Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "1fea998a452ff8e0ea85539fbb082e2d538b244a"
Commit 3977b721 authored by Raul Puri's avatar Raul Puri
Browse files

Merge branch 'staging_gpt2_evals' into 'staging' 

Refactoring evaluate gpt2

See merge request ADLR/megatron-lm!38
parents 6c521eb1 a0bcee94
Hide whitespace changes
Inline Side-by-side
Showing with 413 additions and 648 deletions
evaluate_gpt2.py deleted 100755 → 0
View file @ 6c521eb1
# coding=utf-8
# Copyright (c) 2019, 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.
"""Pretrain BERT"""
import os
import json
import math
import random
import numpy as np
import torch
from arguments import get_args
from configure_data import configure_data
from megatron.fp16 import FP16_Module
from megatron.fp16 import FP16_Optimizer
from megatron.learning_rates import AnnealingLR
from megatron.model import GPT2Model
from megatron.model import DistributedDataParallel as DDP
from megatron import mpu
from apex.optimizers import FusedAdam as Adam
from megatron.utils import Timers
from megatron.utils import load_checkpoint
from megatron.utils import report_memory
from megatron.utils import print_params_min_max_norm
from megatron import print_rank_0
from megatron.data_utils import make_tokenizer
from detokenizer import *
def get_model(args):
"""Build the model."""
print_rank_0('building GPT2 model ...')
model = GPT2Model(num_layers=args.num_layers,
vocab_size=args.vocab_size,
hidden_size=args.hidden_size,
num_attention_heads=args.num_attention_heads,
embedding_dropout_prob=args.hidden_dropout,
attention_dropout_prob=args.attention_dropout,
output_dropout_prob=args.hidden_dropout,
max_sequence_length=args.max_position_embeddings,
checkpoint_activations=args.checkpoint_activations,
checkpoint_num_layers=args.checkpoint_num_layers,
parallel_output=not args.cloze_eval)
print_rank_0(' > number of parameters: {}'.format(
sum([p.nelement() for p in model.parameters()])))
# GPU allocation.
model.cuda(torch.cuda.current_device())
# Fp16 conversion.
if args.fp16:
model = FP16_Module(model)
# Wrap model for distributed training.
model = DDP(model)
return model
def setup_model(args):
"""Setup model and optimizer."""
model = get_model(args)
if args.load is not None:
_ = load_checkpoint(
model, None, None, args)
return model
def get_masks_and_position_ids(data,
eod_token,
reset_position_ids,
reset_attention_mask):
# Extract batch size and sequence length.
batch_size, seq_length = data.size()
# Attention mask (lower triangular).
if reset_attention_mask:
att_mask_batch = batch_size
else:
att_mask_batch = 1
attention_mask = torch.tril(torch.ones(
(att_mask_batch, seq_length, seq_length), device=data.device)).view(
att_mask_batch, 1, seq_length, seq_length)
# Loss mask.
loss_mask = torch.ones(data.size(), dtype=torch.float, device=data.device)
loss_mask[data == eod_token] = 0.0
# Position ids.
position_ids = torch.arange(seq_length, dtype=torch.long,
device=data.device)
position_ids = position_ids.unsqueeze(0).expand_as(data)
# We need to clone as the ids will be modifed based on batch index.
if reset_position_ids:
position_ids = position_ids.clone()
if reset_position_ids or reset_attention_mask:
# Loop through the batches:
for b in range(batch_size):
# Find indecies where EOD token is.
eod_index = position_ids[b, data[b] == eod_token]
# Detach indecies from positions if going to modify positions.
if reset_position_ids:
eod_index = eod_index.clone()
# Loop through EOD indecies:
prev_index = 0
for j in range(eod_index.size()[0]):
i = eod_index[j]
# Mask attention loss.
if reset_attention_mask:
attention_mask[b, 0, (i+1):, :(i+1)] = 0
# Reset positions.
if reset_position_ids:
position_ids[b, (i+1):] -= (i + 1 - prev_index)
prev_index = i + 1
return attention_mask, loss_mask, position_ids
def get_batch(data_iterator, args, timers):
''' get_batch subdivides the source data into chunks of
length args.seq_length. If source is equal to the example
output of the data loading example, with a seq_length limit
of 2, we'd get the following two Variables for i = 0:
┌ a g m s ┐ ┌ b h n t ┐
└ b h n t ┘ └ c i o u ┘
Note that despite the name of the function, the subdivison of data is not
done along the batch dimension (i.e. dimension 1), since that was handled
by the data loader. The chunks are along dimension 0, corresponding
to the seq_len dimension in the LSTM. A Variable representing an appropriate
shard reset mask of the same dimensions is also returned.
'''
# Items and their type.
keys = ['text', 'pad_mask']
datatype = torch.int64
# Broadcast data.
timers('data loader').start()
if data_iterator is not None:
data = next(data_iterator)
else:
data = None
timers('data loader').stop()
data_b = mpu.broadcast_data(keys, data, datatype)
# Unpack.
tokens_ = data_b['text'].long()
lm_labels = tokens_[:, 1:].contiguous()
tokens = tokens_[:, :-1].contiguous()
padding_mask = data_b['pad_mask'].byte()
# Get the masks and postition ids.
attention_mask, loss_mask, position_ids = get_masks_and_position_ids(
tokens,
args.eod_token,
args.reset_position_ids,
args.reset_attention_mask)
# Convert
if args.fp16:
attention_mask = attention_mask.half()
return tokens, lm_labels, attention_mask, position_ids, padding_mask
def forward_step(data_iterator, model, args, timers):
"""Forward step."""
# Get the batch.
timers('batch generator').start()
batch = get_batch(data_iterator, args, timers)
if batch is None:
return None
tokens, lm_labels, attention_mask, position_ids, loss_mask = batch
timers('batch generator').stop()
# Forward model.
if args.eval_hf:
output, _ = model(tokens)
else:
output = model(tokens, position_ids, attention_mask)
if not args.cloze_eval:
#losses = torch.nn.CrossEntropyLoss(reduce=False)(
losses = mpu.vocab_parallel_cross_entropy(
output.contiguous().float(), lm_labels.contiguous())
loss_mask = loss_mask.contiguous()
loss_mask = loss_mask.view(-1)
lm_loss = torch.sum(
losses.view(-1) * loss_mask.float())
else:
outputs = torch.argmax(output, -1)
correct = (outputs == lm_labels).float()
correct[(1-loss_mask).bool()] = 1
correct = correct.prod(-1)
lm_loss = correct.sum()
# loss_mask = loss_mask.contiguous().view(-1).float()
# lm_loss = torch.sum(acc * loss_mask)
return lm_loss
def evaluate(data_loader, model, args, timers,
num_iterations=None):
"""Evaluation."""
# Turn on evaluation mode which disables dropout.
model.eval()
total_lm_loss = 0
if num_iterations is not None:
max_iters = num_iterations
else:
if mpu.get_model_parallel_rank() == 0:
max_iters_gpu = torch.cuda.LongTensor([len(data_loader)])
else:
max_iters_gpu = torch.cuda.LongTensor([0])
torch.distributed.broadcast(max_iters_gpu,
mpu.get_model_parallel_src_rank(),
group=mpu.get_model_parallel_group())
max_iters = max_iters_gpu[0].item()
print_rank_0('global rank: {} | max iters: {}'.format(
torch.distributed.get_rank(), max_iters))
if data_loader is not None:
data_iterator = iter(data_loader)
else:
data_iterator = None
with torch.no_grad():
iteration = 0
while iteration < max_iters:
if iteration % args.log_interval == 0:
print_rank_0('global rank: {} | iteration: {}'.format(
torch.distributed.get_rank(), iteration))
# Forward evaluation.
lm_loss = forward_step(data_iterator, model, args, timers)
if lm_loss is None:
break
# Reduce across processes.
if isinstance(model, DDP):
torch.distributed.all_reduce(lm_loss.data)
if args.cloze_eval:
lm_loss.data = lm_loss.data / args.world_size
else:
lm_loss.data = lm_loss.data / args.model_parallel_size
if not args.cloze_eval:
total_lm_loss += lm_loss.data.detach().float().item()/(args.num_tokenized_tokens-1)
else:
total_lm_loss += lm_loss.data.detach().float().item()
iteration += 1
# Move model back to the train mode.
model.train()
return total_lm_loss
def evaluate_and_print_results(prefix, data_iterator, model,
args, timers, num_iterations=None):
"""Helper function to evaluate and dump results on screen."""
if not args.cloze_eval:
lm_loss = evaluate(data_iterator, model, args, timers, num_iterations)
val_loss = lm_loss
ppl = math.exp(min(20, val_loss))
token_ratio = (args.num_tokenized_tokens-1)/(args.num_original_tokens-1)
adjusted_ppl = math.exp(min(20, val_loss*token_ratio))
print_rank_0('-' * 100)
string = ' validation results on {} | '.format(prefix)
string += 'avg loss: {:.4E} | '.format(val_loss)
string += 'ppl: {:.4E} | '.format(ppl)
string += 'adjusted ppl: {:.4E} | '.format(adjusted_ppl)
string += 'token ratio: {} |'.format(token_ratio)
length = len(string) + 1
print_rank_0('-' * length)
print_rank_0(string)
print_rank_0('-' * length)
return val_loss
else:
num_correct = evaluate(data_iterator, model, args, timers, num_iterations)
acc = num_correct / args.num_examples
print_rank_0('-' * 100)
string = ' validation results on {} | '.format(prefix)
string += 'number correct: {:.4E} | '.format(num_correct)
string += 'total examples: {:.4E} | '.format(args.num_examples)
string += 'avg accuracy: {:.4E}'.format(acc)
length = len(string) + 1
print_rank_0('-' * length)
print_rank_0(string)
print_rank_0('-' * length)
return acc
def initialize_distributed(args):
"""Initialize torch.distributed."""
# Manually set the device ids.
device = args.rank % torch.cuda.device_count()
if args.local_rank is not None:
device = args.local_rank
torch.cuda.set_device(device)
# Call the init process
init_method = 'tcp://'
master_ip = os.getenv('MASTER_ADDR', 'localhost')
master_port = os.getenv('MASTER_PORT', '6000')
init_method += master_ip + ':' + master_port
torch.distributed.init_process_group(
backend=args.distributed_backend,
world_size=args.world_size, rank=args.rank,
init_method=init_method)
# Set the model-parallel / data-parallel communicators.
mpu.initialize_model_parallel(args.model_parallel_size)
def set_random_seed(seed):
"""Set random seed for reproducability."""
if seed is not None and seed > 0:
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
mpu.model_parallel_cuda_manual_seed(seed)
class LM_Eval_Dataset(torch.utils.data.Dataset):
def __init__(self, tokens, seq_len, pad_idx, overalapping_eval=None, **kwargs):
self.tokens = tokens
self.seq_len = seq_len
self.pad_idx = pad_idx
self.overalapping_eval = overalapping_eval
if self.overalapping_eval is None:
self.overalapping_eval = self.seq_len
self.overalapping_eval = max(1, self.overalapping_eval)
self.total_targets = len(self.tokens) - 1
# remove first sequence tokens
targets = max(self.total_targets - self.overalapping_eval, 0)
self.total_sequences = max(math.ceil(targets / self.overalapping_eval)+1, 1)
def __len__(self):
return self.total_sequences
def __getitem__(self, idx):
start_idx = idx * self.overalapping_eval
end_idx = start_idx + self.seq_len
tokens = self.tokens[start_idx:end_idx+1]
num_tokens = len(tokens)
pad_mask = [1]*num_tokens
if num_tokens < self.seq_len+1:
num_pad = (self.seq_len+1-num_tokens)
pad_mask += [0]*(num_pad)
tokens += [self.pad_idx] * num_pad
pad_mask = np.array(pad_mask[1:])
if self.overalapping_eval != self.seq_len and idx!=0:
pad_mask[:-self.overalapping_eval] *= 0
return {'text': np.array(tokens), 'pad_mask': pad_mask}
class Lambada_Eval_Dataset(torch.utils.data.Dataset):
def __init__(self, path, tokenizer, seq_len, strict=False, **kwargs):
self.seq_len = seq_len
self.pad_idx = tokenizer.get_command('pad').Id
self.tokenizer = tokenizer
self.strict = strict
self.tokens = []
self.labels = []
with open(path, 'r') as f:
for line in f.readlines():
text = json.loads(line)['text']
tokens, labels = self.get_tokens(text)
self.tokens.append(tokens)
self.labels.append(labels)
def get_tokens(self, text):
if not self.strict:
tokens = self.tokenizer.EncodeAsIds(text).tokenization
return tokens[:-1], [tokens[-1]]
last_token = text.split()[-1]
start_idx = text.rfind(last_token)
beginning_tokens = self.tokenizer.EncodeAsIds(text[:start_idx].strip()).tokenization
last_token = self.tokenizer.EncodeAsIds(' '+last_token).tokenization
return beginning_tokens, last_token
def __len__(self):
return len(self.tokens)
def __getitem__(self, idx):
tokens = self.tokens[idx]
num_tokens = len(tokens)
pad_mask = [0]*num_tokens
labels = self.labels[idx]
pad_mask += [1]*len(labels)
tokens = tokens+labels
num_tokens = len(tokens)
if num_tokens < self.seq_len+1:
num_pad = (self.seq_len+1-num_tokens)
pad_mask += [0]*(num_pad)
tokens += [self.pad_idx] * num_pad
pad_mask = np.array(pad_mask[1:])
return {'text': np.array(tokens), 'pad_mask': pad_mask}
def get_tokenizer(args):
tokenizer_args = {
'tokenizer_type': args.tokenizer_type,
'corpus': None,
'model_path': args.tokenizer_path,
'vocab_size': args.vocab_size,
'model_type': args.tokenizer_model_type,
'cache_dir': args.cache_dir}
return make_tokenizer(**tokenizer_args)
def get_eval_data(args):
val_dataloader = None
if mpu.get_model_parallel_rank() == 0:
eval_batch_size = args.eval_batch_size
eval_batch_size = args.batch_size if eval_batch_size is None else eval_batch_size
seq_len = args.seq_length
valid_data = args.valid_data
valid_data = valid_data[0] if isinstance(valid_data, list) else valid_data
tokenizer = get_tokenizer(args)
if not args.cloze_eval:
with open(valid_data, "rb") as reader:
entire_data = reader.read().decode('utf-8')
num_original_tokens = len(entire_data.strip().split(" "))
entire_data = get_detokenizer(valid_data)(entire_data)
tokenized_data = tokenizer.EncodeAsIds(entire_data).tokenization
num_tokenized_tokens = len(tokenized_data)
string = 'Original Tokens: %d, Detokenized tokens: %d' % (num_tokenized_tokens, num_original_tokens)
print_rank_0(string)
eod_token = tokenizer.get_command('pad').Id
val_dataset = LM_Eval_Dataset(tokenized_data, seq_len, eod_token,
args.overlapping_eval)
else:
val_dataset = Lambada_Eval_Dataset(valid_data, tokenizer, seq_len, args.strict_lambada)
num_tokenized_tokens = 0
num_original_tokens = 0
val_dataloader = torch.utils.data.DataLoader(
val_dataset, batch_size=eval_batch_size, drop_last=False)
before = tokenizer.num_tokens
after = before
multiple = args.make_vocab_size_divisible_by * \
mpu.get_model_parallel_world_size()
while (after % multiple) != 0:
after += 1
print_rank_0('> padded vocab (size: {}) with {} dummy tokens (new size: {})'.
format(before, after - before, after))
eod_token = tokenizer.get_command('pad').Id
num_examples = len(val_dataset)
token_counts = torch.cuda.LongTensor([after, eod_token, num_examples,
num_original_tokens,
num_tokenized_tokens])
else:
token_counts = torch.cuda.LongTensor([0, 0, 0, 0, 0])
torch.distributed.broadcast(token_counts,
mpu.get_model_parallel_src_rank(),
group=mpu.get_model_parallel_group())
args.vocab_size = token_counts[0].item()
args.eod_token = token_counts[1].item()
args.num_examples = token_counts[2].item()
args.num_original_tokens = token_counts[3].item()
args.num_tokenized_tokens = token_counts[4].item()
print('global rank: {} | vocab size: {} | eod token: {} | '
'num_examples: {} | num_original_tokens: {} | '
'num_tokenized_tokens: {}'.format(
torch.distributed.get_rank(), args.vocab_size,
args.eod_token, args.num_examples, args.num_original_tokens,
args.num_tokenized_tokens ))
return val_dataloader
def main():
"""Main training program."""
print('Evaluate GPT2 model')
# Disable CuDNN.
torch.backends.cudnn.enabled = False
# Timer.
timers = Timers()
# Arguments.
args = get_args()
# Pytorch distributed.
initialize_distributed(args)
# Random seeds for reproducability.
set_random_seed(args.seed)
# Data stuff.
eval_data = get_eval_data(args)
# Model, optimizer, and learning rate.
if args.eval_hf:
from pytorch_pretrained_bert import GPT2LMHeadModel
from pytorch_pretrained_bert import GPT2Model as HFGPT2Model
if args.num_layers == 24:
model_path = args.load
#model_path = '/home/universal-lm-data.cosmos549/repos/gpt2_mp/models/345M'
hfmodel = HFGPT2Model.from_pretrained(model_path, cache_dir='gpt2_weights', from_tf=True).cuda()
model = GPT2LMHeadModel(hfmodel.config)
model.transformer.load_state_dict(hfmodel.state_dict())
model.cuda()
else:
model = GPT2LMHeadModel.from_pretrained('gpt2', cache_dir='gpt2_weights').cuda()
else:
if args.load_openai:
from megatron.utils import move_weights
model_path = args.load
args.load = None
model = setup_model(args)
from pytorch_pretrained_bert import GPT2LMHeadModel
from pytorch_pretrained_bert import GPT2Model as HFGPT2Model
model_path = 'gpt2'
from_tf = False
print('loading openai weights')
model.cpu()
if args.num_layers == 24:
#model_path = '/home/universal-lm-data.cosmos549/repos/gpt2_mp/models/345M'
hfmodel = HFGPT2Model.from_pretrained(model_path, cache_dir='gpt2_weights', from_tf=True)
gpt2model = GPT2LMHeadModel(hfmodel.config)
gpt2model.transformer.load_state_dict(hfmodel.state_dict())
gpt2model
else:
gpt2model = GPT2LMHeadModel.from_pretrained('gpt2', cache_dir='gpt2_weights')
model2fill = model
while isinstance(model2fill, (DDP, FP16_Module)):
model2fill = model2fill.module
move_weights(model2fill, gpt2model)
model.cuda()
else:
model = setup_model(args)
# Run on test data.
prefix = "wiki" #os.path.basename(args.valid_data)
evaluate_and_print_results(prefix, eval_data,
model, args, timers)
if __name__ == "__main__":
main()
megatron/arguments.py
View file @ 3977b721
...@@ -173,7 +173,7 @@ def _add_initialization_args(parser): ...@@ -173,7 +173,7 @@ def _add_initialization_args(parser):
def _add_learning_rate_args(parser): def _add_learning_rate_args(parser):
group = parser.add_argument_group(title='learning rate') group = parser.add_argument_group(title='learning rate')
group.add_argument('--lr', type=float, required=True, group.add_argument('--lr', type=float, default=None,
help='Initial learning rate. Depending on decay style ' help='Initial learning rate. Depending on decay style '
'and initial warmup, the learing rate at each ' 'and initial warmup, the learing rate at each '
'iteration would be different.') 'iteration would be different.')
...@@ -297,7 +297,7 @@ def _add_data_args(parser): ...@@ -297,7 +297,7 @@ def _add_data_args(parser):
' validation, and test split. For example the split ' ' validation, and test split. For example the split '
'`90,5,5` will use 90% of data for training, 5% for ' '`90,5,5` will use 90% of data for training, 5% for '
'validation and 5% for test.') 'validation and 5% for test.')
group.add_argument('--vocab-file', type=str, required=True, group.add_argument('--vocab-file', type=str, default=None,
help='Path to the vocab file.') help='Path to the vocab file.')
group.add_argument('--merge-file', type=str, default=None, group.add_argument('--merge-file', type=str, default=None,
help='Path to the BPE merge file.') help='Path to the BPE merge file.')
...@@ -355,35 +355,7 @@ def _add_gpt2_args(parser): ...@@ -355,35 +355,7 @@ def _add_gpt2_args(parser):
return parser return parser
def add_evaluation_args(parser):
"""Evaluation arguments."""
group = parser.add_argument_group('validation', 'validation configurations')
group.add_argument('--eval-batch-size', type=int, default=None,
help='Data Loader batch size for evaluation datasets.'
'Defaults to `--batch-size`')
group.add_argument('--eval-seq-length', type=int, default=None,
help='Maximum sequence length to process for '
'evaluation. Defaults to `--seq-length`')
group.add_argument('--eval-max-preds-per-seq', type=int, default=None,
help='Maximum number of predictions to use for '
'evaluation. Defaults to '
'math.ceil(`--eval-seq-length`*.15/10)*10')
group.add_argument('--overlapping-eval', type=int, default=32,
help='sliding window for overlapping eval ')
group.add_argument('--cloze-eval', action='store_true',
help='Evaluation dataset from `--valid-data` is a cloze task')
group.add_argument('--strict-lambada', action='store_true',
help='use more difficult formulation of lambada')
group.add_argument('--eval-hf', action='store_true',
help='perform evaluation with huggingface openai model.'
'use `--load` to specify weights path to be loaded')
group.add_argument('--load-openai', action='store_true',
help='load openai weights into our model. Use `--load` '
'to specify weights path to be loaded')
return parser
def add_text_generate_args(parser): def add_text_generate_args(parser):
"""Text generate arguments.""" """Text generate arguments."""
......
megatron/utils.py
View file @ 3977b721
...@@ -119,7 +119,8 @@ def get_ltor_masks_and_position_ids(data, ...@@ -119,7 +119,8 @@ def get_ltor_masks_and_position_ids(data,
eod_token, eod_token,
reset_position_ids, reset_position_ids,
reset_attention_mask, reset_attention_mask,
eod_mask_loss): eod_mask_loss,
fp16):
"""Build masks and position id for left to right model.""" """Build masks and position id for left to right model."""
# Extract batch size and sequence length. # Extract batch size and sequence length.
...@@ -169,4 +170,8 @@ def get_ltor_masks_and_position_ids(data, ...@@ -169,4 +170,8 @@ def get_ltor_masks_and_position_ids(data,
position_ids[b, (i+1):] -= (i + 1 - prev_index) position_ids[b, (i+1):] -= (i + 1 - prev_index)
prev_index = i + 1 prev_index = i + 1
# Convert
if fp16:
attention_mask = attention_mask.half()
return attention_mask, loss_mask, position_ids return attention_mask, loss_mask, position_ids
pretrain_gpt2.py
View file @ 3977b721
...@@ -45,6 +45,7 @@ def model_provider(): ...@@ -45,6 +45,7 @@ def model_provider():
def get_batch(data_iterator): def get_batch(data_iterator):
"""Generate a batch""" """Generate a batch"""
args = get_args() args = get_args()
tokenizer = get_tokenizer()
# Items and their type. # Items and their type.
keys = ['text'] keys = ['text']
...@@ -65,13 +66,11 @@ def get_batch(data_iterator): ...@@ -65,13 +66,11 @@ def get_batch(data_iterator):
# Get the masks and postition ids. # Get the masks and postition ids.
attention_mask, loss_mask, position_ids = get_ltor_masks_and_position_ids( attention_mask, loss_mask, position_ids = get_ltor_masks_and_position_ids(
tokens, tokens,
args.eod_token, tokenizer.eod,
args.reset_position_ids, args.reset_position_ids,
args.reset_attention_mask, args.reset_attention_mask,
args.eod_mask_loss) args.eod_mask_loss,
# Convert args.fp16)
if args.fp16:
attention_mask = attention_mask.half()
return tokens, labels, loss_mask, attention_mask, position_ids return tokens, labels, loss_mask, attention_mask, position_ids
...@@ -160,9 +159,6 @@ def get_train_val_test_data(): ...@@ -160,9 +159,6 @@ def get_train_val_test_data():
args.do_valid = flags[1].item() args.do_valid = flags[1].item()
args.do_test = flags[2].item() args.do_test = flags[2].item()
tokenizer = get_tokenizer()
args.eod_token = tokenizer.eod_id
return train_data, val_data, test_data return train_data, val_data, test_data
......
scripts/run_gpt2_eval.py
View file @ 3977b721
...@@ -30,19 +30,8 @@ parser.add_argument('--cloze-eval', action='store_true', ...@@ -30,19 +30,8 @@ parser.add_argument('--cloze-eval', action='store_true',
help='Run lambada cloze eval instead of perplexity eval.') help='Run lambada cloze eval instead of perplexity eval.')
parser.add_argument('--easy-lambada', action='store_true', parser.add_argument('--easy-lambada', action='store_true',
help='use easier formulation of lambada') help='use easier formulation of lambada')
parser.add_argument('--webtext-eval', action='store_true',
help='Run webtext PPL eval instead of wikitext PPL eval.')
parser.add_argument('--eval-iters', default=5000, type=int,
help='number of iterations to run webtext evaluation')
parser.add_argument('--model-parallel-size', type=int, default=1, parser.add_argument('--model-parallel-size', type=int, default=1,
help='model parallel size to use') help='model parallel size to use')
parser.add_argument('--load-openai', action='store_true',
help='Load weights from saved openai/hf checkpoints')
parser.add_argument('--cache-dir', type=str, default='cache',
help='directory to cache gpt2 tokenizers')
parser.add_argument('--make-vocab-size-divisible-by', type=int, default=128,
help='Pad the vocab size to be divisible by this value.'
'This is added for computational efficieny reasons.')
args = parser.parse_args() args = parser.parse_args()
multinode_args = '' multinode_args = ''
...@@ -54,43 +43,36 @@ CMD = ' --model-parallel-size {model_par} \ ...@@ -54,43 +43,36 @@ CMD = ' --model-parallel-size {model_par} \
--hidden-size {hidden} \ --hidden-size {hidden} \
--log-interval 100 \ --log-interval 100 \
--load {model} \ --load {model} \
--eval-batch-size {batch} \ --batch-size {batch} \
--num-attention-heads {natt} \ --num-attention-heads {natt} \
--seq-length 1024 \ --seq-length 1024 \
--max-position-embeddings 1024 \ --max-position-embeddings 1024 \
--tokenizer-type GPT2BPETokenizer \ --tokenizer-type GPT2BPETokenizer \
--text-key text \
--distributed-backend nccl \ --distributed-backend nccl \
--hidden-dropout 0.1 \ --hidden-dropout 0.1 \
--attention-dropout 0.1 \ --attention-dropout 0.1 \
--fp16 \ --fp16 \
--lr 1 --no-load-optim --no-load-rng --epochs 0 \
--overlapping-eval 32 \ --overlapping-eval 32 \
--make-vocab-size-divisible-by {make_vocab_size_divisible_by} \ --merge-file /home/universal-lm-data.cosmos549/repos/megatron_latest/vocab_cache/merges.txt \
--cache-dir {cache} '.format(model_par=args.model_parallel_size, --vocab-file /home/universal-lm-data.cosmos549/repos/megatron_latest/vocab_cache/vocab.json'.format(model_par=args.model_parallel_size,
nlayers=args.num_layers, nlayers=args.num_layers,
hidden=args.hidden_size, hidden=args.hidden_size,
model=args.model_path, model=args.model_path,
batch=args.batch_size, batch=args.batch_size,
natt=args.num_attention_heads, natt=args.num_attention_heads,)
make_vocab_size_divisible_by=args.make_vocab_size_divisible_by,
cache=args.cache_dir)
if args.load_openai:
CMD += ' --load-openai '
if args.cloze_eval: if args.cloze_eval:
CMD += ' --valid-data {} '.format(args.data_path) CMD += ' --valid-data {} '.format(args.data_path)
CMD += ' --cloze-eval ' CMD += ' --task LAMBADA '
if not args.easy_lambada: if not args.easy_lambada:
CMD += ' --strict-lambada ' CMD += ' --strict-lambada '
CMD = 'evaluate_gpt2.py' + CMD CMD = 'main.py' + CMD
print('Running Lambada Eval Command:', flush=True) print('Running Lambada Eval Command:', flush=True)
elif args.webtext_eval:
CMD += '--train-iters 0 --eval-iters {} --test-data {} --loose-json '.format(args.eval_iters, args.data_path)
CMD = 'pretrain_gpt2.py' + CMD
print('Running Webtext Eval Command:', flush=True)
else: else:
CMD += ' --valid-data {} '.format(args.data_path) CMD += ' --valid-data {} '.format(args.data_path)
CMD = 'evaluate_gpt2.py' + CMD CMD += ' --task WIKITEXT103 '
CMD = 'main.py' + CMD
print('Running PPL Eval Command:', flush=True) print('Running PPL Eval Command:', flush=True)
CMD = 'python3 '+multinode_args+CMD CMD = 'python3 '+multinode_args+CMD
......
tasks/main.py
View file @ 3977b721
...@@ -30,7 +30,7 @@ def get_tasks_args(parser): ...@@ -30,7 +30,7 @@ def get_tasks_args(parser):
group.add_argument('--task', type=str, required=True, group.add_argument('--task', type=str, required=True,
help='Task name.') help='Task name.')
group.add_argument('--epochs', type=int, required=True, group.add_argument('--epochs', type=int, default=None,
help='Number of finetunning epochs. Zero results in ' help='Number of finetunning epochs. Zero results in '
'evaluation only.') 'evaluation only.')
group.add_argument('--pretrained-checkpoint', type=str, default=None, group.add_argument('--pretrained-checkpoint', type=str, default=None,
...@@ -43,6 +43,10 @@ def get_tasks_args(parser): ...@@ -43,6 +43,10 @@ def get_tasks_args(parser):
'for training.') 'for training.')
group.add_argument('--valid-data', nargs='*', default=None, group.add_argument('--valid-data', nargs='*', default=None,
help='path(s) to the validation data.') help='path(s) to the validation data.')
group.add_argument('--overlapping-eval', type=int, default=32,
help='Sliding window for overlapping evaluation.')
group.add_argument('--strict-lambada', action='store_true',
help='Use more difficult formulation of lambada.')
return parser return parser
...@@ -56,6 +60,8 @@ if __name__ == '__main__': ...@@ -56,6 +60,8 @@ if __name__ == '__main__':
from race.finetune import main from race.finetune import main
elif args.task in ['MNLI', 'QQP']: elif args.task in ['MNLI', 'QQP']:
from glue.finetune import main from glue.finetune import main
elif args.task in ['LAMBADA', 'WIKITEXT103']:
from zeroshot_gpt2.evaluate import main
else: else:
raise NotImplementedError('Task {} is not implemented.'.format( raise NotImplementedError('Task {} is not implemented.'.format(
args.task)) args.task))
......
tasks/zeroshot_gpt2/datasets.py 0 → 100644
View file @ 3977b721
# coding=utf-8
# Copyright (c) 2019, 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.
"""Zero-shot datasets."""
import json
import math
import numpy as np
import torch
from megatron import get_args
from megatron import get_tokenizer
from megatron import print_rank_0
from .detokenizer import get_detokenizer
def build_dataset(task):
"""Helper function to select and build dataset."""
if task == 'LAMBADA':
return _build_lambada_dataset()
if task == 'WIKITEXT103':
return _build_wikitext103_dataset()
raise NotImplementedError('dataset for {} task is not '
'implemented.'.format(task))
class _LMDataset(torch.utils.data.Dataset):
def __init__(self, tokens, seq_len, pad_idx, num_original_tokens,
num_tokenized_tokens, overalapping_eval=None):
self.tokens = tokens
self.seq_len = seq_len
self.pad_idx = pad_idx
self.overalapping_eval = overalapping_eval
if self.overalapping_eval is None:
self.overalapping_eval = self.seq_len
self.overalapping_eval = max(1, self.overalapping_eval)
self.num_original_tokens = num_original_tokens
self.num_tokenized_tokens = num_tokenized_tokens
self.total_targets = len(self.tokens) - 1
# remove first sequence tokens
targets = max(self.total_targets - self.overalapping_eval, 0)
self.total_sequences = max(
math.ceil(targets / self.overalapping_eval) + 1, 1)
def __len__(self):
return self.total_sequences
def __getitem__(self, idx):
start_idx = idx * self.overalapping_eval
end_idx = start_idx + self.seq_len
tokens = self.tokens[start_idx:end_idx+1]
num_tokens = len(tokens)
pad_mask = [1]*num_tokens
if num_tokens < self.seq_len+1:
num_pad = (self.seq_len+1-num_tokens)
pad_mask += [0]*(num_pad)
tokens += [self.pad_idx] * num_pad
pad_mask = np.array(pad_mask[1:])
if self.overalapping_eval != self.seq_len and idx != 0:
pad_mask[:-self.overalapping_eval] *= 0
return {'text': np.array(tokens), 'pad_mask': pad_mask}
class _LambadaDataset(torch.utils.data.Dataset):
def __init__(self, path, pad_idx, tokenizer, seq_len, strict=False):
print_rank_0('> building lambada dataset from {} ...'.format(path))
self.seq_len = seq_len
self.pad_idx = pad_idx
self.tokenizer = tokenizer
self.strict = strict
self.tokens = []
self.labels = []
with open(path, 'r') as f:
for line in f.readlines():
text = json.loads(line)['text']
tokens, labels = self.get_tokens(text)
self.tokens.append(tokens)
self.labels.append(labels)
def get_tokens(self, text):
if not self.strict:
tokens = self.tokenizer.tokenize(text)
return tokens[:-1], [tokens[-1]]
last_token = text.split()[-1]
start_idx = text.rfind(last_token)
beginning_tokens = self.tokenizer.tokenize(text[:start_idx].strip())
last_token = self.tokenizer.tokenize(' '+last_token)
return beginning_tokens, last_token
def __len__(self):
return len(self.tokens)
def __getitem__(self, idx):
tokens = self.tokens[idx]
num_tokens = len(tokens)
pad_mask = [0]*num_tokens
labels = self.labels[idx]
pad_mask += [1]*len(labels)
tokens = tokens+labels
num_tokens = len(tokens)
if num_tokens < self.seq_len+1:
num_pad = (self.seq_len+1-num_tokens)
pad_mask += [0]*(num_pad)
tokens += [self.pad_idx] * num_pad
pad_mask = np.array(pad_mask[1:])
return {'text': np.array(tokens), 'pad_mask': pad_mask}
def _build_lambada_dataset():
"""Build lambada dataset."""
args = get_args()
tokenizer = get_tokenizer()
assert len(args.valid_data) == 1
val_dataset = _LambadaDataset(args.valid_data[0], tokenizer.eod, tokenizer,
args.seq_length, args.strict_lambada)
print_rank_0(' > found {} samples.'.format(len(val_dataset)))
return val_dataset
def _build_wikitext103_dataset():
""""""
args = get_args()
tokenizer = get_tokenizer()
assert len(args.valid_data) == 1
with open(args.valid_data[0], "rb") as reader:
entire_data = reader.read().decode('utf-8')
num_original_tokens = len(entire_data.strip().split(" "))
entire_data = get_detokenizer(args.valid_data[0])(entire_data)
tokenized_data = tokenizer.tokenize(entire_data)
num_tokenized_tokens = len(tokenized_data)
val_dataset = _LMDataset(tokenized_data, args.seq_length, tokenizer.eod,
num_original_tokens, num_tokenized_tokens,
args.overlapping_eval)
print_rank_0(' > number of original tokens: {}, number of detokenized '
'tokens: {}'.format(num_original_tokens, num_tokenized_tokens))
return val_dataset
detokenizer.py → tasks/zeroshot_gpt2/detokenizer.py 100755 → 100644
View file @ 3977b721
# coding=utf-8
# Copyright (c) 2019, 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.
"""Detokenization."""
import re import re
def ptb_detokenizer(string): def ptb_detokenizer(string):
string = string.replace(" '", "'") string = string.replace(" '", "'")
string = string.replace(" \n", "\n") string = string.replace(" \n", "\n")
...@@ -44,17 +62,21 @@ def wikitext_detokenizer(string): ...@@ -44,17 +62,21 @@ def wikitext_detokenizer(string):
return string return string
def lambada_detokenizer(string): def lambada_detokenizer(string):
return string return string
def get_detokenizer(path):
for key in DETOKENIZERS.keys():
if key in path:
print(key)
return DETOKENIZERS[key]
DETOKENIZERS = { _DETOKENIZERS = {
'ptb': ptb_detokenizer, 'ptb': ptb_detokenizer,
'wikitext': wikitext_detokenizer, 'wikitext': wikitext_detokenizer,
'lambada': lambada_detokenizer, 'lambada': lambada_detokenizer,
} }
def get_detokenizer(path):
for key in _DETOKENIZERS.keys():
if key in path:
return _DETOKENIZERS[key]
tasks/zeroshot_gpt2/evaluate.py 0 → 100644
View file @ 3977b721
# coding=utf-8
# Copyright (c) 2019, 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.
"""GPT2 zero-shot evaluation."""
import math
import torch
from megatron import get_args
from megatron import get_tokenizer
from megatron import mpu
from megatron import print_rank_0
from megatron.checkpointing import load_checkpoint
from megatron.model import GPT2Model
from megatron.training import get_model
from megatron.utils import get_ltor_masks_and_position_ids
from tasks.finetune_utils import build_data_loader
from .datasets import build_dataset
def get_model_provider(eval_metric):
"""Based on evaluation metric set the parallel-output flag and
return the model provider."""
def model_provider():
"""Build the model."""
if eval_metric == 'loss':
parallel_output = True
elif eval_metric == 'accuracy':
parallel_output = False
else:
raise NotImplementedError('output type for {} evaluation metric '
'is not supported.'.format(eval_metric))
print_rank_0('building GPT2 model ...')
model = GPT2Model(num_tokentypes=0, parallel_output=parallel_output)
return model
return model_provider
def process_batch(batch):
"""Process batch and produce inputs for the model."""
args = get_args()
tokenizer = get_tokenizer()
loss_mask = batch['pad_mask'].long().cuda().contiguous().byte()
tokens_ = batch['text'].long().cuda().contiguous()
labels = tokens_[:, 1:].contiguous()
tokens = tokens_[:, :-1].contiguous()
# Get the masks and postition ids.
attention_mask, _, position_ids = get_ltor_masks_and_position_ids(
tokens,
tokenizer.eod,
args.reset_position_ids,
args.reset_attention_mask,
args.eod_mask_loss,
args.fp16)
return tokens, labels, attention_mask, position_ids, loss_mask
def forward_step(batch, model, eval_metric):
"""Forward step."""
# Get the batch.
tokens, labels, attention_mask, position_ids, loss_mask = process_batch(
batch)
# Forward model.
output = model(tokens, position_ids, attention_mask)
# For loss, return the unreduced loss.
if eval_metric == 'loss':
losses = mpu.vocab_parallel_cross_entropy(
output.contiguous().float(), labels.contiguous())
loss = torch.sum(
losses.view(-1) * loss_mask.contiguous().view(-1).float())
return loss
# For accuracy, return the number of correctly predicted samples.
if eval_metric == 'accuracy':
outputs = torch.argmax(output, -1)
correct = (outputs == labels).float()
correct[(1 - loss_mask).bool()] = 1
correct = correct.prod(-1)
return correct.sum()
raise NotImplementedError('forward method for evaluation metric {} '
'is not implemented.'.format(eval_metric))
def evaluate(data_loader, model, eval_metric):
"""Evaluation."""
args = get_args()
# Turn on evaluation mode which disables dropout.
model.eval()
total_output = 0.0
with torch.no_grad():
# For all the batches in the dataset.
for iteration, batch in enumerate(data_loader):
if iteration % args.log_interval == 0:
print_rank_0('> working on iteration: {}'.format(iteration))
# Forward evaluation.
output = forward_step(batch, model, eval_metric)
# Reduce across processes.
torch.distributed.all_reduce(output,
group=mpu.get_data_parallel_group())
total_output += output
return total_output
def evaluate_and_print_results(task, data_loader, model, eval_metric):
"""Evaluate and print results on screen."""
# Evaluate and get results.
output = evaluate(data_loader, model, eval_metric)
string = ' validation results on {} | '.format(task)
if eval_metric == 'loss':
num_tokenized_tokens = data_loader.dataset.num_tokenized_tokens
num_original_tokens = data_loader.dataset.num_original_tokens
val_loss = output / (num_tokenized_tokens - 1)
ppl = math.exp(min(20, val_loss))
token_ratio = (num_tokenized_tokens - 1) / (num_original_tokens - 1)
adjusted_ppl = math.exp(min(20, val_loss * token_ratio))
string += 'avg loss: {:.4E} | '.format(val_loss)
string += 'ppl: {:.4E} | '.format(ppl)
string += 'adjusted ppl: {:.4E} | '.format(adjusted_ppl)
string += 'token ratio: {} |'.format(token_ratio)
elif eval_metric == 'accuracy':
num_examples = len(data_loader.dataset)
acc = output / num_examples
string += 'number correct: {:.4E} | '.format(output)
string += 'total examples: {:.4E} | '.format(num_examples)
string += 'avg accuracy: {:.4E}'.format(acc)
else:
raise NotImplementedError('evaluation method for {} metric is not '
'implemented yet.'.format(eval_metric))
length = len(string) + 1
print_rank_0('-' * length)
print_rank_0(string)
print_rank_0('-' * length)
def main():
"""Main program."""
args = get_args()
if args.task == 'LAMBADA':
eval_metric = 'accuracy'
elif args.task == 'WIKITEXT103':
eval_metric = 'loss'
else:
raise NotImplementedError('{} task is not implemented.'.format(
args.task))
# Set up model and load checkpoint.
model = get_model(get_model_provider(eval_metric))
if args.load is not None:
_ = load_checkpoint(model, None, None)
# Data stuff.
dataset = build_dataset(args.task)
dataloader = build_data_loader(dataset, args.batch_size,
args.num_workers, drop_last=False)
# Run evaluation.
evaluate_and_print_results(args.task, dataloader, model, eval_metric)
print_rank_0('done :-)')
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment