Merge pull request #116 from FDecaYed/deyuf/fp16_with_apex

Change to use apex for better fp16 and multi-gpu support

Author: thomwolf

README.md

@@ -338,7 +338,7 @@ The optimizer accepts the following arguments:
 - `b1` : Adams b1. Default : `0.9`
 - `b2` : Adams b2. Default : `0.999`
 - `e` : Adams epsilon. Default : `1e-6`
-- `weight_decay_rate:` Weight decay. Default : `0.01`
+- `weight_decay:` Weight decay. Default : `0.01`
 - `max_grad_norm` : Maximum norm for the gradients (`-1` means no clipping). Default : `1.0`
 
 ## Examples

examples/run_classifier.py

@@ -1,5 +1,6 @@
 # coding=utf-8
 # Copyright 2018 The Google AI Language Team Authors and The HugginFace 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.
@@ -35,6 +36,13 @@
 from pytorch_pretrained_bert.optimization import BertAdam
 from pytorch_pretrained_bert.file_utils import PYTORCH_PRETRAINED_BERT_CACHE
 
+try:
+    from apex.optimizers import FP16_Optimizer
+    from apex.optimizers import FusedAdam
+    from apex.parallel import DistributedDataParallel as DDP
+except ImportError:
+    raise ImportError("Please install apex from https://www.github.com/nvidia/apex to run this.")
+
 logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
                     datefmt = '%m/%d/%Y %H:%M:%S',
                     level = logging.INFO)
@@ -295,34 +303,10 @@ def accuracy(out, labels):
     outputs = np.argmax(out, axis=1)
     return np.sum(outputs == labels)
 
-def copy_optimizer_params_to_model(named_params_model, named_params_optimizer):
-    """ Utility function for optimize_on_cpu and 16-bits training.
-        Copy the parameters optimized on CPU/RAM back to the model on GPU
-    """
-    for (name_opti, param_opti), (name_model, param_model) in zip(named_params_optimizer, named_params_model):
-        if name_opti != name_model:
-            logger.error("name_opti != name_model: {} {}".format(name_opti, name_model))
-            raise ValueError
-        param_model.data.copy_(param_opti.data)
-
-def set_optimizer_params_grad(named_params_optimizer, named_params_model, test_nan=False):
-    """ Utility function for optimize_on_cpu and 16-bits training.
-        Copy the gradient of the GPU parameters to the CPU/RAMM copy of the model
-    """
-    is_nan = False
-    for (name_opti, param_opti), (name_model, param_model) in zip(named_params_optimizer, named_params_model):
-        if name_opti != name_model:
-            logger.error("name_opti != name_model: {} {}".format(name_opti, name_model))
-            raise ValueError
-        if param_model.grad is not None:
-            if test_nan and torch.isnan(param_model.grad).sum() > 0:
-                is_nan = True
-            if param_opti.grad is None:
-                param_opti.grad = torch.nn.Parameter(param_opti.data.new().resize_(*param_opti.data.size()))
-            param_opti.grad.data.copy_(param_model.grad.data)
-        else:
-            param_opti.grad = None
-    return is_nan
+def warmup_linear(x, warmup=0.002):
+    if x < warmup:
+        return x/warmup
+    return 1.0 - x
 
 def main():
     parser = argparse.ArgumentParser()
@@ -403,17 +387,15 @@ def main():
                         type=int,
                         default=1,
                         help="Number of updates steps to accumulate before performing a backward/update pass.")
-    parser.add_argument('--optimize_on_cpu',
-                        default=False,
-                        action='store_true',
-                        help="Whether to perform optimization and keep the optimizer averages on CPU")
     parser.add_argument('--fp16',
                         default=False,
                         action='store_true',
                         help="Whether to use 16-bit float precision instead of 32-bit")
     parser.add_argument('--loss_scale',
-                        type=float, default=128,
-                        help='Loss scaling, positive power of 2 values can improve fp16 convergence.')
+                        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()
 
@@ -433,13 +415,11 @@ def main():
         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')
-        if args.fp16:
-            logger.info("16-bits training currently not supported in distributed training")
-            args.fp16 = False # (see https://github.com/pytorch/pytorch/pull/13496)
     logger.info("device %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1))
 
     if args.gradient_accumulation_steps < 1:
@@ -487,32 +467,35 @@ def main():
         model.half()
     model.to(device)
     if args.local_rank != -1:
-        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
-                                                          output_device=args.local_rank)
+        model = DDP(model)
     elif n_gpu > 1:
         model = torch.nn.DataParallel(model)
 
     # Prepare optimizer
-    if args.fp16:
-        param_optimizer = [(n, param.clone().detach().to('cpu').float().requires_grad_()) \
-                            for n, param in model.named_parameters()]
-    elif args.optimize_on_cpu:
-        param_optimizer = [(n, param.clone().detach().to('cpu').requires_grad_()) \
-                            for n, param in model.named_parameters()]
-    else:
-        param_optimizer = list(model.named_parameters())
-    no_decay = ['bias', 'gamma', 'beta']
+    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_rate': 0.01},
-        {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay_rate': 0.0}
+        {'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}
         ]
     t_total = num_train_steps
     if args.local_rank != -1:
         t_total = t_total // torch.distributed.get_world_size()
-    optimizer = BertAdam(optimizer_grouped_parameters,
-                         lr=args.learning_rate,
-                         warmup=args.warmup_proportion,
-                         t_total=t_total)
+    if args.fp16:
+        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 = BertAdam(optimizer_grouped_parameters,
+                             lr=args.learning_rate,
+                             warmup=args.warmup_proportion,
+                             t_total=t_total)
 
     global_step = 0
     if args.do_train:
@@ -543,34 +526,24 @@ def main():
                 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
-                loss.backward()
+
+                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:
-                    if args.fp16 or args.optimize_on_cpu:
-                        if args.fp16 and args.loss_scale != 1.0:
-                            # scale down gradients for fp16 training
-                            for param in model.parameters():
-                                if param.grad is not None:
-                                    param.grad.data = param.grad.data / args.loss_scale
-                        is_nan = set_optimizer_params_grad(param_optimizer, model.named_parameters(), test_nan=True)
-                        if is_nan:
-                            logger.info("FP16 TRAINING: Nan in gradients, reducing loss scaling")
-                            args.loss_scale = args.loss_scale / 2
-                            model.zero_grad()
-                            continue
-                        optimizer.step()
-                        copy_optimizer_params_to_model(model.named_parameters(), param_optimizer)
-                    else:
-                        optimizer.step()
-                    model.zero_grad()
+                    # modify learning rate with special warm up BERT uses
+                    lr_this_step = args.learning_rate * warmup_linear(global_step/t_total, args.warmup_proportion)
+                    for param_group in optimizer.param_groups:
+                        param_group['lr'] = lr_this_step
+                    optimizer.step()
+                    optimizer.zero_grad()
                     global_step += 1
 
     if args.do_eval and (args.local_rank == -1 or torch.distributed.get_rank() == 0):