From 5d74e107fc5696b70e75003c278f8babd17161e8 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期日, 24 三月 2024 00:58:56 +0800
Subject: [PATCH] update

---
 funasr/train_utils/trainer.py |    9 ++++++++-
 1 files changed, 8 insertions(+), 1 deletions(-)

diff --git a/funasr/train_utils/trainer.py b/funasr/train_utils/trainer.py
index 77eee60..23c18d9 100644
--- a/funasr/train_utils/trainer.py
+++ b/funasr/train_utils/trainer.py
@@ -249,6 +249,9 @@
         speed_stats = {}
         time5 = time.perf_counter()
         iterator_stop = torch.tensor(0).to(self.device)
+        dist.barrier()
+        print(f"before iter, iterator_stop: {iterator_stop}\n")
+        dataloader_train.batch_sampler.set_epoch(epoch)
         for batch_idx, batch in enumerate(dataloader_train):
             if self.use_ddp or self.use_fsdp:
                 dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)
@@ -392,9 +395,13 @@
             speed_stats = {}
             time5 = time.perf_counter()
             iterator_stop = torch.tensor(0).to(self.device)
+            dist.barrier()
+            print(f"before iter, iterator_stop: {iterator_stop}\n")
             for batch_idx, batch in enumerate(dataloader_val):
                 if self.use_ddp or self.use_fsdp:
                     dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)
+                    if epoch >= 1:
+                        print(f"iterator_stop: {iterator_stop}\n")
                     if iterator_stop > 0:
                         break
                 time1 = time.perf_counter()
@@ -410,7 +417,7 @@
                     # Apply weighted averaging for loss and stats
                     loss = (loss * weight.type(loss.dtype)).sum()
                     # if distributed, this method can also apply all_reduce()
-                    stats, weight = recursive_average(stats, weight, distributed=True)
+                    # stats, weight = recursive_average(stats, weight, distributed=True)
                     if self.use_ddp or self.use_fsdp:
                         dist.all_reduce(weight, op=dist.ReduceOp.SUM)
                     # Now weight is summation over all workers

--
Gitblit v1.9.1