def main():
args = parser.parse_args()
bps.init()
if args.seed is not None:
random.seed(args.seed)
torch.manual_seed(args.seed)
cudnn.deterministic = True
warnings.warn('You have chosen to seed training. '
'This will turn on the CUDNN deterministic setting, '
'which can slow down your training considerably! '
'You may see unexpected behavior when restarting '
'from checkpoints.')
if args.gpu is not None:
warnings.warn('You have chosen a specific GPU. This will completely '
'disable data parallelism.')
if args.dist_url == "env://" and args.world_size == -1:
args.world_size = int(os.environ["WORLD_SIZE"])
args.distributed = args.world_size > 1 or args.multiprocessing_distributed
ngpus_per_node = torch.cuda.device_count()
if args.multiprocessing_distributed:
# Since we have ngpus_per_node processes per node, the total world_size
# needs to be adjusted accordingly
args.world_size = ngpus_per_node * args.world_size
# Use torch.multiprocessing.spawn to launch distributed processes: the
# main_worker process function
args.gpu = bps.local_rank()
main_worker(args.gpu, ngpus_per_node, args)
else:
# Simply call main_worker function
main_worker(args.gpu, ngpus_per_node, args)
def main():
bps.init()
torch.manual_seed(1)
torch.cuda.manual_seed(1)
torch.cuda.set_device(bps.local_rank())
# parse arguments
args = parse_args()
if args is None:
exit()
# open session
gan = UGATIT(args)
# build graph
gan.build_model()
if args.phase == 'train':
gan.train()
print(" [*] Training finished!")
if args.phase == 'test':
gan.test()
print(" [*] Test finished!")
parser.add_argument('--no-wait',
type=bool,
default=True,
help='wait for other worker request first')
parser.add_argument('--gpu', type=int, default=-1, help='use a specified gpu')
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()
hvd.init()
# Horovod: pin GPU to local rank.
if args.gpu >= 0:
torch.cuda.set_device(args.gpu)
else:
torch.cuda.set_device(hvd.local_rank())
cudnn.benchmark = True
def log(s, nl=True):
if hvd.rank() != 0:
return
print(s, end='\n' if nl else '')
def benchmark(tensor, average, name):
if not args.no_wait and hvd.rank() == 0:
# let other workers submit allreduce request first
time.sleep(0.01)
start = time.time()
action='store_true',
default=False,
help='disables CUDA training')
parser.add_argument('--seed', type=int, default=42, help='random seed')
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()
allreduce_batch_size = args.batch_size * args.batches_per_allreduce
bps.init()
torch.manual_seed(args.seed)
if args.cuda:
# BytePS: pin GPU to local rank.
torch.cuda.set_device(bps.local_rank())
torch.cuda.manual_seed(args.seed)
cudnn.benchmark = True
# If set > 0, will resume training from a given checkpoint.
resume_from_epoch = 0
for try_epoch in range(args.epochs, 0, -1):
if os.path.exists(args.checkpoint_format.format(epoch=try_epoch)):
resume_from_epoch = try_epoch
break
# BytePS: broadcast resume_from_epoch from rank 0 (which will have
# checkpoints) to other ranks.
#resume_from_epoch = bps.broadcast(torch.tensor(resume_from_epoch), root_rank=0,
# name='resume_from_epoch').item()
def log(s, nl=True):
if bps.local_rank() != 0:
return
print(s, end='\n' if nl else '')
sys.stdout.flush()
action='store_true',
default=False,
help='disables profiler')
parser.add_argument('--partition',
type=int,
default=None,
help='partition size')
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()
bps.init()
if args.cuda:
# BytePS: pin GPU to local rank.
torch.cuda.set_device(bps.local_rank())
cudnn.benchmark = True
# Set up standard model.
model = getattr(models, args.model)(num_classes=args.num_classes)
if args.cuda:
# Move model to GPU.
model.cuda()
optimizer = optim.SGD(model.parameters(), lr=0.01)
model = DDP(model, device_ids=[bps.local_rank()])
# BytePS: (optional) compression algorithm.
#compression = bps.Compression.fp16 if args.fp16_pushpull else bps.Compression.none
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument("--train_data_file",
default=None,
type=str,
required=True,
help="The input training data file (a text file).")
parser.add_argument(
"--output_dir",
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written.",
)
parser.add_argument(
"--model_type",
type=str,
required=True,
help="The model architecture to be trained or fine-tuned.",
)
# Other parameters
parser.add_argument(
"--eval_data_file",
default=None,
type=str,
help=
"An optional input evaluation data file to evaluate the perplexity on (a text file).",
)
parser.add_argument(
"--line_by_line",
action="store_true",
help=
"Whether distinct lines of text in the dataset are to be handled as distinct sequences.",
)
parser.add_argument(
"--should_continue",
action="store_true",
help="Whether to continue from latest checkpoint in output_dir")
parser.add_argument(
"--model_name_or_path",
default=None,
type=str,
help=
"The model checkpoint for weights initialization. Leave None if you want to train a model from scratch.",
)
parser.add_argument(
"--mlm",
action="store_true",
help=
"Train with masked-language modeling loss instead of language modeling."
)
parser.add_argument(
"--mlm_probability",
type=float,
default=0.15,
help="Ratio of tokens to mask for masked language modeling loss")
parser.add_argument(
"--config_name",
default=None,
type=str,
help=
"Optional pretrained config name or path if not the same as model_name_or_path. If both are None, initialize a new config.",
)
parser.add_argument(
"--tokenizer_name",
default=None,
type=str,
help=
"Optional pretrained tokenizer name or path if not the same as model_name_or_path. If both are None, initialize a new tokenizer.",
)
parser.add_argument(
"--cache_dir",
default=None,
type=str,
help=
"Optional directory to store the pre-trained models downloaded from s3 (instead of the default one)",
)
parser.add_argument(
"--block_size",
default=-1,
type=int,
help="Optional input sequence length after tokenization."
"The training dataset will be truncated in block of this size for training."
"Default to the model max input length for single sentence inputs (take into account special tokens).",
)
parser.add_argument("--do_train",
action="store_true",
help="Whether to run training.")
parser.add_argument("--do_eval",
action="store_true",
help="Whether to run eval on the dev set.")
parser.add_argument(
"--evaluate_during_training",
action="store_true",
help="Run evaluation during training at each logging step.")
parser.add_argument("--per_gpu_train_batch_size",
default=4,
type=int,
help="Batch size per GPU/CPU for training.")
parser.add_argument("--per_gpu_eval_batch_size",
default=4,
type=int,
help="Batch size per GPU/CPU for evaluation.")
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("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay",
default=0.0,
type=float,
help="Weight decay 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=1.0,
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=500,
help="Log every X updates steps.")
parser.add_argument("--save_steps",
type=int,
default=500,
help="Save checkpoint every X updates steps.")
parser.add_argument(
"--save_total_limit",
type=int,
default=None,
help=
"Limit the total amount of checkpoints, delete the older checkpoints in the output_dir, does not delete by default",
)
parser.add_argument(
"--eval_all_checkpoints",
action="store_true",
help=
"Evaluate all checkpoints starting with the same prefix as model_name_or_path ending and ending with step number",
)
parser.add_argument("--no_cuda",
action="store_true",
help="Avoid using 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(
"--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="For distant debugging.")
parser.add_argument("--server_port",
type=str,
default="",
help="For distant debugging.")
args = parser.parse_args()
# assign local rank
args.local_rank = bps.local_rank()
if args.model_type in ["bert", "roberta", "distilbert", "camembert"
] and not args.mlm:
raise ValueError(
"BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the --mlm "
"flag (masked language modeling).")
if args.eval_data_file is None and args.do_eval:
raise ValueError(
"Cannot do evaluation without an evaluation data file. Either supply a file to --eval_data_file "
"or remove the --do_eval argument.")
if args.should_continue:
sorted_checkpoints = _sorted_checkpoints(args)
if len(sorted_checkpoints) == 0:
raise ValueError(
"Used --should_continue but no checkpoint was found in --output_dir."
)
else:
args.model_name_or_path = sorted_checkpoints[-1]
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")
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)
args.n_gpu = 1
args.device = device
# 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,
)
# Set seed
set_seed(args)
# Load pretrained model and tokenizer
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
if args.config_name:
config = config_class.from_pretrained(args.config_name,
cache_dir=args.cache_dir)
elif args.model_name_or_path:
config = config_class.from_pretrained(args.model_name_or_path,
cache_dir=args.cache_dir)
else:
config = config_class()
if args.tokenizer_name:
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name,
cache_dir=args.cache_dir)
elif args.model_name_or_path:
tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path,
cache_dir=args.cache_dir)
else:
raise ValueError(
"You are instantiating a new {} tokenizer. This is not supported, but you can do it from another script, save it,"
"and load it from here, using --tokenizer_name".format(
tokenizer_class.__name__))
if args.block_size <= 0:
args.block_size = tokenizer.max_len_single_sentence
# Our input block size will be the max possible for the model
else:
args.block_size = min(args.block_size,
tokenizer.max_len_single_sentence)
if args.model_name_or_path:
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir,
)
else:
logger.info("Training new model from scratch")
model = model_class(config=config)
model.to(args.device)
logger.info("Training/evaluation parameters %s", args)
# Training
if args.do_train:
train_dataset = load_and_cache_examples(args,
tokenizer,
evaluate=False)
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step,
tr_loss)
# Saving best-practices: if you use save_pretrained for the model and tokenizer, you can reload them using from_pretrained()
if args.do_train and (args.local_rank == -1 or bps.rank() == 0):
# Create output directory if needed
if args.local_rank in [-1, 0]:
os.makedirs(args.output_dir, exist_ok=True)
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)
# 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 = model_class.from_pretrained(args.output_dir)
tokenizer = tokenizer_class.from_pretrained(args.output_dir)
model.to(args.device)
# Evaluation
results = {}
if args.do_eval and args.local_rank in [-1, 0]:
checkpoints = [args.output_dir]
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("transformers.modeling_utils").setLevel(
logging.WARN) # Reduce logging
logger.info("Evaluate the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
global_step = checkpoint.split(
"-")[-1] if len(checkpoints) > 1 else ""
prefix = checkpoint.split(
"/")[-1] if checkpoint.find("checkpoint") != -1 else ""
model = model_class.from_pretrained(checkpoint)
model.to(args.device)
result = evaluate(args, model, tokenizer, prefix=prefix)
result = dict(
(k + "_{}".format(global_step), v) for k, v in result.items())
results.update(result)
return results
def train(self):
self.genA2B.train(), self.genB2A.train(), self.disGA.train(
), self.disGB.train(), self.disLA.train(), self.disLB.train()
start_iter = 1
if self.resume:
model_list = glob(
os.path.join(self.result_dir, self.dataset, 'model', '*.pt'))
if not len(model_list) == 0:
model_list.sort()
start_iter = int(model_list[-1].split('_')[-1].split('.')[0])
self.load(os.path.join(self.result_dir, self.dataset, 'model'),
start_iter)
print(" [*] Load SUCCESS")
if self.decay_flag and start_iter > (self.iteration // 2):
self.G_optim.param_groups[0]['lr'] -= (
self.lr /
(self.iteration // 2)) * (start_iter -
self.iteration // 2)
self.D_optim.param_groups[0]['lr'] -= (
self.lr /
(self.iteration // 2)) * (start_iter -
self.iteration // 2)
# training loop
print('training start !')
start_time = time.time()
last_time = start_time
for step in range(start_iter, self.iteration + 1):
if self.decay_flag and step > (self.iteration // 2):
self.G_optim.param_groups[0]['lr'] -= (self.lr /
(self.iteration // 2))
self.D_optim.param_groups[0]['lr'] -= (self.lr /
(self.iteration // 2))
try:
real_A, _ = trainA_iter.next()
except:
trainA_iter = iter(self.trainA_loader)
real_A, _ = trainA_iter.next()
try:
real_B, _ = trainB_iter.next()
except:
trainB_iter = iter(self.trainB_loader)
real_B, _ = trainB_iter.next()
real_A, real_B = real_A.to(self.device), real_B.to(self.device)
# Update D
self.D_optim._handles.clear()
self.D_optim.zero_grad()
self.D_optim.set_backward_passes_per_step(1)
self.G_optim.set_backward_passes_per_step(10)
fake_A2B, _, _ = self.genA2B(real_A)
fake_B2A, _, _ = self.genB2A(real_B)
real_GA_logit, real_GA_cam_logit, _ = self.disGA(real_A)
real_LA_logit, real_LA_cam_logit, _ = self.disLA(real_A)
real_GB_logit, real_GB_cam_logit, _ = self.disGB(real_B)
real_LB_logit, real_LB_cam_logit, _ = self.disLB(real_B)
fake_GA_logit, fake_GA_cam_logit, _ = self.disGA(fake_B2A)
fake_LA_logit, fake_LA_cam_logit, _ = self.disLA(fake_B2A)
fake_GB_logit, fake_GB_cam_logit, _ = self.disGB(fake_A2B)
fake_LB_logit, fake_LB_cam_logit, _ = self.disLB(fake_A2B)
D_ad_loss_GA = self.MSE_loss(
real_GA_logit,
torch.ones_like(real_GA_logit).to(
self.device)) + self.MSE_loss(
fake_GA_logit,
torch.zeros_like(fake_GA_logit).to(self.device))
D_ad_cam_loss_GA = self.MSE_loss(
real_GA_cam_logit,
torch.ones_like(real_GA_cam_logit).to(
self.device)) + self.MSE_loss(
fake_GA_cam_logit,
torch.zeros_like(fake_GA_cam_logit).to(self.device))
D_ad_loss_LA = self.MSE_loss(
real_LA_logit,
torch.ones_like(real_LA_logit).to(
self.device)) + self.MSE_loss(
fake_LA_logit,
torch.zeros_like(fake_LA_logit).to(self.device))
D_ad_cam_loss_LA = self.MSE_loss(
real_LA_cam_logit,
torch.ones_like(real_LA_cam_logit).to(
self.device)) + self.MSE_loss(
fake_LA_cam_logit,
torch.zeros_like(fake_LA_cam_logit).to(self.device))
D_ad_loss_GB = self.MSE_loss(
real_GB_logit,
torch.ones_like(real_GB_logit).to(
self.device)) + self.MSE_loss(
fake_GB_logit,
torch.zeros_like(fake_GB_logit).to(self.device))
D_ad_cam_loss_GB = self.MSE_loss(
real_GB_cam_logit,
torch.ones_like(real_GB_cam_logit).to(
self.device)) + self.MSE_loss(
fake_GB_cam_logit,
torch.zeros_like(fake_GB_cam_logit).to(self.device))
D_ad_loss_LB = self.MSE_loss(
real_LB_logit,
torch.ones_like(real_LB_logit).to(
self.device)) + self.MSE_loss(
fake_LB_logit,
torch.zeros_like(fake_LB_logit).to(self.device))
D_ad_cam_loss_LB = self.MSE_loss(
real_LB_cam_logit,
torch.ones_like(real_LB_cam_logit).to(
self.device)) + self.MSE_loss(
fake_LB_cam_logit,
torch.zeros_like(fake_LB_cam_logit).to(self.device))
D_loss_A = self.adv_weight * (D_ad_loss_GA + D_ad_cam_loss_GA +
D_ad_loss_LA + D_ad_cam_loss_LA)
D_loss_B = self.adv_weight * (D_ad_loss_GB + D_ad_cam_loss_GB +
D_ad_loss_LB + D_ad_cam_loss_LB)
Discriminator_loss = D_loss_A + D_loss_B
Discriminator_loss.backward()
self.D_optim.step()
# Update G
self.G_optim._handles.clear()
self.G_optim.zero_grad()
self.D_optim.set_backward_passes_per_step(10)
self.G_optim.set_backward_passes_per_step(1)
fake_A2B, fake_A2B_cam_logit, _ = self.genA2B(real_A)
fake_B2A, fake_B2A_cam_logit, _ = self.genB2A(real_B)
fake_A2B2A, _, _ = self.genB2A(fake_A2B)
fake_B2A2B, _, _ = self.genA2B(fake_B2A)
fake_A2A, fake_A2A_cam_logit, _ = self.genB2A(real_A)
fake_B2B, fake_B2B_cam_logit, _ = self.genA2B(real_B)
fake_GA_logit, fake_GA_cam_logit, _ = self.disGA(fake_B2A)
fake_LA_logit, fake_LA_cam_logit, _ = self.disLA(fake_B2A)
fake_GB_logit, fake_GB_cam_logit, _ = self.disGB(fake_A2B)
fake_LB_logit, fake_LB_cam_logit, _ = self.disLB(fake_A2B)
G_ad_loss_GA = self.MSE_loss(
fake_GA_logit,
torch.ones_like(fake_GA_logit).to(self.device))
G_ad_cam_loss_GA = self.MSE_loss(
fake_GA_cam_logit,
torch.ones_like(fake_GA_cam_logit).to(self.device))
G_ad_loss_LA = self.MSE_loss(
fake_LA_logit,
torch.ones_like(fake_LA_logit).to(self.device))
G_ad_cam_loss_LA = self.MSE_loss(
fake_LA_cam_logit,
torch.ones_like(fake_LA_cam_logit).to(self.device))
G_ad_loss_GB = self.MSE_loss(
fake_GB_logit,
torch.ones_like(fake_GB_logit).to(self.device))
G_ad_cam_loss_GB = self.MSE_loss(
fake_GB_cam_logit,
torch.ones_like(fake_GB_cam_logit).to(self.device))
G_ad_loss_LB = self.MSE_loss(
fake_LB_logit,
torch.ones_like(fake_LB_logit).to(self.device))
G_ad_cam_loss_LB = self.MSE_loss(
fake_LB_cam_logit,
torch.ones_like(fake_LB_cam_logit).to(self.device))
G_recon_loss_A = self.L1_loss(fake_A2B2A, real_A)
G_recon_loss_B = self.L1_loss(fake_B2A2B, real_B)
G_identity_loss_A = self.L1_loss(fake_A2A, real_A)
G_identity_loss_B = self.L1_loss(fake_B2B, real_B)
G_cam_loss_A = self.BCE_loss(
fake_B2A_cam_logit,
torch.ones_like(fake_B2A_cam_logit).to(
self.device)) + self.BCE_loss(
fake_A2A_cam_logit,
torch.zeros_like(fake_A2A_cam_logit).to(self.device))
G_cam_loss_B = self.BCE_loss(
fake_A2B_cam_logit,
torch.ones_like(fake_A2B_cam_logit).to(
self.device)) + self.BCE_loss(
fake_B2B_cam_logit,
torch.zeros_like(fake_B2B_cam_logit).to(self.device))
G_loss_A = self.adv_weight * (
G_ad_loss_GA + G_ad_cam_loss_GA + G_ad_loss_LA +
G_ad_cam_loss_LA
) + self.cycle_weight * G_recon_loss_A + self.identity_weight * G_identity_loss_A + self.cam_weight * G_cam_loss_A
G_loss_B = self.adv_weight * (
G_ad_loss_GB + G_ad_cam_loss_GB + G_ad_loss_LB +
G_ad_cam_loss_LB
) + self.cycle_weight * G_recon_loss_B + self.identity_weight * G_identity_loss_B + self.cam_weight * G_cam_loss_B
Generator_loss = G_loss_A + G_loss_B
Generator_loss.backward()
self.G_optim.step()
# clip parameter of AdaILN and ILN, applied after optimizer step
self.genA2B.apply(self.Rho_clipper)
self.genB2A.apply(self.Rho_clipper)
if bps.local_rank() == 0:
this_time = time.time()
print(
"[%5d/%5d] time: %4.4f speed: %4.4f (sec/step) d_loss: %.8f, g_loss: %.8f"
%
(step, self.iteration, this_time - start_time, this_time -
last_time, Discriminator_loss, Generator_loss))
last_time = this_time
if step % self.print_freq == 0:
train_sample_num = 5
test_sample_num = 5
A2B = np.zeros((self.img_size * 7, 0, 3))
B2A = np.zeros((self.img_size * 7, 0, 3))
self.genA2B.eval(), self.genB2A.eval(), self.disGA.eval(
), self.disGB.eval(), self.disLA.eval(), self.disLB.eval()
for _ in range(train_sample_num):
try:
real_A, _ = trainA_iter.next()
except:
trainA_iter = iter(self.trainA_loader)
real_A, _ = trainA_iter.next()
try:
real_B, _ = trainB_iter.next()
except:
trainB_iter = iter(self.trainB_loader)
real_B, _ = trainB_iter.next()
real_A, real_B = real_A.to(self.device), real_B.to(
self.device)
fake_A2B, _, fake_A2B_heatmap = self.genA2B(real_A)
fake_B2A, _, fake_B2A_heatmap = self.genB2A(real_B)
fake_A2B2A, _, fake_A2B2A_heatmap = self.genB2A(fake_A2B)
fake_B2A2B, _, fake_B2A2B_heatmap = self.genA2B(fake_B2A)
fake_A2A, _, fake_A2A_heatmap = self.genB2A(real_A)
fake_B2B, _, fake_B2B_heatmap = self.genA2B(real_B)
A2B = np.concatenate(
(A2B,
np.concatenate(
(RGB2BGR(tensor2numpy(denorm(real_A[0]))),
cam(tensor2numpy(fake_A2A_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_A2A[0]))),
cam(tensor2numpy(fake_A2B_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_A2B[0]))),
cam(tensor2numpy(fake_A2B2A_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_A2B2A[0])))),
0)), 1)
B2A = np.concatenate(
(B2A,
np.concatenate(
(RGB2BGR(tensor2numpy(denorm(real_B[0]))),
cam(tensor2numpy(fake_B2B_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_B2B[0]))),
cam(tensor2numpy(fake_B2A_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_B2A[0]))),
cam(tensor2numpy(fake_B2A2B_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_B2A2B[0])))),
0)), 1)
for _ in range(test_sample_num):
try:
real_A, _ = testA_iter.next()
except:
testA_iter = iter(self.testA_loader)
real_A, _ = testA_iter.next()
try:
real_B, _ = testB_iter.next()
except:
testB_iter = iter(self.testB_loader)
real_B, _ = testB_iter.next()
real_A, real_B = real_A.to(self.device), real_B.to(
self.device)
fake_A2B, _, fake_A2B_heatmap = self.genA2B(real_A)
fake_B2A, _, fake_B2A_heatmap = self.genB2A(real_B)
fake_A2B2A, _, fake_A2B2A_heatmap = self.genB2A(fake_A2B)
fake_B2A2B, _, fake_B2A2B_heatmap = self.genA2B(fake_B2A)
fake_A2A, _, fake_A2A_heatmap = self.genB2A(real_A)
fake_B2B, _, fake_B2B_heatmap = self.genA2B(real_B)
A2B = np.concatenate(
(A2B,
np.concatenate(
(RGB2BGR(tensor2numpy(denorm(real_A[0]))),
cam(tensor2numpy(fake_A2A_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_A2A[0]))),
cam(tensor2numpy(fake_A2B_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_A2B[0]))),
cam(tensor2numpy(fake_A2B2A_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_A2B2A[0])))),
0)), 1)
B2A = np.concatenate(
(B2A,
np.concatenate(
(RGB2BGR(tensor2numpy(denorm(real_B[0]))),
cam(tensor2numpy(fake_B2B_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_B2B[0]))),
cam(tensor2numpy(fake_B2A_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_B2A[0]))),
cam(tensor2numpy(fake_B2A2B_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_B2A2B[0])))),
0)), 1)
if bps.rank() == 0:
cv2.imwrite(
os.path.join(self.result_dir, self.dataset, 'img',
'A2B_%07d.png' % step), A2B * 255.0)
cv2.imwrite(
os.path.join(self.result_dir, self.dataset, 'img',
'B2A_%07d.png' % step), B2A * 255.0)
self.genA2B.train(), self.genB2A.train(), self.disGA.train(
), self.disGB.train(), self.disLA.train(), self.disLB.train()
if (step % self.save_freq == 0) and (bps.rank() == 0):
self.save(os.path.join(self.result_dir, self.dataset, 'model'),
step)
if (step % 1000 == 0) and (bps.rank() == 0):
params = {}
params['genA2B'] = self.genA2B.state_dict()
params['genB2A'] = self.genB2A.state_dict()
params['disGA'] = self.disGA.state_dict()
params['disGB'] = self.disGB.state_dict()
params['disLA'] = self.disLA.state_dict()
params['disLB'] = self.disLB.state_dict()
torch.save(
params,
os.path.join(self.result_dir,
self.dataset + '_params_latest.pt'))
if (bps.rank() == 0):
self.save(os.path.join(self.result_dir, self.dataset, 'model'),
step)
def main():
# os.system('shutdown -c') # cancel previous shutdown command
log.console(args)
tb.log('sizes/world', bps.size())
# need to index validation directory before we start counting the time
dataloader.sort_ar(args.data + '/validation')
# if args.distributed:
# log.console('Distributed initializing process group')
torch.cuda.set_device(bps.local_rank())
print(f'cuda device set to {bps.local_rank()}')
log.console("cuda initialized (rank=%d)" % (bps.local_rank()))
# dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url, world_size=bps.size())
log.console("Distributed: success (%d/%d)" % (bps.rank(), bps.size()))
log.console("Loading model (rank=%d)" % (bps.rank()))
model = resnet.resnet50(bn0=args.init_bn0).cuda()
# reuse the validate tensor
global validate_tensor, dist_validate_tensor
validate_tensor = torch.tensor([0, 0, 0, 0]).float().cuda()
dist_validate_tensor = torch.tensor([0, 0, 0, 0, 0]).float().cuda()
if args.fp16: model = network_to_half(model)
best_top5 = 93 # only save models over 93%. Otherwise it stops to save every time
global model_params, master_params
if args.fp16: model_params, master_params = prep_param_lists(model)
else: model_params = master_params = model.parameters()
optim_params, name_list = experimental_utils.bnwd_optim_params(
model, model_params, master_params) if args.no_bn_wd else master_params
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(
optim_params,
0,
momentum=args.momentum,
weight_decay=args.weight_decay
) # start with 0 lr. Scheduler will change this later
named_param = []
for p in optim_params:
tensors = p['params']
for tensor in tensors:
named_param.append(tensor)
# create bps_param (tuple)
bps_param = []
for i, tensor in enumerate(named_param):
name = name_list[i]
bps_param.append((name, tensor))
# wrap with byteps optimizer
optimizer = DistributedOptimizer(
optimizer,
named_parameters=bps_param,
backward_passes_per_step=args.batches_per_pushpull,
half=True,
model=model,
fp16_params=model_params,
fp32_params=master_params,
loss_scale=args.loss_scale)
if args.resume:
checkpoint = torch.load(
args.resume,
map_location=lambda storage, loc: storage.cuda(args.local_rank))
model.load_state_dict(checkpoint['state_dict'])
args.start_epoch = checkpoint['epoch']
best_top5 = checkpoint['best_top5']
optimizer.load_state_dict(checkpoint['optimizer'])
log.console(
"Creating data loaders (this could take up to 10 minutes if volume needs to be warmed up)"
)
num_machines = (bps.size() - 1) // 8 + 1
assert (num_machines in schedules)
phases = schedules[num_machines]
dm = DataManager([copy.deepcopy(p) for p in phases if 'bs' in p])
scheduler = Scheduler(optimizer,
[copy.deepcopy(p) for p in phases if 'lr' in p])
# BytePS: broadcast parameters & optimizer state.
broadcast_parameters([(name, p.detach()) for name, p in bps_param],
root_rank=0)
broadcast_optimizer_state(optimizer, root_rank=0)
start_time = datetime.now() # Loading start to after everything is loaded
if args.evaluate:
return validate(dm.val_dl, model, criterion, 0, start_time)
if args.distributed:
log.console('Global Barrier: Syncing machines before training')
tensor = torch.tensor([1.0]).float().cuda()
barrier_handler = push_pull_async_inplace(tensor,
average=True,
name="init.barrier")
while True:
if poll(barrier_handler):
synchronize(barrier_handler)
break
# do broadcast for validate tensor
log.console('Broadcasting validate tensor')
barrier_handler = push_pull_async_inplace(validate_tensor,
average=True,
name="validation_tensor")
while True:
if poll(barrier_handler):
synchronize(barrier_handler)
break
barrier_handler = push_pull_async_inplace(
dist_validate_tensor,
average=True,
name="distributed_validation_tensor")
while True:
if poll(barrier_handler):
synchronize(barrier_handler)
break
log.event("~~epoch\thours\ttop1\ttop5\n")
for epoch in range(args.start_epoch, scheduler.tot_epochs):
dm.set_epoch(epoch)
train(dm.trn_dl, model, criterion, optimizer, scheduler, epoch)
top1, top5 = validate(dm.val_dl, model, criterion, epoch, start_time)
time_diff = (datetime.now() - start_time).total_seconds() / 3600.0
log.event(f'~~{epoch}\t{time_diff:.5f}\t\t{top1:.3f}\t\t{top5:.3f}\n')
is_best = top5 > best_top5
best_top5 = max(top5, best_top5)
if args.local_rank == 0:
if is_best:
save_checkpoint(epoch,
model,
best_top5,
optimizer,
is_best=True,
filename='model_best.pth.tar')
phase = dm.get_phase(epoch)
if phase:
save_checkpoint(
epoch,
model,
best_top5,
optimizer,
filename=f'sz{phase["bs"]}_checkpoint.path.tar')
schedules = {
1: one_machine,
2: two_machine,
4: four_machines,
8: eight_machines,
16: sixteen_machines
}
bps.init()
cudnn.benchmark = True
args = get_parser().parse_args()
# Only want master rank logging to tensorboard
is_master = (not args.distributed) or (bps.rank() == 0)
is_rank0 = bps.local_rank() == 0
tb = TensorboardLogger(args.logdir, is_master=is_master)
log = FileLogger(args.logdir, is_master=is_master, is_rank0=is_rank0)
def main():
# os.system('shutdown -c') # cancel previous shutdown command
log.console(args)
tb.log('sizes/world', bps.size())
# need to index validation directory before we start counting the time
dataloader.sort_ar(args.data + '/validation')
# if args.distributed:
# log.console('Distributed initializing process group')
torch.cuda.set_device(bps.local_rank())
