From 664fd7abc82def968310b9f202a901ac675b901d Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期四, 09 五月 2024 17:54:31 +0800
Subject: [PATCH] total_time/accum_grad
---
funasr/train_utils/trainer.py | 16 ++++++++++------
1 files changed, 10 insertions(+), 6 deletions(-)
diff --git a/funasr/train_utils/trainer.py b/funasr/train_utils/trainer.py
index dd0ac7a..d46a21c 100644
--- a/funasr/train_utils/trainer.py
+++ b/funasr/train_utils/trainer.py
@@ -308,6 +308,7 @@
checkpoint["step_in_epoch"] if "step_in_epoch" in checkpoint else 0
)
self.step_in_epoch = 0 if self.step_in_epoch is None else self.step_in_epoch
+ print(checkpoint["train_acc_avg"])
self.train_acc_avg = (
checkpoint["train_acc_avg"] if "train_acc_avg" in checkpoint else 0
)
@@ -381,8 +382,6 @@
):
torch.cuda.empty_cache()
- time3 = time.perf_counter()
- speed_stats["forward_time"] = f"{time3 - time2:0.3f}"
loss, stats, weight = retval
stats = {k: v for k, v in stats.items() if v is not None}
if self.use_ddp or self.use_fsdp:
@@ -399,12 +398,15 @@
loss *= self.world_size
# Scale the loss since we're not updating for every mini-batch
loss = loss / accum_grad
+
+ time3 = time.perf_counter()
+ speed_stats["forward_time"] = f"{time3 - time2:0.3f}"
if self.use_fp16:
scaler.scale(loss).backward()
else:
loss.backward()
time4 = time.perf_counter()
- speed_stats["backward_time"] = f"{time4 - time3:0.3f}"
+ speed_stats["backward_and_AllReaduce_time"] = f"{time4 - time3:0.3f}"
self.train_loss_avg = (
self.train_loss_avg * (self.step_in_epoch - 1) + loss.detach().cpu().item()
@@ -453,7 +455,7 @@
scheduler.step()
# Clear gradients for the next accumulation stage
optim.zero_grad(set_to_none=True)
- total_time = f"{time.perf_counter() - time5:0.3f}"
+ total_time = f"{(time.perf_counter() - time5)/accum_grad:0.3f}"
time5 = time.perf_counter()
speed_stats["optim_time"] = f"{time5 - time4:0.3f}"
@@ -464,7 +466,8 @@
batch_num_epoch = len(dataloader_train)
self.log(
epoch,
- batch_idx + kwargs.get("start_step", 0),
+ batch_idx,
+ log_step=batch_idx + kwargs.get("start_step", 0),
step_in_epoch=self.step_in_epoch,
batch_num_epoch=batch_num_epoch,
lr=lr,
@@ -633,11 +636,12 @@
tag="train",
data_split_i=0,
data_split_num=1,
+ log_step=None,
**kwargs,
):
if (batch_idx + 1) % self.log_interval == 0:
-
+ batch_idx = log_step if log_step is not None else batch_idx
gpu_info = (
"GPU, memory: usage: {:.3f} GB, "
"peak: {:.3f} GB, "
--
Gitblit v1.9.1