert item < self.corpus_lines
-        if self.on_memory:
-            sample = self.sample_to_doc[item]
-            t1 = self.all_docs[sample["doc_id"]][sample["line"]]
-            t2 = self.all_docs[sample["doc_id"]][sample["line"]+1]
-            # used later to avoid random nextSentence from same doc
-            self.current_doc = sample["doc_id"]
-            return t1, t2
-        else:
-            if self.line_buffer is None:
-                # read first non-empty line of file
-                while t1 == "" :
-                    t1 = next(self.file).strip()
-                    t2 = next(self.file).strip()
-            else:
-                # use t2 from previous iteration as new t1
-                t1 = self.line_buffer
-                t2 = next(self.file).strip()
-                # skip empty rows that are used for separating documents and keep track of current doc id
-                while t2 == "" or t1 == "":
-                    t1 = next(self.file).strip()
-                    t2 = next(self.file).strip()
-                    self.current_doc = self.current_doc+1
-            self.line_buffer = t2
-
-        assert t1 != ""
-        assert t2 != ""
-        return t1, t2
-
-    def get_random_line(self):
-        """
-        Get random line from another document for nextSentence task.
-        :return: str, content of one line
-        """
-        # Similar to original tf repo: This outer loop should rarely go for more than one iteration for large
-        # corpora. However, just to be careful, we try to make sure that
-        # the random document is not the same as the document we're processing.
-        for _ in range(10):
-            if self.on_memory:
-                rand_doc_idx = random.randint(0, len(self.all_docs)-1)
-                rand_doc = self.all_docs[rand_doc_idx]
-                line = rand_doc[random.randrange(len(rand_doc))]
-            else:
-                rand_index = random.randint(1, self.corpus_lines if self.corpus_lines < 1000 else 1000)
-                #pick random line
-                for _ in range(rand_index):
-                    line = self.get_next_line()
-            #check if our picked random line is really from another doc like we want it to be
-            if self.current_random_doc != self.current_doc:
-                break
-        return line
-
-    def get_next_line(self):
-        """ Gets next line of random_file and starts over when reaching end of file"""
-        try:
-            line = next(self.random_file).strip()
-            #keep track of which document we are currently looking at to later avoid having the same doc as t1
-            if line == "":
-                self.current_random_doc = self.current_random_doc + 1
-                line = next(self.random_file).strip()
-        except StopIteration:
-            self.random_file.close()
-            self.random_file = open(self.corpus_path, "r", encoding=self.encoding)
-            line = next(self.random_file).strip()
-        return line
-
-
-class InputExample(object):
-    """A single training/test example for the language model."""
-
-    def __init__(self, guid, tokens_a, tokens_b=None, is_next=None, lm_labels=None):
-        """Constructs a InputExample.
-
-        Args:
-            guid: Unique id for the example.
-            tokens_a: string. The untokenized text of the first sequence. For single
-            sequence tasks, only this sequence must be specified.
-            tokens_b: (Optional) string. The untokenized text of the second sequence.
-            Only must be specified for sequence pair tasks.
-            label: (Optional) string. The label of the example. This should be
-            specified for train and dev examples, but not for test examples.
-        """
-        self.guid = guid
-        self.tokens_a = tokens_a
-        self.tokens_b = tokens_b
-        self.is_next = is_next  # nextSentence
-        self.lm_labels = lm_labels  # masked words for language model
-
-
-class InputFeatures(object):
-    """A single set of features of data."""
-
-    def __init__(self, input_ids, input_mask, segment_ids, is_next, lm_label_ids):
-        self.input_ids = input_ids
-        self.input_mask = input_mask
-        self.segment_ids = segment_ids
-        self.is_next = is_next
-        self.lm_label_ids = lm_label_ids
-
-
-def random_word(tokens, tokenizer):
-    """
-    Masking some random tokens for Language Model task with probabilities as in the original BERT paper.
-    :param tokens: list of str, tokenized sentence.
-    :param tokenizer: Tokenizer, object used for tokenization (we need it's vocab here)
-    :return: (list of str, list of int), masked tokens and related labels for LM prediction
-    """
-    output_label = []
-
-    for i, token in enumerate(tokens):
-        prob = random.random()
-        # mask token with 15% probability
-        if prob < 0.15:
-            prob /= 0.15
-
-            # 80% randomly change token to mask token
-            if prob < 0.8:
-                tokens[i] = "[MASK]"
-
-            # 10% randomly change token to random token
-            elif prob < 0.9:
-                tokens[i] = random.choice(list(tokenizer.vocab.items()))[0]
-
-            # -> rest 10% randomly keep current token
-
-            # append current token to output (we will predict these later)
-            try:
-                output_label.append(tokenizer.vocab[token])
-            except KeyError:
-                # For unknown words (should not occur with BPE vocab)
-                output_label.append(tokenizer.vocab["[UNK]"])
-                logger.warning("Cannot find token '{}' in vocab. Using [UNK] insetad".format(token))
-        else:
-            # no masking token (will be ignored by loss function later)
-            output_label.append(-1)
-
-    return tokens, output_label
-
-
-def convert_example_to_features(example, max_seq_length, tokenizer):
-    """
-    Convert a raw sample (pair of sentences as tokenized strings) into a proper training sample with
-    IDs, LM labels, input_mask, CLS and SEP tokens etc.
-    :param example: InputExample, containing sentence input as strings and is_next label
-    :param max_seq_length: int, maximum length of sequence.
-    :param tokenizer: Tokenizer
-    :return: InputFeatures, containing all inputs and labels of one sample as IDs (as used for model training)
-    """
-    tokens_a = example.tokens_a
-    tokens_b = example.tokens_b
-    # Modifies `tokens_a` and `tokens_b` in place so that the total
-    # length is less than the specified length.
-    # Account for [CLS], [SEP], [SEP] with "- 3"
-    _truncate_seq_pair(tokens_a, tokens_b, max_seq_length - 3)
-
-    tokens_a, t1_label = random_word(tokens_a, tokenizer)
-    tokens_b, t2_label = random_word(tokens_b, tokenizer)
-    # concatenate lm labels and account for CLS, SEP, SEP
-    lm_label_ids = ([-1] + t1_label + [-1] + t2_label + [-1])
-
-    # The convention in BERT is:
-    # (a) For sequence pairs:
-    #  tokens:   [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
-    #  type_ids: 0   0  0    0    0     0       0 0    1  1  1  1   1 1
-    # (b) For single sequences:
-    #  tokens:   [CLS] the dog is hairy . [SEP]
-    #  type_ids: 0   0   0   0  0     0 0
-    #
-    # Where "type_ids" are used to indicate whether this is the first
-    # sequence or the second sequence. The embedding vectors for `type=0` and
-    # `type=1` were learned during pre-training and are added to the wordpiece
-    # embedding vector (and position vector). This is not *strictly* necessary
-    # since the [SEP] token unambigiously separates the sequences, but it makes
-    # it easier for the model to learn the concept of sequences.
-    #
-    # For classification tasks, the first vector (corresponding to [CLS]) is
-    # used as as the "sentence vector". Note that this only makes sense because
-    # the entire model is fine-tuned.
-    tokens = []
-    segment_ids = []
-    tokens.append("[CLS]")
-    segment_ids.append(0)
-    for token in tokens_a:
-        tokens.append(token)
-        segment_ids.append(0)
-    tokens.append("[SEP]")
-    segment_ids.append(0)
-
-    assert len(tokens_b) > 0
-    for token in tokens_b:
-        tokens.append(token)
-        segment_ids.append(1)
-    tokens.append("[SEP]")
-    segment_ids.append(1)
-
-    input_ids = tokenizer.convert_tokens_to_ids(tokens)
-
-    # The mask has 1 for real tokens and 0 for padding tokens. Only real
-    # tokens are attended to.
-    input_mask = [1] * len(input_ids)
-
-    # Zero-pad up to the sequence length.
-    while len(input_ids) < max_seq_length:
-        input_ids.append(0)
-        input_mask.append(0)
-        segment_ids.append(0)
-        lm_label_ids.append(-1)
-
-    assert len(input_ids) == max_seq_length
-    assert len(input_mask) == max_seq_length
-    assert len(segment_ids) == max_seq_length
-    assert len(lm_label_ids) == max_seq_length
-
-    if example.guid < 5:
-        logger.info("*** Example ***")
-        logger.info("guid: %s" % (example.guid))
-        logger.info("tokens: %s" % " ".join(
-                [str(x) for x in tokens]))
-        logger.info("input_ids: %s" % " ".join([str(x) for x in input_ids]))
-        logger.info("input_mask: %s" % " ".join([str(x) for x in input_mask]))
-        logger.info(
-                "segment_ids: %s" % " ".join([str(x) for x in segment_ids]))
-        logger.info("LM label: %s " % (lm_label_ids))
-        logger.info("Is next sentence label: %s " % (example.is_next))
-
-    features = InputFeatures(input_ids=input_ids,
-                             input_mask=input_mask,
-                             segment_ids=segment_ids,
-                             lm_label_ids=lm_label_ids,
-                             is_next=example.is_next)
-    return features
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    ## Required parameters
-    parser.add_argument("--train_corpus",
-                        default=None,
-                        type=str,
-                        required=True,
-                        help="The input train corpus.")
-    parser.add_argument("--bert_model", default=None, type=str, required=True,
-                        help="Bert pre-trained model selected in the list: bert-base-uncased, "
-                             "bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.")
-    parser.add_argument("--output_dir",
-                        default=None,
-                        type=str,
-                        required=True,
-                        help="The output directory where the model checkpoints will be written.")
-
-    ## Other parameters
-    parser.add_argument("--max_seq_length",
-                        default=128,
-                        type=int,
-                        help="The maximum total input sequence length after WordPiece tokenization. \n"
-                             "Sequences longer than this will be truncated, and sequences shorter \n"
-                             "than this will be padded.")
-    parser.add_argument("--do_train",
-                        action='store_true',
-                        help="Whether to run training.")
-    parser.add_argument("--train_batch_size",
-                        default=32,
-                        type=int,
-                        help="Total batch size for training.")
-    parser.add_argument("--learning_rate",
-                        default=3e-5,
-                        type=float,
-                        help="The initial learning rate for Adam.")
-    parser.add_argument("--adam_epsilon", 
-                        default=1e-8, 
-                        type=float,
-                        help="Epsilon for Adam optimizer.")
-    parser.add_argument("--num_train_epochs",
-                        default=3.0,
-                        type=float,
-                        help="Total number of training epochs to perform.")
-    parser.add_argument("--warmup_steps", 
-                        default=0, 
-                        type=int,
-                        help="Linear warmup over warmup_steps.")
-    parser.add_argument("--no_cuda",
-                        action='store_true',
-                        help="Whether not to use CUDA when available")
-    parser.add_argument("--on_memory",
-                        action='store_true',
-                        help="Whether to load train samples into memory or use disk")
-    parser.add_argument("--do_lower_case",
-                        action='store_true',
-                        help="Whether to lower case the input text. True for uncased models, False for cased models.")
-    parser.add_argument("--local_rank",
-                        type=int,
-                        default=-1,
-                        help="local_rank for distributed training on gpus")
-    parser.add_argument('--seed',
-                        type=int,
-                        default=42,
-                        help="random seed for initialization")
-    parser.add_argument('--gradient_accumulation_steps',
-                        type=int,
-                        default=1,
-                        help="Number of updates steps to accumualte before performing a backward/update pass.")
-    parser.add_argument('--fp16',
-                        action='store_true',
-                        help="Whether to use 16-bit float precision instead of 32-bit")
-    parser.add_argument('--loss_scale',
-                        type = float, default = 0,
-                        help = "Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
-                        "0 (default value): dynamic loss scaling.\n"
-                        "Positive power of 2: static loss scaling value.\n")
-
-    args = parser.parse_args()
-
-    if args.local_rank == -1 or args.no_cuda:
-        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        n_gpu = torch.cuda.device_count()
-    else:
-        torch.cuda.set_device(args.local_rank)
-        device = torch.device("cuda", args.local_rank)
-        n_gpu = 1
-        # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
-        torch.distributed.init_process_group(backend='nccl')
-    logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
-        device, n_gpu, bool(args.local_rank != -1), args.fp16))
-
-    if args.gradient_accumulation_steps < 1:
-        raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
-                            args.gradient_accumulation_steps))
-
-    args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
-
-    random.seed(args.seed)
-    np.random.seed(args.seed)
-    torch.manual_seed(args.seed)
-    if n_gpu > 0:
-        torch.cuda.manual_seed_all(args.seed)
-
-    if not args.do_train:
-        raise ValueError("Training is currently the only implemented execution option. Please set `do_train`.")
-
-    if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
-        raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
-    if not os.path.exists(args.output_dir) and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
-        os.makedirs(args.output_dir)
-
-    tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
-
-    #train_examples = None
-    num_train_optimization_steps = None
-    if args.do_train:
-        print("Loading Train Dataset", args.train_corpus)
-        train_dataset = BERTDataset(args.train_corpus, tokenizer, seq_len=args.max_seq_length,
-                                    corpus_lines=None, on_memory=args.on_memory)
-        num_train_optimization_steps = int(
-            len(train_dataset) / args.train_batch_size / args.gradient_accumulation_steps) * args.num_train_epochs
-        if args.local_rank != -1:
-            num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
-
-    # Prepare model
-    model = BertForPreTraining.from_pretrained(args.bert_model)
-    if args.fp16:
-        model.half()
-    model.to(device)
-    if args.local_rank != -1:
-        try:
-            from apex.parallel import DistributedDataParallel as DDP
-        except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
-        model = DDP(model)
-    elif n_gpu > 1:
-        model = torch.nn.DataParallel(model)
-
-    # Prepare optimizer
-    if args.do_train:
-        param_optimizer = list(model.named_parameters())
-        no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
-        optimizer_grouped_parameters = [
-            {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
-            {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
-            ]
-
-        if args.fp16:
-            try:
-                from apex.optimizers import FP16_Optimizer
-                from apex.optimizers import FusedAdam
-            except ImportError:
-                raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
-
-            optimizer = FusedAdam(optimizer_grouped_parameters,
-                                  lr=args.learning_rate,
-                                  bias_correction=False,
-                                  max_grad_norm=1.0)
-            if args.loss_scale == 0:
-                optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
-            else:
-                optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
-
-        else:
-            optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-        scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=num_train_optimization_steps)
-
-    global_step = 0
-    if args.do_train:
-        logger.info("***** Running training *****")
-        logger.info("  Num examples = %d", len(train_dataset))
-        logger.info("  Batch size = %d", args.train_batch_size)
-        logger.info("  Num steps = %d", num_train_optimization_steps)
-
-        if args.local_rank == -1:
-            train_sampler = RandomSampler(train_dataset)
-        else:
-            #TODO: check if this works with current data generator from disk that relies on next(file)
-            # (it doesn't return item back by index)
-            train_sampler = DistributedSampler(train_dataset)
-        train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
-
-        model.train()
-        for _ in trange(int(args.num_train_epochs), desc="Epoch"):
-            tr_loss = 0
-            nb_tr_examples, nb_tr_steps = 0, 0
-            for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")):
-                batch = tuple(t.to(device) for t in batch)
-                input_ids, input_mask, segment_ids, lm_label_ids, is_next = batch
-                outputs = model(input_ids, segment_ids, input_mask, lm_label_ids, is_next)
-                loss = outputs[0]
-                if n_gpu > 1:
-                    loss = loss.mean() # mean() to average on multi-gpu.
-                if args.gradient_accumulation_steps > 1:
-                    loss = loss / args.gradient_accumulation_steps
-                if args.fp16:
-                    optimizer.backward(loss)
-                else:
-                    loss.backward()
-                tr_loss += loss.item()
-                nb_tr_examples += input_ids.size(0)
-                nb_tr_steps += 1
-                if (step + 1) % args.gradient_accumulation_steps == 0:
-                    optimizer.step()
-                    scheduler.step()  # Update learning rate schedule
-                    optimizer.zero_grad()
-                    global_step += 1
-
-        # Save a trained model
-        if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
-            logger.info("** ** * Saving fine - tuned model ** ** * ")
-            model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
-            model_to_save.save_pretrained(args.output_dir)
-            tokenizer.save_pretrained(args.output_dir)
-
-
-def _truncate_seq_pair(tokens_a, tokens_b, max_length):
-    """Truncates a sequence pair in place to the maximum length."""
-
-    # This is a simple heuristic which will always truncate the longer sequence
-    # one token at a time. This makes more sense than truncating an equal percent
-    # of tokens from each, since if one sequence is very short then each token
-    # that's truncated likely contains more information than a longer sequence.
-    while True:
-        total_length = len(tokens_a) + len(tokens_b)
-        if total_length <= max_length:
-            break
-        if len(tokens_a) > len(tokens_b):
-            tokens_a.pop()
-        else:
-            tokens_b.pop()
-
-
-def accuracy(out, labels):
-    outputs = np.argmax(out, axis=1)
-    return np.sum(outputs == labels)
-
-
-if __name__ == "__main__":
-    main()

examples/single_model_scripts/run_multiple_choice.py → examples/run_multiple_choice.py

@@ -158,8 +158,9 @@ def train(args, train_dataset, model, tokenizer):
 
             tr_loss += loss.item()
             if (step + 1) % args.gradient_accumulation_steps == 0:
-                scheduler.step()  # Update learning rate schedule
+
                 optimizer.step()
+                scheduler.step()  # Update learning rate schedule
                 model.zero_grad()
                 global_step += 1
 

pytorch_transformers/modeling_bert.py

@@ -527,7 +527,7 @@ BERT_INPUTS_DOCSTRING = r"""
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
 """
 
-@add_start_docstrings("The bare Bert Model transformer outputing raw hidden-states without any specific head on top.",
+@add_start_docstrings("The bare Bert Model transformer outputting raw hidden-states without any specific head on top.",
                       BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
 class BertModel(BertPreTrainedModel):
     r"""

pytorch_transformers/modeling_distilbert.py

@@ -394,7 +394,7 @@ DISTILBERT_INPUTS_DOCSTRING = r"""
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
 """
 
-@add_start_docstrings("The bare DistilBERT encoder/transformer outputing raw hidden-states without any specific head on top.",
+@add_start_docstrings("The bare DistilBERT encoder/transformer outputting raw hidden-states without any specific head on top.",
                       DISTILBERT_START_DOCSTRING, DISTILBERT_INPUTS_DOCSTRING)
 class DistilBertModel(DistilBertPreTrainedModel):
     r"""

pytorch_transformers/modeling_gpt2.py

@@ -314,7 +314,7 @@ GPT2_INPUTS_DOCSTRING = r"""    Inputs:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
 """
 
-@add_start_docstrings("The bare GPT2 Model transformer outputing raw hidden-states without any specific head on top.",
+@add_start_docstrings("The bare GPT2 Model transformer outputting raw hidden-states without any specific head on top.",
                       GPT2_START_DOCSTRING, GPT2_INPUTS_DOCSTRING)
 class GPT2Model(GPT2PreTrainedModel):
     r"""

pytorch_transformers/modeling_openai.py

@@ -324,7 +324,7 @@ OPENAI_GPT_INPUTS_DOCSTRING = r"""    Inputs:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
 """
 
-@add_start_docstrings("The bare OpenAI GPT transformer model outputing raw hidden-states without any specific head on top.",
+@add_start_docstrings("The bare OpenAI GPT transformer model outputting raw hidden-states without any specific head on top.",
                       OPENAI_GPT_START_DOCSTRING, OPENAI_GPT_INPUTS_DOCSTRING)
 class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
     r"""

pytorch_transformers/modeling_roberta.py

@@ -124,7 +124,7 @@ ROBERTA_INPUTS_DOCSTRING = r"""
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
 """
 
-@add_start_docstrings("The bare RoBERTa Model transformer outputing raw hidden-states without any specific head on top.",
+@add_start_docstrings("The bare RoBERTa Model transformer outputting raw hidden-states without any specific head on top.",
                       ROBERTA_START_DOCSTRING, ROBERTA_INPUTS_DOCSTRING)
 class RobertaModel(BertModel):
     r"""

pytorch_transformers/modeling_transfo_xl.py

@@ -820,7 +820,7 @@ TRANSFO_XL_INPUTS_DOCSTRING = r"""
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
 """
 
-@add_start_docstrings("The bare Bert Model transformer outputing raw hidden-states without any specific head on top.",
+@add_start_docstrings("The bare Bert Model transformer outputting raw hidden-states without any specific head on top.",
                       TRANSFO_XL_START_DOCSTRING, TRANSFO_XL_INPUTS_DOCSTRING)
 class TransfoXLModel(TransfoXLPreTrainedModel):
     r"""

pytorch_transformers/modeling_xlm.py

@@ -313,7 +313,7 @@ XLM_INPUTS_DOCSTRING = r"""
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
 """
 
-@add_start_docstrings("The bare XLM Model transformer outputing raw hidden-states without any specific head on top.",
+@add_start_docstrings("The bare XLM Model transformer outputting raw hidden-states without any specific head on top.",
                       XLM_START_DOCSTRING, XLM_INPUTS_DOCSTRING)
 class XLMModel(XLMPreTrainedModel):
     r"""

pytorch_transformers/modeling_xlnet.py

@@ -546,7 +546,7 @@ XLNET_INPUTS_DOCSTRING = r"""
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
 """
 
-@add_start_docstrings("The bare XLNet Model transformer outputing raw hidden-states without any specific head on top.",
+@add_start_docstrings("The bare XLNet Model transformer outputting raw hidden-states without any specific head on top.",
                       XLNET_START_DOCSTRING, XLNET_INPUTS_DOCSTRING)
 class XLNetModel(XLNetPreTrainedModel):
     r"""
@@ -743,8 +743,9 @@ class XLNetModel(XLNetPreTrainedModel):
 
         if data_mask is not None:
             # all mems can be attended to
-            mems_mask = torch.zeros([data_mask.shape[0], mlen, bsz]).to(data_mask)
-            data_mask = torch.cat([mems_mask, data_mask], dim=1)
+            if mlen > 0:
+                mems_mask = torch.zeros([data_mask.shape[0], mlen, bsz]).to(data_mask)
+                data_mask = torch.cat([mems_mask, data_mask], dim=1)
             if attn_mask is None:
                 attn_mask = data_mask[:, :, :, None]
             else:
@@ -755,7 +756,8 @@ class XLNetModel(XLNetPreTrainedModel):
 
         if attn_mask is not None:
             non_tgt_mask = -torch.eye(qlen).to(attn_mask)
-            non_tgt_mask = torch.cat([torch.zeros([qlen, mlen]).to(attn_mask), non_tgt_mask], dim=-1)
+            if mlen > 0:
+                non_tgt_mask = torch.cat([torch.zeros([qlen, mlen]).to(attn_mask), non_tgt_mask], dim=-1)
             non_tgt_mask = ((attn_mask + non_tgt_mask[:, :, None, None]) > 0).to(attn_mask)
         else:
             non_tgt_mask = None
@@ -775,8 +777,11 @@ class XLNetModel(XLNetPreTrainedModel):
         ##### Segment embedding
         if token_type_ids is not None:
             # Convert `token_type_ids` to one-hot `seg_mat`
-            mem_pad = torch.zeros([mlen, bsz], dtype=torch.long, device=device)
-            cat_ids = torch.cat([mem_pad, token_type_ids], dim=0)
+            if mlen > 0:
+                mem_pad = torch.zeros([mlen, bsz], dtype=torch.long, device=device)
+                cat_ids = torch.cat([mem_pad, token_type_ids], dim=0)
+            else:
+                cat_ids = token_type_ids
 
             # `1` indicates not in the same segment [qlen x klen x bsz]
             seg_mat = (token_type_ids[:, None] != cat_ids[None, :]).long()