speech_recognition/FunASR
1
0
Fork 0
mirror of https://github.com/modelscope/FunASR synced 2025-09-15 14:48:36 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

finetune

Browse Source
This commit is contained in:
游雁 2024-03-24 15:03:54 +08:00
parent a70f5b3edf
commit ed22e34d65
2 changed files with 24 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
funasr/bin/train.py
View File

@ -173,11 +173,11 @@ def main(**kwargs):
except: except:
writer = None writer = None
if use_ddp or use_fsdp: # if use_ddp or use_fsdp:
context = Join([model]) # context = Join([model])
else: # else:
# context = nullcontext()
context = nullcontext() context = nullcontext()
for epoch in range(trainer.start_epoch, trainer.max_epoch + 1): for epoch in range(trainer.start_epoch, trainer.max_epoch + 1):
time1 = time.perf_counter() time1 = time.perf_counter()
with context: with context:

35
funasr/train_utils/trainer.py
View File

@ -250,14 +250,14 @@ class Trainer:
optim.zero_grad() optim.zero_grad()
speed_stats = {} speed_stats = {}
time5 = time.perf_counter() time5 = time.perf_counter()
# iterator_stop = torch.tensor(0).to(self.device) iterator_stop = torch.tensor(0).to(self.device)
dataloader_train.batch_sampler.set_epoch(epoch) dataloader_train.batch_sampler.set_epoch(epoch)
for batch_idx, batch in enumerate(dataloader_train): for batch_idx, batch in enumerate(dataloader_train):
# if self.use_ddp or self.use_fsdp: if self.use_ddp or self.use_fsdp:
# dist.all_reduce(iterator_stop, dist.ReduceOp.SUM) dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)
# if iterator_stop > 0: if iterator_stop > 0:
# break break
self.batch_total += 1 self.batch_total += 1
time1 = time.perf_counter() time1 = time.perf_counter()
speed_stats["data_load"] = f"{time1-time5:0.3f}" speed_stats["data_load"] = f"{time1-time5:0.3f}"
@ -340,7 +340,7 @@ class Trainer:
speed_stats["total_time"] = total_time speed_stats["total_time"] = total_time
lr = scheduler.get_last_lr()[0] lr = scheduler.get_last_lr()[0]
batch_num_epoch = -1 batch_num_epoch = 1
if hasattr(dataloader_train, "__len__"): if hasattr(dataloader_train, "__len__"):
batch_num_epoch = len(dataloader_train) batch_num_epoch = len(dataloader_train)
self.log(epoch, batch_idx, self.log(epoch, batch_idx,
@ -364,14 +364,16 @@ class Trainer:
if (batch_idx+1) % self.save_checkpoint_interval == 0: if (batch_idx+1) % self.save_checkpoint_interval == 0:
self.save_checkpoint(epoch, model=model, optim=optim, scheduler=scheduler, scaler=scaler, step=batch_idx+1) self.save_checkpoint(epoch, model=model, optim=optim, scheduler=scheduler, scaler=scaler, step=batch_idx+1)
# else: else:
# if self.use_ddp or self.use_fsdp: if self.use_ddp or self.use_fsdp:
# iterator_stop.fill_(1) iterator_stop.fill_(1)
# dist.all_reduce(iterator_stop, dist.ReduceOp.SUM) dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)
if self.use_ddp or self.use_fsdp: if self.use_ddp or self.use_fsdp:
dist.barrier() dist.barrier()
iterator_stop = torch.tensor(0).to(self.device)
def validate_epoch(self, def validate_epoch(self,
@ -397,7 +399,7 @@ class Trainer:
speed_stats = {} speed_stats = {}
time5 = time.perf_counter() time5 = time.perf_counter()
# iterator_stop = torch.tensor(0).to(self.device) iterator_stop = torch.tensor(0).to(self.device)
dataloader_val.batch_sampler.set_epoch(epoch) dataloader_val.batch_sampler.set_epoch(epoch)
for batch_idx, batch in enumerate(dataloader_val): for batch_idx, batch in enumerate(dataloader_val):
# if self.use_ddp or self.use_fsdp: # if self.use_ddp or self.use_fsdp:
@ -442,7 +444,7 @@ class Trainer:
self.val_loss_avg = val_loss_avg.detach().cpu().item() / self.world_size self.val_loss_avg = val_loss_avg.detach().cpu().item() / self.world_size
self.val_acc_avg = val_acc_avg.detach().cpu().item() / self.world_size self.val_acc_avg = val_acc_avg.detach().cpu().item() / self.world_size
batch_num_epoch = -1 batch_num_epoch = 1
if hasattr(dataloader_val, "__len__"): if hasattr(dataloader_val, "__len__"):
batch_num_epoch = len(dataloader_val) batch_num_epoch = len(dataloader_val)
self.log(epoch, batch_idx, self.log(epoch, batch_idx,
@ -455,16 +457,17 @@ class Trainer:
tag="val", tag="val",
) )
# else: else:
# if self.use_ddp or self.use_fsdp: if self.use_ddp or self.use_fsdp:
# iterator_stop.fill_(1) iterator_stop.fill_(1)
# dist.all_reduce(iterator_stop, dist.ReduceOp.SUM) dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)
self.val_acc_list.append(self.val_acc_avg) self.val_acc_list.append(self.val_acc_avg)
model.train() model.train()
if self.use_ddp or self.use_fsdp: if self.use_ddp or self.use_fsdp:
dist.barrier() dist.barrier()
iterator_stop = torch.tensor(0).to(self.device)
def log(self, def log(self,