default cmvn compute config; more log of grad clip; diff ds2 cmvn compute and conf; ds2 lr step by epoch;

deepspeech/exps/deepspeech2/model.py

@@ -43,13 +43,11 @@ class DeepSpeech2Trainer(Trainer):
 
     def train_batch(self, batch_index, batch_data, msg):
         start = time.time()
-
         loss = self.model(*batch_data)
         loss.backward()
         layer_tools.print_grads(self.model, print_func=None)
         self.optimizer.step()
         self.optimizer.clear_grad()
-
         iteration_time = time.time() - start
 
         losses_np = {

deepspeech/training/gradclip.py

@@ -31,7 +31,7 @@ class ClipGradByGlobalNormWithLog(paddle.nn.ClipGradByGlobalNorm):
     def _dygraph_clip(self, params_grads):
         params_and_grads = []
         sum_square_list = []
-        for p, g in params_grads:
+        for i, (p, g) in enumerate(params_grads):
             if g is None:
                 continue
             if getattr(p, 'need_clip', True) is False:
@@ -45,7 +45,8 @@ class ClipGradByGlobalNormWithLog(paddle.nn.ClipGradByGlobalNorm):
             sum_square_list.append(sum_square)
 
             # debug log
-            # logger.debug(f"Grad Before Clip: {p.name}: {float(sum_square.sqrt()) }")
+            if i < 10:
+                logger.debug(f"Grad Before Clip: {p.name}: {float(sum_square.sqrt()) }")
 
         # all parameters have been filterd out
         if len(sum_square_list) == 0:
@@ -62,7 +63,7 @@ class ClipGradByGlobalNormWithLog(paddle.nn.ClipGradByGlobalNorm):
         clip_var = layers.elementwise_div(
             x=max_global_norm,
             y=layers.elementwise_max(x=global_norm_var, y=max_global_norm))
-        for p, g in params_grads:
+        for i, (p, g) in enumerate(params_grads):
             if g is None:
                 continue
             if getattr(p, 'need_clip', True) is False:
@@ -72,8 +73,9 @@ class ClipGradByGlobalNormWithLog(paddle.nn.ClipGradByGlobalNorm):
             params_and_grads.append((p, new_grad))
 
             # debug log
-            # logger.debug(
-            #     f"Grad After Clip: {p.name}: {float(merge_grad.square().sum().sqrt())}"
-            # )
+            if i < 10:
+                logger.debug(
+                    f"Grad After Clip: {p.name}: {float(new_grad.square().sum().sqrt())}"
+                )
 
         return params_and_grads

deepspeech/training/trainer.py

@@ -226,6 +226,7 @@ class Trainer():
                               'lr': self.lr_scheduler()}, self.epoch)
 
             self.save(tag=self.epoch, infos={'val_loss': cv_loss})
+            # step lr every epoch
             self.lr_scheduler.step()
             self.new_epoch()
 
@@ -283,7 +284,6 @@ class Trainer():
         """
         # visualizer
         visualizer = SummaryWriter(logdir=str(self.output_dir))
-
         self.visualizer = visualizer
 
     @mp_tools.rank_zero_only
@@ -301,7 +301,6 @@ class Trainer():
         """
         raise NotImplementedError("train_batch should be implemented.")
 
-    @mp_tools.rank_zero_only
     @paddle.no_grad()
     def valid(self):
         """The validation. A subclass should implement this method.

deepspeech/utils/layer_tools.py

@@ -33,7 +33,7 @@ def summary(layer: nn.Layer, print_func=print):
     if print_func:
         num_elements = num_elements / 1024**2
         print_func(
-            f"Total parameters: {num_params}, {num_elements:.2f} M elements.")
+            f"Total parameters: {num_params}, {num_elements:.2f}M elements.")
 
 
 def print_grads(model, print_func=print):
@@ -57,7 +57,7 @@ def print_params(model, print_func=print):
             print_func(msg)
     if print_func:
         total = total / 1024**2
-        print_func(f"Total parameters: {num_params}, {total:.2f} M elements.")
+        print_func(f"Total parameters: {num_params}, {total:.2f}M elements.")
 
 
 def gradient_norm(layer: nn.Layer):

examples/aishell/s0/conf/deepspeech2.yaml

@@ -10,9 +10,9 @@ data:
   min_input_len: 0.0
   max_input_len: 27.0 # second
   min_output_len: 0.0
-  max_output_len: 400.0
-  min_output_input_ratio: 0.05
-  max_output_input_ratio: 10.0
+  max_output_len: .inf
+  min_output_input_ratio: 0.00
+  max_output_input_ratio: .inf
   specgram_type: linear
   target_sample_rate: 16000
   max_freq: None
@@ -41,7 +41,7 @@ training:
   lr: 2e-3
   lr_decay: 0.83
   weight_decay: 1e-06
-  global_grad_clip: 5.0
+  global_grad_clip: 3.0
   log_interval: 100
   
 decoding:

examples/aishell/s0/local/data.sh

@@ -32,7 +32,7 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     --unit_type="char" \
     --count_threshold=0 \
     --vocab_path="data/vocab.txt" \
-    --manifest_paths "data/manifest.train.raw"
+    --manifest_paths "data/manifest.train.raw" "data/manifest.dev.raw"
 
     if [ $? -ne 0 ]; then
         echo "Build vocabulary failed. Terminated."
@@ -51,8 +51,8 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     --stride_ms=10.0 \
     --window_ms=20.0 \
     --sample_rate=16000 \
-    --use_dB_normalization=False \
-    --num_samples=-1 \
+    --use_dB_normalization=True \
+    --num_samples=2000 \
     --num_workers=${num_workers} \
     --output_path="data/mean_std.json"
 

tests/mask_test.py

@@ -26,25 +26,27 @@ class TestU2Model(unittest.TestCase):
         paddle.set_device('cpu')
         self.lengths = paddle.to_tensor([5, 3, 2])
         self.masks = np.array([
-            [1, 1, 1, 1, 1],
-            [1, 1, 1, 0, 0],
-            [1, 1, 0, 0, 0],
+            [True, True, True, True, True],
+            [True, True, True, False, False],
+            [True, True, False, False, False],
         ])
         self.pad_masks = np.array([
-            [0, 0, 0, 0, 0],
-            [0, 0, 0, 1, 1],
-            [0, 0, 1, 1, 1],
+            [False, False, False, False, False],
+            [False, False, False, True, True],
+            [False, False, True, True, True],
         ])
 
     def test_sequence_mask(self):
-        res = sequence_mask(self.lengths)
+        res = sequence_mask(self.lengths, dtype='bool')
         self.assertSequenceEqual(res.numpy().tolist(), self.masks.tolist())
 
     def test_make_non_pad_mask(self):
         res = make_non_pad_mask(self.lengths)
-        res1 = sequence_mask(self.lengths)
+        res1 = sequence_mask(self.lengths, dtype='bool')
+        res2 = make_pad_mask(self.lengths).logical_not()
         self.assertSequenceEqual(res.numpy().tolist(), self.masks.tolist())
         self.assertSequenceEqual(res.numpy().tolist(), res1.numpy().tolist())
+        self.assertSequenceEqual(res.numpy().tolist(), res2.numpy().tolist())
 
     def test_make_pad_mask(self):
         res = make_pad_mask(self.lengths)

utils/compute_mean_std.py

@@ -24,7 +24,7 @@ from deepspeech.utils.utility import print_arguments
 parser = argparse.ArgumentParser(description=__doc__)
 add_arg = functools.partial(add_arguments, argparser=parser)
 # yapf: disable
-add_arg('num_samples',      int,    -1,    "# of samples to for statistics.")
+add_arg('num_samples',      int,    2000,    "# of samples to for statistics.")
 
 add_arg('specgram_type',    str,
         'linear',
@@ -35,7 +35,7 @@ add_arg('delta_delta', bool,  False, "Audio feature with delta delta.")
 add_arg('stride_ms', float, 10.0,  "stride length in ms.")
 add_arg('window_ms', float, 20.0,  "stride length in ms.")
 add_arg('sample_rate',  int, 16000,  "target sample rate.")
-add_arg('use_dB_normalization', bool, False, "do dB normalization.")
+add_arg('use_dB_normalization', bool, True, "do dB normalization.")
 add_arg('target_dB',   int, -20,  "target dB.")
 
 add_arg('manifest_path',    str,