From 543d900522403eccb4e387cbc41c5dce24091d1d Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期四, 22 二月 2024 23:53:10 +0800
Subject: [PATCH] Merge branch 'main' of github.com:alibaba-damo-academy/FunASR merge
---
funasr/train_utils/trainer.py | 37 ++++++++++++++++++++-----------------
1 files changed, 20 insertions(+), 17 deletions(-)
diff --git a/funasr/train_utils/trainer.py b/funasr/train_utils/trainer.py
index 4b85a66..d175fbe 100644
--- a/funasr/train_utils/trainer.py
+++ b/funasr/train_utils/trainer.py
@@ -3,6 +3,7 @@
import torch
import logging
from tqdm import tqdm
+from datetime import datetime
import torch.distributed as dist
from contextlib import nullcontext
# from torch.utils.tensorboard import SummaryWriter
@@ -107,14 +108,10 @@
filename = os.path.join(self.output_dir, f'model.pt.ep{epoch}')
torch.save(state, filename)
- print(f'Checkpoint saved to {filename}')
+ print(f'\nCheckpoint saved to {filename}\n')
latest = Path(os.path.join(self.output_dir, f'model.pt'))
- try:
- latest.unlink()
- except:
- pass
+ torch.save(state, latest)
- latest.symlink_to(filename)
def _resume_checkpoint(self, resume_path):
"""
@@ -160,7 +157,7 @@
self._resume_checkpoint(self.output_dir)
for epoch in range(self.start_epoch, self.max_epoch + 1):
-
+ time1 = time.perf_counter()
self._train_epoch(epoch)
@@ -182,6 +179,9 @@
self.scheduler.step()
+ time2 = time.perf_counter()
+ time_escaped = (time2 - time1)/3600.0
+ print(f"\nrank: {self.local_rank}, time_escaped_epoch: {time_escaped:.3f} hours, estimated to finish {self.max_epoch} epoch: {(self.max_epoch-epoch)*time_escaped:.3f} hours\n")
if self.rank == 0:
average_checkpoints(self.output_dir, self.avg_nbest_model)
@@ -287,10 +287,13 @@
torch.cuda.max_memory_reserved()/1024/1024/1024,
)
lr = self.scheduler.get_last_lr()[0]
+ time_now = datetime.now()
+ time_now = time_now.strftime("%Y-%m-%d %H:%M:%S")
description = (
+ f"{time_now}, "
f"rank: {self.local_rank}, "
f"epoch: {epoch}/{self.max_epoch}, "
- f"step: {batch_idx+1}/{len(self.dataloader_train)}, total: {self.batch_total}, "
+ f"step: {batch_idx+1}/{len(self.dataloader_train)}, total step: {self.batch_total}, "
f"(loss: {loss.detach().cpu().item():.3f}), "
f"(lr: {lr:.3e}), "
f"{[(k, round(v.cpu().item(), 3)) for k, v in stats.items()]}, "
@@ -299,17 +302,14 @@
)
pbar.set_description(description)
if self.writer:
- self.writer.add_scalar(f'rank{self.local_rank}_Loss/train', loss.item(),
- epoch*len(self.dataloader_train) + batch_idx)
+ self.writer.add_scalar(f'rank{self.local_rank}_Loss/train', loss.item(), self.batch_total)
+ self.writer.add_scalar(f'rank{self.local_rank}_lr/train', lr, self.batch_total)
for key, var in stats.items():
- self.writer.add_scalar(f'rank{self.local_rank}_{key}/train', var.item(),
- epoch * len(self.dataloader_train) + batch_idx)
+ self.writer.add_scalar(f'rank{self.local_rank}_{key}/train', var.item(), self.batch_total)
for key, var in speed_stats.items():
- self.writer.add_scalar(f'rank{self.local_rank}_{key}/train', eval(var),
- epoch * len(self.dataloader_train) + batch_idx)
-
- # if batch_idx == 2:
- # break
+ self.writer.add_scalar(f'rank{self.local_rank}_{key}/train', eval(var), self.batch_total)
+
+
pbar.close()
def _validate_epoch(self, epoch):
@@ -353,7 +353,10 @@
if (batch_idx+1) % self.log_interval == 0 or (batch_idx+1) == len(self.dataloader_val):
pbar.update(self.log_interval)
+ time_now = datetime.now()
+ time_now = time_now.strftime("%Y-%m-%d %H:%M:%S")
description = (
+ f"{time_now}, "
f"rank: {self.local_rank}, "
f"validation epoch: {epoch}/{self.max_epoch}, "
f"step: {batch_idx+1}/{len(self.dataloader_val)}, "
--
Gitblit v1.9.1