Merge pull request #1273 from zh794390558/batch_sampler

[s2t] Fix Batch sampler set epoch

paddlespeech/s2t/exps/u2/model.py

@@ -238,7 +238,9 @@ class U2Trainer(Trainer):
                 preprocess_conf=config.preprocess_config,
                 n_iter_processes=config.num_workers,
                 subsampling_factor=1,
-                num_encs=1)
+                num_encs=1,
+                dist_sampler=False,
+                shortest_first=False)
 
             self.valid_loader = BatchDataLoader(
                 json_file=config.dev_manifest,
@@ -257,7 +259,9 @@ class U2Trainer(Trainer):
                 preprocess_conf=config.preprocess_config,
                 n_iter_processes=config.num_workers,
                 subsampling_factor=1,
-                num_encs=1)
+                num_encs=1,
+                dist_sampler=False,
+                shortest_first=False)
             logger.info("Setup train/valid Dataloader!")
         else:
             decode_batch_size = config.get('decode', dict()).get(

paddlespeech/s2t/io/dataloader.py

@@ -78,7 +78,8 @@ class BatchDataLoader():
                  load_aux_input: bool=False,
                  load_aux_output: bool=False,
                  num_encs: int=1,
-                 dist_sampler: bool=False):
+                 dist_sampler: bool=False,
+                 shortest_first: bool=False):
         self.json_file = json_file
         self.train_mode = train_mode
         self.use_sortagrad = sortagrad == -1 or sortagrad > 0
@@ -97,6 +98,7 @@ class BatchDataLoader():
         self.load_aux_input = load_aux_input
         self.load_aux_output = load_aux_output
         self.dist_sampler = dist_sampler
+        self.shortest_first = shortest_first
 
         # read json data
         with jsonlines.open(json_file, 'r') as reader:
@@ -113,7 +115,7 @@ class BatchDataLoader():
             maxlen_out,
             minibatches,  # for debug
             min_batch_size=mini_batch_size,
-            shortest_first=self.use_sortagrad,
+            shortest_first=self.shortest_first or self.use_sortagrad,
             count=batch_count,
             batch_bins=batch_bins,
             batch_frames_in=batch_frames_in,
@@ -149,13 +151,13 @@ class BatchDataLoader():
                                         self.reader)
 
         if self.dist_sampler:
-            self.sampler = DistributedBatchSampler(
+            self.batch_sampler = DistributedBatchSampler(
                 dataset=self.dataset,
                 batch_size=1,
                 shuffle=not self.use_sortagrad if self.train_mode else False,
                 drop_last=False, )
         else:
-            self.sampler = BatchSampler(
+            self.batch_sampler = BatchSampler(
                 dataset=self.dataset,
                 batch_size=1,
                 shuffle=not self.use_sortagrad if self.train_mode else False,
@@ -163,7 +165,7 @@ class BatchDataLoader():
 
         self.dataloader = DataLoader(
             dataset=self.dataset,
-            batch_sampler=self.sampler,
+            batch_sampler=self.batch_sampler,
             collate_fn=batch_collate,
             num_workers=self.n_iter_processes, )
 
@@ -194,5 +196,6 @@ class BatchDataLoader():
         echo += f"load_aux_input: {self.load_aux_input}, "
         echo += f"load_aux_output: {self.load_aux_output}, "
         echo += f"dist_sampler: {self.dist_sampler}, "
+        echo += f"shortest_first: {self.shortest_first}, "
         echo += f"file: {self.json_file}"
         return echo

paddlespeech/s2t/modules/ctc.py

@@ -39,10 +39,6 @@ except ImportError:
     except Exception as e:
         logger.info("paddlespeech_ctcdecoders not installed!")
 
-#try:
-#except Exception as e:
-#    logger.info("ctcdecoder not installed!")
-
 __all__ = ['CTCDecoder']
 
 

paddlespeech/s2t/training/scheduler.py

@@ -67,9 +67,10 @@ class WarmupLR(LRScheduler):
         super().__init__(learning_rate, last_epoch, verbose)
 
     def __repr__(self):
-        return f"{self.__class__.__name__}(warmup_steps={self.warmup_steps})"
+        return f"{self.__class__.__name__}(warmup_steps={self.warmup_steps}, lr={self.base_lr}, last_epoch={self.last_epoch})"
 
     def get_lr(self):
+        # self.last_epoch start from zero
         step_num = self.last_epoch + 1
         return self.base_lr * self.warmup_steps**0.5 * min(
             step_num**-0.5, step_num * self.warmup_steps**-1.5)

paddlespeech/s2t/training/trainer.py

@@ -222,7 +222,7 @@ class Trainer():
             batch_sampler = self.train_loader.batch_sampler
             if isinstance(batch_sampler, paddle.io.DistributedBatchSampler):
                 logger.debug(
-                    f"train_loader.batch_sample set epoch: {self.epoch}")
+                    f"train_loader.batch_sample.set_epoch: {self.epoch}")
                 batch_sampler.set_epoch(self.epoch)
 
     def before_train(self):