Merge pull request #867 from LittleChenCc/develop

update the results of TIMIT and Ted-ST

deepspeech/models/u2_st.py

@@ -113,7 +113,8 @@ class U2STBaseModel(nn.Layer):
                  asr_weight: float=0.0,
                  ignore_id: int=IGNORE_ID,
                  lsm_weight: float=0.0,
-                 length_normalized_loss: bool=False):
+                 length_normalized_loss: bool=False,
+                 **kwargs):
         assert 0.0 <= ctc_weight <= 1.0, ctc_weight
 
         super().__init__()
@@ -650,7 +651,7 @@ class U2STModel(U2STBaseModel):
                 odim=vocab_size,
                 enc_n_units=encoder.output_size(),
                 blank_id=0,
-                dropout_rate=model_conf['ctc_dropout_rate'],
+                dropout_rate=model_conf['ctc_dropoutrate'],
                 reduction=True,  # sum
                 batch_average=True,  # sum / batch_size
                 grad_norm_type=model_conf['ctc_grad_norm_type'])

examples/ted_en_zh/t0/README.md

@@ -6,5 +6,10 @@
 | Data Subset | Duration in Seconds |
 | --- | --- |
 | data/manifest.train | 0.942 ~ 60   |
-| data/manifest.dev   | 1.151 ~ 39   |  
+| data/manifest.dev   | 1.151 ~ 39   |
 | data/manifest.test  | 1.1 ~ 42.746 |
+
+## Transformer
+| Model | Params | Config | Char-BLEU |
+| --- | --- | --- | --- |
+| Transformer+ASR MTL | 50.26M | conf/transformer_joint_noam.yaml | 17.38 |

examples/timit/s1/README.md

@@ -1,3 +1,11 @@
 # TIMIT
 
-Results will be organized and updated soon.
+
+
+
+### Transformer
+| Model | Params | Config | Decode method          | PER    |
+| --- | --- | --- | --- | --- |
+| transformer | 5.17M | conf/transformer.yaml | attention              | 0.5531 |
+| transformer | 5.17M | conf/transformer.yaml | ctc_greedy_search      | 0.3922 |
+| transformer | 5.17M | conf/transformer.yaml | ctc_prefix_beam_search | 0.3768 |

examples/timit/s1/conf/transformer.yaml

@@ -3,12 +3,12 @@ data:
   train_manifest: data/manifest.train
   dev_manifest: data/manifest.dev
   test_manifest: data/manifest.test
-  min_input_len: 0.5  # second
+  min_input_len: 0.0  # second
-  max_input_len: 30.0 # second
+  max_input_len: 10.0 # second
   min_output_len: 0.0 # tokens
-  max_output_len: 400.0 # tokens
+  max_output_len: 150.0 # tokens
-  min_output_input_ratio: 0.05
+  min_output_input_ratio: 0.005
-  max_output_input_ratio: 100.0
+  max_output_input_ratio: 1000.0
 
 collator:
   vocab_filepath: data/vocab.txt
@@ -42,10 +42,10 @@ model:
     # encoder related
     encoder: transformer
     encoder_conf:
-        output_size: 256    # dimension of attention
+        output_size: 128    # dimension of attention
         attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
+        linear_units: 1024  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
+        num_blocks: 6      # the number of encoder blocks
         dropout_rate: 0.1
         positional_dropout_rate: 0.1
         attention_dropout_rate: 0.0
@@ -56,7 +56,7 @@ model:
     decoder: transformer
     decoder_conf:
         attention_heads: 4
-        linear_units: 2048
+        linear_units: 1024
         num_blocks: 6
         dropout_rate: 0.1
         positional_dropout_rate: 0.1
@@ -65,26 +65,26 @@ model:
 
     # hybrid CTC/attention
     model_conf:
-        ctc_weight: 0.3
+        ctc_weight: 0.5
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: batch
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
 
 training:
-  n_epoch: 120
+  n_epoch: 200
   accum_grad: 2
   global_grad_clip: 5.0
   optim: adam
   optim_conf:
-    lr: 0.002
+    lr: 0.004
     weight_decay: 1e-06
   scheduler: warmuplr     # pytorch v1.1.0+ required
   scheduler_conf:
-    warmup_steps: 400
+    warmup_steps: 2000
     lr_decay: 1.0
-  log_interval: 100
+  log_interval: 10
   checkpoint:
     kbest_n: 50
     latest_n: 5