From 6427c834dfd97b1f05c6659cdc7ccf010bf82fe1 Mon Sep 17 00:00:00 2001
From: 嘉渊 <wangjiaming.wjm@alibaba-inc.com>
Date: 星期一, 24 四月 2023 19:50:07 +0800
Subject: [PATCH] update

---
 funasr/train/trainer.py |   33 +++++++++++++++++++++++++++------
 1 files changed, 27 insertions(+), 6 deletions(-)

diff --git a/funasr/train/trainer.py b/funasr/train/trainer.py
index efe2009..4f2e28c 100644
--- a/funasr/train/trainer.py
+++ b/funasr/train/trainer.py
@@ -39,7 +39,7 @@
 from funasr.torch_utils.device_funcs import to_device
 from funasr.torch_utils.recursive_op import recursive_average
 from funasr.torch_utils.set_all_random_seed import set_all_random_seed
-from funasr.train.abs_espnet_model import AbsESPnetModel
+from funasr.models.base_model import FunASRModel
 from funasr.train.distributed_utils import DistributedOption
 from funasr.train.reporter import Reporter
 from funasr.train.reporter import SubReporter
@@ -94,7 +94,7 @@
     wandb_model_log_interval: int
     use_pai: bool
     oss_bucket: Union[oss2.Bucket, None]
-
+    batch_interval: int
 
 class Trainer:
     """Trainer having a optimizer.
@@ -165,7 +165,7 @@
     @classmethod
     def run(
         cls,
-        model: AbsESPnetModel,
+        model: FunASRModel,
         optimizers: Sequence[torch.optim.Optimizer],
         schedulers: Sequence[Optional[AbsScheduler]],
         train_iter_factory: AbsIterFactory,
@@ -186,7 +186,10 @@
                 logging.warning("No keep_nbest_models is given. Change to [1]")
                 trainer_options.keep_nbest_models = [1]
             keep_nbest_models = trainer_options.keep_nbest_models
-
+     
+        #assert batch_interval is set and >0
+        assert trainer_options.batch_interval > 0
+ 
         output_dir = Path(trainer_options.output_dir)
         reporter = Reporter()
         if trainer_options.use_amp:
@@ -560,13 +563,31 @@
         # [For distributed] Because iteration counts are not always equals between
         # processes, send stop-flag to the other processes if iterator is finished
         iterator_stop = torch.tensor(0).to("cuda" if ngpu > 0 else "cpu")
-
+        
+        #get the rank
+        rank = distributed_option.dist_rank
+        #get the num batch updates
+        num_batch_updates = 0
+        #ouput dir
+        output_dir = Path(options.output_dir)
+        #batch interval
+        batch_interval = options.batch_interval       
+        assert batch_interval > 0
+ 
         start_time = time.perf_counter()
         for iiter, (_, batch) in enumerate(
             reporter.measure_iter_time(iterator, "iter_time"), 1
         ):
             assert isinstance(batch, dict), type(batch)
 
+            if rank == 0:
+                if hasattr(model, "num_updates") or (hasattr(model, "module") and hasattr(model.module, "num_updates")):
+                    num_batch_updates = model.get_num_updates() if hasattr(model,"num_updates") else model.module.get_num_updates()
+                if (num_batch_updates%batch_interval == 0) and (options.oss_bucket is not None) and options.use_pai:
+                    buffer = BytesIO()
+                    torch.save(model.state_dict(), buffer)
+                    options.oss_bucket.put_object(os.path.join(output_dir, f"{num_batch_updates}batch.pth"), buffer.getvalue())
+ 
             if distributed:
                 torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
                 if iterator_stop > 0:
@@ -811,4 +832,4 @@
         else:
             if distributed:
                 iterator_stop.fill_(1)
-                torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
\ No newline at end of file
+                torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)

--
Gitblit v1.9.1