def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument(
"--model_name_or_path",
default=None,
type=str,
required=True,
help="Path to pre-trained model or shortcut name selected in the list: "
+ ", ".join(ALL_MODELS))
parser.add_argument(
"--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train selected in the list: " +
", ".join(processors.keys()))
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(
"--config_name",
default="",
type=str,
help=
"Pretrained config name or path if not the same as model_name_or_path")
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
help=
"Pretrained tokenizer name or path if not the same as model_name_or_path"
)
parser.add_argument(
"--cache_dir",
default="",
type=str,
help=
"Where do you want to store the pre-trained models downloaded from s3")
parser.add_argument(
"--data_subset",
type=int,
default=-1,
help="If > 0: limit the data to a subset of data_subset instances.")
parser.add_argument("--overwrite_output_dir",
action='store_true',
help="Whether to overwrite data in output directory")
parser.add_argument(
'--overwrite_cache',
action='store_true',
help="Overwrite the cached training and evaluation sets")
parser.add_argument("--dont_normalize_importance_by_layer",
action='store_true',
help="Don't normalize importance score by layers")
parser.add_argument(
"--dont_normalize_global_importance",
action='store_true',
help="Don't normalize all importance scores between 0 and 1")
parser.add_argument(
"--try_masking",
action='store_true',
help="Whether to try to mask head until a threshold of accuracy.")
parser.add_argument(
"--masking_threshold",
default=0.9,
type=float,
help=
"masking threshold in term of metrics (stop masking when metric < threshold * original metric value)."
)
parser.add_argument(
"--masking_amount",
default=0.1,
type=float,
help="Amount to heads to masking at each masking step.")
parser.add_argument("--metric_name",
default="acc",
type=str,
help="Metric to use for head masking.")
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, sequences shorter padded."
)
parser.add_argument("--batch_size",
default=1,
type=int,
help="Batch size.")
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--server_ip',
type=str,
default='',
help="Can be used for distant debugging.")
parser.add_argument('--server_port',
type=str,
default='',
help="Can be used for distant debugging.")
args = parser.parse_args()
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 devices and distributed training
if args.local_rank == -1 or args.no_cuda:
args.device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
args.device = torch.device("cuda", args.local_rank)
args.n_gpu = 1
torch.distributed.init_process_group(
backend='nccl') # Initializes the distributed backend
# Setup logging
logging.basicConfig(
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.info("device: {} n_gpu: {}, distributed: {}".format(
args.device, args.n_gpu, bool(args.local_rank != -1)))
# Set seeds
set_seed(args)
# Prepare GLUE task
args.task_name = args.task_name.lower()
if args.task_name not in processors:
raise ValueError("Task not found: %s" % (args.task_name))
processor = processors[args.task_name]()
args.output_mode = output_modes[args.task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
# Load pretrained model and tokenizer
if args.local_rank not in [-1, 0]:
torch.distributed.barrier(
) # Make sure only the first process in distributed training will download model & vocab
args.model_type = ""
for key in MODEL_CLASSES:
if key in args.model_name_or_path.lower():
args.model_type = key # take the first match in model types
break
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,
num_labels=num_labels,
finetuning_task=args.task_name,
output_attentions=True,
cache_dir=args.cache_dir if args.cache_dir else None)
tokenizer = tokenizer_class.from_pretrained(
args.tokenizer_name
if args.tokenizer_name else args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else None)
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool('.ckpt' in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None)
if args.local_rank == 0:
torch.distributed.barrier(
) # Make sure only the first process in distributed training will download model & vocab
# Distributed and parallel training
model.to(args.device)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
elif args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Print/save training arguments
torch.save(args, os.path.join(args.output_dir, 'run_args.bin'))
logger.info("Training/evaluation parameters %s", args)
# Prepare dataset for the GLUE task
eval_data = load_and_cache_examples(args,
args.task_name,
tokenizer,
evaluate=True)
if args.data_subset > 0:
eval_data = Subset(eval_data,
list(range(min(args.data_subset, len(eval_data)))))
eval_sampler = SequentialSampler(
eval_data) if args.local_rank == -1 else DistributedSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.batch_size)
# Compute head entropy and importance score
compute_heads_importance(args, model, eval_dataloader)
# Try head masking (set heads to zero until the score goes under a threshole)
# and head pruning (remove masked heads and see the effect on the network)
if args.try_masking and args.masking_threshold > 0.0 and args.masking_threshold < 1.0:
head_mask = mask_heads(args, model, eval_dataloader)
prune_heads(args, model, eval_dataloader, head_mask)
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task.",
)
parser.add_argument(
"--model_type",
default=None,
type=str,
required=True,
help="Model type selected in the list: " +
", ".join(MODEL_CLASSES.keys()),
)
parser.add_argument(
"--model_name_or_path",
default=None,
type=str,
required=True,
help="Path to pre-trained model or shortcut name selected in the list: "
+ ", ".join(ALL_MODELS),
)
parser.add_argument(
"--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train selected in the list: " +
", ".join(processors.keys()),
)
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(
"--config_name",
default="",
type=str,
help=
"Pretrained config name or path if not the same as model_name_or_path",
)
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
help=
"Pretrained tokenizer name or path if not the same as model_name_or_path",
)
parser.add_argument(
"--cache_dir",
default="",
type=str,
help=
"Where do you want to store the pre-trained models downloaded from s3",
)
parser.add_argument(
"--data_subset",
type=int,
default=-1,
help="If > 0: limit the data to a subset of data_subset instances.")
parser.add_argument("--overwrite_output_dir",
action="store_true",
help="Whether to overwrite data in output directory")
parser.add_argument(
"--overwrite_cache",
action="store_true",
help="Overwrite the cached training and evaluation sets")
parser.add_argument("--dont_normalize_importance_by_layer",
action="store_true",
help="Don't normalize importance score by layers")
parser.add_argument(
"--dont_normalize_global_importance",
action="store_true",
help="Don't normalize all importance scores between 0 and 1",
)
parser.add_argument("--use_train_data",
action="store_true",
help="Use training set for computing masks")
parser.add_argument(
"--masking_threshold",
default=0.9,
type=float,
help=
"masking threshold in term of metrics (stop masking when metric < threshold * original metric value).",
)
parser.add_argument(
"--masking_amount",
default=0.1,
type=float,
help="Amount to heads to masking at each masking step.")
parser.add_argument("--metric_name",
default=None,
type=str,
help="Metric to use for head masking.")
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, sequences shorter padded.",
)
parser.add_argument("--batch_size",
default=1,
type=int,
help="Batch size.")
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument("--no_cuda",
action="store_true",
help="Whether not to use CUDA when available")
parser.add_argument("--server_ip",
type=str,
default="",
help="Can be used for distant debugging.")
parser.add_argument("--server_port",
type=str,
default="",
help="Can be used for distant debugging.")
args = parser.parse_args()
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 devices and distributed training
if args.local_rank == -1 or args.no_cuda:
args.device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
args.device = torch.device("cuda", args.local_rank)
args.n_gpu = 1
torch.distributed.init_process_group(
backend="nccl") # Initializes the distributed backend
# Setup logging
logging.basicConfig(
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.info("device: {} n_gpu: {}, distributed: {}".format(
args.device, args.n_gpu, bool(args.local_rank != -1)))
# Set seeds
set_seed(args)
tracker = ImpactTracker(args.output_dir)
tracker.launch_impact_monitor()
# Prepare GLUE task
args.task_name = args.task_name.lower()
if args.task_name not in processors:
raise ValueError("Task not found: %s" % (args.task_name))
if args.metric_name is None:
args.metric_name = {
"cola": "mcc",
"mnli": "acc",
"mnli-mm": "acc",
"mrpc": "acc",
"sst-2": "acc",
"sts-b": "corr",
"qqp": "acc",
"qnli": "acc",
"rte": "acc",
"wnli": "acc",
"hans": "acc",
"mnli_two": "acc",
"hans_mnli": "acc"
}[args.task_name]
processor = processors[args.task_name]()
args.output_mode = output_modes[args.task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
# Load pretrained model and tokenizer
if args.local_rank not in [-1, 0]:
torch.distributed.barrier(
) # Make sure only the first process in distributed training will download model & vocab
MODEL_CLASSES["bert"] = (BertConfig, BertForSequenceClassification,
MODEL_CLASSES["bert"][2])
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
if args.model_type != "bert":
raise NotImplemented("Not implemented for non BERT classes yet.")
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_labels,
finetuning_task=args.task_name,
output_attentions=True,
cache_dir=args.cache_dir if args.cache_dir else None,
)
tokenizer = tokenizer_class.from_pretrained(
args.tokenizer_name
if args.tokenizer_name else args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else None,
)
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None,
)
if args.local_rank == 0:
torch.distributed.barrier(
) # Make sure only the first process in distributed training will download model & vocab
# Distributed and parallel training
model.to(args.device)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
elif args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Print/save training arguments
torch.save(args, os.path.join(args.output_dir, "run_args.bin"))
logger.info("Training/evaluation parameters %s", args)
# Prepare dataset for the GLUE task
eval_data = load_and_cache_examples(args,
args.task_name,
tokenizer,
evaluate=True)
true_eval_len = len(eval_data)
if args.use_train_data:
train_data = load_and_cache_examples(args,
args.task_name,
tokenizer,
evaluate=False)
eval_data = random_split(
train_data,
[true_eval_len, len(train_data) - true_eval_len])[0]
if args.data_subset > 0:
eval_data = Subset(eval_data,
list(range(min(args.data_subset, len(eval_data)))))
eval_sampler = SequentialSampler(
eval_data) if args.local_rank == -1 else DistributedSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.batch_size)
_, masked_amount = mask_global(args, model, eval_dataloader)
# Generating a random mask of equal size
fresh_model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None,
)
mask = prune_model(fresh_model, masked_amount, prune.RandomUnstructured)
torch.save(mask, os.path.join(args.output_dir, "random_mask.p"))
fresh_model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None,
)
mask = prune_model(fresh_model, masked_amount, L1UnstructuredInvert)
torch.save(mask, os.path.join(args.output_dir, "bad_mask.pt"))
def main():
parser = argparse.ArgumentParser(description="Training")
## Required parameters
parser.add_argument("--data_dir", default=None, type=str, required=True,
help="The input data dir. Should contain the .tsv files (or other data files) for the task.")
parser.add_argument("--output_dir", type=str, required=True,
help="The output directory (log, checkpoints, parameters, etc.)")
# parser.add_argument("--data_file", type=str, required=True,
# help="The binarized file (tokenized + tokens_to_ids) and grouped by sequence.")
# parser.add_argument("--token_counts", type=str, required=True,
# help="The token counts in the data_file for MLM.")
parser.add_argument("--force", action='store_true',
help="Overwrite output_dir if it already exists.")
parser.add_argument("--task_name", default=None, type=str, required=True,
help="The name of the task to train selected in the list: " + ", ".join(processors.keys()))
parser.add_argument("--do_lower_case", action='store_true',
help="Set this flag if you are using an uncased model.")
# parser.add_argument("--output_dir", default=None, type=str, required=True,
# help="The output directory where the model predictions and checkpoints will be written.")
parser.add_argument("--vocab_size", default=30522, type=int,
help="The vocabulary size.")
parser.add_argument("--max_position_embeddings", default=512, type=int,
help="Maximum sequence length we can model (including [CLS] and [SEP]).")
parser.add_argument("--sinusoidal_pos_embds", action='store_false',
help="If true, the position embeddings are simply fixed with sinusoidal embeddings.")
parser.add_argument("--n_layers", default=6, type=int,
help="Number of Transformer blocks.")
parser.add_argument("--n_heads", default=12, type=int,
help="Number of heads in the self-attention module.")
parser.add_argument("--dim", default=768, type=int,
help="Dimension through the network. Must be divisible by n_heads")
parser.add_argument("--hidden_dim", default=3072, type=int,
help="Intermediate dimension in the FFN.")
parser.add_argument("--dropout", default=0.1, type=float,
help="Dropout.")
parser.add_argument("--attention_dropout", default=0.1, type=float,
help="Dropout in self-attention.")
parser.add_argument("--activation", default='gelu', type=str,
help="Activation to use in self-attention")
parser.add_argument("--tie_weights_", action='store_false',
help="If true, we tie the embeddings matrix with the projection over the vocabulary matrix. Default is true.")
# parser.add_argument("--from_pretrained_weights", default=None, type=str,
# help="Load student initialization checkpoint.")
# parser.add_argument("--from_pretrained_config", default=None, type=str,
# help="Load student initialization architecture config.")
parser.add_argument("--teacher_name", default="bert-base-uncased", type=str,
help="The teacher model.")
parser.add_argument("--temperature", default=2., type=float,
help="Temperature for the softmax temperature.")
parser.add_argument("--alpha_ce", default=1.0, type=float,
help="Linear weight for the distillation loss. Must be >=0.")
# parser.add_argument("--alpha_mlm", default=0.5, type=float,
# help="Linear weight for the MLM loss. Must be >=0.")
parser.add_argument("--alpha_mse", default=0.0, type=float,
help="Linear weight of the MSE loss. Must be >=0.")
parser.add_argument("--alpha_cos", default=0.0, type=float,
help="Linear weight of the cosine embedding loss. Must be >=0.")
# parser.add_argument("--mlm_mask_prop", default=0.15, type=float,
# help="Proportion of tokens for which we need to make a prediction.")
# parser.add_argument("--word_mask", default=0.8, type=float,
# help="Proportion of tokens to mask out.")
# parser.add_argument("--word_keep", default=0.1, type=float,
# help="Proportion of tokens to keep.")
# parser.add_argument("--word_rand", default=0.1, type=float,
# help="Proportion of tokens to randomly replace.")
# parser.add_argument("--mlm_smoothing", default=0.7, type=float,
# help="Smoothing parameter to emphasize more rare tokens (see XLM, similar to word2vec).")
# parser.add_argument("--restrict_ce_to_mask", action='store_true',
# help="If true, compute the distilation loss only the [MLM] prediction distribution.")
parser.add_argument("--n_epoch", type=int, default=3,
help="Number of pass on the whole dataset.")
parser.add_argument("--batch_size", type=int, default=5,
help="Batch size (for each process).")
parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
help="Batch size per GPU/CPU for evaluation.")
# parser.add_argument("--tokens_per_batch", type=int, default=-1,
# help="If specified, modify the batches so that they have approximately this number of tokens.")
# parser.add_argument("--shuffle", action='store_false',
# help="If true, shuffle the sequence order. Default is true.")
parser.add_argument("--group_by_size", action='store_false',
help="If true, group sequences that have similar length into the same batch. Default is true.")
parser.add_argument("--gradient_accumulation_steps", type=int, default=50,
help="Gradient accumulation for larger training batches.")
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_prop", default=0.05, type=float,
help="Linear warmup proportion.")
parser.add_argument("--weight_decay", default=0.0, type=float,
help="Weight deay if we apply some.")
parser.add_argument("--learning_rate", default=5e-4, type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--adam_epsilon", default=1e-6, type=float,
help="Epsilon for Adam optimizer.")
parser.add_argument("--max_grad_norm", default=5.0, type=float,
help="Max gradient norm.")
parser.add_argument("--initializer_range", default=0.02, type=float,
help="Random initialization range.")
# 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("--n_gpu", type=int, default=1,
help="Number of GPUs in the node.")
parser.add_argument("--local_rank", type=int, default=-1,
help="Distributed training - Local rank")
parser.add_argument("--seed", type=int, default=56,
help="Random seed")
parser.add_argument("--log_interval", type=int, default=10,
help="Tensorboard logging interval.")
parser.add_argument('--log_examples', action='store_false',
help="Show input examples on the command line during evaluation. Enabled by default.")
parser.add_argument("--evaluate_during_training", action='store_true',
help="Rul evaluation during training at each logging step.")
parser.add_argument("--checkpoint_interval", type=int, default=100,
help="Checkpoint interval.")
parser.add_argument("--no_cuda", type=bool, default=False,
help="Avoid using CUDA when available")
parser.add_argument("--toy_mode", action='store_true', help="Toy mode for development.")
parser.add_argument("--rich_eval", action='store_true', help="Rich evaluation (more metrics + mistake reporting).")
args = parser.parse_args()
args.no_cuda = False
print("NO CUDA", args.no_cuda)
## ARGS ##
init_gpu_params(args)
set_seed(args)
if args.is_master:
if os.path.exists(args.output_dir):
if not args.force:
raise ValueError(f'Serialization dir {args.output_dir} already exists, but you have not precised wheter to overwrite it'
'Use `--force` if you want to overwrite it')
else:
shutil.rmtree(args.output_dir)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
logger.info(f'Experiment will be dumped and logged in {args.output_dir}')
### SAVE PARAMS ###
logger.info(f'Param: {args}')
with open(os.path.join(args.output_dir, 'parameters.json'), 'w') as f:
json.dump(vars(args), f, indent=4)
# git_log(args.output_dir)
print(args.local_rank)
print(args.no_cuda)
# 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()
args.local_rank = -1
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
device = torch.device("cuda", args.local_rank)
args.device = device
logger.info("DEVICE: {}".format(args.device))
logger.info("N_GPU: {}".format(args.n_gpu))
# exit(0)
print(args.local_rank)
### TOKENIZER ###
# if args.teacher_type == 'bert':
tokenizer = BertTokenizer.from_pretrained(args.teacher_name, do_lower_case=args.do_lower_case)
# elif args.teacher_type == 'roberta':
# tokenizer = RobertaTokenizer.from_pretrained(args.teacher_name)
special_tok_ids = {}
for tok_name, tok_symbol in tokenizer.special_tokens_map.items():
idx = tokenizer.all_special_tokens.index(tok_symbol)
special_tok_ids[tok_name] = tokenizer.all_special_ids[idx]
logger.info(f'Special tokens {special_tok_ids}')
args.special_tok_ids = special_tok_ids
# Prepare GLUE task
args.task_name = args.task_name.lower()
if args.task_name not in processors:
raise ValueError("Task not found: %s" % (args.task_name))
args.model_name_or_path = args.teacher_name
args.max_seq_length = args.max_position_embeddings
args.model_type = "bert"
args.output_mode = output_modes[args.task_name]
train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
## DATA LOADER ##
# logger.info(f'Loading data from {args.data_file}')
# with open(args.data_file, 'rb') as fp:
# data = pickle.load(fp)
# assert os.path.isfile(args.token_counts)
# logger.info(f'Loading token counts from {args.token_counts} (already pre-computed)')
# with open(args.token_counts, 'rb') as fp:
# counts = pickle.load(fp)
# assert len(counts) == args.vocab_size
# token_probs = np.maximum(counts, 1) ** -args.mlm_smoothing
# for idx in special_tok_ids.values():
# token_probs[idx] = 0. # do not predict special tokens
# token_probs = torch.from_numpy(token_probs)
# train_dataloader = Dataset(params=args, data=data)
train_sampler = RandomSampler(train_dataset)
train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.batch_size)
logger.info(f'Data loader created.')
## STUDENT ##
# if args.from_pretrained_weights is not None:
# assert os.path.isfile(args.from_pretrained_weights)
# assert os.path.isfile(args.from_pretrained_config)
# logger.info(f'Loading pretrained weights from {args.from_pretrained_weights}')
# logger.info(f'Loading pretrained config from {args.from_pretrained_config}')
# stu_architecture_config = DistilBertConfig.from_json_file(args.from_pretrained_config)
# stu_architecture_config.output_hidden_states = True
# student = DistilBertForMaskedLM.from_pretrained(args.from_pretrained_weights,
# config=stu_architecture_config)
# else:
# args.vocab_size_or_config_json_file = args.vocab_size
# stu_architecture_config = DistilBertConfig(**vars(args), output_hidden_states=True)
# student = DistilBertForMaskedLM(stu_architecture_config)
student_config = BertConfig(
vocab_size_or_config_json_file=args.vocab_size,
hidden_size=args.dim,
num_hidden_layers=args.n_layers,
num_attention_heads=args.n_heads,
intermediate_size=args.hidden_dim,
hidden_dropout_prob=args.dropout,
attention_probs_dropout_prob=args.attention_dropout,
max_position_embeddings=args.max_position_embeddings,
hidden_act=args.activation,
initializer_range=0.02)
student = BertForSequenceClassification(student_config)
if args.n_gpu > 0:
student.to(f'cuda:{args.local_rank}')
logger.info(f'Student loaded.')
## TEACHER ##
teacher = BertForSequenceClassification.from_pretrained(args.teacher_name) # take outputs[1] for the logits
if args.n_gpu > 0:
teacher.to(f'cuda:{args.local_rank}')
logger.info(f'Teacher loaded from {args.teacher_name}.')
## DISTILLER ##
torch.cuda.empty_cache()
distiller = Distiller(params=args,
dataloader=train_dataloader,
# token_probs=token_probs,
student=student,
teacher=teacher,
tokenizer=tokenizer)
distiller.train()
logger.info("Let's go get some drinks.")
args["model_name_or_path"] = 'bert-multilingual-base'
args["max_seq_length"] = 128
args["overwrite_cache"] = True
args["data_dir"] = "/data/misc/cc/2_pairing/glue_data/MRPC"
args["model_type"] = "bert"
args["task_name"] = "mrpc"
args["output_dir"] = "/tmp/mrpc_eval_outputs"
args["per_gpu_eval_batch_size"] = 2
args["n_gpu"] = 1
args["device"] = "cuda"
args["output_mode"] = "classification"
tokenizer = BertTokenizer.from_pretrained('bert-base-multilingual-cased')
train_dataset = load_and_cache_examples(args,
"mrpc",
tokenizer,
evaluate=False)
def split_token_pair(ids_tensor, sep_token_id=102, target_len=128):
ids_np = ids_tensor.numpy()
split_idxs = np.argwhere(ids_np == sep_token_id).flatten()
out = [ids_np[:split_idxs[0]], ids_np[split_idxs[0] + 1:split_idxs[1]]]
out = [
np.append(part, np.zeros(target_len - len(part)), axis=0)
for part in out
]
return torch.tensor(out)
train_tokens = torch.stack([
