CIF Related

examples/aishell/cif/README.md

@@ -0,0 +1,53 @@
+# Performance Record
+
+## Conformer CIF DecoderSAN + PredictV1 Result
+
+* Feature info: using fbank feature, dither, cmvn, online speed perturb
+* Training info: lr 0.002, batch size 16, 4 gpu, acc_grad 4, 240 epochs, dither 0.1
+* Decoding info: ctc_weight 0.5, average_num 20
+
+| decoding mode          | Dev CER | Test CER |
+|------------------------| --- | ---- |
+| ctc greedy search      | 4.65 | 5.24 |
+| ctc prefix beam search | 4.65 | 5.24 |
+| cif greedy search      | 4.41 | 4.92 |
+| cif beam search        | 4.35 | 4.86 |
+
+## Conformer CIF DecoderSAN + PredictV2 Result
+
+* Feature info: using fbank feature, dither, cmvn, online speed perturb
+* Training info: lr 0.002, batch size 16, 4 gpu, acc_grad 4, 240 epochs, dither 0.1
+* Decoding info: ctc_weight 0.5, average_num 20
+
+| decoding mode          | Dev CER | Test CER |
+|------------------------| ---- | ---- |
+| ctc greedy search      | 5.35 | 5.98 |
+| ctc prefix beam search | 5.35 | 5.98 |
+| cif greedy search      | 4.77 | 5.32 |
+| cif beam search        | 4.71 | 5.25 |
+
+## Conformer CIF DecoderSANM + PredictV1 Result
+
+* Feature info: using fbank feature, dither, cmvn, online speed perturb
+* Training info: lr 0.002, batch size 16, 4 gpu, acc_grad 4, 240 epochs, dither 0.1
+* Decoding info: ctc_weight 0.5, average_num 20
+
+| decoding mode          | Dev CER | Test CER |
+|------------------------|------| ---- |
+| ctc greedy search      | 4.86 | 5.46 |
+| ctc prefix beam search | 4.86 | 5.46 |
+| cif greedy search      | 4.34 | 4.81 |
+| cif beam search        | 4.27 | 4.75 |
+
+## Conformer CIF DecoderSANM + PredictV2 Result
+
+* Feature info: using fbank feature, dither, cmvn, online speed perturb
+* Training info: lr 0.002, batch size 16, 4 gpu, acc_grad 4, 240 epochs, dither 0.1
+* Decoding info: ctc_weight 0.5, average_num 20
+
+| decoding mode          | Dev CER | Test CER |
+|------------------------| ---- | ---- |
+| ctc greedy search      | 5.60 | 6.15 |
+| ctc prefix beam search | 5.60 | 6.14 |
+| cif greedy search      | 4.85 | 5.29 |
+| cif beam search        | 4.77 | 5.16 |

examples/aishell/cif/conf/train_cif_conformer_prev1.yaml

@@ -0,0 +1,86 @@
+# network architecture
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
+    cnn_module_kernel: 15
+    use_cnn_module: True
+    activation_type: 'swish'
+    pos_enc_layer_type: 'rel_pos'
+    selfattention_layer_type: 'rel_selfattn'
+
+# decoder related
+decoder: cif_decoder_san
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
+
+cif_predictor: predictor_v1
+cif_predictor_conf:
+    idim: 256
+    threshold: 1.0
+    l_order: 1
+    r_order: 1
+    tail_threshold: 0.45
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+    predictor_weight: 1.0
+
+dataset_conf:
+    filter_conf:
+        max_length: 40960
+        min_length: 0
+        token_max_length: 200
+        token_min_length: 1
+    resample_conf:
+        resample_rate: 16000
+    speed_perturb: true
+    fbank_conf:
+        num_mel_bins: 80
+        frame_shift: 10
+        frame_length: 25
+        dither: 0.1
+    spec_aug: true
+    spec_aug_conf:
+        num_t_mask: 2
+        num_f_mask: 2
+        max_t: 50
+        max_f: 10
+    shuffle: true
+    shuffle_conf:
+        shuffle_size: 1500
+    sort: true
+    sort_conf:
+        sort_size: 500  # sort_size should be less than shuffle_size
+    batch_conf:
+        batch_type: 'static' # static or dynamic
+        batch_size: 16
+
+grad_clip: 5
+accum_grad: 4
+max_epoch: 240
+log_interval: 100
+
+optim: adam
+optim_conf:
+    lr: 0.002
+scheduler: warmuplr     # pytorch v1.1.0+ required
+scheduler_conf:
+    warmup_steps: 25000

examples/aishell/cif/conf/train_cif_conformer_prev2.yaml

@@ -0,0 +1,86 @@
+# network architecture
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
+    cnn_module_kernel: 15
+    use_cnn_module: True
+    activation_type: 'swish'
+    pos_enc_layer_type: 'rel_pos'
+    selfattention_layer_type: 'rel_selfattn'
+
+# decoder related
+decoder: cif_decoder_san
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
+
+cif_predictor: predictor_v2
+cif_predictor_conf:
+    idim: 256
+    threshold: 1.0
+    l_order: 1
+    r_order: 1
+    tail_threshold: 0.45
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+    predictor_weight: 1.0
+
+dataset_conf:
+    filter_conf:
+        max_length: 40960
+        min_length: 0
+        token_max_length: 200
+        token_min_length: 1
+    resample_conf:
+        resample_rate: 16000
+    speed_perturb: true
+    fbank_conf:
+        num_mel_bins: 80
+        frame_shift: 10
+        frame_length: 25
+        dither: 0.1
+    spec_aug: true
+    spec_aug_conf:
+        num_t_mask: 2
+        num_f_mask: 2
+        max_t: 50
+        max_f: 10
+    shuffle: true
+    shuffle_conf:
+        shuffle_size: 1500
+    sort: true
+    sort_conf:
+        sort_size: 500  # sort_size should be less than shuffle_size
+    batch_conf:
+        batch_type: 'static' # static or dynamic
+        batch_size: 16
+
+grad_clip: 5
+accum_grad: 4
+max_epoch: 240
+log_interval: 100
+
+optim: adam
+optim_conf:
+    lr: 0.002
+scheduler: warmuplr     # pytorch v1.1.0+ required
+scheduler_conf:
+    warmup_steps: 25000

examples/aishell/cif/conf/train_cif_conformer_sanm_prev1.yaml

@@ -0,0 +1,87 @@
+# network architecture
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
+    cnn_module_kernel: 15
+    use_cnn_module: True
+    activation_type: 'swish'
+    pos_enc_layer_type: 'rel_pos'
+    selfattention_layer_type: 'rel_selfattn'
+
+# decoder related
+decoder: cif_decoder_sanm
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
+    input_layer: linear # Options (embed, linear)
+
+cif_predictor: predictor_v1
+cif_predictor_conf:
+    idim: 256
+    threshold: 1.0
+    l_order: 1
+    r_order: 1
+    tail_threshold: 0.45
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+    predictor_weight: 1.0
+
+dataset_conf:
+    filter_conf:
+        max_length: 40960
+        min_length: 0
+        token_max_length: 200
+        token_min_length: 1
+    resample_conf:
+        resample_rate: 16000
+    speed_perturb: true
+    fbank_conf:
+        num_mel_bins: 80
+        frame_shift: 10
+        frame_length: 25
+        dither: 0.1
+    spec_aug: true
+    spec_aug_conf:
+        num_t_mask: 2
+        num_f_mask: 2
+        max_t: 50
+        max_f: 10
+    shuffle: true
+    shuffle_conf:
+        shuffle_size: 1500
+    sort: true
+    sort_conf:
+        sort_size: 500  # sort_size should be less than shuffle_size
+    batch_conf:
+        batch_type: 'static' # static or dynamic
+        batch_size: 16
+
+grad_clip: 5
+accum_grad: 4
+max_epoch: 240
+log_interval: 100
+
+optim: adam
+optim_conf:
+    lr: 0.002
+scheduler: warmuplr     # pytorch v1.1.0+ required
+scheduler_conf:
+    warmup_steps: 25000