ctc decoding weight 0.5 (#614)

* ctc decoding weight 0.5

* tiny decoding conf

* more label of mergify

* format doc

.mergify.yml

@@ -39,6 +39,18 @@ pull_request_rules:
     actions:
       label:
         remove: ["conflicts"]
+  - name: "auto add label=enhancement"
+    conditions:
+      - files~=^deepspeech/
+    actions:
+      label:
+        add: ["enhancement"]
+  - name: "auto add label=Example"
+    conditions:
+      - files~=^examples/
+    actions:
+      label:
+        add: ["Example"]
   - name: "auto add label=README"
     conditions:
       - files~=README.md

examples/aishell/s1/conf/conformer.yaml

@@ -74,13 +74,13 @@ model:
 
 
 training:
-  n_epoch: 300
+  n_epoch: 240 
   accum_grad: 2
   global_grad_clip: 5.0
   optim: adam
   optim_conf:
     lr: 0.002
-    weight_decay: 1e-06
+    weight_decay: 1e-6
   scheduler: warmuplr     # pytorch v1.1.0+ required
   scheduler_conf:
     warmup_steps: 25000
@@ -99,7 +99,7 @@ decoding:
   cutoff_prob: 1.0
   cutoff_top_n: 0
   num_proc_bsearch: 8
-  ctc_weight: 0.0 # ctc weight for attention rescoring decode mode.
+  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
   decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
       # <0: for decoding, use full chunk.
       # >0: for decoding, use fixed chunk size as set.

examples/librispeech/s1/conf/chunk_confermer.yaml

@@ -104,7 +104,7 @@ decoding:
   cutoff_prob: 1.0
   cutoff_top_n: 0
   num_proc_bsearch: 8
-  ctc_weight: 0.0 # ctc weight for attention rescoring decode mode.
+  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
   decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
       # <0: for decoding, use full chunk.
       # >0: for decoding, use fixed chunk size as set.

examples/librispeech/s1/conf/chunk_transformer.yaml

@@ -97,7 +97,7 @@ decoding:
   cutoff_prob: 1.0
   cutoff_top_n: 0
   num_proc_bsearch: 8
-  ctc_weight: 0.0 # ctc weight for attention rescoring decode mode.
+  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
   decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
       # <0: for decoding, use full chunk.
       # >0: for decoding, use fixed chunk size as set.

examples/librispeech/s1/conf/conformer.yaml

@@ -8,7 +8,7 @@ data:
   spm_model_prefix: 'data/bpe_unigram_5000'
   mean_std_filepath: ""
   augmentation_config: conf/augmentation.json
-  batch_size: 64
+  batch_size: 16
   min_input_len: 0.5  # seconds
   max_input_len: 20.0 # seconds
   min_output_len: 0.0 # tokens
@@ -76,7 +76,7 @@ model:
 
 training:
   n_epoch: 120
-  accum_grad: 2
+  accum_grad: 8
   global_grad_clip: 5.0
   optim: adam
   optim_conf:
@@ -100,7 +100,7 @@ decoding:
   cutoff_prob: 1.0
   cutoff_top_n: 0
   num_proc_bsearch: 8
-  ctc_weight: 0.0 # ctc weight for attention rescoring decode mode.
+  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
   decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
       # <0: for decoding, use full chunk.
       # >0: for decoding, use fixed chunk size as set.

examples/librispeech/s1/conf/transformer.yaml

@@ -95,7 +95,7 @@ decoding:
   cutoff_prob: 1.0
   cutoff_top_n: 0
   num_proc_bsearch: 8
-  ctc_weight: 0.0 # ctc weight for attention rescoring decode mode.
+  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
   decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
       # <0: for decoding, use full chunk.
       # >0: for decoding, use fixed chunk size as set.

examples/tiny/s1/conf/chunk_confermer.yaml

@@ -104,7 +104,7 @@ decoding:
   cutoff_prob: 1.0
   cutoff_top_n: 0
   num_proc_bsearch: 8
-  ctc_weight: 0.0 # ctc weight for attention rescoring decode mode.
+  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
   decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
       # <0: for decoding, use full chunk.
       # >0: for decoding, use fixed chunk size as set.

examples/tiny/s1/conf/chunk_transformer.yaml

@@ -97,7 +97,7 @@ decoding:
   cutoff_prob: 1.0
   cutoff_top_n: 0
   num_proc_bsearch: 8
-  ctc_weight: 0.0 # ctc weight for attention rescoring decode mode.
+  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
   decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
       # <0: for decoding, use full chunk.
       # >0: for decoding, use fixed chunk size as set.

examples/tiny/s1/conf/conformer.yaml

@@ -100,7 +100,7 @@ decoding:
   cutoff_prob: 1.0
   cutoff_top_n: 0
   num_proc_bsearch: 8
-  ctc_weight: 0.0 # ctc weight for attention rescoring decode mode.
+  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
   decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
       # <0: for decoding, use full chunk.
       # >0: for decoding, use fixed chunk size as set.

examples/tiny/s1/conf/transformer.yaml

@@ -95,7 +95,7 @@ decoding:
   cutoff_prob: 1.0
   cutoff_top_n: 0
   num_proc_bsearch: 8
-  ctc_weight: 0.0 # ctc weight for attention rescoring decode mode.
+  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
   decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
       # <0: for decoding, use full chunk.
       # >0: for decoding, use fixed chunk size as set.