与原论文一致。
| baseline | |
|---|---|
| log file | /home/panxie/workspace/sign-lang/baseline/log/reimp-conv/train_seed8_log.txt, train-2_seed8_log.txt |
| best wer | epoch26,wer:29.2 |
| lr schedule | halve at 40/60 epoch |
| seed | 8 |
| model | baseline |
|---|---|
| log file | /home/panxie/workspace/sign-lang/unlikeli_ctc/log/reimp-conv/train_origin_log.txt, train_origin_2_log.txt |
| best wer | epoch: 26, wer:27.2 |
| lr schedule | halve at 40/60 epoch |
| baseline | |
|---|---|
| log file | train_video_len-2_seed8_log.txt |
| best wer | epoch26,wer:27.5 |
| lr schedule | halve at 40/60 epoch |
| seed | 8 |
| len_video的计算有问题。。。改进之后总算能得到不错的效果了。 |
| model | full_conv_v3 |
|---|---|
| module | full_conv + Encoder(1 layer) |
| log file | train_v3_pre0.00001_attn0.0001_seed8_log.txt |
| best wer | epoch: 6, wer: 27.7 |
| local attn layer | 1 layer |
| pretrain | load backbone and other module. and their lr is 1e-5. lr of the attention layer is 1e-4. |
可以试试 from scratch 的情况
full_model_v5: full_conv + dynamic framing + rnn
| model | full_model_v5, with residual connection |
|---|---|
| log file | train_v5_load_backbone_2_seed8_log.txt, train_v5_load_backbone_2-2_seed8_log.txt |
| best wer | epoch16, 29.4 |
| pretrain | load backbone and freeze. |
| model | full_model_v5, with residual connection |
|---|---|
| module | full_conv + dynamic framing + rnn |
| log file | train_v5_scratch_seed8_log.txt |
| best wer | epoch: 69, 28.0 |
| pretrain | from scratch |
| model | full_model_v5, without residual connection |
|---|---|
| log file | train_v5_scratch_no_residual_seed8_log.txt |
| best wer | epoch: 25, 36.2. Abanbon! |
| pretrain | from scratch |
| model | full_model_v5, with residual connection, len_video |
|---|---|
| module | full_conv + dynamic framing + rnn |
| log file | train_v5_video_len_seed8_log.txt |
| best wer | epoch: 59, 26.3 |
| pretrain | from scratch |
| model | full_conv_v6, random sample after framing |
|---|---|
| module | TemporalAttention4 + conv1d + conv1d(no pooling) |
| log file | train_v6_dev_uniform_seed8_log.txt |
| best wer | epoch 86, wer: 29.9 |
| model | full_conv_v6 + residual, random sample after framing |
|---|---|
| module | TemporalAttention4 + conv1d + conv1d(no pooling) |
| log file | train_v6_dev_uniform_residual_seed8_log.txt |
| best wer | epoch 17, wer: 38.7, Abanbon! overfitting. |
| model | full_conv_v7, full_conv_v6 + tcn |
|---|---|
| module | TemporalAttention4 + tcn + tcn(no pooling) |
| log file | .txt |
| best wer | epoch , wer: |
| model | full_conv_v8, full_conv_v7 + hash_map |
|---|---|
| module | TemporalAttention4 + tcn + tcn(no pooling) |
| log file | train_v8_seed8_log.txt |
| best wer | epoch 16, wer: 44.2. Abanbon! |