From 9c6a48c8c3e73ed3272c07f897930c17d5b0286d Mon Sep 17 00:00:00 2001 From: Matej Svejda Date: Sun, 27 Jan 2019 14:07:24 +0100 Subject: [PATCH] fix learning rate/fp16 and warmup problem for all examples --- examples/run_classifier.py | 8 +++++++- examples/run_lm_finetuning.py | 20 ++++++++------------ examples/run_squad.py | 19 ++++++++----------- examples/run_squad2.py | 19 ++++++++----------- examples/run_swag.py | 19 ++++++++----------- 5 files changed, 39 insertions(+), 46 deletions(-) diff --git a/examples/run_classifier.py b/examples/run_classifier.py index 7adcf1097c..0b60eb66ed 100644 --- a/examples/run_classifier.py +++ b/examples/run_classifier.py @@ -33,7 +33,7 @@ from torch.utils.data.distributed import DistributedSampler from pytorch_pretrained_bert.tokenization import BertTokenizer from pytorch_pretrained_bert.modeling import BertForSequenceClassification -from pytorch_pretrained_bert.optimization import BertAdam +from pytorch_pretrained_bert.optimization import BertAdam, warmup_linear from pytorch_pretrained_bert.file_utils import PYTORCH_PRETRAINED_BERT_CACHE logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s', @@ -536,6 +536,12 @@ def main(): nb_tr_examples += input_ids.size(0) nb_tr_steps += 1 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(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 diff --git a/examples/run_lm_finetuning.py b/examples/run_lm_finetuning.py index 35a2f797c7..2e26842c14 100644 --- a/examples/run_lm_finetuning.py +++ b/examples/run_lm_finetuning.py @@ -31,7 +31,7 @@ from torch.utils.data.distributed import DistributedSampler from pytorch_pretrained_bert.tokenization import BertTokenizer from pytorch_pretrained_bert.modeling import BertForPreTraining -from pytorch_pretrained_bert.optimization import BertAdam +from pytorch_pretrained_bert.optimization import BertAdam, warmup_linear from torch.utils.data import Dataset import random @@ -42,12 +42,6 @@ logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message logger = logging.getLogger(__name__) -def warmup_linear(x, warmup=0.002): - if x < warmup: - return x/warmup - return 1.0 - x - - class BERTDataset(Dataset): def __init__(self, corpus_path, tokenizer, seq_len, encoding="utf-8", corpus_lines=None, on_memory=True): self.vocab = tokenizer.vocab @@ -527,7 +521,7 @@ def main(): train_dataset = BERTDataset(args.train_file, tokenizer, seq_len=args.max_seq_length, corpus_lines=None, on_memory=args.on_memory) num_train_steps = int( - len(train_dataset) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs) + len(train_dataset) / args.train_batch_size * args.num_train_epochs) # Prepare model model = BertForPreTraining.from_pretrained(args.bert_model) @@ -607,10 +601,12 @@ def main(): nb_tr_examples += input_ids.size(0) nb_tr_steps += 1 if (step + 1) % args.gradient_accumulation_steps == 0: - # modify learning rate with special warm up BERT uses - lr_this_step = args.learning_rate * warmup_linear(global_step/num_train_steps, args.warmup_proportion) - for param_group in optimizer.param_groups: - param_group['lr'] = lr_this_step + 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(global_step/num_train_steps, args.warmup_proportion) + for param_group in optimizer.param_groups: + param_group['lr'] = lr_this_step optimizer.step() optimizer.zero_grad() global_step += 1 diff --git a/examples/run_squad.py b/examples/run_squad.py index 39e9c50199..0881e82aba 100644 --- a/examples/run_squad.py +++ b/examples/run_squad.py @@ -36,7 +36,7 @@ from torch.utils.data.distributed import DistributedSampler from pytorch_pretrained_bert.tokenization import whitespace_tokenize, BasicTokenizer, BertTokenizer from pytorch_pretrained_bert.modeling import BertForQuestionAnswering -from pytorch_pretrained_bert.optimization import BertAdam +from pytorch_pretrained_bert.optimization import BertAdam, warmup_linear from pytorch_pretrained_bert.file_utils import PYTORCH_PRETRAINED_BERT_CACHE logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s', @@ -670,11 +670,6 @@ def _compute_softmax(scores): probs.append(score / total_sum) return probs -def warmup_linear(x, warmup=0.002): - if x < warmup: - return x/warmup - return 1.0 - x - def main(): parser = argparse.ArgumentParser() @@ -794,7 +789,7 @@ def main(): train_examples = read_squad_examples( input_file=args.train_file, is_training=True) num_train_steps = int( - len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs) + len(train_examples) / args.train_batch_size * args.num_train_epochs) # Prepare model model = BertForQuestionAnswering.from_pretrained(args.bert_model, @@ -905,10 +900,12 @@ def main(): else: loss.backward() if (step + 1) % args.gradient_accumulation_steps == 0: - # 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 + 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(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 diff --git a/examples/run_squad2.py b/examples/run_squad2.py index 558b24764e..ad5e820db8 100644 --- a/examples/run_squad2.py +++ b/examples/run_squad2.py @@ -36,7 +36,7 @@ from torch.utils.data.distributed import DistributedSampler from pytorch_pretrained_bert.tokenization import whitespace_tokenize, BasicTokenizer, BertTokenizer from pytorch_pretrained_bert.modeling import BertForQuestionAnswering -from pytorch_pretrained_bert.optimization import BertAdam +from pytorch_pretrained_bert.optimization import BertAdam, warmup_linear from pytorch_pretrained_bert.file_utils import PYTORCH_PRETRAINED_BERT_CACHE logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s', @@ -759,11 +759,6 @@ def _compute_softmax(scores): probs.append(score / total_sum) return probs -def warmup_linear(x, warmup=0.002): - if x < warmup: - return x/warmup - return 1.0 - x - def main(): parser = argparse.ArgumentParser() @@ -887,7 +882,7 @@ def main(): train_examples = read_squad_examples( input_file=args.train_file, is_training=True) num_train_steps = int( - len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs) + len(train_examples) / args.train_batch_size * args.num_train_epochs) # Prepare model model = BertForQuestionAnswering.from_pretrained(args.bert_model, @@ -999,10 +994,12 @@ def main(): else: loss.backward() if (step + 1) % args.gradient_accumulation_steps == 0: - # 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 + 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(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 diff --git a/examples/run_swag.py b/examples/run_swag.py index 3fb87ae3e7..597b093a26 100644 --- a/examples/run_swag.py +++ b/examples/run_swag.py @@ -29,7 +29,7 @@ from torch.utils.data.distributed import DistributedSampler from pytorch_pretrained_bert.tokenization import BertTokenizer from pytorch_pretrained_bert.modeling import BertForMultipleChoice -from pytorch_pretrained_bert.optimization import BertAdam +from pytorch_pretrained_bert.optimization import BertAdam, warmup_linear from pytorch_pretrained_bert.file_utils import PYTORCH_PRETRAINED_BERT_CACHE logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s', @@ -233,11 +233,6 @@ def select_field(features, field): for feature in features ] -def warmup_linear(x, warmup=0.002): - if x < warmup: - return x/warmup - return 1.0 - x - def main(): parser = argparse.ArgumentParser() @@ -358,7 +353,7 @@ def main(): if args.do_train: train_examples = read_swag_examples(os.path.join(args.data_dir, 'train.csv'), is_training = True) num_train_steps = int( - len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs) + len(train_examples) / args.train_batch_size * args.num_train_epochs) # Prepare model model = BertForMultipleChoice.from_pretrained(args.bert_model, @@ -457,10 +452,12 @@ def main(): else: loss.backward() if (step + 1) % args.gradient_accumulation_steps == 0: - # 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 + 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(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