def lsuv_init(model,
xb,
layers=None,
types=(nn.Linear, nn.Conv2d),
unfreeze=True):
"""LSUV init of the model.
"""
output = []
# If no layers are specified we will grab the CNN and Linear layers
if not layers:
layers = flatten_model(model)
layers = [layer for layer in layers if isinstance(layer, types)]
requires_grad(model, False)
print("Freezing all layers")
with Hooks(layers, register_stats) as hooks:
for i, hook in enumerate(hooks):
# We first get the module
model(xb)
m = hook.stored["m"]
while model(xb) is not None and abs(hook.stored["mean"]) > 1e-3:
m.bias.data -= hook.stored["mean"]
while model(xb) is not None and abs(hook.stored["std"] - 1) > 1e-3:
m.weight.data /= hook.stored["std"]
output.append(
f"Layer {i} named {str(m)} with m={hook.stored['mean']},std={hook.stored['std']}"
)
if unfreeze:
print("Unfreezing all layers")
requires_grad(model, True)
return output
def freeze_to(self, n: int) -> None:
"Freeze layers up to layer group `n`."
for g in self.layer_groups[:n]:
for l in g:
if not isinstance(l, bn_types): requires_grad(l, False)
for g in self.layer_groups[n:]:
requires_grad(g, True)
def create_loss(device_type='cuda'):
if device_type == 'cuda':
vgg_m = vgg16_bn(True).features.cuda().eval()
else:
vgg_m = vgg16_bn(True).features.eval()
requires_grad(vgg_m, False)
blocks = [
i - 1 for i, o in enumerate(children(vgg_m))
if isinstance(o, nn.MaxPool2d)
]
blocks, [vgg_m[i] for i in blocks]
feat_loss = FeatureLoss(vgg_m, blocks[2:5], [5, 15, 2])
return feat_loss
def __init__(self, layer_wgts=[20, 70, 10]):
super().__init__()
self.m_feat = models.vgg16_bn(True).features.cuda().eval()
requires_grad(self.m_feat, False)
blocks = [
i - 1 for i, o in enumerate(children(self.m_feat))
if isinstance(o, nn.MaxPool2d)
]
layer_ids = blocks[2:5]
self.loss_features = [self.m_feat[i] for i in layer_ids]
self.hooks = hook_outputs(self.loss_features, detach=False)
self.wgts = layer_wgts
self.metric_names = ['pixel'] + \
[f'feat_{i}' for i in range(len(layer_ids))]
self.base_loss = F.l1_loss
vgg_m = vgg16_bn(True).features
conv1 = getattr(vgg_m, '0')
conv_new = nn.Conv2d(31, 64, kernel_size=3, stride=1, padding=1)
weight_old = conv1.weight
weight_new = conv_new.weight
for i in range(10):
weight_new[:, i:i + 3] = weight_old.data.clone()
weight_new[:, 30] = weight_old.data[:, 0].clone()
conv_new.weight = nn.Parameter(weight_new)
setattr(vgg_m, '0', conv_new)
vgg_m = vgg_m.cuda().eval()
requires_grad(vgg_m, False)
blocks = [
i - 1 for i, o in enumerate(children(vgg_m))
if isinstance(o, nn.MaxPool2d)
]
blocks, [vgg_m[i] for i in blocks]
class FeatureLoss2(nn.Module):
def __init__(self, m_feat, layer_ids, layer_wgts, mrse=True):
super().__init__()
self.m_feat = m_feat
self.loss_features = [self.m_feat[i] for i in layer_ids]
self.hooks = hook_outputs(self.loss_features, detach=False)
self.wgts = layer_wgts
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",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
## 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("--model_type",
default="bert",
type=str,
help="The model architecture to be fine-tuned.")
parser.add_argument(
"--model_name_or_path",
default="bert-base-cased",
type=str,
help="The model checkpoint for weights initialization.")
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="",
type=str,
help=
"Optional pretrained config name or path if not the same as model_name_or_path"
)
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
help=
"Optional pretrained tokenizer name or path if not the same as model_name_or_path"
)
parser.add_argument("--tokenizer_class",
default="",
type=str,
help="Optional pretrained tokenizer clas")
parser.add_argument(
"--cache_dir",
default="",
type=str,
help=
"Optional directory to store the pre-trained models downloaded from s3 (instread 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('--eval_steps',
type=int,
default=100,
help="Evaluate every X updates steps.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
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 deay if we apply some.")
parser.add_argument("--adam_epsilon",
default=1e-6,
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_samples",
default=0,
type=int,
help="Linear warmup over warmup_samples.")
parser.add_argument("--lr_decay",
action='store_true',
help="Decay LR using get_linear_schedule_with_warmup.")
parser.add_argument(
"--unfreeze_level",
default=-1,
type=int,
help="If > 0: freeze all layers except few first and last.")
parser.add_argument('--logging_steps',
type=int,
default=50,
help="Log every X updates steps.")
parser.add_argument('--save_steps',
type=int,
default=50,
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("--local_rank",
type=int,
default=-1,
help="For distributed training: local_rank")
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()
if args.model_type in ["bert", "roberta", "distilbert"] and not args.mlm:
raise ValueError(
"BERT and RoBERTa 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 os.path.exists(args.output_dir) and os.listdir(
args.output_dir
) and args.do_train and not args.overwrite_output_dir:
raise ValueError(
f"Output directory ({args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
# 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)
torch.distributed.init_process_group(backend='nccl')
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, args.device, args.n_gpu, bool(args.local_rank != -1),
args.fp16)
# Set seed
set_seed(args)
# Load pretrained model and tokenizer
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
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path)
if args.tokenizer_class: tokenizer_class = globals()[args.tokenizer_class]
# TODO check tokenizer here
tokenizer = tokenizer_class.from_pretrained(
args.tokenizer_name
if args.tokenizer_name else args.model_name_or_path,
do_lower_case=args.do_lower_case)
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
args.block_size = min(args.block_size, tokenizer.max_len_single_sentence)
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool('.ckpt' in args.model_name_or_path),
config=config)
model.to(args.device)
print(200 * '/')
print(
len([
param for item in flatten_model(model)
for param in item.parameters() if param.requires_grad
])) # freeze all layers but few first and last
if args.unfreeze_level >= 0:
flat = flatten_model(model)
flat = [item for item in flat if list(item.parameters())]
i_start = 3
i_end = 1
need_grads = set(flat[:i_start + args.unfreeze_level * 3]) | set(
flat[-(i_end + args.unfreeze_level * 3):])
for item in flat:
requires_grad(item, item in need_grads)
print(200 * '/')
print(
len([
param for item in flatten_model(model)
for param in item.parameters() if param.requires_grad
]))
if args.local_rank == 0:
torch.distributed.barrier(
) # End of barrier to make sure only the first process in distributed training download model & vocab
logger.info("Training/evaluation parameters %s", args)
# Training
if args.do_train:
if args.local_rank not in [-1, 0]:
torch.distributed.barrier(
) # Barrier to make sure only the first process in distributed training process the dataset, and the others will use the cache
args.train_batch_size = args.per_gpu_train_batch_size * max(
1, args.n_gpu)
train_dataset = load_and_cache_examples(args,
tokenizer,
evaluate=False)
if args.local_rank == 0:
torch.distributed.barrier()
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 torch.distributed.get_rank() == 0):
save_state(args, model, tokenizer, global_step)
# 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, do_lower_case=args.do_lower_case)
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 ""
model = model_class.from_pretrained(checkpoint)
model.to(args.device)
result = evaluate(args, model, tokenizer, prefix=global_step)
result = dict(
(k + '_{}'.format(global_step), v) for k, v in result.items())
results.update(result)
return results
