File size: 11,317 Bytes
c2c125c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 |
# coding=utf-8
# Copyright (c) 2020, 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.
"""Finetune utilities."""
import torch
import torch.nn.functional as F
from megatron import get_args
from megatron import print_rank_0
from megatron import get_timers
from megatron import mpu, utils
from megatron.checkpointing import load_checkpoint
from megatron.checkpointing import save_checkpoint
from megatron.training import evaluate_and_print_results
from megatron.training import setup_model_and_optimizer
from megatron.training import train_step
from megatron.training import training_log
from megatron.utils import check_adlr_autoresume_termination
from megatron.utils import average_losses_across_data_parallel_group, print_params_min_max_norm
from torch.nn.parallel.distributed import DistributedDataParallel as torchDDP
from megatron.model import DistributedDataParallel as LocalDDP
from megatron.model import Float16Module, ModelType
def process_batch(batch):
"""Process batch and produce inputs for the model."""
images = batch[0].cuda().contiguous()
labels = batch[1].cuda().contiguous()
return images, labels
def build_data_loader(dataset, micro_batch_size,
num_workers, drop_last, shuffle):
"""Data loader. Note that batch-size is the local (per GPU) batch-size."""
# Sampler.
world_size = mpu.get_data_parallel_world_size()
rank = mpu.get_data_parallel_rank()
sampler = torch.utils.data.distributed.DistributedSampler(
dataset, num_replicas=world_size, rank=rank,
drop_last=drop_last, shuffle=shuffle
)
# Data loader. Note that batch size is the per GPU batch size.
data_loader = torch.utils.data.DataLoader(
dataset,
batch_size=micro_batch_size,
sampler=sampler,
shuffle=False,
num_workers=num_workers,
drop_last=drop_last,
pin_memory=True,
)
return data_loader
def _build_infinite_size_dataloader(dataloader):
"""Build a looped dataloader with infinite size."""
iterator = dataloader.__iter__()
while True:
try:
yield iterator.__next__()
except StopIteration:
iterator = dataloader.__iter__()
def _build_train_valid_dataloaders(train_dataset, valid_dataset):
"""Traing and validation dataloaders."""
args = get_args()
print_rank_0('building train and validation dataloaders ...')
# Training dataset.
train_dataloader = build_data_loader(train_dataset, args.micro_batch_size,
args.num_workers, False, True)
# Set the training iterations.
args.train_iters_per_epoch = len(train_dataloader)
args.train_iters = args.epochs * args.train_iters_per_epoch
# Validation dataset. For this dataset, we do not need to set up
# shuffling so we can just use a simple infinite loop.
valid_dataloader_ = build_data_loader(valid_dataset, args.micro_batch_size,
args.num_workers, True, False)
valid_dataloader = _build_infinite_size_dataloader(valid_dataloader_)
# Now that we've built the data loaders, set batch_size arguments
# to the actual batch size the model will see for this dataset.
# This is necessary so pipeline transfers know what size they are
# and the LR schedule, which is based on samples seen, gets set
# correctly.
args.orig_micro_batch_size = args.micro_batch_size
args.orig_global_batch_size = args.global_batch_size
return train_dataloader, valid_dataloader
def _train(
model,
optimizer,
opt_param_scheduler,
forward_step,
train_dataloader,
valid_dataloader,
end_of_epoch_callback,
process_non_loss_data_func=None
):
"""Train the model."""
args = get_args()
timers = get_timers()
# Turn on training mode which enables dropout.
for m in model:
m.train()
# Tracking loss.
losses_dict_sum = {}
# Starting epoch and iteration
start_epoch = args.iteration // args.train_iters_per_epoch
start_iteration = args.iteration % args.train_iters_per_epoch
iteration = args.iteration
# Memory reporting flag.
report_memory_flag = True
# For each remaining epoch
timers("interval-time").start()
for epoch in range(start_epoch, args.epochs):
print_rank_0("working on epoch {} ...".format(epoch + 1))
# Set the data loader epoch to shuffle the index iterator.
train_dataloader.sampler.set_epoch(args.seed + epoch)
train_dataloader.dataset.set_epoch(epoch)
# For all the batches in the dataset.
for iteration_, batch in enumerate(train_dataloader):
# Ignore the iterations before starting value
if iteration_ < start_iteration:
continue
# Set to zero so the next epoch does not skip any batches.
start_iteration = 0
# Train for one step.
losses_dict, skipped_iter, grad_norm, num_zeros_in_grad = train_step(
forward_step, batch, model, optimizer, opt_param_scheduler
)
iteration += 1
# Logging.
params_norm = None
report_memory_flag = training_log(
losses_dict,
losses_dict_sum,
optimizer.param_groups[0]["lr"],
iteration,
optimizer.get_loss_scale().item(),
report_memory_flag,
skipped_iter,
grad_norm,
params_norm,
num_zeros_in_grad
)
# Autoresume
if args.adlr_autoresume and \
iteration % args.adlr_autoresume_interval == 0:
check_adlr_autoresume_termination(iteration, model, optimizer,
opt_param_scheduler)
# Checkpointing
if args.save and args.save_interval and \
iteration % args.save_interval == 0:
save_checkpoint(iteration, model, optimizer,
opt_param_scheduler)
# Evaluation
if args.eval_interval and iteration % args.eval_interval == 0:
prefix = "iteration {}".format(iteration)
evaluate_and_print_results(
prefix,
forward_step,
valid_dataloader,
model,
iteration,
process_non_loss_data_func,
False,
)
# Callback at the end of each epoch.
if end_of_epoch_callback is not None:
end_of_epoch_callback(model, epoch)
def finetune(
train_valid_datasets_provider,
model_provider,
forward_step,
model_type=ModelType.encoder_or_decoder,
process_non_loss_data_func=None,
end_of_epoch_callback_provider=None,
):
"""Main finetune function used across all tasks."""
args = get_args()
timers = get_timers()
# Train and validation data loaders.
timers("train/valid/test dataset/dataloder").start()
if args.epochs > 0:
train_dataset, valid_dataset = train_valid_datasets_provider()
train_dataloader, valid_dataloader = _build_train_valid_dataloaders(
train_dataset, valid_dataset
)
timers("train/valid/test dataset/dataloder").stop()
# Build calback function.
timers("callback function").start()
end_of_epoch_callback = None
if end_of_epoch_callback_provider is not None:
end_of_epoch_callback = end_of_epoch_callback_provider()
timers("callback function").stop()
# Build model, optimizer and learning rate scheduler.
timers("model and optimizer").start()
model, optimizer, opt_param_scheduler = \
setup_model_and_optimizer(
model_provider,
model_type,
scale_lr_cond=lambda name, param: ".head." in name,
lr_mult=args.head_lr_mult)
timers("model and optimizer").stop()
# If pretrained checkpoint is provided and we have not trained for
# any iteration (i.e., iteration is zero), then load the pretrained
# checkpoint.
timers("pretrained checkpoint").start()
if args.iteration == 0 and args.pretrained_checkpoint is not None:
if args.pretrained_checkpoint_type == 'default':
original_load = args.load
args.load = args.pretrained_checkpoint
_ = load_checkpoint(model, None, None, strict=False)
args.load = original_load
elif args.pretrained_checkpoint_type == 'external':
unwrap_model = utils.unwrap_model(model)
state_dict = torch.load(args.pretrained_checkpoint,
map_location="cpu")
unwrap_model[0].module.backbone.load_state_dict(state_dict,
strict=False)
elif args.pretrained_checkpoint_type == 'constrastive':
unwrap_model = utils.unwrap_model(model)
state_dict = torch.load(args.pretrained_checkpoint,
map_location="cpu")
state_dict = state_dict["model"]
state_dict = {k.replace("teacher.backbone.", ""): v
for k, v in state_dict.items()
if k.startswith("teacher.backbone.")}
unwrap_model[0].module.backbone.load_state_dict(state_dict,
strict=False)
else:
raise Exception("pretrained checkpoint type {} not supported".format(args.pretrained_checkpoint_type))
# This is critical when only model is loaded. We should make sure
# master parameters are also updated.
optimizer.reload_model_params()
timers("pretrained checkpoint").stop()
# Print setup timing.
print_rank_0("done with setups ...")
timers.log(
[
"train/valid/test dataset/dataloder",
"callback function",
"model and optimizer",
"pretrained checkpoint",
]
)
print_rank_0("training ...")
# Finetune the model.
if args.epochs > 0:
_train(
model,
optimizer,
opt_param_scheduler,
forward_step,
train_dataloader,
valid_dataloader,
end_of_epoch_callback,
process_non_loss_data_func,
)
# Or just evaluate.
else:
if end_of_epoch_callback is not None:
print_rank_0("evaluation only mode, setting epoch to -1")
end_of_epoch_callback(model, epoch=-1)
print_rank_0("done :-)")
|