refactoring old run_swag. This script is mainly refatored from run_squad in pytorch_transformers

examples/single_model_scripts/run_swag.py

@@ -13,17 +13,18 @@
 # 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
+"""BERT finetuning runner.
+   Finetuning the library models for multiple choice on SWAG (Bert).
+"""
+from __future__ import absolute_import, division, print_function
 
 import argparse
-import csv
 import logging
+import csv
 import os
 import random
 import sys
-from io import open
+import glob
 
 import numpy as np
 import torch
@@ -32,16 +33,21 @@ from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
 from torch.utils.data.distributed import DistributedSampler
 from tqdm import tqdm, trange
 
-from pytorch_transformers.file_utils import PYTORCH_PRETRAINED_BERT_CACHE, WEIGHTS_NAME, CONFIG_NAME
-from pytorch_transformers.modeling_bert import BertForMultipleChoice, BertConfig
-from pytorch_transformers.optimization import AdamW, WarmupLinearSchedule
-from pytorch_transformers.tokenization_bert import BertTokenizer
+from tensorboardX import SummaryWriter
+
+from pytorch_transformers import (WEIGHTS_NAME, BertConfig,
+                                  BertForMultipleChoice, BertTokenizer)
+
+from pytorch_transformers 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__)
 
+ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) \
+                  for conf in [BertConfig]), ())
+
+MODEL_CLASSES = {
+    'bert': (BertConfig, BertForMultipleChoice, BertTokenizer),
+}
 
 class SwagExample(object):
     """A single training/test example for the SWAG dataset."""
@@ -84,7 +90,6 @@ class SwagExample(object):
 
         return ", ".join(l)
 
-
 class InputFeatures(object):
     def __init__(self,
                  example_id,
@@ -103,8 +108,7 @@ class InputFeatures(object):
         ]
         self.label = label
 
-
-def read_swag_examples(input_file, is_training):
+def read_swag_examples(input_file, is_training=True):
     with open(input_file, 'r', encoding='utf-8') as f:
         reader = csv.reader(f)
         lines = []
@@ -156,7 +160,7 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
     # final decision of the model, we will run a softmax over these 4
     # outputs.
     features = []
-    for example_index, example in enumerate(examples):
+    for example_index, example in tqdm(enumerate(examples)):
         context_tokens = tokenizer.tokenize(example.context_sentence)
         start_ending_tokens = tokenizer.tokenize(example.start_ending)
 
@@ -242,314 +246,428 @@ def select_field(features, field):
         for feature in features
     ]
 
+
+def set_seed(args):
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if args.n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)
+
+def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
+    if args.local_rank not in [-1, 0]:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
+    # Load data features from cache or dataset file
+    input_file = args.predict_file if evaluate else args.train_file
+    cached_features_file = os.path.join(os.path.dirname(input_file), 'cached_{}_{}_{}'.format(
+        'dev' if evaluate else 'train',
+        list(filter(None, args.model_name_or_path.split('/'))).pop(),
+        str(args.max_seq_length)))
+    if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples:
+        logger.info("Loading features from cached file %s", cached_features_file)
+        features = torch.load(cached_features_file)
+    else:
+        logger.info("Creating features from dataset file at %s", input_file)
+        examples = read_swag_examples(input_file)
+        features = convert_examples_to_features(
+            examples, tokenizer, args.max_seq_length, not evaluate)
+
+        if args.local_rank in [-1, 0]:
+            logger.info("Saving features into cached file %s", cached_features_file)
+            torch.save(features, cached_features_file)
+
+    if args.local_rank == 0:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
+    # Convert to Tensors and build dataset
+    all_input_ids = torch.tensor(select_field(features, 'input_ids'), dtype=torch.long)
+    all_input_mask = torch.tensor(select_field(features, 'input_mask'), dtype=torch.long)
+    all_segment_ids = torch.tensor(select_field(features, 'segment_ids'), dtype=torch.long)
+    all_label = torch.tensor([f.label for f in features], dtype=torch.long)
+
+    if evaluate:
+        dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
+                                all_label)
+    else:
+        dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
+                                all_label)
+
+    if output_examples:
+        return dataset, examples, features
+    return dataset
+def train(args, train_dataset, model, tokenizer):
+    """ Train the model """
+    if args.local_rank in [-1, 0]:
+        tb_writer = SummaryWriter()
+
+    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
+    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
+    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
+
+    if args.max_steps > 0:
+        t_total = args.max_steps
+        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
+    else:
+        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
+
+    # Prepare optimizer and schedule (linear warmup and decay)
+    no_decay = ['bias', 'LayerNorm.weight']
+    optimizer_grouped_parameters = [
+        {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
+        {'params': [p for n, p in model.named_parameters() 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=t_total)
+    if args.fp16:
+        try:
+            from apex import amp
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
+
+    # multi-gpu training (should be after apex fp16 initialization)
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
+    # Distributed training (should be after apex fp16 initialization)
+    if args.local_rank != -1:
+        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
+                                                          output_device=args.local_rank,
+                                                          find_unused_parameters=True)
+
+    # Train!
+    logger.info("***** Running training *****")
+    logger.info("  Num examples = %d", len(train_dataset))
+    logger.info("  Num Epochs = %d", args.num_train_epochs)
+    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
+    logger.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
+                   args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
+    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
+    logger.info("  Total optimization steps = %d", t_total)
+
+    global_step = 0
+    tr_loss, logging_loss = 0.0, 0.0
+    model.zero_grad()
+    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
+    for _ in train_iterator:
+        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
+        for step, batch in enumerate(epoch_iterator):
+            model.train()
+            batch = tuple(t.to(args.device) for t in batch)
+            inputs = {'input_ids':       batch[0],
+                      'attention_mask':  batch[1],
+                      #'token_type_ids':  None if args.model_type == 'xlm' else batch[2],
+                      'token_type_ids': batch[2],
+                      'labels':         batch[3]}
+            # if args.model_type in ['xlnet', 'xlm']:
+            #     inputs.update({'cls_index': batch[5],
+            #                    'p_mask':       batch[6]})
+            outputs = model(**inputs)
+            loss = outputs[0]  # model outputs are always tuple in pytorch-transformers (see doc)
+
+            if args.n_gpu > 1:
+                loss = loss.mean() # mean() to average on multi-gpu parallel (not distributed) training
+            if args.gradient_accumulation_steps > 1:
+                loss = loss / args.gradient_accumulation_steps
+
+            if args.fp16:
+                with amp.scale_loss(loss, optimizer) as scaled_loss:
+                    scaled_loss.backward()
+                torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
+            else:
+                loss.backward()
+                torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
+
+            tr_loss += loss.item()
+            if (step + 1) % args.gradient_accumulation_steps == 0:
+                optimizer.step()
+                scheduler.step()  # Update learning rate schedule
+                model.zero_grad()
+                global_step += 1
+
+                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
+                    # Log metrics
+                    if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
+                        results = evaluate(args, model, tokenizer)
+                        for key, value in results.items():
+                            tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
+                    tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
+                    tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
+                    logging_loss = tr_loss
+
+                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
+                    # Save model checkpoint
+                    output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
+                    if not os.path.exists(output_dir):
+                        os.makedirs(output_dir)
+                    model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
+                    model_to_save.save_pretrained(output_dir)
+                    tokenizer.save_vocabulary(output_dir)
+                    torch.save(args, os.path.join(output_dir, 'training_args.bin'))
+                    logger.info("Saving model checkpoint to %s", output_dir)
+
+            if args.max_steps > 0 and global_step > args.max_steps:
+                epoch_iterator.close()
+                break
+        if args.max_steps > 0 and global_step > args.max_steps:
+            train_iterator.close()
+            break
+
+    if args.local_rank in [-1, 0]:
+        tb_writer.close()
+
+    return global_step, tr_loss / global_step
+
+def evaluate(args, model, tokenizer, prefix=""):
+    dataset, examples, features = load_and_cache_examples(args, tokenizer, evaluate=True, output_examples=True)
+
+    if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+        os.makedirs(args.output_dir)
+
+    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
+    # Note that DistributedSampler samples randomly
+    eval_sampler = SequentialSampler(dataset) if args.local_rank == -1 else DistributedSampler(dataset)
+    eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
+
+    # Eval!
+    logger.info("***** Running evaluation {} *****".format(prefix))
+    logger.info("  Num examples = %d", len(dataset))
+    logger.info("  Batch size = %d", args.eval_batch_size)
+
+
+    eval_loss, eval_accuracy = 0, 0
+    nb_eval_steps, nb_eval_examples = 0, 0
+
+    for batch in tqdm(eval_dataloader, desc="Evaluating"):
+        model.eval()
+        batch = tuple(t.to(args.device) for t in batch)
+        with torch.no_grad():
+            inputs = {'input_ids':      batch[0],
+                      'attention_mask': batch[1],
+                      # 'token_type_ids': None if args.model_type == 'xlm' else batch[2]  # XLM don't use segment_ids
+                      'token_type_ids': batch[2],
+                      'labels':         batch[3]}
+
+            # if args.model_type in ['xlnet', 'xlm']:
+            #     inputs.update({'cls_index': batch[4],
+            #                    'p_mask':    batch[5]})
+            outputs = model(**inputs)
+            tmp_eval_loss, logits = outputs[:2]
+            eval_loss += tmp_eval_loss.mean().item()
+
+        logits = logits.detach().cpu().numpy()
+        label_ids = inputs['labels'].to('cpu').numpy()
+        tmp_eval_accuracy = accuracy(logits, label_ids)
+        eval_accuracy += tmp_eval_accuracy
+
+        nb_eval_steps += 1
+        nb_eval_examples += inputs['input_ids'].size(0)
+
+    eval_loss = eval_loss / nb_eval_steps
+    eval_accuracy = eval_accuracy / nb_eval_examples
+    result = {'eval_loss': eval_loss,
+              'eval_accuracy': eval_accuracy}
+
+    output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
+    with open(output_eval_file, "w") as writer:
+        logger.info("***** Eval results *****")
+        for key in sorted(result.keys()):
+            logger.info("%s = %s", key, str(result[key]))
+            writer.write("%s = %s\n" % (key, str(result[key])))
+
+    return result
+
 def main():
     parser = argparse.ArgumentParser()
 
     ## Required parameters
-    parser.add_argument("--data_dir",
-                        default=None,
-                        type=str,
-                        required=True,
-                        help="The input data dir. Should contain the .csv files (or other data files) for the task.")
-    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-large-cased, bert-base-multilingual-uncased, "
-                        "bert-base-multilingual-cased, 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.")
+    parser.add_argument("--train_file", default=None, type=str, required=True,
+                        help="SWAG csv for training. E.g., train.csv")
+    parser.add_argument("--predict_file", default=None, type=str, required=True,
+                        help="SWAG csv for predictions. E.g., val.csv or test.csv")
+    parser.add_argument("--model_type", default=None, type=str, required=True,
+                        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
+    parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
+                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
+    parser.add_argument("--output_dir", default=None, type=str, required=True,
+                        help="The output directory where the model checkpoints and predictions 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',
+    parser.add_argument("--config_name", default="", type=str,
+                        help="Pretrained config name or path if not the same as model_name")
+    parser.add_argument("--tokenizer_name", default="", type=str,
+                        help="Pretrained tokenizer name or path if not the same as model_name")
+    parser.add_argument("--max_seq_length", default=384, type=int,
+                        help="The maximum total input sequence length after tokenization. Sequences "
+                             "longer than this will be truncated, and sequences shorter than this will be padded.")
+    parser.add_argument("--do_train", action='store_true',
                         help="Whether to run training.")
-    parser.add_argument("--do_eval",
-                        action='store_true',
+    parser.add_argument("--do_eval", action='store_true',
                         help="Whether to run eval on the dev set.")
-    parser.add_argument("--do_lower_case",
-                        action='store_true',
+    parser.add_argument("--evaluate_during_training", action='store_true',
+                        help="Rul evaluation during training at each logging step.")
+    parser.add_argument("--do_lower_case", action='store_true',
                         help="Set this flag if you are using an uncased model.")
-    parser.add_argument("--train_batch_size",
-                        default=32,
-                        type=int,
-                        help="Total batch size for training.")
-    parser.add_argument("--eval_batch_size",
-                        default=8,
-                        type=int,
-                        help="Total batch size for eval.")
-    parser.add_argument("--learning_rate",
-                        default=5e-5,
-                        type=float,
-                        help="The initial learning rate for Adam.")
-    parser.add_argument("--num_train_epochs",
-                        default=3.0,
-                        type=float,
-                        help="Total number of training epochs to perform.")
-    parser.add_argument("--warmup_proportion",
-                        default=0.1,
-                        type=float,
-                        help="Proportion of training to perform linear learning rate warmup for. "
-                             "E.g., 0.1 = 10%% of training.")
-    parser.add_argument("--no_cuda",
-                        action='store_true',
-                        help="Whether not to use CUDA when available")
-    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 accumulate 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")
 
+    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int,
+                        help="Batch size per GPU/CPU for training.")
+    parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
+                        help="Batch size per GPU/CPU for evaluation.")
+    parser.add_argument("--learning_rate", default=5e-5, type=float,
+                        help="The initial learning rate for Adam.")
+    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("--weight_decay", default=0.0, type=float,
+                        help="Weight deay if we apply some.")
+    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
+                        help="Epsilon for Adam optimizer.")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float,
+                        help="Max gradient norm.")
+    parser.add_argument("--num_train_epochs", default=3.0, type=float,
+                        help="Total number of training epochs to perform.")
+    parser.add_argument("--max_steps", default=-1, type=int,
+                        help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
+    parser.add_argument("--warmup_steps", default=0, type=int,
+                        help="Linear warmup over warmup_steps.")
+
+    parser.add_argument('--logging_steps', type=int, default=50,
+                        help="Log every X updates steps.")
+    parser.add_argument('--save_steps', type=int, default=50,
+                        help="Save checkpoint every X updates steps.")
+    parser.add_argument("--eval_all_checkpoints", action='store_true',
+                        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
+    parser.add_argument("--no_cuda", action='store_true',
+                        help="Whether not to use CUDA when available")
+    parser.add_argument('--overwrite_output_dir', action='store_true',
+                        help="Overwrite the content of the output directory")
+    parser.add_argument('--overwrite_cache', action='store_true',
+                        help="Overwrite the cached training and evaluation sets")
+    parser.add_argument('--seed', type=int, default=42,
+                        help="random seed for initialization")
+
+    parser.add_argument("--local_rank", type=int, default=-1,
+                        help="local_rank for distributed training on gpus")
+    parser.add_argument('--fp16', action='store_true',
+                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
+    parser.add_argument('--fp16_opt_level', type=str, default='O1',
+                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
+                             "See details at https://nvidia.github.io/apex/amp.html")
+    parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
+    parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
     args = parser.parse_args()
 
+    if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
+        raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
+
+    # Setup distant debugging if needed
+    if args.server_ip and args.server_port:
+        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
+        import ptvsd
+        print("Waiting for debugger attach")
+        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
+        ptvsd.wait_for_attach()
+
+    # Setup CUDA, GPU & distributed training
     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:
+        args.n_gpu = torch.cuda.device_count()
+    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
         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))
+        args.n_gpu = 1
+    args.device = device
 
-    if args.gradient_accumulation_steps < 1:
-        raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
-                            args.gradient_accumulation_steps))
+    # Setup logging
+    logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
+                        datefmt = '%m/%d/%Y %H:%M:%S',
+                        level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
+    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
+                    args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
 
-    args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
+    # Set seed
+    set_seed(args)
 
-    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)
+    # Load pretrained model and tokenizer
+    if args.local_rank not in [-1, 0]:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
 
-    if not args.do_train and not args.do_eval:
-        raise ValueError("At least one of `do_train` or `do_eval` must be True.")
+    args.model_type = args.model_type.lower()
+    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path)
+    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
+    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
 
-    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):
-        os.makedirs(args.output_dir)
+    if args.local_rank == 0:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
 
-    tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
+    model.to(args.device)
 
-    # Prepare model
-    model = BertForMultipleChoice.from_pretrained(args.bert_model,
-        cache_dir=os.path.join(str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(args.local_rank)),
-        num_choices=4)
-    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)
+    logger.info("Training/evaluation parameters %s", args)
 
+    # Training
     if args.do_train:
-
-        # Prepare data loader
-
-        train_examples = read_swag_examples(os.path.join(args.data_dir, 'train.csv'), is_training = True)
-        train_features = convert_examples_to_features(
-            train_examples, tokenizer, args.max_seq_length, True)
-        all_input_ids = torch.tensor(select_field(train_features, 'input_ids'), dtype=torch.long)
-        all_input_mask = torch.tensor(select_field(train_features, 'input_mask'), dtype=torch.long)
-        all_segment_ids = torch.tensor(select_field(train_features, 'segment_ids'), dtype=torch.long)
-        all_label = torch.tensor([f.label for f in train_features], dtype=torch.long)
-        train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label)
-        if args.local_rank == -1:
-            train_sampler = RandomSampler(train_data)
-        else:
-            train_sampler = DistributedSampler(train_data)
-        train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
-
-        num_train_optimization_steps = len(train_dataloader) // 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 optimizer
-
-        param_optimizer = list(model.named_parameters())
-
-        # hack to remove pooler, which is not used
-        # thus it produce None grad that break apex
-        param_optimizer = [n for n in param_optimizer]
-
-        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)
-            warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion,
-                                                 t_total=num_train_optimization_steps)
-        else:
-            optimizer = BertAdam(optimizer_grouped_parameters,
-                                 lr=args.learning_rate,
-                                 warmup=args.warmup_proportion,
-                                 t_total=num_train_optimization_steps)
-
-        global_step = 0
-
-        logger.info("***** Running training *****")
-        logger.info("  Num examples = %d", len(train_examples))
-        logger.info("  Batch size = %d", args.train_batch_size)
-        logger.info("  Num steps = %d", num_train_optimization_steps)
-
-        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, label_ids = batch
-                loss = model(input_ids, segment_ids, input_mask, label_ids)
-                if n_gpu > 1:
-                    loss = loss.mean() # mean() to average on multi-gpu.
-                if args.fp16 and args.loss_scale != 1.0:
-                    # rescale loss for fp16 training
-                    # see https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html
-                    loss = loss * args.loss_scale
-                if args.gradient_accumulation_steps > 1:
-                    loss = loss / args.gradient_accumulation_steps
-                tr_loss += loss.item()
-                nb_tr_examples += input_ids.size(0)
-                nb_tr_steps += 1
-
-                if args.fp16:
-                    optimizer.backward(loss)
-                else:
-                    loss.backward()
-                if (step + 1) % args.gradient_accumulation_steps == 0:
-                    if args.fp16:
-                        # modify learning rate with special warm up BERT uses
-                        # if args.fp16 is False, BertAdam is used that handles this automatically
-                        lr_this_step = args.learning_rate * warmup_linear.get_lr(global_step, args.warmup_proportion)
-                        for param_group in optimizer.param_groups:
-                            param_group['lr'] = lr_this_step
-                    optimizer.step()
-                    optimizer.zero_grad()
-                    global_step += 1
+        train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
+        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
+        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
 
 
-    if args.do_train:
-        # Save a trained model, configuration and tokenizer
-        model_to_save = model.module if hasattr(model, 'module') else model  # Only save the model it-self
+    # Save the trained model and the tokenizer
+    if args.local_rank == -1 or torch.distributed.get_rank() == 0:
+        # Create output directory if needed
+        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(args.output_dir)
 
-        # If we save using the predefined names, we can load using `from_pretrained`
-        output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
-        output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
+        logger.info("Saving model checkpoint to %s", args.output_dir)
+        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
+        # They can then be reloaded using `from_pretrained()`
+        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)
 
-        torch.save(model_to_save.state_dict(), output_model_file)
-        model_to_save.config.to_json_file(output_config_file)
-        tokenizer.save_vocabulary(args.output_dir)
+        # Good practice: save your training arguments together with the trained model
+        torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
 
         # Load a trained model and vocabulary that you have fine-tuned
-        model = BertForMultipleChoice.from_pretrained(args.output_dir, num_choices=4)
-        tokenizer = BertTokenizer.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
-    else:
-        model = BertForMultipleChoice.from_pretrained(args.bert_model, num_choices=4)
-    model.to(device)
+        model = model_class.from_pretrained(args.output_dir)
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir)
+        model.to(args.device)
 
 
-    if args.do_eval and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
-        eval_examples = read_swag_examples(os.path.join(args.data_dir, 'val.csv'), is_training = True)
-        eval_features = convert_examples_to_features(
-            eval_examples, tokenizer, args.max_seq_length, True)
-        logger.info("***** Running evaluation *****")
-        logger.info("  Num examples = %d", len(eval_examples))
-        logger.info("  Batch size = %d", args.eval_batch_size)
-        all_input_ids = torch.tensor(select_field(eval_features, 'input_ids'), dtype=torch.long)
-        all_input_mask = torch.tensor(select_field(eval_features, 'input_mask'), dtype=torch.long)
-        all_segment_ids = torch.tensor(select_field(eval_features, 'segment_ids'), dtype=torch.long)
-        all_label = torch.tensor([f.label for f in eval_features], dtype=torch.long)
-        eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label)
-        # Run prediction for full data
-        eval_sampler = SequentialSampler(eval_data)
-        eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)
+    # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
+    results = {}
+    if args.do_eval and args.local_rank in [-1, 0]:
+        if args.do_train:
+            checkpoints = [args.output_dir]
+        else:
+            # if do_train is False and do_eval is true, load model directly from pretrained.
+            checkpoints = [args.model_name_or_path]
 
-        model.eval()
-        eval_loss, eval_accuracy = 0, 0
-        nb_eval_steps, nb_eval_examples = 0, 0
-        for input_ids, input_mask, segment_ids, label_ids in tqdm(eval_dataloader, desc="Evaluating"):
-            input_ids = input_ids.to(device)
-            input_mask = input_mask.to(device)
-            segment_ids = segment_ids.to(device)
-            label_ids = label_ids.to(device)
+        if args.eval_all_checkpoints:
+            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
+            logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN)  # Reduce model loading logs
 
-            with torch.no_grad():
-                tmp_eval_loss = model(input_ids, segment_ids, input_mask, label_ids)
-                logits = model(input_ids, segment_ids, input_mask)
+        logger.info("Evaluate the following checkpoints: %s", checkpoints)
 
-            logits = logits.detach().cpu().numpy()
-            label_ids = label_ids.to('cpu').numpy()
-            tmp_eval_accuracy = accuracy(logits, label_ids)
+        for checkpoint in checkpoints:
+            # Reload the model
+            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
+            model = model_class.from_pretrained(checkpoint)
+            tokenizer = tokenizer_class.from_pretrained(checkpoint)
+            model.to(args.device)
 
-            eval_loss += tmp_eval_loss.mean().item()
-            eval_accuracy += tmp_eval_accuracy
+            # Evaluate
+            result = evaluate(args, model, tokenizer, prefix=global_step)
 
-            nb_eval_examples += input_ids.size(0)
-            nb_eval_steps += 1
+            result = dict((k + ('_{}'.format(global_step) if global_step else ''), v) for k, v in result.items())
+            results.update(result)
 
-        eval_loss = eval_loss / nb_eval_steps
-        eval_accuracy = eval_accuracy / nb_eval_examples
+    logger.info("Results: {}".format(results))
 
-        result = {'eval_loss': eval_loss,
-                  'eval_accuracy': eval_accuracy,
-                  'global_step': global_step,
-                  'loss': tr_loss/global_step}
-
-        output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
-        with open(output_eval_file, "w") as writer:
-            logger.info("***** Eval results *****")
-            for key in sorted(result.keys()):
-                logger.info("  %s = %s", key, str(result[key]))
-                writer.write("%s = %s\n" % (key, str(result[key])))
+    return results
 
 
 if __name__ == "__main__":
-    main()
+    main()