match = tf.equal(predictions_z, tf.cast(labels_z, predictions_z.dtype))
metrics_dict['test_accuracy'](tf.reduce_mean(tf.cast(match, tf.float32)))
from collections import defaultdict
scheduled_parameters = defaultdict(lambda: {})
if p_type == "schedule":
for epoch in range(0, 6):
scheduled_parameters[epoch] = {"p_schedule": tf.constant(1.)}
for epoch in range(6, 12):
scheduled_parameters[epoch] = {"p_schedule": tf.constant(2.)}
for epoch in range(12, 18):
scheduled_parameters[epoch] = {"p_schedule": tf.constant(4.)}
for epoch in range(18, EPOCHS):
scheduled_parameters[epoch] = {"p_schedule": tf.constant(6.)}
if p_type == "1":
for epoch in range(EPOCHS):
scheduled_parameters[epoch] = {"p_schedule": tf.constant(1.)}
if p_type == "2":
for epoch in range(EPOCHS):
scheduled_parameters[epoch] = {"p_schedule": tf.constant(2.)}
commons.train(EPOCHS,
metrics_dict,
ds_train,
ds_test,
train_step,
test_step,
csv_path=csv_path,
scheduled_parameters=scheduled_parameters)
sat = axioms(data, labels)
loss = 1. - sat
gradients = tape.gradient(loss, A.trainable_variables)
optimizer.apply_gradients(zip(gradients, A.trainable_variables))
metrics_dict['train_sat'](sat)
# accuracy
predictions = A.model(data)
metrics_dict['train_accuracy'](labels, predictions)
@tf.function
def test_step(data, labels):
# sat and update
sat = axioms(data, labels)
metrics_dict['test_sat'](sat)
# accuracy
predictions = A.model(data)
metrics_dict['test_accuracy'](labels, predictions)
import commons
commons.train(EPOCHS,
metrics_dict,
ds_train,
ds_test,
train_step,
test_step,
csv_path=csv_path,
track_metrics=20)
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model checkpoints and predictions will be written."
)
parser.add_argument("--mode",
default='transfer',
choices=['transfer', 'meta', 'lrcmeta', 'svimeta'],
help="Modes.")
parser.add_argument("--meta_emb_dim",
default=100,
type=int,
help="Dimensionality of task and language embeddings.")
parser.add_argument("--n_samples",
default=3,
type=int,
help="Number of samples in the Bayesian mode.")
parser.add_argument(
"--scaling",
default='uniform',
type=str,
choices=['uniform', 'linear_annealing', 'logistic_annealing'],
help="Scaling for KL term in VI.")
parser.add_argument("--max_patience",
default=10,
type=int,
help="Maximum patience for early stopping.")
parser.add_argument(
"--weight_by_size",
action='store_true',
help=
"Sample task-language example according to data size, weight lr accordingly"
)
parser.add_argument(
"--num_hidden_layers",
default=6,
type=int,
help="Number of hidden layers for the functions psi and phi")
parser.add_argument(
"--rank_cov",
default=0,
type=int,
help="Rank of the factored covariance matrix. Diagonal if < 1")
parser.add_argument(
"--typ_dist",
default="src/typ_feats.tab",
type=str,
help=
"File containing pre-computed typological distances between languages")
parser.add_argument(
"--largest_source",
action='store_true',
help=
"Always choose the source language with the largest number of examples for transfer"
)
parser.add_argument(
"--model_averaging",
action='store_true',
help="Predict through model averaging rather than pluggin in the mean")
# Experiment
parser.add_argument("--model_name_or_path",
default="bert-base-multilingual-cased",
type=str,
help="Path to pre-trained model or shortcut name")
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("--no_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--no_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--debug",
action='store_true',
help="Whether to debug gradient flow.")
parser.add_argument(
"--max_seq_length",
default=250,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. Sequences "
"longer than this will be truncated, and sequences shorter than this will be padded."
)
parser.add_argument("--per_gpu_train_batch_size",
default=8,
type=int,
help="Batch size per GPU/CPU for training.")
parser.add_argument("--per_gpu_eval_batch_size",
default=8,
type=int,
help="Batch size per GPU/CPU for evaluation.")
parser.add_argument("--learning_rate",
default=5e-6,
type=float,
help="The initial learning rate for Adam.")
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("--weight_decay",
default=0.0,
type=float,
help="Weight deay if we apply some.")
parser.add_argument("--adam_epsilon",
default=1e-8,
type=float,
help="Epsilon for Adam optimizer.")
parser.add_argument("--max_grad_norm",
default=1.0,
type=float,
help="Max gradient norm.")
parser.add_argument("--num_train_epochs",
default=8.0,
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_proportion",
default=0.,
type=float,
help="Linear warmup over warmup_steps.")
parser.add_argument('--logging_steps',
type=int,
default=2500,
help="Log every X updates steps.")
parser.add_argument('--save_steps',
type=int,
default=2500,
help="Save checkpoint every X updates steps.")
parser.add_argument(
"--eval_all_checkpoints",
action='store_true',
help=
"Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number"
)
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use 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("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
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")
args = parser.parse_args()
if args.largest_source:
assert args.mode == "transfer"
args.tasks = ['pos', 'ner']
args.languages = sorted([
'aii', 'am', 'ar', 'bm', 'cy', 'et', 'eu', 'fi', 'fo', 'gl', 'gun',
'he', 'hsb', 'hu', 'hy', 'id', 'kk', 'kmr', 'ko', 'kpv', 'mt', 'myv',
'sme', 'ta', 'te', 'th', 'tl', 'tr', 'ug', 'vi', 'wo', 'yo', 'yue'
])
# NER changes zh-yue -> yue, gn -> gun, sme -> sm, arc -> aii, ku -> kmr, kv -> kpv,
# find_overlapping_languages(args.tasks)
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir
) and not args.no_train and not args.overwrite_output_dir:
raise ValueError(
"Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome."
.format(args.output_dir))
# Setup 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, device, args.n_gpu, bool(args.local_rank != -1),
args.fp16)
# Set seed
set_seed(args)
# Creating partitions of observed and held-out data for each iteration
cartesian_product = list(
itertools.product(sorted(args.tasks), sorted(args.languages)))
partitions = {
ti: sorted([
cartesian_product[pi]
for pi in range(ti, len(cartesian_product), len(args.tasks))
])
for ti in range(len(args.tasks))
}
for partition, heldout_pairs in partitions.items():
logger.info("Partition: {}".format(partition))
logger.info("Held-out task-language pairs: {}".format(heldout_pairs))
observed_pairs = sorted(
list(set(cartesian_product) - set(heldout_pairs)))
lang_nns = find_nearest_languages(heldout_pairs, observed_pairs,
args.typ_dist)
# 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
config_class, tokenizer_class, model_class = BertConfig, BertTokenizer, MultiTaskBert
# Data and model
config = config_class.from_pretrained(args.model_name_or_path)
tokenizer = tokenizer_class.from_pretrained(
args.model_name_or_path, do_lower_case=args.do_lower_case)
dataloaders, num_batches = load_and_cache_examples(
args, tokenizer, observed_pairs if not args.largest_source else
[('pos', 'en'), ('ner', 'en_{}'.format(partition))], partition)
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool('.ckpt' in args.model_name_or_path),
config=config,
mode=args.mode,
languages=args.languages,
n_classes={t: len(c)
for t, c in CLASSES_PER_TASK.items()},
emb_dim=args.meta_emb_dim,
n_samples=args.n_samples,
num_hidden_layers=args.num_hidden_layers,
rank_cov=args.rank_cov,
largest_source=args.largest_source)
if args.local_rank == 0:
torch.distributed.barrier(
) # Make sure only the first process in distributed training will download model & vocab
model.to(args.device)
logger.info("Training/evaluation parameters %s", args)
# Training
if not args.no_train:
global_step, tr_loss = train(args, dataloaders, model, tokenizer,
num_batches, observed_pairs,
heldout_pairs, partition, lang_nns)
logger.info(" global_step = %s, average loss = %s", global_step,
tr_loss)
model = model_class.from_pretrained(
os.path.join(args.output_dir,
'checkpoint-best-{}'.format(partition)),
from_tf=bool('.ckpt' in args.model_name_or_path),
config=config,
mode=args.mode,
languages=args.languages,
n_classes={t: len(c)
for t, c in CLASSES_PER_TASK.items()},
emb_dim=args.meta_emb_dim,
n_samples=args.n_samples,
num_hidden_layers=args.num_hidden_layers,
rank_cov=args.rank_cov,
largest_source=args.largest_source)
# Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
if not args.no_eval and args.local_rank in [-1, 0]:
# Evaluate
with torch.no_grad():
result = evaluate(args,
dataloaders,
model,
tokenizer,
'test',
heldout_pairs,
lang_nns,
partition,
prefix=partition,
sample=args.model_averaging)
logger.info("Results: {}".format(result))
del model
gc.collect()
torch.cuda.empty_cache()
