From 9be30f99dd09cfe0de929266ec43c1b95abb6d96 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 10 五月 2024 10:16:28 +0800
Subject: [PATCH] update avg slice
---
funasr/train_utils/trainer.py | 8 +++++---
1 files changed, 5 insertions(+), 3 deletions(-)
diff --git a/funasr/train_utils/trainer.py b/funasr/train_utils/trainer.py
index d46a21c..33dd351 100644
--- a/funasr/train_utils/trainer.py
+++ b/funasr/train_utils/trainer.py
@@ -396,6 +396,7 @@
# Multiply world_size because DistributedDataParallel
# automatically normalizes the gradient by world_size.
loss *= self.world_size
+ # loss *= self.world_size
# Scale the loss since we're not updating for every mini-batch
loss = loss / accum_grad
@@ -457,6 +458,7 @@
optim.zero_grad(set_to_none=True)
total_time = f"{(time.perf_counter() - time5)/accum_grad:0.3f}"
time5 = time.perf_counter()
+
speed_stats["optim_time"] = f"{time5 - time4:0.3f}"
speed_stats["total_time"] = total_time
@@ -663,9 +665,9 @@
f"data_slice: {data_split_i}/{data_split_num}, "
f"step_in_slice: {batch_idx + 1}/{batch_num_epoch}, step_in_epoch: {step_in_epoch}, total step: {self.batch_total}, "
f"(loss_avg_rank: {loss:.3f}), "
- f"(loss_avg_epoch: {loss_avg_epoch:.3f}), "
- f"(ppl_avg_epoch: {math.exp(loss_avg_epoch):.3e}), "
- f"(acc_avg_epoch: {acc_avg_epoch:.3f}), "
+ f"(loss_avg_slice: {loss_avg_epoch:.3f}), "
+ f"(ppl_avg_slice: {math.exp(loss_avg_epoch):.3e}), "
+ f"(acc_avg_slice: {acc_avg_epoch:.3f}), "
f"(lr: {lr:.3e}), "
f"{[(k, round(v.detach().cpu().item(), 3)) for k, v in stats.items()]}, "
f"{speed_stats}, "
--
Gitblit v1.9.1