def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--bert_model",
default="bert-base-uncased",
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.",
)
parser.add_argument(
"--from_pretrained",
default="bert-base-uncased",
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.",
)
parser.add_argument(
"--output_dir",
default="save",
type=str,
help=
"The output directory where the model checkpoints will be written.",
)
parser.add_argument(
"--config_file",
default="config/bert_config.json",
type=str,
help="The config file which specified the model details.",
)
parser.add_argument("--learning_rate",
default=2e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument(
"--num_train_epochs",
default=20,
type=int,
help="Total number of training epochs to perform.",
)
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.",
)
parser.add_argument("--no_cuda",
action="store_true",
help="Whether not to use CUDA when available")
parser.add_argument(
"--do_lower_case",
default=True,
type=bool,
help=
"Whether to lower case the input text. True for uncased models, False for cased models.",
)
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument("--seed",
type=int,
default=0,
help="random seed for initialization")
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help=
"Number of updates steps to accumualte before performing a backward/update pass.",
)
parser.add_argument(
"--fp16",
action="store_true",
help="Whether to use 16-bit float precision instead of 32-bit",
)
parser.add_argument(
"--loss_scale",
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n",
)
parser.add_argument("--num_workers",
type=int,
default=16,
help="Number of workers in the dataloader.")
parser.add_argument(
"--save_name",
default='',
type=str,
help="save name for training.",
)
parser.add_argument("--use_chunk",
default=0,
type=float,
help="whether use chunck for parallel training.")
parser.add_argument("--in_memory",
default=False,
type=bool,
help="whether use chunck for parallel training.")
parser.add_argument("--optimizer",
default='BertAdam',
type=str,
help="whether use chunck for parallel training.")
parser.add_argument("--tasks",
default='',
type=str,
help="1-2-3... training task separate by -")
parser.add_argument(
"--freeze",
default=-1,
type=int,
help="till which layer of textual stream of vilbert need to fixed.")
parser.add_argument("--vision_scratch",
action="store_true",
help="whether pre-trained the image or not.")
parser.add_argument("--evaluation_interval",
default=1,
type=int,
help="evaluate very n epoch.")
parser.add_argument("--lr_scheduler",
default='mannul',
type=str,
help="whether use learning rate scheduler.")
parser.add_argument("--baseline",
action="store_true",
help="whether use single stream baseline.")
parser.add_argument("--compact",
action="store_true",
help="whether use compact vilbert model.")
parser.add_argument("--debug",
action="store_true",
help="whether in debug mode.")
parser.add_argument(
"--tensorboard_dir",
default="tensorboard_log",
type=str,
help="The output directory where tensorboard log will be written.",
)
parser.add_argument(
"--batch_size",
default=-1,
type=int,
help="Custom Batch size for task.",
)
parser.add_argument(
"--data_root",
default="",
type=str,
help="The data root of the task.",
)
args = parser.parse_args()
with open('vlbert_tasks.yml', 'r') as f:
task_cfg = edict(yaml.load(f))
# random.seed(args.seed)
# np.random.seed(args.seed)
# torch.manual_seed(args.seed)
if args.baseline:
from pytorch_pretrained_bert.modeling import BertConfig
from vilbert.basebert import BaseBertForVLTasks
elif args.compact:
from vilbert.vilbert_compact import BertConfig
from vilbert.vilbert_compact import VILBertForVLTasks
else:
from vilbert.vilbert import BertConfig
from vilbert.vilbert import VILBertForVLTasks
task_names = []
task_lr = []
for i, task_id in enumerate(args.tasks.split('-')):
task = 'TASK' + task_id
name = task_cfg[task]['name']
task_names.append(name)
task_lr.append(task_cfg[task]['lr'])
base_lr = min(task_lr)
loss_scale = {}
for i, task_id in enumerate(args.tasks.split('-')):
task = 'TASK' + task_id
loss_scale[task] = task_lr[i] / base_lr
if args.save_name:
prefix = '-' + args.save_name
else:
prefix = ''
timeStamp = '-'.join(task_names) + '_' + args.config_file.split(
'/')[1].split('.')[0] + prefix
savePath = os.path.join(args.output_dir, timeStamp)
logPath = os.path.join(args.tensorboard_dir, timeStamp)
# removes everything in that directory
if os.path.isdir(logPath):
logger.error('Tensorboard Log path exists. Overwriting.')
bert_weight_name = json.load(
open("config/" + args.bert_model + "_weight_name.json", "r"))
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")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend="nccl")
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
default_gpu = False
if dist.is_available() and args.local_rank != -1:
rank = dist.get_rank()
if rank == 0:
default_gpu = True
else:
default_gpu = True
if default_gpu:
if not os.path.exists(savePath):
os.makedirs(savePath)
config = BertConfig.from_json_file(args.config_file)
if default_gpu:
# save all the hidden parameters.
with open(os.path.join(savePath, 'command.txt'), 'w') as f:
print(args, file=f) # Python 3.x
print('\n', file=f)
print(config, file=f)
if args.batch_size != -1:
for i, task_id in enumerate(args.tasks.split('-')):
task = 'TASK' + task_id
task_cfg[task]['batch_size'] = args.batch_size
if args.data_root != "":
for i, task_id in enumerate(args.tasks.split('-')):
data_root = args.data_root
task = 'TASK' + task_id
task_cfg[task]['dataroot'] = data_root
task_cfg[task]['features_h5path1'] = os.path.join(
data_root, task_cfg[task]['features_h5path1'].split('/')[-1])
task_cfg[task]['features_h5path2'] = os.path.join(
data_root, task_cfg[task]['features_h5path2'].split('/')[-1])
task_cfg[task]['train_annotations_jsonpath'] = os.path.join(
data_root,
task_cfg[task]['train_annotations_jsonpath'].split('/')[-1])
task_cfg[task]['val_annotations_jsonpath'] = os.path.join(
data_root,
task_cfg[task]['val_annotations_jsonpath'].split('/')[-1])
# Done it for VCR Dataset only, need to put this train_100.jsonl for other datasets
if args.debug:
for i, task_id in enumerate(args.tasks.split('-')):
task = 'TASK' + task_id
task_cfg[task]['train_annotations_jsonpath'] = '/'.join(
task_cfg[task]['train_annotations_jsonpath'].split('/')[:-1] +
['train_100.jsonl'])
task_cfg[task]['val_annotations_jsonpath'] = '/'.join(
task_cfg[task]['val_annotations_jsonpath'].split('/')[:-1] +
['val_100.jsonl'])
task_cfg[task]['batch_size'] = 90
if args.local_rank not in [-1, 0]:
torch.distributed.barrier(
) # Barrier to make sure only the first process in distributed training download model & vocab
gpt2_tokenizer = GPT2Tokenizer.from_pretrained('gpt2', do_lower_case=True)
# Have added args.debug to only VCR Datasets (vcr_dataset.py) will need to add it to other dataset too.
task_batch_size, task_num_iters, task_ids, task_datasets_train, task_datasets_val, \
task_dataloader_train, task_dataloader_val = LoadDatasets(args, task_cfg, gpt2_tokenizer, args.tasks.split('-'), args.debug)
if args.local_rank == 0:
torch.distributed.barrier(
) # End of barrier to make sure only the first process in distributed training download model & vocab
tbLogger = utils.tbLogger(logPath, savePath, task_names, task_ids,
task_num_iters, args.gradient_accumulation_steps)
# if n_gpu > 0:
# torch.cuda.manual_seed_all(args.seed)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
num_train_optimization_steps = max(task_num_iters.values(
)) * args.num_train_epochs // args.gradient_accumulation_steps
num_labels = max(
[dataset.num_labels for dataset in task_datasets_train.values()])
if args.local_rank not in [-1, 0]:
torch.distributed.barrier(
) # Barrier to make sure only the first process in distributed training download model & vocab
if args.baseline:
vil_model = BaseBertForVLTasks.from_pretrained(args.from_pretrained,
config,
num_labels=num_labels,
default_gpu=default_gpu)
else:
vil_model = VILBertForVLTasks.from_pretrained(args.from_pretrained,
config,
num_labels=num_labels,
default_gpu=default_gpu)
model = ViLBertGPT2(vil_model,
gpt2_tokenizer,
gpt2_embed_dim=768,
config=config)
PATH = 'save/trained_models/vilbert_gpt2_loss_3_1.bin'
model.load_state_dict(torch.load(PATH))
model.to(device)
if args.local_rank == 0:
torch.distributed.barrier(
) # End of barrier to make sure only the first process in distributed training download model & vocab
task_losses = LoadLosses(args, task_cfg, args.tasks.split('-'))
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model, delay_allreduce=True)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
model.eval()
# when run evaluate, we run each task sequentially.
epochId = 0
for task_id in task_ids:
num_batch_10 = int(0.1 * len(task_dataloader_val[task_id]))
if args.debug:
num_batch_10 = 1
for i, batch in enumerate(task_dataloader_val[task_id]):
# generate
if i % num_batch_10 == 0:
generate = True
loss_vl, gpt2_loss, score, batch_size, bleu_score = ForwardModelsVal(
args,
task_cfg,
device,
task_id,
batch,
model,
task_losses,
generate=generate)
else:
generate = False
loss_vl, gpt2_loss, score, batch_size = ForwardModelsVal(
args,
task_cfg,
device,
task_id,
batch,
model,
task_losses,
generate=generate)
loss = loss_vl + gpt2_loss
tbLogger.step_val(epochId, float(loss), float(loss_vl),
float(gpt2_loss), float(score), bleu_score,
task_id, batch_size, 'val')
if default_gpu:
sys.stdout.write('%d/%d\r' %
(i, len(task_dataloader_val[task_id])))
sys.stdout.flush()
epochId += 1
ave_score = tbLogger.showLossVal()
tbLogger.txt_close()
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--bert_model",
default="bert-base-uncased",
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.",
)
parser.add_argument(
"--from_pretrained",
default="bert-base-uncased",
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.",
)
parser.add_argument(
"--output_dir",
default="save",
type=str,
help=
"The output directory where the model checkpoints will be written.",
)
parser.add_argument(
"--config_file",
default="config/bert_config.json",
type=str,
help="The config file which specified the model details.",
)
parser.add_argument("--learning_rate",
default=2e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument(
"--num_train_epochs",
default=20,
type=int,
help="Total number of training epochs to perform.",
)
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.",
)
parser.add_argument("--no_cuda",
action="store_true",
help="Whether not to use CUDA when available")
parser.add_argument(
"--do_lower_case",
default=True,
type=bool,
help=
"Whether to lower case the input text. True for uncased models, False for cased models.",
)
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument("--seed",
type=int,
default=0,
help="random seed for initialization")
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help=
"Number of updates steps to accumualte before performing a backward/update pass.",
)
parser.add_argument(
"--fp16",
action="store_true",
help="Whether to use 16-bit float precision instead of 32-bit",
)
parser.add_argument(
"--loss_scale",
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n",
)
parser.add_argument("--num_workers",
type=int,
default=16,
help="Number of workers in the dataloader.")
parser.add_argument(
"--save_name",
default='',
type=str,
help="save name for training.",
)
parser.add_argument("--use_chunk",
default=0,
type=float,
help="whether use chunck for parallel training.")
parser.add_argument("--in_memory",
default=False,
type=bool,
help="whether use chunck for parallel training.")
parser.add_argument("--optimizer",
default='BertAdam',
type=str,
help="whether use chunck for parallel training.")
parser.add_argument("--tasks",
default='',
type=str,
help="1-2-3... training task separate by -")
parser.add_argument(
"--freeze",
default=-1,
type=int,
help="till which layer of textual stream of vilbert need to fixed.")
parser.add_argument("--vision_scratch",
action="store_true",
help="whether pre-trained the image or not.")
parser.add_argument("--evaluation_interval",
default=1,
type=int,
help="evaluate very n epoch.")
parser.add_argument("--lr_scheduler",
default='mannul',
type=str,
help="whether use learning rate scheduler.")
parser.add_argument("--baseline",
action="store_true",
help="whether use single stream baseline.")
parser.add_argument("--compact",
action="store_true",
help="whether use compact vilbert model.")
parser.add_argument("--debug",
action="store_true",
help="whether in debug mode.")
parser.add_argument(
"--tensorboard_dir",
default="tensorboard_log",
type=str,
help="The output directory where tensorboard log will be written.",
)
parser.add_argument(
"--batch_size",
default=-1,
type=int,
help="Custom Batch size for task.",
)
parser.add_argument(
"--data_root",
default="",
type=str,
help="The data root of the task.",
)
args = parser.parse_args()
with open('vlbert_tasks.yml', 'r') as f:
task_cfg = edict(yaml.load(f))
# random.seed(args.seed)
# np.random.seed(args.seed)
# torch.manual_seed(args.seed)
if args.baseline:
from pytorch_pretrained_bert.modeling import BertConfig
from vilbert.basebert import BaseBertForVLTasks
elif args.compact:
from vilbert.vilbert_compact import BertConfig
from vilbert.vilbert_compact import VILBertForVLTasks
else:
from vilbert.vilbert import BertConfig
from vilbert.vilbert import VILBertForVLTasks
task_names = []
task_lr = []
for i, task_id in enumerate(args.tasks.split('-')):
task = 'TASK' + task_id
name = task_cfg[task]['name']
task_names.append(name)
task_lr.append(task_cfg[task]['lr'])
base_lr = min(task_lr)
loss_scale = {}
for i, task_id in enumerate(args.tasks.split('-')):
task = 'TASK' + task_id
loss_scale[task] = task_lr[i] / base_lr
if args.save_name:
prefix = '-' + args.save_name
else:
prefix = ''
timeStamp = '-'.join(task_names) + '_' + args.config_file.split(
'/')[1].split('.')[0] + prefix
savePath = os.path.join(args.output_dir, timeStamp)
logPath = os.path.join(args.tensorboard_dir, timeStamp)
bert_weight_name = json.load(
open("config/" + args.bert_model + "_weight_name.json", "r"))
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")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend="nccl")
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
default_gpu = False
if dist.is_available() and args.local_rank != -1:
rank = dist.get_rank()
if rank == 0:
default_gpu = True
else:
default_gpu = True
if default_gpu:
if not os.path.exists(savePath):
os.makedirs(savePath)
config = BertConfig.from_json_file(args.config_file)
if default_gpu:
# save all the hidden parameters.
with open(os.path.join(savePath, 'command.txt'), 'w') as f:
print(args, file=f) # Python 3.x
print('\n', file=f)
print(config, file=f)
if args.batch_size != -1:
for i, task_id in enumerate(args.tasks.split('-')):
task = 'TASK' + task_id
task_cfg[task]['batch_size'] = args.batch_size
if args.data_root != "":
for i, task_id in enumerate(args.tasks.split('-')):
data_root = args.data_root
task = 'TASK' + task_id
task_cfg[task]['dataroot'] = data_root
task_cfg[task]['features_h5path1'] = os.path.join(
data_root, task_cfg[task]['features_h5path1'].split('/')[-1])
task_cfg[task]['features_h5path2'] = os.path.join(
data_root, task_cfg[task]['features_h5path2'].split('/')[-1])
task_cfg[task]['train_annotations_jsonpath'] = os.path.join(
data_root,
task_cfg[task]['train_annotations_jsonpath'].split('/')[-1])
task_cfg[task]['val_annotations_jsonpath'] = os.path.join(
data_root,
task_cfg[task]['val_annotations_jsonpath'].split('/')[-1])
# Done it for VCR Dataset only, need to put this train_100.jsonl for other datasets
if args.debug:
for i, task_id in enumerate(args.tasks.split('-')):
task = 'TASK' + task_id
task_cfg[task]['train_annotations_jsonpath'] = '/'.join(
task_cfg[task]['train_annotations_jsonpath'].split('/')[:-1] +
['train_100.jsonl'])
task_cfg[task]['val_annotations_jsonpath'] = '/'.join(
task_cfg[task]['val_annotations_jsonpath'].split('/')[:-1] +
['val_100.jsonl'])
task_cfg[task]['batch_size'] = 2
# Have added args.debug to only VCR Datasets (vcr_dataset.py) will need to add it to other dataset too.
task_batch_size, task_num_iters, task_ids, task_datasets_train, task_datasets_val, \
task_dataloader_train, task_dataloader_val = LoadDatasets(args, task_cfg, args.tasks.split('-'), args.debug)
tbLogger = utils.tbLogger(logPath, savePath, task_names, task_ids,
task_num_iters, args.gradient_accumulation_steps)
# if n_gpu > 0:
# torch.cuda.manual_seed_all(args.seed)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
num_train_optimization_steps = max(task_num_iters.values(
)) * args.num_train_epochs // args.gradient_accumulation_steps
num_labels = max(
[dataset.num_labels for dataset in task_datasets_train.values()])
task_start_iter = {}
task_interval = {}
for task_id, num_iter in task_num_iters.items():
task_start_iter[task_id] = num_train_optimization_steps - (
task_cfg[task]['num_epoch'] * num_iter //
args.gradient_accumulation_steps)
task_interval[task_id] = num_train_optimization_steps // (
task_cfg[task]['num_epoch'] * num_iter //
args.gradient_accumulation_steps)
if args.baseline:
model = BaseBertForVLTasks.from_pretrained(args.from_pretrained,
config,
num_labels=num_labels,
default_gpu=default_gpu)
else:
model = VILBertForVLTasks.from_pretrained(args.from_pretrained,
config,
num_labels=num_labels,
default_gpu=default_gpu)
task_losses = LoadLosses(args, task_cfg, args.tasks.split('-'))
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model, delay_allreduce=True)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
no_decay = ["bias", "LayerNorm.bias", "LayerNorm.weight"]
if args.freeze != -1:
bert_weight_name_filtered = []
for name in bert_weight_name:
if 'embeddings' in name:
bert_weight_name_filtered.append(name)
elif 'encoder' in name:
layer_num = name.split('.')[2]
if int(layer_num) <= args.freeze:
bert_weight_name_filtered.append(name)
optimizer_grouped_parameters = []
for key, value in dict(model.named_parameters()).items():
if key[12:] in bert_weight_name_filtered:
value.requires_grad = False
if default_gpu:
print("filtered weight")
print(bert_weight_name_filtered)
optimizer_grouped_parameters = []
lr = args.learning_rate
for key, value in dict(model.named_parameters()).items():
if value.requires_grad:
if 'vil_prediction' in key:
# if args.learning_rate <= 2e-5:
lr = 1e-4
else:
if args.vision_scratch:
if key[12:] in bert_weight_name:
lr = args.learning_rate
else:
lr = 1e-4
else:
lr = args.learning_rate
if any(nd in key for nd in no_decay):
optimizer_grouped_parameters += [{
"params": [value],
"lr": lr,
"weight_decay": 0.01
}]
if not any(nd in key for nd in no_decay):
optimizer_grouped_parameters += [{
"params": [value],
"lr": lr,
"weight_decay": 0.0
}]
if default_gpu:
print(len(list(model.named_parameters())),
len(optimizer_grouped_parameters))
max_num_iter = max(task_num_iters.values())
max_batch_size = max(task_batch_size.values())
if args.optimizer == 'BertAdam':
optimizer = BertAdam(
optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps,
schedule='warmup_constant',
)
elif args.optimizer == 'Adam':
optimizer = Adam(
optimizer_grouped_parameters,
lr=base_lr,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps,
schedule='warmup_constant',
)
elif args.optimizer == 'Adamax':
optimizer = Adamax(
optimizer_grouped_parameters,
lr=base_lr,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps,
schedule='warmup_constant',
)
if args.lr_scheduler == 'automatic':
lr_scheduler = ReduceLROnPlateau(optimizer, \
mode='max',
factor=0.2,
patience=1,
cooldown=1,
threshold=0.001)
elif args.lr_scheduler == 'mannul':
lr_reduce_list = np.array([12, 16])
# lr_reduce_list = np.array([6, 8, 10])
def lr_lambda_fun(epoch):
return pow(0.1, np.sum(lr_reduce_list <= epoch))
lr_scheduler = LambdaLR(optimizer, lr_lambda=lr_lambda_fun)
if default_gpu:
print("***** Running training *****")
print(" Num Iters: ", task_num_iters)
print(" Batch size: ", task_batch_size)
print(" Num steps: %d" % num_train_optimization_steps)
startIterID = 0
# initialize the data iteration.
task_iter_train = {name: None for name in task_ids}
task_count = {name: 0 for name in task_ids}
for epochId in tqdm(range(args.num_train_epochs), desc="Epoch"):
model.train()
for step in range(max_num_iter):
iterId = startIterID + step + (epochId * max_num_iter)
for task_id in task_ids:
if iterId >= task_start_iter[task_id]:
# if iterId % task_interval[task_id] == 0:
loss, score = ForwardModelsTrain(args, task_cfg, device,
task_id, task_count,
task_iter_train,
task_dataloader_train,
model, task_losses,
task_start_iter)
loss = loss * loss_scale[task_id]
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step()
model.zero_grad()
if default_gpu:
tbLogger.step_train(epochId, iterId, float(loss),
float(score),
optimizer.show_lr(), task_id,
'train')
if step % (20 * args.gradient_accumulation_steps
) == 0 and step != 0 and default_gpu:
tbLogger.showLossTrain()
model.eval()
# when run evaluate, we run each task sequentially.
for task_id in task_ids:
for i, batch in enumerate(task_dataloader_val[task_id]):
loss, score, batch_size = ForwardModelsVal(
args, task_cfg, device, task_id, batch, model, task_losses)
tbLogger.step_val(epochId, float(loss), float(score), task_id,
batch_size, 'val')
if default_gpu:
sys.stdout.write('%d/%d\r' %
(i, len(task_dataloader_val[task_id])))
sys.stdout.flush()
ave_score = tbLogger.showLossVal()
if args.lr_scheduler == 'automatic':
lr_scheduler.step(ave_score)
logger.info("best average score is %3f" % lr_scheduler.best)
else:
lr_scheduler.step()
if default_gpu:
# Save a trained model
logger.info("** ** * Saving fine - tuned model on " + logPath +
"** ** * ")
model_to_save = (
model.module if hasattr(model, "module") else model
) # Only save the model it-self
if not os.path.exists(savePath):
os.makedirs(savePath)
output_model_file = os.path.join(
savePath, "pytorch_model_" + str(epochId) + ".bin")
torch.save(model_to_save.state_dict(), output_model_file)
tbLogger.txt_close()
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--bert_model",
default="bert-base-uncased",
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.",
)
parser.add_argument(
"--from_pretrained",
default="bert-base-uncased",
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.",
)
parser.add_argument(
"--output_dir",
default="results",
type=str,
help=
"The output directory where the model checkpoints will be written.",
)
parser.add_argument(
"--config_file",
default="config/bert_config.json",
type=str,
help="The config file which specified the model details.",
)
parser.add_argument("--no_cuda",
action="store_true",
help="Whether not to use CUDA when available")
parser.add_argument(
"--do_lower_case",
default=True,
type=bool,
help=
"Whether to lower case the input text. True for uncased models, False for cased models.",
)
parser.add_argument(
"--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus",
)
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 float precision instead of 32-bit",
)
parser.add_argument(
"--loss_scale",
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n",
)
parser.add_argument(
"--num_workers",
type=int,
default=0,
help="Number of workers in the dataloader.",
)
parser.add_argument("--save_name",
default="",
type=str,
help="save name for training.")
parser.add_argument(
"--use_chunk",
default=0,
type=float,
help="whether use chunck for parallel training.",
)
parser.add_argument("--batch_size",
default=1,
type=int,
help="what is the batch size?")
parser.add_argument("--tasks",
default="",
type=str,
help="1-2-3... training task separate by -")
parser.add_argument(
"--in_memory",
default=False,
type=bool,
help="whether use chunck for parallel training.",
)
parser.add_argument("--baseline",
action="store_true",
help="whether use single stream baseline.")
parser.add_argument("--split",
default="",
type=str,
help="which split to use.")
parser.add_argument(
"--dynamic_attention",
action="store_true",
help="whether use dynamic attention.",
)
parser.add_argument(
"--clean_train_sets",
default=False,
type=bool,
help="whether clean train sets for multitask data.",
)
parser.add_argument(
"--visual_target",
default=0,
type=int,
help="which target to use for visual branch. \
0: soft label, \
1: regress the feature, \
2: NCE loss.",
)
parser.add_argument(
"--task_specific_tokens",
action="store_true",
help="whether to use task specific tokens for the multi-task learning.",
)
args = parser.parse_args()
with open("vilbert_tasks.yml", "r") as f:
task_cfg = edict(yaml.safe_load(f))
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if args.baseline:
from pytorch_transformers.modeling_bert import BertConfig
from vilbert.basebert import BaseBertForVLTasks
else:
from vilbert.vilbert import BertConfig
from vilbert.vilbert import VILBertForVLTasks
task_names = []
for i, task_id in enumerate(args.tasks.split("-")):
task = "TASK" + task_id
name = task_cfg[task]["name"]
task_names.append(name)
# if args.task_specific_tokens:
# config.task_specific_tokens = True
# timeStamp = '-'.join(task_names) + '_' + args.config_file.split('/')[1].split('.')[0]
timeStamp = args.from_pretrained.split("/")[-1] + "-" + args.save_name
savePath = os.path.join(args.output_dir, timeStamp)
config = BertConfig.from_json_file(args.config_file)
if args.task_specific_tokens:
config.task_specific_tokens = True
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")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend="nccl")
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
default_gpu = False
if dist.is_available() and args.local_rank != -1:
rank = dist.get_rank()
if rank == 0:
default_gpu = True
else:
default_gpu = True
if default_gpu and not os.path.exists(savePath):
os.makedirs(savePath)
task_batch_size, task_num_iters, task_ids, task_datasets_val, task_dataloader_val = LoadDatasetEval(
args, task_cfg, args.tasks.split("-"))
tbLogger = utils.tbLogger(
timeStamp,
savePath,
task_names,
task_ids,
task_num_iters,
1,
save_logger=False,
txt_name="eval.txt",
)
num_labels = max(
[dataset.num_labels for dataset in task_datasets_val.values()])
if args.dynamic_attention:
config.dynamic_attention = True
if "roberta" in args.bert_model:
config.model = "roberta"
if args.visual_target == 0:
config.v_target_size = 1601
config.visual_target = args.visual_target
else:
config.v_target_size = 2048
config.visual_target = args.visual_target
if args.task_specific_tokens:
config.task_specific_tokens = True
# set visualization to true
config.visualization = True
# uncomment this to use sum fusion
# config.fusion_method = "sum"
if args.baseline:
model = BaseBertForVLTasks.from_pretrained(
args.from_pretrained,
config=config,
num_labels=num_labels,
default_gpu=default_gpu,
)
else:
model = VILBertForVLTasks.from_pretrained(
args.from_pretrained,
config=config,
num_labels=num_labels,
default_gpu=default_gpu,
)
task_losses = LoadLosses(args, task_cfg, args.tasks.split("-"))
model.to(device)
# if args.local_rank != -1:
# try:
# from apex.parallel import DistributedDataParallel as DDP
# except ImportError:
# raise ImportError(
# "Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
# )
# model = DDP(model, delay_allreduce=True)
#
# elif n_gpu > 1:
# model = nn.DataParallel(model)
print("***** Running evaluation *****")
print(" Num Iters: ", task_num_iters)
print(" Batch size: ", task_batch_size)
print(task_ids)
model.eval()
# when run evaluate, we run each task sequentially.
for task_id in task_ids:
results = []
others = []
for i, batch in enumerate(task_dataloader_val[task_id]):
loss, score, batch_size, results, others = EvaluatingModel(
args, task_cfg, device, task_id, batch, model,
task_dataloader_val, task_losses, results, others, i)
tbLogger.step_val(0, float(loss), float(score), task_id,
batch_size, "val")
sys.stdout.write("%d/%d\r" %
(i, len(task_dataloader_val[task_id])))
sys.stdout.flush()
# save the result or evaluate the result.
ave_score = tbLogger.showLossVal(task_id)
if args.split:
json_path = os.path.join(savePath, args.split)
else:
json_path = os.path.join(savePath, task_cfg[task_id]["val_split"])
print(json_path)
json.dump(results, open(json_path + "_result.json", "w"))
json.dump(others, open(json_path + "_others.json", "w"))
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--file_path",
default="data/conceptual_caption/",
type=str,
help="The input train corpus.",
)
parser.add_argument(
"--from_pretrained",
default="bert-base-uncased",
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-base-uncased, roberta-base, roberta-large, ",
)
parser.add_argument(
"--bert_model",
default="bert-base-uncased",
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, roberta-base",
)
parser.add_argument(
"--output_dir",
default="save",
type=str,
# required=True,
help=
"The output directory where the model checkpoints will be written.",
)
parser.add_argument(
"--config_file",
type=str,
default="config/bert_base_6layer_6conect.json",
help="The config file which specified the model details.",
)
## Other parameters
parser.add_argument(
"--max_seq_length",
default=36,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.",
)
parser.add_argument(
"--train_batch_size",
default=512,
type=int,
help="Total batch size for training.",
)
parser.add_argument(
"--learning_rate",
default=1e-4,
type=float,
help="The initial learning rate for Adam.",
)
parser.add_argument(
"--num_train_epochs",
default=10.0,
type=float,
help="Total number of training epochs to perform.",
)
parser.add_argument(
"--start_epoch",
default=0,
type=float,
help="Total number of training epochs to perform.",
)
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.",
)
parser.add_argument("--img_weight",
default=1,
type=float,
help="weight for image loss")
parser.add_argument("--no_cuda",
action="store_true",
help="Whether not to use CUDA when available")
parser.add_argument(
"--on_memory",
action="store_true",
help="Whether to load train samples into memory or use disk",
)
parser.add_argument(
"--do_lower_case",
type=bool,
default=True,
help=
"Whether to lower case the input text. True for uncased models, False for cased models.",
)
parser.add_argument(
"--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus",
)
parser.add_argument("--seed",
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help=
"Number of updates steps to accumualte before performing a backward/update pass.",
)
parser.add_argument(
"--fp16",
action="store_true",
help="Whether to use 16-bit float precision instead of 32-bit",
)
parser.add_argument(
"--loss_scale",
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n",
)
parser.add_argument(
"--dynamic_attention",
action="store_true",
help="whether use dynamic attention.",
)
parser.add_argument(
"--num_workers",
type=int,
default=25,
help="Number of workers in the dataloader.",
)
parser.add_argument("--save_name",
default="",
type=str,
help="save name for training.")
parser.add_argument(
"--baseline",
action="store_true",
help="Wheter to use the baseline model (single bert).",
)
parser.add_argument(
"--freeze",
default=-1,
type=int,
help="till which layer of textual stream of vilbert need to fixed.",
)
parser.add_argument(
"--distributed",
action="store_true",
help="whether use chunck for parallel training.",
)
parser.add_argument("--without_coattention",
action="store_true",
help="whether pair loss.")
parser.add_argument(
"--visual_target",
default=0,
type=int,
help="which target to use for visual branch. \
0: soft label, \
1: regress the feature, \
2: NCE loss.",
)
parser.add_argument(
"--objective",
default=0,
type=int,
help="which objective to use \
0: with ICA loss, \
1: with ICA loss, for the not aligned pair, no masking objective, \
2: without ICA loss, do not sample negative pair.",
)
parser.add_argument("--num_negative",
default=255,
type=int,
help="num of negative to use")
parser.add_argument("--resume_file",
default="",
type=str,
help="Resume from checkpoint")
parser.add_argument("--adam_epsilon",
default=1e-8,
type=float,
help="Epsilon for Adam optimizer.")
args = parser.parse_args()
if args.baseline:
from pytorch_pretrained_bert.modeling import BertConfig
from vilbert.basebert import BertForMultiModalPreTraining
else:
from vilbert.vilbert import BertForMultiModalPreTraining, BertConfig
if args.save_name:
prefix = "-" + args.save_name
else:
prefix = ""
timeStamp = args.config_file.split("/")[1].split(".")[0] + prefix
savePath = os.path.join(args.output_dir, timeStamp)
bert_weight_name = json.load(
open("config/" + args.from_pretrained + "_weight_name.json", "r"))
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")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend="nccl")
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
default_gpu = False
if dist.is_available() and args.local_rank != -1:
rank = dist.get_rank()
if rank == 0:
default_gpu = True
else:
default_gpu = True
if default_gpu:
if not os.path.exists(savePath):
os.makedirs(savePath)
config = BertConfig.from_json_file(args.config_file)
if default_gpu:
# save all the hidden parameters.
with open(os.path.join(savePath, "command.txt"), "w") as f:
print(args, file=f) # Python 3.x
print("\n", file=f)
print(config, file=f)
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
cache = 5000
if dist.is_available() and args.local_rank != -1:
num_replicas = dist.get_world_size()
args.train_batch_size = args.train_batch_size // num_replicas
args.num_workers = args.num_workers // num_replicas
cache = cache // num_replicas
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
if "roberta" in args.bert_model:
tokenizer = RobertaTokenizer.from_pretrained(
args.bert_model, do_lower_case=args.do_lower_case)
else:
tokenizer = BertTokenizer.from_pretrained(
args.bert_model, do_lower_case=args.do_lower_case)
num_train_optimization_steps = None
train_dataset = ConceptCapLoaderTrain(
args.file_path,
tokenizer,
args.bert_model,
seq_len=args.max_seq_length,
batch_size=args.train_batch_size,
visual_target=args.visual_target,
num_workers=args.num_workers,
local_rank=args.local_rank,
objective=args.objective,
cache=cache,
)
validation_dataset = ConceptCapLoaderVal(
args.file_path,
tokenizer,
args.bert_model,
seq_len=args.max_seq_length,
batch_size=args.train_batch_size,
visual_target=args.visual_target,
num_workers=2,
objective=args.objective,
)
num_train_optimization_steps = int(
train_dataset.num_dataset / args.train_batch_size /
args.gradient_accumulation_steps) * (args.num_train_epochs -
args.start_epoch)
task_names = ["Conceptual_Caption"]
task_ids = ["TASK0"]
task_num_iters = {
"TASK0": train_dataset.num_dataset / args.train_batch_size
}
logdir = os.path.join("logs", timeStamp)
if default_gpu:
tbLogger = utils.tbLogger(
logdir,
savePath,
task_names,
task_ids,
task_num_iters,
args.gradient_accumulation_steps,
)
if args.visual_target == 0:
config.v_target_size = 1601
config.visual_target = args.visual_target
else:
config.v_target_size = 2048
config.visual_target = args.visual_target
if "roberta" in args.bert_model:
config.model = "roberta"
if args.freeze > config.t_biattention_id[0]:
config.fixed_t_layer = config.t_biattention_id[0]
if args.without_coattention:
config.with_coattention = False
if args.dynamic_attention:
config.dynamic_attention = True
if args.from_pretrained:
model = BertForMultiModalPreTraining.from_pretrained(
args.from_pretrained, config=config, default_gpu=default_gpu)
else:
model = BertForMultiModalPreTraining(config)
no_decay = ["bias", "LayerNorm.bias", "LayerNorm.weight"]
if args.freeze != -1:
bert_weight_name_filtered = []
for name in bert_weight_name:
if "embeddings" in name:
bert_weight_name_filtered.append(name)
elif "encoder" in name:
layer_num = name.split(".")[2]
if int(layer_num) <= args.freeze:
bert_weight_name_filtered.append(name)
optimizer_grouped_parameters = []
for key, value in dict(model.named_parameters()).items():
if key[12:] in bert_weight_name_filtered:
value.requires_grad = False
if default_gpu:
print("filtered weight")
print(bert_weight_name_filtered)
if not args.from_pretrained:
param_optimizer = list(model.named_parameters())
optimizer_grouped_parameters = [
{
"params": [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
"weight_decay":
0.01,
},
{
"params": [
p for n, p in param_optimizer
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0,
},
]
else:
optimizer_grouped_parameters = []
for key, value in dict(model.named_parameters()).items():
if value.requires_grad:
if key[12:] in bert_weight_name:
lr = args.learning_rate * 0.1
else:
lr = args.learning_rate
if any(nd in key for nd in no_decay):
optimizer_grouped_parameters += [{
"params": [value],
"lr": lr,
"weight_decay": 0.0
}]
if not any(nd in key for nd in no_decay):
optimizer_grouped_parameters += [{
"params": [value],
"lr": lr,
"weight_decay": 0.01
}]
if default_gpu:
print(len(list(model.named_parameters())),
len(optimizer_grouped_parameters))
# set different parameters for vision branch and lanugage branch.
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(
optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0,
)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = AdamW(
optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon,
betas=(0.9, 0.98),
)
scheduler = WarmupLinearSchedule(
optimizer,
warmup_steps=args.warmup_proportion * num_train_optimization_steps,
t_total=num_train_optimization_steps,
)
startIterID = 0
global_step = 0
if args.resume_file != "" and os.path.exists(args.resume_file):
checkpoint = torch.load(args.resume_file, map_location="cpu")
new_dict = {}
for attr in checkpoint["model_state_dict"]:
if attr.startswith("module."):
new_dict[attr.replace(
"module.", "", 1)] = checkpoint["model_state_dict"][attr]
else:
new_dict[attr] = checkpoint["model_state_dict"][attr]
model.load_state_dict(new_dict)
scheduler.load_state_dict(checkpoint["scheduler_state_dict"])
optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
global_step = checkpoint["global_step"]
del checkpoint
model.cuda()
for state in optimizer.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.cuda()
if args.fp16:
model.half()
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
if default_gpu:
logger.info("***** Running training *****")
logger.info(" Num examples = %d", train_dataset.num_dataset)
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
for epochId in range(int(args.start_epoch), int(args.num_train_epochs)):
model.train()
for step, batch in enumerate(train_dataset):
iterId = startIterID + step + (epochId * len(train_dataset))
image_ids = batch[-1]
batch = tuple(
t.cuda(device=device, non_blocking=True) for t in batch[:-1])
input_ids, input_mask, segment_ids, lm_label_ids, is_next, image_feat, image_loc, image_target, image_label, image_mask = (
batch)
if args.objective == 1:
image_label = image_label * (is_next == 0).long().unsqueeze(1)
image_label[image_label == 0] = -1
lm_label_ids = lm_label_ids * (is_next
== 0).long().unsqueeze(1)
lm_label_ids[lm_label_ids == 0] = -1
masked_loss_t, masked_loss_v, next_sentence_loss = model(
input_ids,
image_feat,
image_loc,
segment_ids,
input_mask,
image_mask,
lm_label_ids,
image_label,
image_target,
is_next,
)
if args.objective == 2:
next_sentence_loss = next_sentence_loss * 0
masked_loss_v = masked_loss_v * args.img_weight
loss = masked_loss_t + masked_loss_v + next_sentence_loss
if n_gpu > 1:
loss = loss.mean()
masked_loss_t = masked_loss_t.mean()
masked_loss_v = masked_loss_v.mean()
next_sentence_loss = next_sentence_loss.mean()
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion,
)
for param_group in optimizer.param_groups:
param_group["lr"] = lr_this_step
scheduler.step()
optimizer.step()
optimizer.zero_grad()
global_step += 1
if default_gpu:
tbLogger.step_train_CC(
epochId,
iterId,
float(masked_loss_t),
float(masked_loss_v),
float(next_sentence_loss),
optimizer.param_groups[0]["lr"],
"TASK0",
"train",
)
if (step % (20 * args.gradient_accumulation_steps) == 0
and step != 0 and default_gpu):
tbLogger.showLossTrainCC()
# Do the evaluation
torch.set_grad_enabled(False)
numBatches = len(validation_dataset)
model.eval()
for step, batch in enumerate(validation_dataset):
image_ids = batch[-1]
batch = tuple(
t.cuda(device=device, non_blocking=True) for t in batch[:-1])
input_ids, input_mask, segment_ids, lm_label_ids, is_next, image_feat, image_loc, image_target, image_label, image_mask = (
batch)
batch_size = input_ids.size(0)
masked_loss_t, masked_loss_v, next_sentence_loss = model(
input_ids,
image_feat,
image_loc,
segment_ids,
input_mask,
image_mask,
lm_label_ids,
image_label,
image_target,
is_next,
)
masked_loss_v = masked_loss_v * args.img_weight
loss = masked_loss_t + masked_loss_v + next_sentence_loss
if n_gpu > 1:
loss = loss.mean()
masked_loss_t = masked_loss_t.mean()
masked_loss_v = masked_loss_v.mean()
next_sentence_loss = next_sentence_loss.mean()
if default_gpu:
tbLogger.step_val_CC(
epochId,
float(masked_loss_t),
float(masked_loss_v),
float(next_sentence_loss),
"TASK0",
batch_size,
"val",
)
sys.stdout.write("%d / %d \r" % (step, numBatches))
sys.stdout.flush()
if default_gpu:
ave_score = tbLogger.showLossValCC()
torch.set_grad_enabled(True)
if default_gpu:
# Save a trained model
logger.info("** ** * Saving fine - tuned model ** ** * ")
model_to_save = (
model.module if hasattr(model, "module") else model
) # Only save the model it-self
output_model_file = os.path.join(
savePath, "pytorch_model_" + str(epochId) + ".bin")
output_checkpoint = os.path.join(
savePath, "pytorch_ckpt_" + str(epochId) + ".tar")
torch.save(model_to_save.state_dict(), output_model_file)
torch.save(
{
"model_state_dict": model_to_save.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
"scheduler_state_dict": scheduler.state_dict(),
"global_step": global_step,
},
output_checkpoint,
)
if default_gpu:
tbLogger.txt_close()
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--bert_model",
default="bert-base-uncased",
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.",
)
parser.add_argument(
"--from_pretrained",
default="bert-base-uncased",
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.",
)
parser.add_argument(
"--output_dir",
default="results",
type=str,
help=
"The output directory where the model checkpoints will be written.",
)
parser.add_argument(
"--config_file",
default="config/bert_config.json",
type=str,
help="The config file which specified the model details.",
)
parser.add_argument("--no_cuda",
action="store_true",
help="Whether not to use CUDA when available")
parser.add_argument(
"--do_lower_case",
default=True,
type=bool,
help=
"Whether to lower case the input text. True for uncased models, False for cased models.",
)
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
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 float precision instead of 32-bit",
)
parser.add_argument(
"--loss_scale",
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n",
)
parser.add_argument("--num_workers",
type=int,
default=10,
help="Number of workers in the dataloader.")
parser.add_argument(
"--save_name",
default='',
type=str,
help="save name for training.",
)
parser.add_argument("--batch_size",
default=1000,
type=int,
help="what is the batch size?")
parser.add_argument("--tasks",
default='',
type=str,
help="1-2-3... training task separate by -")
parser.add_argument("--in_memory",
default=False,
type=bool,
help="whether use chunck for parallel training.")
parser.add_argument("--baseline",
action="store_true",
help="whether use single stream baseline.")
parser.add_argument("--split",
default="",
type=str,
help="which split to use.")
args = parser.parse_args()
with open('vlbert_tasks.yml', 'r') as f:
task_cfg = edict(yaml.safe_load(f))
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if args.baseline:
from pytorch_pretrained_bert.modeling import BertConfig
from vilbert.basebert import BaseBertForVLTasks
else:
from vilbert.vilbert import BertConfig
from vilbert.vilbert import VILBertForVLTasks
task_names = []
for i, task_id in enumerate(args.tasks.split('-')):
task = 'TASK' + task_id
name = task_cfg[task]['name']
task_names.append(name)
# timeStamp = '-'.join(task_names) + '_' + args.config_file.split('/')[1].split('.')[0]
timeStamp = args.from_pretrained.split('/')[1] + '-' + args.save_name
savePath = os.path.join(args.output_dir, timeStamp)
config = BertConfig.from_json_file(args.config_file)
bert_weight_name = json.load(
open("config/" + args.bert_model + "_weight_name.json", "r"))
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")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
torch.distributed.init_process_group(backend="nccl")
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
default_gpu = False
if dist.is_available() and args.local_rank != -1:
rank = dist.get_rank()
if rank == 0:
default_gpu = True
else:
default_gpu = True
if default_gpu and not os.path.exists(savePath):
os.makedirs(savePath)
task_batch_size, task_num_iters, task_ids, task_datasets_val, task_dataloader_val \
= LoadDatasetEval(args, task_cfg, args.tasks.split('-'))
tbLogger = utils.tbLogger(timeStamp,
savePath,
task_names,
task_ids,
task_num_iters,
1,
save_logger=False,
txt_name='eval.txt')
num_labels = max(
[dataset.num_labels for dataset in task_datasets_val.values()])
if args.baseline:
model = BaseBertForVLTasks.from_pretrained(args.from_pretrained,
config,
num_labels=num_labels,
default_gpu=default_gpu)
else:
model = VILBertForVLTasks.from_pretrained(args.from_pretrained,
config,
num_labels=num_labels,
default_gpu=default_gpu)
task_losses = LoadLosses(args, task_cfg, args.tasks.split('-'))
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model, delay_allreduce=True)
elif n_gpu > 1:
model = nn.DataParallel(model)
no_decay = ["bias", "LayerNorm.bias", "LayerNorm.weight"]
print(" Num Iters: ", task_num_iters)
print(" Batch size: ", task_batch_size)
model.eval()
for task_id in task_ids:
results = []
others = []
for i, batch in enumerate(
tqdm(task_dataloader_val[task_id])
): #, total=len(task_dataloader_val[task_id]), position=0, leave=True):
loss, score, batch_size, results, others = EvaluatingModel(args, task_cfg, device, \
task_id, batch, model, task_dataloader_val, task_losses, results, others)
tbLogger.step_val(0, float(loss), float(score), task_id,
batch_size, 'val')
# sys.stdout.write('%d/%d\r' % (i, len(task_dataloader_val[task_id])))
# sys.stdout.flush()
# save the result or evaluate the result.
ave_score = tbLogger.showLossVal()
if args.split:
json_path = os.path.join(savePath, args.split)
else:
json_path = os.path.join(savePath, task_cfg[task_id]['val_split'])
json.dump(results, open(json_path + '_result.json', 'w'))
json.dump(others, open(json_path + '_others.json', 'w'))
if rank == 0:
default_gpu = True
else:
default_gpu = True
if default_gpu and not os.path.exists(savePath):
os.makedirs(savePath)
task_batch_size, task_num_iters, task_ids, task_datasets_val, task_dataloader_val = LoadDatasetEval(
args, task_cfg, args.tasks.split("-"))
tbLogger = utils.tbLogger(
timeStamp,
savePath,
task_names,
task_ids,
task_num_iters,
1,
save_logger=False,
txt_name="eval.txt",
)
# num_labels = max([dataset.num_labels for dataset in task_datasets_val.values()])
if args.dynamic_attention:
config.dynamic_attention = True
if "roberta" in args.bert_model:
config.model = "roberta"
if args.visual_target == 0:
config.v_target_size = 1601
config.visual_target = args.visual_target
else:
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--bert_model",
default="bert-base-uncased",
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.",
)
parser.add_argument(
"--from_pretrained",
default="bert-base-uncased",
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.",
)
parser.add_argument(
"--output_dir",
default="save",
type=str,
help="The output directory where the model checkpoints will be written.",
)
parser.add_argument(
"--config_file",
default="config/bert_base_6layer_6conect.json",
type=str,
help="The config file which specified the model details.",
)
parser.add_argument(
"--num_train_epochs",
default=20,
type=int,
help="Total number of training epochs to perform.",
)
parser.add_argument(
"--train_iter_multiplier",
default=1.0,
type=float,
help="multiplier for the multi-task training.",
)
parser.add_argument(
"--train_iter_gap",
default=4,
type=int,
help="forward every n iteration is the validation score is not improving over the last 3 epoch, -1 means will stop",
)
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help="Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.",
)
parser.add_argument(
"--no_cuda", action="store_true", help="Whether not to use CUDA when available"
)
parser.add_argument(
"--do_lower_case",
default=True,
type=bool,
help="Whether to lower case the input text. True for uncased models, False for cased models.",
)
parser.add_argument(
"--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus",
)
parser.add_argument(
"--seed", type=int, default=0, help="random seed for initialization"
)
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumualte before performing a backward/update pass.",
)
parser.add_argument(
"--fp16",
action="store_true",
help="Whether to use 16-bit float precision instead of 32-bit",
)
parser.add_argument(
"--loss_scale",
type=float,
default=0,
help="Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n",
)
parser.add_argument(
"--num_workers",
type=int,
default=16,
help="Number of workers in the dataloader.",
)
parser.add_argument(
"--save_name", default="", type=str, help="save name for training."
)
parser.add_argument(
"--in_memory",
default=False,
type=bool,
help="whether use chunck for parallel training.",
)
parser.add_argument(
"--optim", default="AdamW", type=str, help="what to use for the optimization."
)
parser.add_argument(
"--tasks", default="", type=str, help="1-2-3... training task separate by -"
)
parser.add_argument(
"--freeze",
default=-1,
type=int,
help="till which layer of textual stream of vilbert need to fixed.",
)
parser.add_argument(
"--vision_scratch",
action="store_true",
help="whether pre-trained the image or not.",
)
parser.add_argument(
"--evaluation_interval", default=1, type=int, help="evaluate very n epoch."
)
parser.add_argument(
"--lr_scheduler",
default="mannul",
type=str,
help="whether use learning rate scheduler.",
)
parser.add_argument(
"--baseline", action="store_true", help="whether use single stream baseline."
)
parser.add_argument(
"--resume_file", default="", type=str, help="Resume from checkpoint"
)
parser.add_argument(
"--dynamic_attention",
action="store_true",
help="whether use dynamic attention.",
)
parser.add_argument(
"--clean_train_sets",
default=True,
type=bool,
help="whether clean train sets for multitask data.",
)
parser.add_argument(
"--visual_target",
default=0,
type=int,
help="which target to use for visual branch. \
0: soft label, \
1: regress the feature, \
2: NCE loss.",
)
parser.add_argument(
"--task_specific_tokens",
action="store_true",
help="whether to use task specific tokens for the multi-task learning.",
)
args = parser.parse_args()
with open("vilbert_tasks.yml", "r") as f:
task_cfg = edict(yaml.safe_load(f))
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if args.baseline:
from pytorch_transformers.modeling_bert import BertConfig
from vilbert.basebert import BaseBertForVLTasks
else:
from vilbert.vilbert import BertConfig
from vilbert.vilbert import VILBertForVLTasks
task_names = []
task_lr = []
for i, task_id in enumerate(args.tasks.split("-")):
task = "TASK" + task_id
name = task_cfg[task]["name"]
task_names.append(name)
task_lr.append(task_cfg[task]["lr"])
base_lr = min(task_lr)
loss_scale = {}
for i, task_id in enumerate(args.tasks.split("-")):
task = "TASK" + task_id
loss_scale[task] = task_lr[i] / base_lr
if args.save_name:
prefix = "-" + args.save_name
else:
prefix = ""
timeStamp = (
"-".join(task_names)
+ "_"
+ args.config_file.split("/")[1].split(".")[0]
+ prefix
)
savePath = os.path.join(args.output_dir, timeStamp)
bert_weight_name = json.load(
open("config/" + args.bert_model + "_weight_name.json", "r")
)
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"
)
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
torch.distributed.init_process_group(backend="nccl")
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
device, n_gpu, bool(args.local_rank != -1), args.fp16
)
)
default_gpu = False
if dist.is_available() and args.local_rank != -1:
rank = dist.get_rank()
if rank == 0:
default_gpu = True
else:
default_gpu = True
if default_gpu:
if not os.path.exists(savePath):
os.makedirs(savePath)
config = BertConfig.from_json_file(args.config_file)
if default_gpu:
# save all the hidden parameters.
with open(os.path.join(savePath, "command.txt"), "w") as f:
print(args, file=f) # Python 3.x
print("\n", file=f)
print(config, file=f)
task_batch_size, task_num_iters, task_ids, task_datasets_train, task_datasets_val, task_dataloader_train, task_dataloader_val = LoadDatasets(
args, task_cfg, args.tasks.split("-")
)
logdir = os.path.join(savePath, "logs")
tbLogger = utils.tbLogger(
logdir,
savePath,
task_names,
task_ids,
task_num_iters,
args.gradient_accumulation_steps,
)
if args.visual_target == 0:
config.v_target_size = 1601
config.visual_target = args.visual_target
else:
config.v_target_size = 2048
config.visual_target = args.visual_target
if args.task_specific_tokens:
config.task_specific_tokens = True
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
task_ave_iter = {}
task_stop_controller = {}
for task_id, num_iter in task_num_iters.items():
task_ave_iter[task_id] = int(
task_cfg[task]["num_epoch"]
* num_iter
* args.train_iter_multiplier
/ args.num_train_epochs
)
task_stop_controller[task_id] = utils.MultiTaskStopOnPlateau(
mode="max",
patience=1,
continue_threshold=0.005,
cooldown=1,
threshold=0.001,
)
task_ave_iter_list = sorted(task_ave_iter.values())
median_num_iter = task_ave_iter_list[-1]
num_train_optimization_steps = (
median_num_iter * args.num_train_epochs // args.gradient_accumulation_steps
)
num_labels = max([dataset.num_labels for dataset in task_datasets_train.values()])
if args.dynamic_attention:
config.dynamic_attention = True
if "roberta" in args.bert_model:
config.model = "roberta"
if args.baseline:
model = BaseBertForVLTasks.from_pretrained(
args.from_pretrained,
config=config,
num_labels=num_labels,
default_gpu=default_gpu,
)
else:
model = VILBertForVLTasks.from_pretrained(
args.from_pretrained,
config=config,
num_labels=num_labels,
default_gpu=default_gpu,
)
task_losses = LoadLosses(args, task_cfg, args.tasks.split("-"))
no_decay = ["bias", "LayerNorm.bias", "LayerNorm.weight"]
if args.freeze != -1:
bert_weight_name_filtered = []
for name in bert_weight_name:
if "embeddings" in name:
bert_weight_name_filtered.append(name)
elif "encoder" in name:
layer_num = name.split(".")[2]
if int(layer_num) <= args.freeze:
bert_weight_name_filtered.append(name)
optimizer_grouped_parameters = []
for key, value in dict(model.named_parameters()).items():
if key[12:] in bert_weight_name_filtered:
value.requires_grad = False
if default_gpu:
print("filtered weight")
print(bert_weight_name_filtered)
optimizer_grouped_parameters = []
for key, value in dict(model.named_parameters()).items():
if value.requires_grad:
if "vil_" in key:
lr = 1e-4
else:
if args.vision_scratch:
if key[12:] in bert_weight_name:
lr = base_lr
else:
lr = 1e-4
else:
lr = base_lr
if any(nd in key for nd in no_decay):
optimizer_grouped_parameters += [
{"params": [value], "lr": lr, "weight_decay": 0.0}
]
if not any(nd in key for nd in no_decay):
optimizer_grouped_parameters += [
{"params": [value], "lr": lr, "weight_decay": 0.01}
]
if default_gpu:
print(len(list(model.named_parameters())), len(optimizer_grouped_parameters))
if args.optim == "AdamW":
optimizer = AdamW(optimizer_grouped_parameters, lr=base_lr, correct_bias=False)
elif args.optim == "RAdam":
optimizer = RAdam(optimizer_grouped_parameters, lr=base_lr)
warmpu_steps = args.warmup_proportion * num_train_optimization_steps
if args.lr_scheduler == "warmup_linear":
warmup_scheduler = WarmupLinearSchedule(
optimizer, warmup_steps=warmpu_steps, t_total=num_train_optimization_steps
)
else:
warmup_scheduler = WarmupConstantSchedule(optimizer, warmup_steps=warmpu_steps)
lr_reduce_list = np.array([5, 7])
if args.lr_scheduler == "automatic":
lr_scheduler = ReduceLROnPlateau(
optimizer, mode="max", factor=0.2, patience=1, cooldown=1, threshold=0.001
)
elif args.lr_scheduler == "cosine":
lr_scheduler = CosineAnnealingLR(
optimizer, T_max=median_num_iter * args.num_train_epochs
)
elif args.lr_scheduler == "cosine_warm":
lr_scheduler = CosineAnnealingWarmRestarts(
optimizer, T_0=median_num_iter * args.num_train_epochs
)
elif args.lr_scheduler == "mannul":
def lr_lambda_fun(epoch):
return pow(0.2, np.sum(lr_reduce_list <= epoch))
lr_scheduler = LambdaLR(optimizer, lr_lambda=lr_lambda_fun)
startIterID = 0
global_step = 0
start_epoch = 0
if args.resume_file != "" and os.path.exists(args.resume_file):
checkpoint = torch.load(args.resume_file, map_location="cpu")
new_dict = {}
for attr in checkpoint["model_state_dict"]:
if attr.startswith("module."):
new_dict[attr.replace("module.", "", 1)] = checkpoint[
"model_state_dict"
][attr]
else:
new_dict[attr] = checkpoint["model_state_dict"][attr]
model.load_state_dict(new_dict)
warmup_scheduler.load_state_dict(checkpoint["warmup_scheduler_state_dict"])
# lr_scheduler.load_state_dict(checkpoint['lr_scheduler_state_dict'])
optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
global_step = checkpoint["global_step"]
start_epoch = int(checkpoint["epoch_id"]) + 1
task_stop_controller = checkpoint["task_stop_controller"]
tbLogger = checkpoint["tb_logger"]
del checkpoint
model.to(device)
for state in optimizer.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.cuda()
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model, delay_allreduce=True)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
if default_gpu:
print("***** Running training *****")
print(" Num Iters: ", task_num_iters)
print(" Batch size: ", task_batch_size)
print(" Num steps: %d" % num_train_optimization_steps)
task_iter_train = {name: None for name in task_ids}
task_count = {name: 0 for name in task_ids}
for epochId in tqdm(range(start_epoch, args.num_train_epochs), desc="Epoch"):
model.train()
for step in range(median_num_iter):
iterId = startIterID + step + (epochId * median_num_iter)
first_task = True
for task_id in task_ids:
is_forward = False
if (not task_stop_controller[task_id].in_stop) or (
iterId % args.train_iter_gap == 0
):
is_forward = True
if is_forward:
loss, score = ForwardModelsTrain(
args,
task_cfg,
device,
task_id,
task_count,
task_iter_train,
task_dataloader_train,
model,
task_losses,
)
loss = loss * loss_scale[task_id]
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion,
)
for param_group in optimizer.param_groups:
param_group["lr"] = lr_this_step
if first_task and (
global_step < warmpu_steps
or args.lr_scheduler == "warmup_linear"
):
warmup_scheduler.step()
optimizer.step()
model.zero_grad()
if first_task:
global_step += 1
first_task = False
if default_gpu:
tbLogger.step_train(
epochId,
iterId,
float(loss),
float(score),
optimizer.param_groups[0]["lr"],
task_id,
"train",
)
if "cosine" in args.lr_scheduler and global_step > warmpu_steps:
lr_scheduler.step()
if (
step % (20 * args.gradient_accumulation_steps) == 0
and step != 0
and default_gpu
):
tbLogger.showLossTrain()
# decided whether to evaluate on each tasks.
for task_id in task_ids:
if (iterId != 0 and iterId % task_num_iters[task_id] == 0) or (
epochId == args.num_train_epochs - 1 and step == median_num_iter - 1
):
evaluate(
args,
task_dataloader_val,
task_stop_controller,
task_cfg,
device,
task_id,
model,
task_losses,
epochId,
default_gpu,
tbLogger,
)
if args.lr_scheduler == "automatic":
lr_scheduler.step(sum(val_scores.values()))
logger.info("best average score is %3f" % lr_scheduler.best)
elif args.lr_scheduler == "mannul":
lr_scheduler.step()
if epochId in lr_reduce_list:
for task_id in task_ids:
# reset the task_stop_controller once the lr drop
task_stop_controller[task_id]._reset()
if default_gpu:
# Save a trained model
logger.info("** ** * Saving fine - tuned model ** ** * ")
model_to_save = (
model.module if hasattr(model, "module") else model
) # Only save the model it-self
output_model_file = os.path.join(
savePath, "pytorch_model_" + str(epochId) + ".bin"
)
output_checkpoint = os.path.join(savePath, "pytorch_ckpt_latest.tar")
torch.save(model_to_save.state_dict(), output_model_file)
torch.save(
{
"model_state_dict": model_to_save.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
"warmup_scheduler_state_dict": warmup_scheduler.state_dict(),
# 'lr_scheduler_state_dict': lr_scheduler.state_dict(),
"global_step": global_step,
"epoch_id": epochId,
"task_stop_controller": task_stop_controller,
"tb_logger": tbLogger,
},
output_checkpoint,
)
tbLogger.txt_close()
def main():
# os.environ['C UDA_VISIBLE_DEVICES'] = "0,1"
batch_size = 64
parser = argparse.ArgumentParser()
parser.add_argument(
"--bert_model",
default="bert-base-uncased",
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.",
)
parser.add_argument(
"--from_pretrained",
default="bert-base-uncased",
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.",
)
parser.add_argument(
"--output_dir",
default="save",
type=str,
help=
"The output directory where the model checkpoints will be written.",
)
parser.add_argument(
"--config_file",
default="config/bert_base_6layer_6conect.json",
type=str,
help="The config file which specified the model details.",
)
parser.add_argument(
"--num_train_epochs",
default=20,
type=int,
help="Total number of training epochs to perform.",
)
parser.add_argument(
"--train_iter_multiplier",
default=1.0,
type=float,
help="multiplier for the multi-task training.",
)
parser.add_argument(
"--train_iter_gap",
default=4,
type=int,
help=
"forward every n iteration is the validation score is not improving over the last 3 epoch, -1 means will stop",
)
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.",
)
parser.add_argument("--no_cuda",
action="store_true",
help="Whether not to use CUDA when available")
parser.add_argument(
"--do_lower_case",
default=True,
type=bool,
help=
"Whether to lower case the input text. True for uncased models, False for cased models.",
)
parser.add_argument(
"--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus",
)
parser.add_argument("--seed",
type=int,
default=0,
help="random seed for initialization")
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help=
"Number of updates steps to accumualte before performing a backward/update pass.",
)
parser.add_argument(
"--fp16",
action="store_true",
help="Whether to use 16-bit float precision instead of 32-bit",
)
parser.add_argument(
"--loss_scale",
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n",
)
parser.add_argument(
"--num_workers",
type=int,
default=16,
help="Number of workers in the dataloader.",
)
parser.add_argument("--save_name",
default="",
type=str,
help="save name for training.")
parser.add_argument(
"--in_memory",
default=False,
type=bool,
help="whether use chunck for parallel training.",
)
parser.add_argument("--optim",
default="AdamW",
type=str,
help="what to use for the optimization.")
parser.add_argument("--tasks",
default="0",
type=str,
help="discourse : TASK0")
parser.add_argument(
"--freeze",
default=-1,
type=int,
help="till which layer of textual stream of vilbert need to fixed.",
)
parser.add_argument(
"--vision_scratch",
action="store_true",
help="whether pre-trained the image or not.",
)
parser.add_argument("--evaluation_interval",
default=1,
type=int,
help="evaluate very n epoch.")
parser.add_argument(
"--lr_scheduler",
default="mannul",
type=str,
help="whether use learning rate scheduler.",
)
parser.add_argument("--baseline",
action="store_true",
help="whether use single stream baseline.")
parser.add_argument("--resume_file",
default="",
type=str,
help="Resume from checkpoint")
parser.add_argument(
"--dynamic_attention",
action="store_true",
help="whether use dynamic attention.",
)
parser.add_argument(
"--clean_train_sets",
default=True,
type=bool,
help="whether clean train sets for multitask data.",
)
parser.add_argument(
"--visual_target",
default=0,
type=int,
help="which target to use for visual branch. \
0: soft label, \
1: regress the feature, \
2: NCE loss.",
)
parser.add_argument(
"--task_specific_tokens",
action="store_true",
default=False,
help="whether to use task specific tokens for the multi-task learning.",
)
# todo
args = parser.parse_args()
with open("vilbert_tasks.yml", "r") as f:
task_cfg = edict(yaml.safe_load(f))
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
if args.baseline:
from pytorch_transformers.modeling_bert import BertConfig
from vilbert.basebert import BaseBertForVLTasks
else:
from vilbert.vilbert import BertConfig
from vilbert.vilbert import VILBertForVLTasks
task_names = []
task_lr = []
task_id = 1
for i, task_id in enumerate(args.tasks.split("-")):
task_id = str(1)
task = "TASK" + task_id
name = task_cfg[task]["name"]
task_names.append(name)
task_lr.append(task_cfg[task]["lr"])
base_lr = min(task_lr)
loss_scale = {}
for i, task_id in enumerate(args.tasks.split("-")):
task = "TASK" + task_id
loss_scale[task] = task_lr[i] / base_lr
if args.save_name:
prefix = "-" + args.save_name
else:
prefix = ""
timeStamp = ("-".join("discourse") + "_" +
args.config_file.split("/")[1].split(".")[0] + prefix)
savePath = os.path.join(args.output_dir, timeStamp)
bert_weight_name = json.load(
open("config/" + args.bert_model + "_weight_name.json", "r"))
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")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
torch.distributed.init_process_group(backend="nccl")
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
default_gpu = False
if dist.is_available() and args.local_rank != -1:
rank = dist.get_rank()
if rank == 0:
default_gpu = True
else:
default_gpu = True
if default_gpu:
if not os.path.exists(savePath):
os.makedirs(savePath)
config = BertConfig.from_json_file(args.config_file)
if default_gpu:
# save all the hidden parameters.
with open(os.path.join(savePath, "command.txt"), "w") as f:
print(args, file=f) # Python 3.x
print("\n", file=f)
print(config, file=f)
# task_batch_size, task_num_iters, task_ids, task_datasets_train, task_datasets_val, task_dataloader_train, task_dataloader_val = LoadDatasets(
# args, task_cfg, args.tasks.split("-"),'train'
# )
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
labels = [
"Visible", 'Subjective', 'Action', 'Story', 'Meta', 'Irrelevant',
'Other'
]
train_dataset = DiscourseRelationDataset(
labels,
task_cfg[task]["dataroot"],
tokenizer,
args.bert_model,
task_cfg[task]["max_seq_length"],
encoding="utf-8",
visual_target=0,
batch_size=batch_size,
shuffle=False,
num_workers=4,
cache=5000,
drop_last=False,
cuda=False,
objective=0,
visualization=False,
)
train_sampler = RandomSampler(train_dataset)
train_loader = DataLoader(
train_dataset,
sampler=train_sampler,
batch_size=batch_size,
num_workers=0,
pin_memory=True,
)
# for i in train_loader:
# print("hello")
# todo task_ids , task_num_tiers
task_ids = ['TASK0']
task_num_iters = [100]
task_batch_size = task_cfg['TASK0']["batch_size"]
print("task_batch_size")
print(task_batch_size)
logdir = os.path.join(savePath, "logs")
tbLogger = utils.tbLogger(
logdir,
savePath,
task_names,
task_ids,
task_num_iters,
args.gradient_accumulation_steps,
)
if args.visual_target == 0:
config.v_target_size = 1601
config.visual_target = args.visual_target
else:
config.v_target_size = 2048
config.visual_target = args.visual_target
if args.task_specific_tokens:
print("*********** config.task_specific_tokens = True ************")
config.task_specific_tokens = True
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
task_ave_iter = {}
task_stop_controller = {}
# for task_id, num_iter in task_num_iters.items():
# task_ave_iter[task_id] = int(
# task_cfg[task]["num_epoch"]
# * num_iter
# * args.train_iter_multiplier
# / args.num_train_epochs
# )
# task_stop_controller[task_id] = utils.MultiTaskStopOnPlateau(
# mode="max",
# patience=1,
# continue_threshold=0.005,
# cooldown=1,
# threshold=0.001,
# )
# task_ave_iter_list = sorted(task_ave_iter.values())
# median_num_iter = task_ave_iter_list[-1]
# num_train_optimization_steps = (
# median_num_iter * args.num_train_epochs // args.gradient_accumulation_steps
# )
# num_labels = max([dataset.num_labels for dataset in task_datasets_train.values()])
# num_train_optimization_steps = int(
# train_dataset.num_dataset
# / task_batch_size
# / args.gradient_accumulation_steps
# ) * (args.num_train_epochs - args.start_epoch)
# num_train_optimization_steps = int(
# train_dataset.num_dataset
# / task_batch_size
# / args.gradient_accumulation_steps
# ) * (args.num_train_epochs - args.start_epoch)
num_train_optimization_steps = 10
num_labels = len(labels)
if args.dynamic_attention:
config.dynamic_attention = True
if "roberta" in args.bert_model:
config.model = "roberta"
if args.baseline:
model = BaseBertForVLTasks.from_pretrained(
args.from_pretrained,
config=config,
num_labels=num_labels,
default_gpu=default_gpu,
)
else:
model = VILBertForVLTasks.from_pretrained(
args.from_pretrained,
config=config,
num_labels=num_labels,
default_gpu=default_gpu,
)
model.double()
model = model.to(device)
task_losses = LoadLosses(args, task_cfg, args.tasks.split("-"))
no_decay = ["bias", "LayerNorm.bias", "LayerNorm.weight"]
if args.freeze != -1:
bert_weight_name_filtered = []
for name in bert_weight_name:
if "embeddings" in name:
bert_weight_name_filtered.append(name)
elif "encoder" in name:
layer_num = name.split(".")[2]
if int(layer_num) <= args.freeze:
bert_weight_name_filtered.append(name)
optimizer_grouped_parameters = []
for key, value in dict(model.named_parameters()).items():
if key[12:] in bert_weight_name_filtered:
value.requires_grad = False
if default_gpu:
print("filtered weight")
print(bert_weight_name_filtered)
optimizer_grouped_parameters = []
for key, value in dict(model.named_parameters()).items():
if value.requires_grad:
if "vil_" in key:
lr = 1e-4
else:
if args.vision_scratch:
if key[12:] in bert_weight_name:
lr = base_lr
else:
lr = 1e-4
else:
lr = base_lr
if any(nd in key for nd in no_decay):
optimizer_grouped_parameters += [{
"params": [value],
"lr": lr,
"weight_decay": 0.0
}]
if not any(nd in key for nd in no_decay):
optimizer_grouped_parameters += [{
"params": [value],
"lr": lr,
"weight_decay": 0.01
}]
if default_gpu:
print(len(list(model.named_parameters())),
len(optimizer_grouped_parameters))
if args.optim == "AdamW":
optimizer = AdamW(optimizer_grouped_parameters,
lr=base_lr,
correct_bias=False,
weight_decay=1e-4)
elif args.optim == "RAdam":
optimizer = RAdam(optimizer_grouped_parameters,
lr=base_lr,
weight_decay=1e-4)
# warmpu_steps = args.warmup_proportion * num_train_optimization_steps
# if args.lr_scheduler == "warmup_linear":
# warmup_scheduler = WarmupLinearSchedule(
# optimizer, warmup_steps=warmpu_steps, t_total=num_train_optimization_steps
# )
# else:
# warmup_scheduler = WarmupConstantSchedule(optimizer, warmup_steps=warmpu_steps)
#
# lr_reduce_list = np.array([5, 7])
# if args.lr_scheduler == "automatic":
# lr_scheduler = ReduceLROnPlateau(
# optimizer, mode="max", factor=0.2, patience=1, cooldown=1, threshold=0.001
# )
# elif args.lr_scheduler == "cosine":
# lr_scheduler = CosineAnnealingLR(
# optimizer, T_max=median_num_iter * args.num_train_epochs
# )
# elif args.lr_scheduler == "cosine_warm":
# lr_scheduler = CosineAnnealingWarmRestarts(
# # optimizer, T_0=median_num_iter * args.num_train_epochs
# )
# elif args.lr_scheduler == "mannul":
#
# def lr_lambda_fun(epoch):
# return pow(0.2, np.sum(lr_reduce_list <= epoch))
#
# lr_scheduler = LambdaLR(optimizer, lr_lambda=lr_lambda_fun)
startIterID = 0
global_step = 0
start_epoch = 0
if args.resume_file != "" and os.path.exists(args.resume_file):
checkpoint = torch.load(args.resume_file, map_location="cpu")
new_dict = {}
for attr in checkpoint["model_state_dict"]:
if attr.startswith("module."):
new_dict[attr.replace(
"module.", "", 1)] = checkpoint["model_state_dict"][attr]
else:
new_dict[attr] = checkpoint["model_state_dict"][attr]
model.load_state_dict(new_dict)
# warmup_scheduler.load_state_dict(checkpoint["warmup_scheduler_state_dict"])
# lr_scheduler.load_state_dict(checkpoint['lr_scheduler_state_dict'])
optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
global_step = checkpoint["global_step"]
start_epoch = int(checkpoint["epoch_id"]) + 1
task_stop_controller = checkpoint["task_stop_controller"]
tbLogger = checkpoint["tb_logger"]
del checkpoint
model.to(device)
for state in optimizer.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.cuda()
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model, delay_allreduce=True)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
if default_gpu:
print("***** Running training *****")
print(" Num Iters: ", task_num_iters)
print(" Batch size: ", batch_size)
print(" Num steps: %d" % num_train_optimization_steps)
task_iter_train = {name: None for name in task_ids}
task_count = {name: 0 for name in task_ids}
# for epochId in tqdm(range(start_epoch, args.num_train_epochs), desc="Epoch"):
# model.train()
# torch.autograd.set_detect_anomaly(True)
# # for step in range(median_num_iter):
# for step in range(1)
# # iterId = startIterID + step + (epochId * median_num_iter)
# first_task = True
# for task_id in task_ids:
# is_forward = False
# # if (not task_stop_controller[task_id].in_stop) or (
# # iterId % args.train_iter_gap == 0
# # ):
# args['start_epoch'] = 0
# args.num_train_epochs
criterion = nn.BCEWithLogitsLoss()
target_path = os.path.join(task_cfg[task]["dataroot"],
"all_targets_json.json")
all_targets = json.load(open(target_path, "r"))
model = model.to(device)
print(next(model.parameters()).is_cuda)
for epochId in range(int(start_epoch), int(args.num_train_epochs)):
model.train()
is_forward = True
if is_forward:
# print("beforeLoop")
# loss, score = ForwardModelsTrain(
# args,
# task_cfg,
# device,
# task_id,
# task_count,
# task_iter_train,
# train_dataset,
# model,
# task_losses,
# )
for step, batch in enumerate(train_loader):
batch = tuple(
t.to(device=device, non_blocking=True) if type(t) ==
torch.Tensor else t for t in batch)
input_ids, input_mask, segment_ids, image_feat, image_loc, image_mask, image_id = (
batch)
true_targets = []
for id in image_id:
true_targets.append(
np.fromiter(all_targets[id].values(), dtype=np.double))
true_targets = torch.from_numpy(np.array(true_targets))
true_targets = true_targets.to(device)
model.double()
model = model.to(device)
discourse_prediction, vil_prediction, vil_prediction_gqa, vil_logit, vil_binary_prediction, vil_tri_prediction, vision_prediction, vision_logit, linguisic_prediction, linguisic_logit, _ \
= model(
True,
input_ids,
image_feat,
image_loc,
segment_ids,
input_mask,
image_mask
)
loss = criterion(discourse_prediction,
true_targets.type(torch.double))
loss.backward()
optimizer.step()
model.zero_grad()
print("train train train done")
#
print("*********** ITERATION {} ***********".format(epochId))
print("*********** TRAIN PERFORMANCE ***********")
print(loss)
print(
compute_score(discourse_prediction.to('cpu'),
true_targets.type(torch.float).to('cpu'), 0.5))
print("*********** TEST PERFORMANCE ***********")
evaluate(model, device, task_cfg, tokenizer, args, labels)
