mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-26 18:24:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add files via upload

Browse Source
This commit is contained in:
Mingshuang Luo 2021-09-29 19:23:54 +08:00 committed by GitHub
parent bdd890bab9
commit 279dc74b4e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 14 additions and 39 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

53
egs/librispeech/ASR/tdnn_lstm_ctc/train.py
View File

@ -58,10 +58,7 @@ def get_parser():
) )
parser.add_argument( parser.add_argument(
"--world-size", "--world-size", type=int, default=1, help="Number of GPUs for DDP training.",
type=int,
default=1,
help="Number of GPUs for DDP training.",
) )
parser.add_argument( parser.add_argument(
@ -79,10 +76,7 @@ def get_parser():
) )
parser.add_argument( parser.add_argument(
"--num-epochs", "--num-epochs", type=int, default=20, help="Number of epochs to train.",
type=int,
default=20,
help="Number of epochs to train.",
) )
parser.add_argument( parser.add_argument(
@ -209,10 +203,7 @@ def load_checkpoint_if_available(
filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
saved_params = load_checkpoint( saved_params = load_checkpoint(
filename, filename, model=model, optimizer=optimizer, scheduler=scheduler,
model=model,
optimizer=optimizer,
scheduler=scheduler,
) )
keys = [ keys = [
@ -312,9 +303,7 @@ def compute_loss(
decoding_graph = graph_compiler.compile(texts) decoding_graph = graph_compiler.compile(texts)
dense_fsa_vec = k2.DenseFsaVec( dense_fsa_vec = k2.DenseFsaVec(
nnet_output, nnet_output, supervision_segments, allow_truncate=params.subsampling_factor - 1,
supervision_segments,
allow_truncate=params.subsampling_factor - 1,
) )
loss = k2.ctc_loss( loss = k2.ctc_loss(
@ -328,8 +317,8 @@ def compute_loss(
assert loss.requires_grad == is_training assert loss.requires_grad == is_training
info = LossRecord() info = LossRecord()
info['frames'] = supervision_segments[:, 2].sum().item() info["frames"] = supervision_segments[:, 2].sum().item()
info['loss'] = loss.detach().cpu().item() info["loss"] = loss.detach().cpu().item()
return loss, info return loss, info
@ -363,7 +352,7 @@ def compute_validation_loss(
if world_size > 1: if world_size > 1:
tot_loss.reduce(loss.device) tot_loss.reduce(loss.device)
loss_value = tot_loss['loss'] / tot_loss['frames'] loss_value = tot_loss["loss"] / tot_loss["frames"]
if loss_value < params.best_valid_loss: if loss_value < params.best_valid_loss:
params.best_valid_epoch = params.cur_epoch params.best_valid_epoch = params.cur_epoch
@ -439,15 +428,9 @@ def train_one_epoch(
if tb_writer is not None: if tb_writer is not None:
loss_info.write_summary( loss_info.write_summary(
tb_writer, tb_writer, "train/current_", params.batch_idx_train
"train/current_",
params.batch_idx_train
)
tot_loss.write_summary(
tb_writer,
"train/tot_",
params.batch_idx_train
) )
tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
if batch_idx > 0 and batch_idx % params.valid_interval == 0: if batch_idx > 0 and batch_idx % params.valid_interval == 0:
valid_info = compute_validation_loss( valid_info = compute_validation_loss(
@ -458,17 +441,13 @@ def train_one_epoch(
world_size=world_size, world_size=world_size,
) )
model.train() model.train()
logging.info( logging.info(f"Epoch {params.cur_epoch}, validation {valid_info}")
f"Epoch {params.cur_epoch}, validation {valid_info}"
)
if tb_writer is not None: if tb_writer is not None:
valid_info.write_summary( valid_info.write_summary(
tb_writer, tb_writer, "train/valid_", params.batch_idx_train,
"train/valid_",
params.batch_idx_train,
) )
loss_value = tot_loss['loss'] / tot_loss['frames'] loss_value = tot_loss["loss"] / tot_loss["frames"]
params.train_loss = loss_value params.train_loss = loss_value
if params.train_loss < params.best_train_loss: if params.train_loss < params.best_train_loss:
@ -526,9 +505,7 @@ def run(rank, world_size, args):
model = DDP(model, device_ids=[rank]) model = DDP(model, device_ids=[rank])
optimizer = optim.AdamW( optimizer = optim.AdamW(
model.parameters(), model.parameters(), lr=params.lr, weight_decay=params.weight_decay,
lr=params.lr,
weight_decay=params.weight_decay,
) )
scheduler = StepLR(optimizer, step_size=8, gamma=0.1) scheduler = StepLR(optimizer, step_size=8, gamma=0.1)
@ -548,9 +525,7 @@ def run(rank, world_size, args):
if tb_writer is not None: if tb_writer is not None:
tb_writer.add_scalar( tb_writer.add_scalar(
"train/lr", "train/lr", scheduler.get_last_lr()[0], params.batch_idx_train,
scheduler.get_last_lr()[0],
params.batch_idx_train,
) )
tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)