From 497edf4c9d6c1565a4bcf1a3edfcd47ffec8c10d Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期三, 21 二月 2024 11:30:59 +0800
Subject: [PATCH] bugfix
---
funasr/train_utils/trainer.py | 29 ++++++++++++++++++++++-------
1 files changed, 22 insertions(+), 7 deletions(-)
diff --git a/funasr/train_utils/trainer.py b/funasr/train_utils/trainer.py
index 10f7f80..4b85a66 100644
--- a/funasr/train_utils/trainer.py
+++ b/funasr/train_utils/trainer.py
@@ -128,7 +128,20 @@
if os.path.isfile(ckpt):
checkpoint = torch.load(ckpt)
self.start_epoch = checkpoint['epoch'] + 1
- self.model.load_state_dict(checkpoint['state_dict'])
+ # self.model.load_state_dict(checkpoint['state_dict'])
+ src_state = checkpoint['state_dict']
+ dst_state = self.model.state_dict()
+ for k in dst_state.keys():
+ if not k.startswith("module.") and "module."+k in src_state.keys():
+ k_ddp = "module."+k
+ else:
+ k_ddp = k
+ if k_ddp in src_state.keys():
+ dst_state[k] = src_state[k_ddp]
+ else:
+ print(f"Miss key in ckpt: model: {k}, ckpt: {k_ddp}")
+
+ self.model.load_state_dict(dst_state)
self.optim.load_state_dict(checkpoint['optimizer'])
self.scheduler.load_state_dict(checkpoint['scheduler'])
print(f"Checkpoint loaded successfully from '{ckpt}'")
@@ -188,7 +201,7 @@
epoch (int): The current epoch number.
"""
self.model.train()
- pbar = tqdm(colour="blue", desc=f"Training Epoch: {epoch + 1}", total=len(self.dataloader_train),
+ pbar = tqdm(colour="blue", desc=f"rank: {self.local_rank}, Training Epoch: {epoch + 1}", total=len(self.dataloader_train),
dynamic_ncols=True)
# Set the number of steps for gradient accumulation
@@ -273,12 +286,14 @@
torch.cuda.memory_reserved()/1024/1024/1024,
torch.cuda.max_memory_reserved()/1024/1024/1024,
)
+ lr = self.scheduler.get_last_lr()[0]
description = (
f"rank: {self.local_rank}, "
f"epoch: {epoch}/{self.max_epoch}, "
- f"step: {batch_idx}/{len(self.dataloader_train)}, total: {self.batch_total}, "
+ f"step: {batch_idx+1}/{len(self.dataloader_train)}, total: {self.batch_total}, "
f"(loss: {loss.detach().cpu().item():.3f}), "
- f"{[(k, round(v.cpu().item(), 3)) for k, v in stats.items()]}"
+ f"(lr: {lr:.3e}), "
+ f"{[(k, round(v.cpu().item(), 3)) for k, v in stats.items()]}, "
f"{speed_stats}, "
f"{gpu_info}"
)
@@ -307,7 +322,7 @@
"""
self.model.eval()
with torch.no_grad():
- pbar = tqdm(colour="red", desc=f"Training Epoch: {epoch + 1}", total=len(self.dataloader_val),
+ pbar = tqdm(colour="red", desc=f"rank: {self.local_rank}, Validation Epoch: {epoch + 1}", total=len(self.dataloader_val),
dynamic_ncols=True)
speed_stats = {}
time5 = time.perf_counter()
@@ -341,9 +356,9 @@
description = (
f"rank: {self.local_rank}, "
f"validation epoch: {epoch}/{self.max_epoch}, "
- f"step: {batch_idx}/{len(self.dataloader_val)}, "
+ f"step: {batch_idx+1}/{len(self.dataloader_val)}, "
f"(loss: {loss.detach().cpu().item():.3f}), "
- f"{[(k, round(v.cpu().item(), 3)) for k, v in stats.items()]}"
+ f"{[(k, round(v.cpu().item(), 3)) for k, v in stats.items()]}, "
f"{speed_stats}, "
)
pbar.set_description(description)
--
Gitblit v1.9.1