Upload model
Browse files- config.json +5 -0
- hf_wrapper.py +1974 -0
- pytorch_model.bin +2 -2
config.json
CHANGED
|
@@ -1,8 +1,13 @@
|
|
| 1 |
{
|
|
|
|
| 2 |
"architectures": [
|
| 3 |
"Effb2TrmCaptioningModel"
|
| 4 |
],
|
| 5 |
"attn_emb_dim": 1408,
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
"decoder_dropout": 0.2,
|
| 7 |
"decoder_emb_dim": 256,
|
| 8 |
"decoder_n_layers": 2,
|
|
|
|
| 1 |
{
|
| 2 |
+
"_name_or_path": "/mnt/cloudstorfs/sjtu_home/xuenan.xu/hf_cache/hub/models--wsntxxn--effb2-trm-clotho-captioning/snapshots/a6295f4d7e0f2314bd3fc02b03512814c63fc132/",
|
| 3 |
"architectures": [
|
| 4 |
"Effb2TrmCaptioningModel"
|
| 5 |
],
|
| 6 |
"attn_emb_dim": 1408,
|
| 7 |
+
"auto_map": {
|
| 8 |
+
"AutoConfig": "hf_wrapper.Effb2TrmConfig",
|
| 9 |
+
"AutoModel": "hf_wrapper.Effb2TrmCaptioningModel"
|
| 10 |
+
},
|
| 11 |
"decoder_dropout": 0.2,
|
| 12 |
"decoder_emb_dim": 256,
|
| 13 |
"decoder_n_layers": 2,
|
hf_wrapper.py
ADDED
|
@@ -0,0 +1,1974 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Dict, Callable, Union, List
|
| 2 |
+
import random
|
| 3 |
+
import math
|
| 4 |
+
import sys
|
| 5 |
+
|
| 6 |
+
import numpy as np
|
| 7 |
+
import torch
|
| 8 |
+
import torch.nn as nn
|
| 9 |
+
import torch.nn.functional as F
|
| 10 |
+
from torch.nn.utils.rnn import PackedSequence, pack_padded_sequence, pad_packed_sequence
|
| 11 |
+
from torchaudio import transforms
|
| 12 |
+
from efficientnet_pytorch import EfficientNet
|
| 13 |
+
from efficientnet_pytorch import utils as efficientnet_utils
|
| 14 |
+
from einops import rearrange, reduce
|
| 15 |
+
from transformers import PretrainedConfig, PreTrainedModel
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
def sort_pack_padded_sequence(input, lengths):
|
| 19 |
+
sorted_lengths, indices = torch.sort(lengths, descending=True)
|
| 20 |
+
tmp = pack_padded_sequence(input[indices], sorted_lengths.cpu(), batch_first=True)
|
| 21 |
+
inv_ix = indices.clone()
|
| 22 |
+
inv_ix[indices] = torch.arange(0,len(indices)).type_as(inv_ix)
|
| 23 |
+
return tmp, inv_ix
|
| 24 |
+
|
| 25 |
+
def pad_unsort_packed_sequence(input, inv_ix):
|
| 26 |
+
tmp, _ = pad_packed_sequence(input, batch_first=True)
|
| 27 |
+
tmp = tmp[inv_ix]
|
| 28 |
+
return tmp
|
| 29 |
+
|
| 30 |
+
def pack_wrapper(module, attn_feats, attn_feat_lens):
|
| 31 |
+
packed, inv_ix = sort_pack_padded_sequence(attn_feats, attn_feat_lens)
|
| 32 |
+
if isinstance(module, torch.nn.RNNBase):
|
| 33 |
+
return pad_unsort_packed_sequence(module(packed)[0], inv_ix)
|
| 34 |
+
else:
|
| 35 |
+
return pad_unsort_packed_sequence(PackedSequence(module(packed[0]), packed[1]), inv_ix)
|
| 36 |
+
|
| 37 |
+
def embedding_pooling(x, lens, pooling="mean"):
|
| 38 |
+
if pooling == "max":
|
| 39 |
+
fc_embs = max_with_lens(x, lens)
|
| 40 |
+
elif pooling == "mean":
|
| 41 |
+
fc_embs = mean_with_lens(x, lens)
|
| 42 |
+
elif pooling == "mean+max":
|
| 43 |
+
x_mean = mean_with_lens(x, lens)
|
| 44 |
+
x_max = max_with_lens(x, lens)
|
| 45 |
+
fc_embs = x_mean + x_max
|
| 46 |
+
elif pooling == "last":
|
| 47 |
+
indices = (lens - 1).reshape(-1, 1, 1).repeat(1, 1, x.size(-1))
|
| 48 |
+
# indices: [N, 1, hidden]
|
| 49 |
+
fc_embs = torch.gather(x, 1, indices).squeeze(1)
|
| 50 |
+
else:
|
| 51 |
+
raise Exception(f"pooling method {pooling} not support")
|
| 52 |
+
return fc_embs
|
| 53 |
+
|
| 54 |
+
def interpolate(x, ratio):
|
| 55 |
+
"""Interpolate data in time domain. This is used to compensate the
|
| 56 |
+
resolution reduction in downsampling of a CNN.
|
| 57 |
+
|
| 58 |
+
Args:
|
| 59 |
+
x: (batch_size, time_steps, classes_num)
|
| 60 |
+
ratio: int, ratio to interpolate
|
| 61 |
+
|
| 62 |
+
Returns:
|
| 63 |
+
upsampled: (batch_size, time_steps * ratio, classes_num)
|
| 64 |
+
"""
|
| 65 |
+
(batch_size, time_steps, classes_num) = x.shape
|
| 66 |
+
upsampled = x[:, :, None, :].repeat(1, 1, ratio, 1)
|
| 67 |
+
upsampled = upsampled.reshape(batch_size, time_steps * ratio, classes_num)
|
| 68 |
+
return upsampled
|
| 69 |
+
|
| 70 |
+
def pad_framewise_output(framewise_output, frames_num):
|
| 71 |
+
"""Pad framewise_output to the same length as input frames. The pad value
|
| 72 |
+
is the same as the value of the last frame.
|
| 73 |
+
|
| 74 |
+
Args:
|
| 75 |
+
framewise_output: (batch_size, frames_num, classes_num)
|
| 76 |
+
frames_num: int, number of frames to pad
|
| 77 |
+
|
| 78 |
+
Outputs:
|
| 79 |
+
output: (batch_size, frames_num, classes_num)
|
| 80 |
+
"""
|
| 81 |
+
pad = framewise_output[:, -1 :, :].repeat(1, frames_num - framewise_output.shape[1], 1)
|
| 82 |
+
"""tensor for padding"""
|
| 83 |
+
|
| 84 |
+
output = torch.cat((framewise_output, pad), dim=1)
|
| 85 |
+
"""(batch_size, frames_num, classes_num)"""
|
| 86 |
+
|
| 87 |
+
return output
|
| 88 |
+
|
| 89 |
+
def find_contiguous_regions(activity_array):
|
| 90 |
+
"""Find contiguous regions from bool valued numpy.array.
|
| 91 |
+
Copy of https://dcase-repo.github.io/dcase_util/_modules/dcase_util/data/decisions.html#DecisionEncoder
|
| 92 |
+
|
| 93 |
+
Reason is:
|
| 94 |
+
1. This does not belong to a class necessarily
|
| 95 |
+
2. Import DecisionEncoder requires sndfile over some other imports..which causes some problems on clusters
|
| 96 |
+
|
| 97 |
+
"""
|
| 98 |
+
|
| 99 |
+
# Find the changes in the activity_array
|
| 100 |
+
change_indices = np.logical_xor(activity_array[1:],
|
| 101 |
+
activity_array[:-1]).nonzero()[0]
|
| 102 |
+
|
| 103 |
+
# Shift change_index with one, focus on frame after the change.
|
| 104 |
+
change_indices += 1
|
| 105 |
+
|
| 106 |
+
if activity_array[0]:
|
| 107 |
+
# If the first element of activity_array is True add 0 at the beginning
|
| 108 |
+
change_indices = np.r_[0, change_indices]
|
| 109 |
+
|
| 110 |
+
if activity_array[-1]:
|
| 111 |
+
# If the last element of activity_array is True, add the length of the array
|
| 112 |
+
change_indices = np.r_[change_indices, activity_array.size]
|
| 113 |
+
|
| 114 |
+
# Reshape the result into two columns
|
| 115 |
+
return change_indices.reshape((-1, 2))
|
| 116 |
+
|
| 117 |
+
def double_threshold(x, high_thres, low_thres, n_connect=1):
|
| 118 |
+
"""double_threshold
|
| 119 |
+
Helper function to calculate double threshold for n-dim arrays
|
| 120 |
+
|
| 121 |
+
:param x: input array
|
| 122 |
+
:param high_thres: high threshold value
|
| 123 |
+
:param low_thres: Low threshold value
|
| 124 |
+
:param n_connect: Distance of <= n clusters will be merged
|
| 125 |
+
"""
|
| 126 |
+
assert x.ndim <= 3, "Whoops something went wrong with the input ({}), check if its <= 3 dims".format(
|
| 127 |
+
x.shape)
|
| 128 |
+
if x.ndim == 3:
|
| 129 |
+
apply_dim = 1
|
| 130 |
+
elif x.ndim < 3:
|
| 131 |
+
apply_dim = 0
|
| 132 |
+
# x is assumed to be 3d: (batch, time, dim)
|
| 133 |
+
# Assumed to be 2d : (time, dim)
|
| 134 |
+
# Assumed to be 1d : (time)
|
| 135 |
+
# time axis is therefore at 1 for 3d and 0 for 2d (
|
| 136 |
+
return np.apply_along_axis(lambda x: _double_threshold(
|
| 137 |
+
x, high_thres, low_thres, n_connect=n_connect),
|
| 138 |
+
axis=apply_dim,
|
| 139 |
+
arr=x)
|
| 140 |
+
|
| 141 |
+
def _double_threshold(x, high_thres, low_thres, n_connect=1, return_arr=True):
|
| 142 |
+
"""_double_threshold
|
| 143 |
+
Computes a double threshold over the input array
|
| 144 |
+
|
| 145 |
+
:param x: input array, needs to be 1d
|
| 146 |
+
:param high_thres: High threshold over the array
|
| 147 |
+
:param low_thres: Low threshold over the array
|
| 148 |
+
:param n_connect: Postprocessing, maximal distance between clusters to connect
|
| 149 |
+
:param return_arr: By default this function returns the filtered indiced, but if return_arr = True it returns an array of tsame size as x filled with ones and zeros.
|
| 150 |
+
"""
|
| 151 |
+
assert x.ndim == 1, "Input needs to be 1d"
|
| 152 |
+
high_locations = np.where(x > high_thres)[0]
|
| 153 |
+
locations = x > low_thres
|
| 154 |
+
encoded_pairs = find_contiguous_regions(locations)
|
| 155 |
+
|
| 156 |
+
filtered_list = list(
|
| 157 |
+
filter(
|
| 158 |
+
lambda pair:
|
| 159 |
+
((pair[0] <= high_locations) & (high_locations <= pair[1])).any(),
|
| 160 |
+
encoded_pairs))
|
| 161 |
+
|
| 162 |
+
filtered_list = connect_(filtered_list, n_connect)
|
| 163 |
+
if return_arr:
|
| 164 |
+
zero_one_arr = np.zeros_like(x, dtype=int)
|
| 165 |
+
for sl in filtered_list:
|
| 166 |
+
zero_one_arr[sl[0]:sl[1]] = 1
|
| 167 |
+
return zero_one_arr
|
| 168 |
+
return filtered_list
|
| 169 |
+
|
| 170 |
+
def connect_(pairs, n=1):
|
| 171 |
+
"""connect_
|
| 172 |
+
Connects two adjacent clusters if their distance is <= n
|
| 173 |
+
|
| 174 |
+
:param pairs: Clusters of iterateables e.g., [(1,5),(7,10)]
|
| 175 |
+
:param n: distance between two clusters
|
| 176 |
+
"""
|
| 177 |
+
if len(pairs) == 0:
|
| 178 |
+
return []
|
| 179 |
+
start_, end_ = pairs[0]
|
| 180 |
+
new_pairs = []
|
| 181 |
+
for i, (next_item, cur_item) in enumerate(zip(pairs[1:], pairs[0:])):
|
| 182 |
+
end_ = next_item[1]
|
| 183 |
+
if next_item[0] - cur_item[1] <= n:
|
| 184 |
+
pass
|
| 185 |
+
else:
|
| 186 |
+
new_pairs.append((start_, cur_item[1]))
|
| 187 |
+
start_ = next_item[0]
|
| 188 |
+
new_pairs.append((start_, end_))
|
| 189 |
+
return new_pairs
|
| 190 |
+
|
| 191 |
+
def segments_to_temporal_tag(segments, thre=0.5):
|
| 192 |
+
after_flag, while_flag = 0, 0
|
| 193 |
+
for j in range(len(segments)):
|
| 194 |
+
for k in range(len(segments)):
|
| 195 |
+
if segments[j][0] == segments[k][0]:
|
| 196 |
+
continue
|
| 197 |
+
min_duration = min(segments[j][2] - segments[j][1], segments[k][2] - segments[k][1])
|
| 198 |
+
overlap = segments[j][2] - segments[k][1]
|
| 199 |
+
if overlap < thre * min_duration:
|
| 200 |
+
after_flag = 2
|
| 201 |
+
if segments[j][1] < segments[k][1] and overlap > thre * min_duration:
|
| 202 |
+
while_flag = 1
|
| 203 |
+
return after_flag + while_flag
|
| 204 |
+
|
| 205 |
+
def decode_with_timestamps(labels, time_resolution):
|
| 206 |
+
batch_results = []
|
| 207 |
+
for lab in labels:
|
| 208 |
+
segments = []
|
| 209 |
+
for i, label_column in enumerate(lab.T):
|
| 210 |
+
change_indices = find_contiguous_regions(label_column)
|
| 211 |
+
# append [onset, offset] in the result list
|
| 212 |
+
for row in change_indices:
|
| 213 |
+
segments.append((i, row[0] * time_resolution, row[1] * time_resolution))
|
| 214 |
+
temporal_tag = segments_to_temporal_tag(segments)
|
| 215 |
+
batch_results.append(temporal_tag)
|
| 216 |
+
return batch_results
|
| 217 |
+
|
| 218 |
+
class _EffiNet(nn.Module):
|
| 219 |
+
"""A proxy for efficient net models"""
|
| 220 |
+
def __init__(self,
|
| 221 |
+
blocks_args=None,
|
| 222 |
+
global_params=None,
|
| 223 |
+
) -> None:
|
| 224 |
+
super().__init__()
|
| 225 |
+
self.eff_net = EfficientNet(blocks_args=blocks_args,
|
| 226 |
+
global_params=global_params)
|
| 227 |
+
|
| 228 |
+
|
| 229 |
+
def forward(self, x: torch.Tensor):
|
| 230 |
+
x = rearrange(x, 'b f t -> b 1 f t')
|
| 231 |
+
x = self.eff_net.extract_features(x)
|
| 232 |
+
return reduce(x, 'b c f t -> b t c', 'mean')
|
| 233 |
+
|
| 234 |
+
|
| 235 |
+
def get_effb2_model() -> _EffiNet:
|
| 236 |
+
blocks_args, global_params = efficientnet_utils.get_model_params(
|
| 237 |
+
'efficientnet-b2', {'include_top': False})
|
| 238 |
+
model = _EffiNet(blocks_args=blocks_args,
|
| 239 |
+
global_params=global_params)
|
| 240 |
+
model.eff_net._change_in_channels(1)
|
| 241 |
+
return model
|
| 242 |
+
|
| 243 |
+
def merge_load_state_dict(state_dict,
|
| 244 |
+
model: torch.nn.Module,
|
| 245 |
+
output_fn: Callable = sys.stdout.write):
|
| 246 |
+
model_dict = model.state_dict()
|
| 247 |
+
pretrained_dict = {}
|
| 248 |
+
mismatch_keys = []
|
| 249 |
+
for key, value in state_dict.items():
|
| 250 |
+
if key in model_dict and model_dict[key].shape == value.shape:
|
| 251 |
+
pretrained_dict[key] = value
|
| 252 |
+
else:
|
| 253 |
+
mismatch_keys.append(key)
|
| 254 |
+
output_fn(f"Loading pre-trained model, with mismatched keys {mismatch_keys}\n")
|
| 255 |
+
model_dict.update(pretrained_dict)
|
| 256 |
+
model.load_state_dict(model_dict, strict=True)
|
| 257 |
+
return pretrained_dict.keys()
|
| 258 |
+
|
| 259 |
+
|
| 260 |
+
class EfficientNetB2(nn.Module):
|
| 261 |
+
|
| 262 |
+
def __init__(self,
|
| 263 |
+
n_mels: int = 64,
|
| 264 |
+
win_length: int = 32,
|
| 265 |
+
hop_length: int = 10,
|
| 266 |
+
f_min: int = 0,
|
| 267 |
+
freeze: bool = False,):
|
| 268 |
+
super().__init__()
|
| 269 |
+
sample_rate = 16000
|
| 270 |
+
self.melspec_extractor = transforms.MelSpectrogram(
|
| 271 |
+
sample_rate=sample_rate,
|
| 272 |
+
n_fft=win_length * sample_rate // 1000,
|
| 273 |
+
win_length=win_length * sample_rate // 1000,
|
| 274 |
+
hop_length=hop_length * sample_rate // 1000,
|
| 275 |
+
f_min=f_min,
|
| 276 |
+
n_mels=n_mels,
|
| 277 |
+
)
|
| 278 |
+
self.hop_length = 10 * sample_rate // 1000
|
| 279 |
+
self.db_transform = transforms.AmplitudeToDB(top_db=120)
|
| 280 |
+
self.backbone = get_effb2_model()
|
| 281 |
+
self.fc_emb_size = self.backbone.eff_net._conv_head.out_channels
|
| 282 |
+
self.downsample_ratio = 32
|
| 283 |
+
if freeze:
|
| 284 |
+
for param in self.parameters():
|
| 285 |
+
param.requires_grad = False
|
| 286 |
+
|
| 287 |
+
def forward(self, input_dict):
|
| 288 |
+
|
| 289 |
+
waveform = input_dict["wav"]
|
| 290 |
+
wave_length = input_dict["wav_len"]
|
| 291 |
+
specaug = input_dict["specaug"]
|
| 292 |
+
x = self.melspec_extractor(waveform)
|
| 293 |
+
x = self.db_transform(x) # (batch_size, mel_bins, time_steps)
|
| 294 |
+
|
| 295 |
+
x = rearrange(x, 'b f t -> b 1 t f')
|
| 296 |
+
if self.training and specaug:
|
| 297 |
+
x = self.spec_augmenter(x)
|
| 298 |
+
x = rearrange(x, 'b 1 t f -> b f t')
|
| 299 |
+
|
| 300 |
+
x = self.backbone(x)
|
| 301 |
+
attn_emb = x
|
| 302 |
+
|
| 303 |
+
wave_length = torch.as_tensor(wave_length)
|
| 304 |
+
feat_length = torch.div(wave_length, self.hop_length,
|
| 305 |
+
rounding_mode="floor") + 1
|
| 306 |
+
feat_length = torch.div(feat_length, self.downsample_ratio,
|
| 307 |
+
rounding_mode="floor")
|
| 308 |
+
fc_emb = mean_with_lens(attn_emb, feat_length)
|
| 309 |
+
|
| 310 |
+
output_dict = {
|
| 311 |
+
'fc_emb': fc_emb,
|
| 312 |
+
'attn_emb': attn_emb,
|
| 313 |
+
'attn_emb_len': feat_length
|
| 314 |
+
}
|
| 315 |
+
return output_dict
|
| 316 |
+
|
| 317 |
+
|
| 318 |
+
def generate_length_mask(lens, max_length=None):
|
| 319 |
+
lens = torch.as_tensor(lens)
|
| 320 |
+
N = lens.size(0)
|
| 321 |
+
if max_length is None:
|
| 322 |
+
max_length = max(lens)
|
| 323 |
+
if isinstance(max_length, torch.Tensor):
|
| 324 |
+
max_length = max_length.item()
|
| 325 |
+
idxs = torch.arange(max_length).repeat(N).view(N, max_length)
|
| 326 |
+
idxs = idxs.to(lens.device)
|
| 327 |
+
mask = (idxs < lens.view(-1, 1))
|
| 328 |
+
return mask
|
| 329 |
+
|
| 330 |
+
def mean_with_lens(features, lens):
|
| 331 |
+
"""
|
| 332 |
+
features: [N, T, ...] (assume the second dimension represents length)
|
| 333 |
+
lens: [N,]
|
| 334 |
+
"""
|
| 335 |
+
lens = torch.as_tensor(lens)
|
| 336 |
+
if max(lens) != features.size(1):
|
| 337 |
+
max_length = features.size(1)
|
| 338 |
+
mask = generate_length_mask(lens, max_length)
|
| 339 |
+
else:
|
| 340 |
+
mask = generate_length_mask(lens)
|
| 341 |
+
mask = mask.to(features.device) # [N, T]
|
| 342 |
+
|
| 343 |
+
while mask.ndim < features.ndim:
|
| 344 |
+
mask = mask.unsqueeze(-1)
|
| 345 |
+
feature_mean = features * mask
|
| 346 |
+
feature_mean = feature_mean.sum(1)
|
| 347 |
+
while lens.ndim < feature_mean.ndim:
|
| 348 |
+
lens = lens.unsqueeze(1)
|
| 349 |
+
feature_mean = feature_mean / lens.to(features.device)
|
| 350 |
+
# feature_mean = features * mask.unsqueeze(-1)
|
| 351 |
+
# feature_mean = feature_mean.sum(1) / lens.unsqueeze(1).to(features.device)
|
| 352 |
+
return feature_mean
|
| 353 |
+
|
| 354 |
+
def max_with_lens(features, lens):
|
| 355 |
+
"""
|
| 356 |
+
features: [N, T, ...] (assume the second dimension represents length)
|
| 357 |
+
lens: [N,]
|
| 358 |
+
"""
|
| 359 |
+
lens = torch.as_tensor(lens)
|
| 360 |
+
if max(lens) != features.size(1):
|
| 361 |
+
max_length = features.size(1)
|
| 362 |
+
mask = generate_length_mask(lens, max_length)
|
| 363 |
+
else:
|
| 364 |
+
mask = generate_length_mask(lens)
|
| 365 |
+
mask = mask.to(features.device) # [N, T]
|
| 366 |
+
|
| 367 |
+
feature_max = features.clone()
|
| 368 |
+
feature_max[~mask] = float("-inf")
|
| 369 |
+
feature_max, _ = feature_max.max(1)
|
| 370 |
+
return feature_max
|
| 371 |
+
|
| 372 |
+
def repeat_tensor(x, n):
|
| 373 |
+
return x.unsqueeze(0).repeat(n, *([1] * len(x.shape)))
|
| 374 |
+
|
| 375 |
+
|
| 376 |
+
class CaptionMetaMixin:
|
| 377 |
+
pad_idx = 0
|
| 378 |
+
start_idx = 1
|
| 379 |
+
end_idx = 2
|
| 380 |
+
max_length = 20
|
| 381 |
+
|
| 382 |
+
@classmethod
|
| 383 |
+
def set_index(cls, start_idx, end_idx, pad_idx):
|
| 384 |
+
cls.start_idx = start_idx
|
| 385 |
+
cls.end_idx = end_idx
|
| 386 |
+
cls.pad_idx = pad_idx
|
| 387 |
+
|
| 388 |
+
|
| 389 |
+
class CaptionModel(nn.Module, CaptionMetaMixin):
|
| 390 |
+
"""
|
| 391 |
+
Encoder-decoder captioning model.
|
| 392 |
+
"""
|
| 393 |
+
|
| 394 |
+
def __init__(self, encoder: nn.Module, decoder: nn.Module, **kwargs):
|
| 395 |
+
super().__init__()
|
| 396 |
+
self.encoder = encoder
|
| 397 |
+
self.decoder = decoder
|
| 398 |
+
self.vocab_size = decoder.vocab_size
|
| 399 |
+
self.train_forward_keys = ["cap", "cap_len", "ss_ratio"]
|
| 400 |
+
self.inference_forward_keys = ["sample_method", "max_length", "temp"]
|
| 401 |
+
freeze_encoder = kwargs.get("freeze_encoder", False)
|
| 402 |
+
if freeze_encoder:
|
| 403 |
+
for param in self.encoder.parameters():
|
| 404 |
+
param.requires_grad = False
|
| 405 |
+
self.check_decoder_compatibility()
|
| 406 |
+
|
| 407 |
+
def check_decoder_compatibility(self):
|
| 408 |
+
compatible_decoders = [x.__class__.__name__ for x in self.compatible_decoders]
|
| 409 |
+
assert isinstance(self.decoder, self.compatible_decoders), \
|
| 410 |
+
f"{self.decoder.__class__.__name__} is incompatible with " \
|
| 411 |
+
f"{self.__class__.__name__}, please use decoder in {compatible_decoders} "
|
| 412 |
+
|
| 413 |
+
def forward(self, input_dict: Dict):
|
| 414 |
+
"""
|
| 415 |
+
input_dict: {
|
| 416 |
+
(required)
|
| 417 |
+
mode: train/inference,
|
| 418 |
+
[spec, spec_len],
|
| 419 |
+
[fc],
|
| 420 |
+
[attn, attn_len],
|
| 421 |
+
[wav, wav_len],
|
| 422 |
+
[sample_method: greedy],
|
| 423 |
+
[temp: 1.0] (in case of no teacher forcing)
|
| 424 |
+
|
| 425 |
+
(optional, mode=train)
|
| 426 |
+
cap,
|
| 427 |
+
cap_len,
|
| 428 |
+
ss_ratio,
|
| 429 |
+
|
| 430 |
+
(optional, mode=inference)
|
| 431 |
+
sample_method: greedy/beam,
|
| 432 |
+
max_length,
|
| 433 |
+
temp,
|
| 434 |
+
beam_size (optional, sample_method=beam),
|
| 435 |
+
n_best (optional, sample_method=beam),
|
| 436 |
+
}
|
| 437 |
+
"""
|
| 438 |
+
encoder_output_dict = self.encoder(input_dict)
|
| 439 |
+
output = self.forward_decoder(input_dict, encoder_output_dict)
|
| 440 |
+
return output
|
| 441 |
+
|
| 442 |
+
def forward_decoder(self, input_dict: Dict, encoder_output_dict: Dict):
|
| 443 |
+
if input_dict["mode"] == "train":
|
| 444 |
+
forward_dict = {
|
| 445 |
+
"mode": "train", "sample_method": "greedy", "temp": 1.0
|
| 446 |
+
}
|
| 447 |
+
for key in self.train_forward_keys:
|
| 448 |
+
forward_dict[key] = input_dict[key]
|
| 449 |
+
forward_dict.update(encoder_output_dict)
|
| 450 |
+
output = self.train_forward(forward_dict)
|
| 451 |
+
elif input_dict["mode"] == "inference":
|
| 452 |
+
forward_dict = {"mode": "inference"}
|
| 453 |
+
default_args = { "sample_method": "greedy", "max_length": self.max_length, "temp": 1.0 }
|
| 454 |
+
for key in self.inference_forward_keys:
|
| 455 |
+
if key in input_dict:
|
| 456 |
+
forward_dict[key] = input_dict[key]
|
| 457 |
+
else:
|
| 458 |
+
forward_dict[key] = default_args[key]
|
| 459 |
+
|
| 460 |
+
if forward_dict["sample_method"] == "beam":
|
| 461 |
+
forward_dict["beam_size"] = input_dict.get("beam_size", 3)
|
| 462 |
+
forward_dict["n_best"] = input_dict.get("n_best", False)
|
| 463 |
+
forward_dict["n_best_size"] = input_dict.get("n_best_size", forward_dict["beam_size"])
|
| 464 |
+
elif forward_dict["sample_method"] == "dbs":
|
| 465 |
+
forward_dict["beam_size"] = input_dict.get("beam_size", 6)
|
| 466 |
+
forward_dict["group_size"] = input_dict.get("group_size", 3)
|
| 467 |
+
forward_dict["diversity_lambda"] = input_dict.get("diversity_lambda", 0.5)
|
| 468 |
+
forward_dict["group_nbest"] = input_dict.get("group_nbest", True)
|
| 469 |
+
|
| 470 |
+
forward_dict.update(encoder_output_dict)
|
| 471 |
+
output = self.inference_forward(forward_dict)
|
| 472 |
+
else:
|
| 473 |
+
raise Exception("mode should be either 'train' or 'inference'")
|
| 474 |
+
output.update(encoder_output_dict)
|
| 475 |
+
return output
|
| 476 |
+
|
| 477 |
+
def prepare_output(self, input_dict):
|
| 478 |
+
output = {}
|
| 479 |
+
batch_size = input_dict["fc_emb"].size(0)
|
| 480 |
+
if input_dict["mode"] == "train":
|
| 481 |
+
max_length = input_dict["cap"].size(1) - 1
|
| 482 |
+
elif input_dict["mode"] == "inference":
|
| 483 |
+
max_length = input_dict["max_length"]
|
| 484 |
+
else:
|
| 485 |
+
raise Exception("mode should be either 'train' or 'inference'")
|
| 486 |
+
device = input_dict["fc_emb"].device
|
| 487 |
+
output["seq"] = torch.full((batch_size, max_length), self.end_idx,
|
| 488 |
+
dtype=torch.long)
|
| 489 |
+
output["logit"] = torch.empty(batch_size, max_length,
|
| 490 |
+
self.vocab_size).to(device)
|
| 491 |
+
output["sampled_logprob"] = torch.zeros(batch_size, max_length)
|
| 492 |
+
output["embed"] = torch.empty(batch_size, max_length,
|
| 493 |
+
self.decoder.d_model).to(device)
|
| 494 |
+
return output
|
| 495 |
+
|
| 496 |
+
def train_forward(self, input_dict):
|
| 497 |
+
if input_dict["ss_ratio"] != 1: # scheduled sampling training
|
| 498 |
+
input_dict["mode"] = "train"
|
| 499 |
+
return self.stepwise_forward(input_dict)
|
| 500 |
+
output = self.seq_forward(input_dict)
|
| 501 |
+
self.train_process(output, input_dict)
|
| 502 |
+
return output
|
| 503 |
+
|
| 504 |
+
def seq_forward(self, input_dict):
|
| 505 |
+
raise NotImplementedError
|
| 506 |
+
|
| 507 |
+
def train_process(self, output, input_dict):
|
| 508 |
+
pass
|
| 509 |
+
|
| 510 |
+
def inference_forward(self, input_dict):
|
| 511 |
+
if input_dict["sample_method"] == "beam":
|
| 512 |
+
return self.beam_search(input_dict)
|
| 513 |
+
elif input_dict["sample_method"] == "dbs":
|
| 514 |
+
return self.diverse_beam_search(input_dict)
|
| 515 |
+
return self.stepwise_forward(input_dict)
|
| 516 |
+
|
| 517 |
+
def stepwise_forward(self, input_dict):
|
| 518 |
+
"""Step-by-step decoding"""
|
| 519 |
+
output = self.prepare_output(input_dict)
|
| 520 |
+
max_length = output["seq"].size(1)
|
| 521 |
+
# start sampling
|
| 522 |
+
for t in range(max_length):
|
| 523 |
+
input_dict["t"] = t
|
| 524 |
+
self.decode_step(input_dict, output)
|
| 525 |
+
if input_dict["mode"] == "inference": # decide whether to stop when sampling
|
| 526 |
+
unfinished_t = output["seq"][:, t] != self.end_idx
|
| 527 |
+
if t == 0:
|
| 528 |
+
unfinished = unfinished_t
|
| 529 |
+
else:
|
| 530 |
+
unfinished *= unfinished_t
|
| 531 |
+
output["seq"][:, t][~unfinished] = self.end_idx
|
| 532 |
+
if unfinished.sum() == 0:
|
| 533 |
+
break
|
| 534 |
+
self.stepwise_process(output)
|
| 535 |
+
return output
|
| 536 |
+
|
| 537 |
+
def decode_step(self, input_dict, output):
|
| 538 |
+
"""Decoding operation of timestep t"""
|
| 539 |
+
decoder_input = self.prepare_decoder_input(input_dict, output)
|
| 540 |
+
# feed to the decoder to get logit
|
| 541 |
+
output_t = self.decoder(decoder_input)
|
| 542 |
+
logit_t = output_t["logit"]
|
| 543 |
+
# assert logit_t.ndim == 3
|
| 544 |
+
if logit_t.size(1) == 1:
|
| 545 |
+
logit_t = logit_t.squeeze(1)
|
| 546 |
+
embed_t = output_t["embed"].squeeze(1)
|
| 547 |
+
elif logit_t.size(1) > 1:
|
| 548 |
+
logit_t = logit_t[:, -1, :]
|
| 549 |
+
embed_t = output_t["embed"][:, -1, :]
|
| 550 |
+
else:
|
| 551 |
+
raise Exception("no logit output")
|
| 552 |
+
# sample the next input word and get the corresponding logit
|
| 553 |
+
sampled = self.sample_next_word(logit_t,
|
| 554 |
+
method=input_dict["sample_method"],
|
| 555 |
+
temp=input_dict["temp"])
|
| 556 |
+
|
| 557 |
+
output_t.update(sampled)
|
| 558 |
+
output_t["t"] = input_dict["t"]
|
| 559 |
+
output_t["logit"] = logit_t
|
| 560 |
+
output_t["embed"] = embed_t
|
| 561 |
+
self.stepwise_process_step(output, output_t)
|
| 562 |
+
|
| 563 |
+
def prepare_decoder_input(self, input_dict, output):
|
| 564 |
+
"""Prepare the inp ut dict for the decoder"""
|
| 565 |
+
raise NotImplementedError
|
| 566 |
+
|
| 567 |
+
def stepwise_process_step(self, output, output_t):
|
| 568 |
+
"""Postprocessing (save output values) after each timestep t"""
|
| 569 |
+
t = output_t["t"]
|
| 570 |
+
output["logit"][:, t, :] = output_t["logit"]
|
| 571 |
+
output["seq"][:, t] = output_t["word"]
|
| 572 |
+
output["sampled_logprob"][:, t] = output_t["probs"]
|
| 573 |
+
output["embed"][:, t, :] = output_t["embed"]
|
| 574 |
+
|
| 575 |
+
def stepwise_process(self, output):
|
| 576 |
+
"""Postprocessing after the whole step-by-step autoregressive decoding"""
|
| 577 |
+
pass
|
| 578 |
+
|
| 579 |
+
def sample_next_word(self, logit, method, temp):
|
| 580 |
+
"""Sample the next word, given probs output by the decoder"""
|
| 581 |
+
logprob = torch.log_softmax(logit, dim=1)
|
| 582 |
+
if method == "greedy":
|
| 583 |
+
sampled_logprob, word = torch.max(logprob.detach(), 1)
|
| 584 |
+
elif method == "gumbel":
|
| 585 |
+
def sample_gumbel(shape, eps=1e-20):
|
| 586 |
+
U = torch.rand(shape).to(logprob.device)
|
| 587 |
+
return -torch.log(-torch.log(U + eps) + eps)
|
| 588 |
+
def gumbel_softmax_sample(logit, temperature):
|
| 589 |
+
y = logit + sample_gumbel(logit.size())
|
| 590 |
+
return torch.log_softmax(y / temperature, dim=-1)
|
| 591 |
+
_logprob = gumbel_softmax_sample(logprob, temp)
|
| 592 |
+
_, word = torch.max(_logprob.data, 1)
|
| 593 |
+
sampled_logprob = logprob.gather(1, word.unsqueeze(-1))
|
| 594 |
+
else:
|
| 595 |
+
logprob = logprob / temp
|
| 596 |
+
if method.startswith("top"):
|
| 597 |
+
top_num = float(method[3:])
|
| 598 |
+
if 0 < top_num < 1: # top-p sampling
|
| 599 |
+
probs = torch.softmax(logit, dim=1)
|
| 600 |
+
sorted_probs, sorted_indices = torch.sort(probs, descending=True, dim=1)
|
| 601 |
+
_cumsum = sorted_probs.cumsum(1)
|
| 602 |
+
mask = _cumsum < top_num
|
| 603 |
+
mask = torch.cat([torch.ones_like(mask[:,:1]), mask[:,:-1]], 1)
|
| 604 |
+
sorted_probs = sorted_probs * mask.to(sorted_probs)
|
| 605 |
+
sorted_probs = sorted_probs / sorted_probs.sum(1, keepdim=True)
|
| 606 |
+
logprob.scatter_(1, sorted_indices, sorted_probs.log())
|
| 607 |
+
else: # top-k sampling
|
| 608 |
+
k = int(top_num)
|
| 609 |
+
tmp = torch.empty_like(logprob).fill_(float('-inf'))
|
| 610 |
+
topk, indices = torch.topk(logprob, k, dim=1)
|
| 611 |
+
tmp = tmp.scatter(1, indices, topk)
|
| 612 |
+
logprob = tmp
|
| 613 |
+
word = torch.distributions.Categorical(logits=logprob.detach()).sample()
|
| 614 |
+
sampled_logprob = logprob.gather(1, word.unsqueeze(-1)).squeeze(1)
|
| 615 |
+
word = word.detach().long()
|
| 616 |
+
# sampled_logprob: [N,], word: [N,]
|
| 617 |
+
return {"word": word, "probs": sampled_logprob}
|
| 618 |
+
|
| 619 |
+
def beam_search(self, input_dict):
|
| 620 |
+
output = self.prepare_output(input_dict)
|
| 621 |
+
max_length = input_dict["max_length"]
|
| 622 |
+
beam_size = input_dict["beam_size"]
|
| 623 |
+
if input_dict["n_best"]:
|
| 624 |
+
n_best_size = input_dict["n_best_size"]
|
| 625 |
+
batch_size, max_length = output["seq"].size()
|
| 626 |
+
output["seq"] = torch.full((batch_size, n_best_size, max_length),
|
| 627 |
+
self.end_idx, dtype=torch.long)
|
| 628 |
+
|
| 629 |
+
temp = input_dict["temp"]
|
| 630 |
+
# instance by instance beam seach
|
| 631 |
+
for i in range(output["seq"].size(0)):
|
| 632 |
+
output_i = self.prepare_beamsearch_output(input_dict)
|
| 633 |
+
input_dict["sample_idx"] = i
|
| 634 |
+
for t in range(max_length):
|
| 635 |
+
input_dict["t"] = t
|
| 636 |
+
output_t = self.beamsearch_step(input_dict, output_i)
|
| 637 |
+
#######################################
|
| 638 |
+
# merge with previous beam and select the current max prob beam
|
| 639 |
+
#######################################
|
| 640 |
+
logit_t = output_t["logit"]
|
| 641 |
+
if logit_t.size(1) == 1:
|
| 642 |
+
logit_t = logit_t.squeeze(1)
|
| 643 |
+
elif logit_t.size(1) > 1:
|
| 644 |
+
logit_t = logit_t[:, -1, :]
|
| 645 |
+
else:
|
| 646 |
+
raise Exception("no logit output")
|
| 647 |
+
logprob_t = torch.log_softmax(logit_t, dim=1)
|
| 648 |
+
logprob_t = torch.log_softmax(logprob_t / temp, dim=1)
|
| 649 |
+
logprob_t = output_i["topk_logprob"].unsqueeze(1) + logprob_t
|
| 650 |
+
if t == 0: # for the first step, all k seq will have the same probs
|
| 651 |
+
topk_logprob, topk_words = logprob_t[0].topk(
|
| 652 |
+
beam_size, 0, True, True)
|
| 653 |
+
else: # unroll and find top logprob, and their unrolled indices
|
| 654 |
+
topk_logprob, topk_words = logprob_t.view(-1).topk(
|
| 655 |
+
beam_size, 0, True, True)
|
| 656 |
+
topk_words = topk_words.cpu()
|
| 657 |
+
output_i["topk_logprob"] = topk_logprob
|
| 658 |
+
# output_i["prev_words_beam"] = topk_words // self.vocab_size # [beam_size,]
|
| 659 |
+
output_i["prev_words_beam"] = torch.div(topk_words, self.vocab_size,
|
| 660 |
+
rounding_mode='trunc')
|
| 661 |
+
output_i["next_word"] = topk_words % self.vocab_size # [beam_size,]
|
| 662 |
+
if t == 0:
|
| 663 |
+
output_i["seq"] = output_i["next_word"].unsqueeze(1)
|
| 664 |
+
else:
|
| 665 |
+
output_i["seq"] = torch.cat([
|
| 666 |
+
output_i["seq"][output_i["prev_words_beam"]],
|
| 667 |
+
output_i["next_word"].unsqueeze(1)], dim=1)
|
| 668 |
+
|
| 669 |
+
# add finished beams to results
|
| 670 |
+
is_end = output_i["next_word"] == self.end_idx
|
| 671 |
+
if t == max_length - 1:
|
| 672 |
+
is_end.fill_(1)
|
| 673 |
+
|
| 674 |
+
for beam_idx in range(beam_size):
|
| 675 |
+
if is_end[beam_idx]:
|
| 676 |
+
final_beam = {
|
| 677 |
+
"seq": output_i["seq"][beam_idx].clone(),
|
| 678 |
+
"score": output_i["topk_logprob"][beam_idx].item()
|
| 679 |
+
}
|
| 680 |
+
final_beam["score"] = final_beam["score"] / (t + 1)
|
| 681 |
+
output_i["done_beams"].append(final_beam)
|
| 682 |
+
output_i["topk_logprob"][is_end] -= 1000
|
| 683 |
+
|
| 684 |
+
self.beamsearch_process_step(output_i, output_t)
|
| 685 |
+
|
| 686 |
+
if len(output_i["done_beams"]) == beam_size:
|
| 687 |
+
break
|
| 688 |
+
|
| 689 |
+
self.beamsearch_process(output, output_i, input_dict)
|
| 690 |
+
return output
|
| 691 |
+
|
| 692 |
+
def prepare_beamsearch_output(self, input_dict):
|
| 693 |
+
beam_size = input_dict["beam_size"]
|
| 694 |
+
device = input_dict["fc_emb"].device
|
| 695 |
+
output = {
|
| 696 |
+
"topk_logprob": torch.zeros(beam_size).to(device),
|
| 697 |
+
"seq": None,
|
| 698 |
+
"prev_words_beam": None,
|
| 699 |
+
"next_word": None,
|
| 700 |
+
"done_beams": [],
|
| 701 |
+
}
|
| 702 |
+
return output
|
| 703 |
+
|
| 704 |
+
def beamsearch_step(self, input_dict, output_i):
|
| 705 |
+
decoder_input = self.prepare_beamsearch_decoder_input(input_dict, output_i)
|
| 706 |
+
output_t = self.decoder(decoder_input)
|
| 707 |
+
output_t["t"] = input_dict["t"]
|
| 708 |
+
return output_t
|
| 709 |
+
|
| 710 |
+
def prepare_beamsearch_decoder_input(self, input_dict, output_i):
|
| 711 |
+
raise NotImplementedError
|
| 712 |
+
|
| 713 |
+
def beamsearch_process_step(self, output_i, output_t):
|
| 714 |
+
pass
|
| 715 |
+
|
| 716 |
+
def beamsearch_process(self, output, output_i, input_dict):
|
| 717 |
+
i = input_dict["sample_idx"]
|
| 718 |
+
done_beams = sorted(output_i["done_beams"], key=lambda x: -x["score"])
|
| 719 |
+
if input_dict["n_best"]:
|
| 720 |
+
done_beams = done_beams[:input_dict["n_best_size"]]
|
| 721 |
+
for out_idx, done_beam in enumerate(done_beams):
|
| 722 |
+
seq = done_beam["seq"]
|
| 723 |
+
output["seq"][i][out_idx, :len(seq)] = seq
|
| 724 |
+
else:
|
| 725 |
+
seq = done_beams[0]["seq"]
|
| 726 |
+
output["seq"][i][:len(seq)] = seq
|
| 727 |
+
|
| 728 |
+
def diverse_beam_search(self, input_dict):
|
| 729 |
+
|
| 730 |
+
def add_diversity(seq_table, logprob, t, divm, diversity_lambda, bdash):
|
| 731 |
+
local_time = t - divm
|
| 732 |
+
unaug_logprob = logprob.clone()
|
| 733 |
+
|
| 734 |
+
if divm > 0:
|
| 735 |
+
change = torch.zeros(logprob.size(-1))
|
| 736 |
+
for prev_choice in range(divm):
|
| 737 |
+
prev_decisions = seq_table[prev_choice][..., local_time]
|
| 738 |
+
for prev_labels in range(bdash):
|
| 739 |
+
change.scatter_add_(0, prev_decisions[prev_labels], change.new_ones(1))
|
| 740 |
+
|
| 741 |
+
change = change.to(logprob.device)
|
| 742 |
+
logprob = logprob - repeat_tensor(change, bdash) * diversity_lambda
|
| 743 |
+
|
| 744 |
+
return logprob, unaug_logprob
|
| 745 |
+
|
| 746 |
+
output = self.prepare_output(input_dict)
|
| 747 |
+
group_size = input_dict["group_size"]
|
| 748 |
+
batch_size = output["seq"].size(0)
|
| 749 |
+
beam_size = input_dict["beam_size"]
|
| 750 |
+
bdash = beam_size // group_size
|
| 751 |
+
input_dict["bdash"] = bdash
|
| 752 |
+
diversity_lambda = input_dict["diversity_lambda"]
|
| 753 |
+
device = input_dict["fc_emb"].device
|
| 754 |
+
max_length = input_dict["max_length"]
|
| 755 |
+
temp = input_dict["temp"]
|
| 756 |
+
group_nbest = input_dict["group_nbest"]
|
| 757 |
+
batch_size, max_length = output["seq"].size()
|
| 758 |
+
if group_nbest:
|
| 759 |
+
output["seq"] = torch.full((batch_size, beam_size, max_length),
|
| 760 |
+
self.end_idx, dtype=torch.long)
|
| 761 |
+
else:
|
| 762 |
+
output["seq"] = torch.full((batch_size, group_size, max_length),
|
| 763 |
+
self.end_idx, dtype=torch.long)
|
| 764 |
+
|
| 765 |
+
|
| 766 |
+
for i in range(batch_size):
|
| 767 |
+
input_dict["sample_idx"] = i
|
| 768 |
+
seq_table = [torch.LongTensor(bdash, 0) for _ in range(group_size)] # group_size x [bdash, 0]
|
| 769 |
+
logprob_table = [torch.zeros(bdash).to(device) for _ in range(group_size)]
|
| 770 |
+
done_beams_table = [[] for _ in range(group_size)]
|
| 771 |
+
|
| 772 |
+
output_i = {
|
| 773 |
+
"prev_words_beam": [None for _ in range(group_size)],
|
| 774 |
+
"next_word": [None for _ in range(group_size)],
|
| 775 |
+
"state": [None for _ in range(group_size)]
|
| 776 |
+
}
|
| 777 |
+
|
| 778 |
+
for t in range(max_length + group_size - 1):
|
| 779 |
+
input_dict["t"] = t
|
| 780 |
+
for divm in range(group_size):
|
| 781 |
+
input_dict["divm"] = divm
|
| 782 |
+
if t >= divm and t <= max_length + divm - 1:
|
| 783 |
+
local_time = t - divm
|
| 784 |
+
decoder_input = self.prepare_dbs_decoder_input(input_dict, output_i)
|
| 785 |
+
output_t = self.decoder(decoder_input)
|
| 786 |
+
output_t["divm"] = divm
|
| 787 |
+
logit_t = output_t["logit"]
|
| 788 |
+
if logit_t.size(1) == 1:
|
| 789 |
+
logit_t = logit_t.squeeze(1)
|
| 790 |
+
elif logit_t.size(1) > 1:
|
| 791 |
+
logit_t = logit_t[:, -1, :]
|
| 792 |
+
else:
|
| 793 |
+
raise Exception("no logit output")
|
| 794 |
+
logprob_t = torch.log_softmax(logit_t, dim=1)
|
| 795 |
+
logprob_t = torch.log_softmax(logprob_t / temp, dim=1)
|
| 796 |
+
logprob_t, unaug_logprob_t = add_diversity(seq_table, logprob_t, t, divm, diversity_lambda, bdash)
|
| 797 |
+
logprob_t = logprob_table[divm].unsqueeze(-1) + logprob_t
|
| 798 |
+
if local_time == 0: # for the first step, all k seq will have the same probs
|
| 799 |
+
topk_logprob, topk_words = logprob_t[0].topk(
|
| 800 |
+
bdash, 0, True, True)
|
| 801 |
+
else: # unroll and find top logprob, and their unrolled indices
|
| 802 |
+
topk_logprob, topk_words = logprob_t.view(-1).topk(
|
| 803 |
+
bdash, 0, True, True)
|
| 804 |
+
topk_words = topk_words.cpu()
|
| 805 |
+
logprob_table[divm] = topk_logprob
|
| 806 |
+
output_i["prev_words_beam"][divm] = topk_words // self.vocab_size # [bdash,]
|
| 807 |
+
output_i["next_word"][divm] = topk_words % self.vocab_size # [bdash,]
|
| 808 |
+
if local_time > 0:
|
| 809 |
+
seq_table[divm] = seq_table[divm][output_i["prev_words_beam"][divm]]
|
| 810 |
+
seq_table[divm] = torch.cat([
|
| 811 |
+
seq_table[divm],
|
| 812 |
+
output_i["next_word"][divm].unsqueeze(-1)], -1)
|
| 813 |
+
|
| 814 |
+
is_end = seq_table[divm][:, t-divm] == self.end_idx
|
| 815 |
+
assert seq_table[divm].shape[-1] == t - divm + 1
|
| 816 |
+
if t == max_length + divm - 1:
|
| 817 |
+
is_end.fill_(1)
|
| 818 |
+
for beam_idx in range(bdash):
|
| 819 |
+
if is_end[beam_idx]:
|
| 820 |
+
final_beam = {
|
| 821 |
+
"seq": seq_table[divm][beam_idx].clone(),
|
| 822 |
+
"score": logprob_table[divm][beam_idx].item()
|
| 823 |
+
}
|
| 824 |
+
final_beam["score"] = final_beam["score"] / (t - divm + 1)
|
| 825 |
+
done_beams_table[divm].append(final_beam)
|
| 826 |
+
logprob_table[divm][is_end] -= 1000
|
| 827 |
+
self.dbs_process_step(output_i, output_t)
|
| 828 |
+
done_beams_table = [sorted(done_beams_table[divm], key=lambda x: -x["score"])[:bdash] for divm in range(group_size)]
|
| 829 |
+
if group_nbest:
|
| 830 |
+
done_beams = sum(done_beams_table, [])
|
| 831 |
+
else:
|
| 832 |
+
done_beams = [group_beam[0] for group_beam in done_beams_table]
|
| 833 |
+
for _, done_beam in enumerate(done_beams):
|
| 834 |
+
output["seq"][i, _, :len(done_beam["seq"])] = done_beam["seq"]
|
| 835 |
+
|
| 836 |
+
return output
|
| 837 |
+
|
| 838 |
+
def prepare_dbs_decoder_input(self, input_dict, output_i):
|
| 839 |
+
raise NotImplementedError
|
| 840 |
+
|
| 841 |
+
def dbs_process_step(self, output_i, output_t):
|
| 842 |
+
pass
|
| 843 |
+
|
| 844 |
+
|
| 845 |
+
class TransformerModel(CaptionModel):
|
| 846 |
+
|
| 847 |
+
def __init__(self, encoder: nn.Module, decoder: nn.Module, **kwargs):
|
| 848 |
+
if not hasattr(self, "compatible_decoders"):
|
| 849 |
+
self.compatible_decoders = (
|
| 850 |
+
TransformerDecoder,
|
| 851 |
+
)
|
| 852 |
+
super().__init__(encoder, decoder, **kwargs)
|
| 853 |
+
|
| 854 |
+
def seq_forward(self, input_dict):
|
| 855 |
+
cap = input_dict["cap"]
|
| 856 |
+
cap_padding_mask = (cap == self.pad_idx).to(cap.device)
|
| 857 |
+
cap_padding_mask = cap_padding_mask[:, :-1]
|
| 858 |
+
output = self.decoder(
|
| 859 |
+
{
|
| 860 |
+
"word": cap[:, :-1],
|
| 861 |
+
"attn_emb": input_dict["attn_emb"],
|
| 862 |
+
"attn_emb_len": input_dict["attn_emb_len"],
|
| 863 |
+
"cap_padding_mask": cap_padding_mask
|
| 864 |
+
}
|
| 865 |
+
)
|
| 866 |
+
return output
|
| 867 |
+
|
| 868 |
+
def prepare_decoder_input(self, input_dict, output):
|
| 869 |
+
decoder_input = {
|
| 870 |
+
"attn_emb": input_dict["attn_emb"],
|
| 871 |
+
"attn_emb_len": input_dict["attn_emb_len"]
|
| 872 |
+
}
|
| 873 |
+
t = input_dict["t"]
|
| 874 |
+
|
| 875 |
+
###############
|
| 876 |
+
# determine input word
|
| 877 |
+
################
|
| 878 |
+
if input_dict["mode"] == "train" and random.random() < input_dict["ss_ratio"]: # training, scheduled sampling
|
| 879 |
+
word = input_dict["cap"][:, :t+1]
|
| 880 |
+
else:
|
| 881 |
+
start_word = torch.tensor([self.start_idx,] * input_dict["attn_emb"].size(0)).unsqueeze(1).long()
|
| 882 |
+
if t == 0:
|
| 883 |
+
word = start_word
|
| 884 |
+
else:
|
| 885 |
+
word = torch.cat((start_word, output["seq"][:, :t]), dim=-1)
|
| 886 |
+
# word: [N, T]
|
| 887 |
+
decoder_input["word"] = word
|
| 888 |
+
|
| 889 |
+
cap_padding_mask = (word == self.pad_idx).to(input_dict["attn_emb"].device)
|
| 890 |
+
decoder_input["cap_padding_mask"] = cap_padding_mask
|
| 891 |
+
return decoder_input
|
| 892 |
+
|
| 893 |
+
def prepare_beamsearch_decoder_input(self, input_dict, output_i):
|
| 894 |
+
decoder_input = {}
|
| 895 |
+
t = input_dict["t"]
|
| 896 |
+
i = input_dict["sample_idx"]
|
| 897 |
+
beam_size = input_dict["beam_size"]
|
| 898 |
+
###############
|
| 899 |
+
# prepare attn embeds
|
| 900 |
+
################
|
| 901 |
+
if t == 0:
|
| 902 |
+
attn_emb = repeat_tensor(input_dict["attn_emb"][i], beam_size)
|
| 903 |
+
attn_emb_len = repeat_tensor(input_dict["attn_emb_len"][i], beam_size)
|
| 904 |
+
output_i["attn_emb"] = attn_emb
|
| 905 |
+
output_i["attn_emb_len"] = attn_emb_len
|
| 906 |
+
decoder_input["attn_emb"] = output_i["attn_emb"]
|
| 907 |
+
decoder_input["attn_emb_len"] = output_i["attn_emb_len"]
|
| 908 |
+
###############
|
| 909 |
+
# determine input word
|
| 910 |
+
################
|
| 911 |
+
start_word = torch.tensor([self.start_idx,] * beam_size).unsqueeze(1).long()
|
| 912 |
+
if t == 0:
|
| 913 |
+
word = start_word
|
| 914 |
+
else:
|
| 915 |
+
word = torch.cat((start_word, output_i["seq"]), dim=-1)
|
| 916 |
+
decoder_input["word"] = word
|
| 917 |
+
cap_padding_mask = (word == self.pad_idx).to(input_dict["attn_emb"].device)
|
| 918 |
+
decoder_input["cap_padding_mask"] = cap_padding_mask
|
| 919 |
+
|
| 920 |
+
return decoder_input
|
| 921 |
+
|
| 922 |
+
|
| 923 |
+
class BaseDecoder(nn.Module):
|
| 924 |
+
"""
|
| 925 |
+
Take word/audio embeddings and output the next word probs
|
| 926 |
+
"""
|
| 927 |
+
def __init__(self, emb_dim, vocab_size, fc_emb_dim,
|
| 928 |
+
attn_emb_dim, dropout=0.2, tie_weights=False):
|
| 929 |
+
super().__init__()
|
| 930 |
+
self.emb_dim = emb_dim
|
| 931 |
+
self.vocab_size = vocab_size
|
| 932 |
+
self.fc_emb_dim = fc_emb_dim
|
| 933 |
+
self.attn_emb_dim = attn_emb_dim
|
| 934 |
+
self.tie_weights = tie_weights
|
| 935 |
+
self.word_embedding = nn.Embedding(vocab_size, emb_dim)
|
| 936 |
+
self.in_dropout = nn.Dropout(dropout)
|
| 937 |
+
|
| 938 |
+
def forward(self, x):
|
| 939 |
+
raise NotImplementedError
|
| 940 |
+
|
| 941 |
+
def load_word_embedding(self, weight, freeze=True):
|
| 942 |
+
embedding = np.load(weight)
|
| 943 |
+
assert embedding.shape[0] == self.vocab_size, "vocabulary size mismatch"
|
| 944 |
+
assert embedding.shape[1] == self.emb_dim, "embed size mismatch"
|
| 945 |
+
|
| 946 |
+
# embeddings = torch.as_tensor(embeddings).float()
|
| 947 |
+
# self.word_embeddings.weight = nn.Parameter(embeddings)
|
| 948 |
+
# for para in self.word_embeddings.parameters():
|
| 949 |
+
# para.requires_grad = tune
|
| 950 |
+
self.word_embedding = nn.Embedding.from_pretrained(embedding,
|
| 951 |
+
freeze=freeze)
|
| 952 |
+
|
| 953 |
+
|
| 954 |
+
class PositionalEncoding(nn.Module):
|
| 955 |
+
|
| 956 |
+
def __init__(self, d_model, dropout=0.1, max_len=100):
|
| 957 |
+
super(PositionalEncoding, self).__init__()
|
| 958 |
+
self.dropout = nn.Dropout(p=dropout)
|
| 959 |
+
|
| 960 |
+
pe = torch.zeros(max_len, d_model)
|
| 961 |
+
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
|
| 962 |
+
div_term = torch.exp(torch.arange(0, d_model, 2).float() * \
|
| 963 |
+
(-math.log(10000.0) / d_model))
|
| 964 |
+
pe[:, 0::2] = torch.sin(position * div_term)
|
| 965 |
+
pe[:, 1::2] = torch.cos(position * div_term)
|
| 966 |
+
pe = pe.unsqueeze(0).transpose(0, 1)
|
| 967 |
+
# self.register_buffer("pe", pe)
|
| 968 |
+
self.register_parameter("pe", nn.Parameter(pe, requires_grad=False))
|
| 969 |
+
|
| 970 |
+
def forward(self, x):
|
| 971 |
+
# x: [T, N, E]
|
| 972 |
+
x = x + self.pe[:x.size(0), :]
|
| 973 |
+
return self.dropout(x)
|
| 974 |
+
|
| 975 |
+
|
| 976 |
+
class TransformerDecoder(BaseDecoder):
|
| 977 |
+
|
| 978 |
+
def __init__(self,
|
| 979 |
+
emb_dim,
|
| 980 |
+
vocab_size,
|
| 981 |
+
fc_emb_dim,
|
| 982 |
+
attn_emb_dim,
|
| 983 |
+
dropout,
|
| 984 |
+
freeze=False,
|
| 985 |
+
tie_weights=False,
|
| 986 |
+
**kwargs):
|
| 987 |
+
super().__init__(emb_dim, vocab_size, fc_emb_dim, attn_emb_dim,
|
| 988 |
+
dropout=dropout, tie_weights=tie_weights)
|
| 989 |
+
self.d_model = emb_dim
|
| 990 |
+
self.nhead = kwargs.get("nhead", self.d_model // 64)
|
| 991 |
+
self.nlayers = kwargs.get("nlayers", 2)
|
| 992 |
+
self.dim_feedforward = kwargs.get("dim_feedforward", self.d_model * 4)
|
| 993 |
+
|
| 994 |
+
self.pos_encoder = PositionalEncoding(self.d_model, dropout)
|
| 995 |
+
layer = nn.TransformerDecoderLayer(d_model=self.d_model,
|
| 996 |
+
nhead=self.nhead,
|
| 997 |
+
dim_feedforward=self.dim_feedforward,
|
| 998 |
+
dropout=dropout)
|
| 999 |
+
self.model = nn.TransformerDecoder(layer, self.nlayers)
|
| 1000 |
+
self.classifier = nn.Linear(self.d_model, vocab_size, bias=False)
|
| 1001 |
+
if tie_weights:
|
| 1002 |
+
self.classifier.weight = self.word_embedding.weight
|
| 1003 |
+
self.attn_proj = nn.Sequential(
|
| 1004 |
+
nn.Linear(self.attn_emb_dim, self.d_model),
|
| 1005 |
+
nn.ReLU(),
|
| 1006 |
+
nn.Dropout(dropout),
|
| 1007 |
+
nn.LayerNorm(self.d_model)
|
| 1008 |
+
)
|
| 1009 |
+
self.init_params()
|
| 1010 |
+
|
| 1011 |
+
self.freeze = freeze
|
| 1012 |
+
if freeze:
|
| 1013 |
+
for p in self.parameters():
|
| 1014 |
+
p.requires_grad = False
|
| 1015 |
+
|
| 1016 |
+
def init_params(self):
|
| 1017 |
+
for p in self.parameters():
|
| 1018 |
+
if p.dim() > 1:
|
| 1019 |
+
nn.init.xavier_uniform_(p)
|
| 1020 |
+
|
| 1021 |
+
def load_pretrained(self, pretrained, output_fn):
|
| 1022 |
+
checkpoint = torch.load(pretrained, map_location="cpu")
|
| 1023 |
+
|
| 1024 |
+
if "model" in checkpoint:
|
| 1025 |
+
checkpoint = checkpoint["model"]
|
| 1026 |
+
if next(iter(checkpoint)).startswith("decoder."):
|
| 1027 |
+
state_dict = {}
|
| 1028 |
+
for k, v in checkpoint.items():
|
| 1029 |
+
state_dict[k[8:]] = v
|
| 1030 |
+
|
| 1031 |
+
loaded_keys = merge_load_state_dict(state_dict, self, output_fn)
|
| 1032 |
+
if self.freeze:
|
| 1033 |
+
for name, param in self.named_parameters():
|
| 1034 |
+
if name in loaded_keys:
|
| 1035 |
+
param.requires_grad = False
|
| 1036 |
+
else:
|
| 1037 |
+
param.requires_grad = True
|
| 1038 |
+
|
| 1039 |
+
|
| 1040 |
+
def generate_square_subsequent_mask(self, max_length):
|
| 1041 |
+
mask = (torch.triu(torch.ones(max_length, max_length)) == 1).transpose(0, 1)
|
| 1042 |
+
mask = mask.float().masked_fill(mask == 0, float('-inf')).masked_fill(mask == 1, float(0.0))
|
| 1043 |
+
return mask
|
| 1044 |
+
|
| 1045 |
+
def forward(self, input_dict):
|
| 1046 |
+
word = input_dict["word"]
|
| 1047 |
+
attn_emb = input_dict["attn_emb"]
|
| 1048 |
+
attn_emb_len = input_dict["attn_emb_len"]
|
| 1049 |
+
cap_padding_mask = input_dict["cap_padding_mask"]
|
| 1050 |
+
|
| 1051 |
+
p_attn_emb = self.attn_proj(attn_emb)
|
| 1052 |
+
p_attn_emb = p_attn_emb.transpose(0, 1) # [T_src, N, emb_dim]
|
| 1053 |
+
word = word.to(attn_emb.device)
|
| 1054 |
+
embed = self.in_dropout(self.word_embedding(word)) * math.sqrt(self.emb_dim) # [N, T, emb_dim]
|
| 1055 |
+
embed = embed.transpose(0, 1) # [T, N, emb_dim]
|
| 1056 |
+
embed = self.pos_encoder(embed)
|
| 1057 |
+
|
| 1058 |
+
tgt_mask = self.generate_square_subsequent_mask(embed.size(0)).to(attn_emb.device)
|
| 1059 |
+
memory_key_padding_mask = ~generate_length_mask(attn_emb_len, attn_emb.size(1)).to(attn_emb.device)
|
| 1060 |
+
output = self.model(embed, p_attn_emb, tgt_mask=tgt_mask,
|
| 1061 |
+
tgt_key_padding_mask=cap_padding_mask,
|
| 1062 |
+
memory_key_padding_mask=memory_key_padding_mask)
|
| 1063 |
+
output = output.transpose(0, 1)
|
| 1064 |
+
output = {
|
| 1065 |
+
"embed": output,
|
| 1066 |
+
"logit": self.classifier(output),
|
| 1067 |
+
}
|
| 1068 |
+
return output
|
| 1069 |
+
|
| 1070 |
+
|
| 1071 |
+
class ContraEncoderKdWrapper(nn.Module, CaptionMetaMixin):
|
| 1072 |
+
|
| 1073 |
+
def __init__(self,
|
| 1074 |
+
model: nn.Module,
|
| 1075 |
+
shared_dim: int,
|
| 1076 |
+
tchr_dim: int,
|
| 1077 |
+
):
|
| 1078 |
+
super().__init__()
|
| 1079 |
+
self.model = model
|
| 1080 |
+
self.tchr_dim = tchr_dim
|
| 1081 |
+
if hasattr(model, "encoder"):
|
| 1082 |
+
self.stdnt_proj = nn.Linear(model.encoder.fc_emb_size,
|
| 1083 |
+
shared_dim)
|
| 1084 |
+
else:
|
| 1085 |
+
self.stdnt_proj = nn.Linear(model.fc_emb_size,
|
| 1086 |
+
shared_dim)
|
| 1087 |
+
self.tchr_proj = nn.Linear(tchr_dim, shared_dim)
|
| 1088 |
+
self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
|
| 1089 |
+
|
| 1090 |
+
def forward(self, input_dict: Dict):
|
| 1091 |
+
unsup = input_dict.get("unsup", False)
|
| 1092 |
+
if unsup is False:
|
| 1093 |
+
output_dict = self.model(input_dict)
|
| 1094 |
+
else:
|
| 1095 |
+
output_dict = self.model.encoder(input_dict)
|
| 1096 |
+
if "tchr_output" in input_dict:
|
| 1097 |
+
stdnt_emb = output_dict["fc_emb"]
|
| 1098 |
+
stdnt_emb = self.stdnt_proj(stdnt_emb)
|
| 1099 |
+
tchr_emb = input_dict["tchr_output"]["embedding"]
|
| 1100 |
+
thcr_emb = self.tchr_proj(tchr_emb)
|
| 1101 |
+
|
| 1102 |
+
stdnt_emb = F.normalize(stdnt_emb, dim=-1)
|
| 1103 |
+
thcr_emb = F.normalize(thcr_emb, dim=-1)
|
| 1104 |
+
|
| 1105 |
+
unscaled_logit = stdnt_emb @ thcr_emb.transpose(0, 1)
|
| 1106 |
+
logit = self.logit_scale * unscaled_logit
|
| 1107 |
+
label = torch.arange(logit.shape[0]).to(logit.device)
|
| 1108 |
+
loss1 = F.cross_entropy(logit, label)
|
| 1109 |
+
loss2 = F.cross_entropy(logit.transpose(0, 1), label)
|
| 1110 |
+
loss = (loss1 + loss2) / 2
|
| 1111 |
+
output_dict["enc_kd_loss"] = loss
|
| 1112 |
+
return output_dict
|
| 1113 |
+
|
| 1114 |
+
|
| 1115 |
+
class Effb2TrmConfig(PretrainedConfig):
|
| 1116 |
+
|
| 1117 |
+
def __init__(
|
| 1118 |
+
self,
|
| 1119 |
+
sample_rate: int = 16000,
|
| 1120 |
+
tchr_dim: int = 768,
|
| 1121 |
+
shared_dim: int = 1024,
|
| 1122 |
+
fc_emb_dim: int = 1408,
|
| 1123 |
+
attn_emb_dim: int = 1408,
|
| 1124 |
+
decoder_n_layers: int = 2,
|
| 1125 |
+
decoder_we_tie_weights: bool = True,
|
| 1126 |
+
decoder_emb_dim: int = 256,
|
| 1127 |
+
decoder_dropout: float = 0.2,
|
| 1128 |
+
vocab_size: int = 4981,
|
| 1129 |
+
**kwargs
|
| 1130 |
+
):
|
| 1131 |
+
self.sample_rate = sample_rate
|
| 1132 |
+
self.tchr_dim = tchr_dim
|
| 1133 |
+
self.shared_dim = shared_dim
|
| 1134 |
+
self.fc_emb_dim = fc_emb_dim
|
| 1135 |
+
self.attn_emb_dim = attn_emb_dim
|
| 1136 |
+
self.decoder_n_layers = decoder_n_layers
|
| 1137 |
+
self.decoder_we_tie_weights = decoder_we_tie_weights
|
| 1138 |
+
self.decoder_emb_dim = decoder_emb_dim
|
| 1139 |
+
self.decoder_dropout = decoder_dropout
|
| 1140 |
+
self.vocab_size = vocab_size
|
| 1141 |
+
super().__init__(**kwargs)
|
| 1142 |
+
|
| 1143 |
+
|
| 1144 |
+
class Effb2TrmCaptioningModel(PreTrainedModel):
|
| 1145 |
+
config_class = Effb2TrmConfig
|
| 1146 |
+
|
| 1147 |
+
def __init__(self, config):
|
| 1148 |
+
super().__init__(config)
|
| 1149 |
+
encoder = EfficientNetB2()
|
| 1150 |
+
decoder = TransformerDecoder(
|
| 1151 |
+
emb_dim=config.decoder_emb_dim,
|
| 1152 |
+
vocab_size=config.vocab_size,
|
| 1153 |
+
fc_emb_dim=config.fc_emb_dim,
|
| 1154 |
+
attn_emb_dim=config.attn_emb_dim,
|
| 1155 |
+
dropout=config.decoder_dropout,
|
| 1156 |
+
nlayers=config.decoder_n_layers,
|
| 1157 |
+
tie_weights=config.decoder_we_tie_weights
|
| 1158 |
+
)
|
| 1159 |
+
model = TransformerModel(encoder, decoder)
|
| 1160 |
+
self.model = ContraEncoderKdWrapper(model, config.shared_dim, config.tchr_dim)
|
| 1161 |
+
|
| 1162 |
+
def forward(self,
|
| 1163 |
+
audio: torch.Tensor,
|
| 1164 |
+
audio_length: Union[List, np.ndarray, torch.Tensor],
|
| 1165 |
+
sample_method: str = "beam",
|
| 1166 |
+
beam_size: int = 3,
|
| 1167 |
+
max_length: int = 20,
|
| 1168 |
+
temp: float = 1.0,):
|
| 1169 |
+
device = self.device
|
| 1170 |
+
input_dict = {
|
| 1171 |
+
"wav": audio.to(device),
|
| 1172 |
+
"wav_len": audio_length,
|
| 1173 |
+
"specaug": False,
|
| 1174 |
+
"mode": "inference",
|
| 1175 |
+
"sample_method": sample_method,
|
| 1176 |
+
"max_length": max_length,
|
| 1177 |
+
"temp": temp,
|
| 1178 |
+
}
|
| 1179 |
+
if sample_method == "beam":
|
| 1180 |
+
input_dict["beam_size"] = beam_size
|
| 1181 |
+
return self.model(input_dict)["seq"].cpu()
|
| 1182 |
+
|
| 1183 |
+
|
| 1184 |
+
class ConvBlock(nn.Module):
|
| 1185 |
+
|
| 1186 |
+
def __init__(self, in_channels, out_channels):
|
| 1187 |
+
|
| 1188 |
+
super(ConvBlock, self).__init__()
|
| 1189 |
+
|
| 1190 |
+
self.conv1 = nn.Conv2d(in_channels=in_channels,
|
| 1191 |
+
out_channels=out_channels,
|
| 1192 |
+
kernel_size=(3, 3), stride=(1, 1),
|
| 1193 |
+
padding=(1, 1), bias=False)
|
| 1194 |
+
|
| 1195 |
+
self.conv2 = nn.Conv2d(in_channels=out_channels,
|
| 1196 |
+
out_channels=out_channels,
|
| 1197 |
+
kernel_size=(3, 3), stride=(1, 1),
|
| 1198 |
+
padding=(1, 1), bias=False)
|
| 1199 |
+
|
| 1200 |
+
self.bn1 = nn.BatchNorm2d(out_channels)
|
| 1201 |
+
self.bn2 = nn.BatchNorm2d(out_channels)
|
| 1202 |
+
|
| 1203 |
+
def forward(self, input, pool_size=(2, 2), pool_type='avg'):
|
| 1204 |
+
|
| 1205 |
+
x = input
|
| 1206 |
+
x = F.relu_(self.bn1(self.conv1(x)))
|
| 1207 |
+
x = F.relu_(self.bn2(self.conv2(x)))
|
| 1208 |
+
if pool_type == 'max':
|
| 1209 |
+
x = F.max_pool2d(x, kernel_size=pool_size)
|
| 1210 |
+
elif pool_type == 'avg':
|
| 1211 |
+
x = F.avg_pool2d(x, kernel_size=pool_size)
|
| 1212 |
+
elif pool_type == 'avg+max':
|
| 1213 |
+
x1 = F.avg_pool2d(x, kernel_size=pool_size)
|
| 1214 |
+
x2 = F.max_pool2d(x, kernel_size=pool_size)
|
| 1215 |
+
x = x1 + x2
|
| 1216 |
+
else:
|
| 1217 |
+
raise Exception('Incorrect argument!')
|
| 1218 |
+
|
| 1219 |
+
return x
|
| 1220 |
+
|
| 1221 |
+
|
| 1222 |
+
class Cnn14Encoder(nn.Module):
|
| 1223 |
+
|
| 1224 |
+
def __init__(self, sample_rate=32000):
|
| 1225 |
+
super().__init__()
|
| 1226 |
+
sr_to_fmax = {
|
| 1227 |
+
32000: 14000,
|
| 1228 |
+
16000: 8000
|
| 1229 |
+
}
|
| 1230 |
+
# Logmel spectrogram extractor
|
| 1231 |
+
self.melspec_extractor = transforms.MelSpectrogram(
|
| 1232 |
+
sample_rate=sample_rate,
|
| 1233 |
+
n_fft=32 * sample_rate // 1000,
|
| 1234 |
+
win_length=32 * sample_rate // 1000,
|
| 1235 |
+
hop_length=10 * sample_rate // 1000,
|
| 1236 |
+
f_min=50,
|
| 1237 |
+
f_max=sr_to_fmax[sample_rate],
|
| 1238 |
+
n_mels=64,
|
| 1239 |
+
norm="slaney",
|
| 1240 |
+
mel_scale="slaney"
|
| 1241 |
+
)
|
| 1242 |
+
self.hop_length = 10 * sample_rate // 1000
|
| 1243 |
+
self.db_transform = transforms.AmplitudeToDB()
|
| 1244 |
+
|
| 1245 |
+
self.bn0 = nn.BatchNorm2d(64)
|
| 1246 |
+
|
| 1247 |
+
self.conv_block1 = ConvBlock(in_channels=1, out_channels=64)
|
| 1248 |
+
self.conv_block2 = ConvBlock(in_channels=64, out_channels=128)
|
| 1249 |
+
self.conv_block3 = ConvBlock(in_channels=128, out_channels=256)
|
| 1250 |
+
self.conv_block4 = ConvBlock(in_channels=256, out_channels=512)
|
| 1251 |
+
self.conv_block5 = ConvBlock(in_channels=512, out_channels=1024)
|
| 1252 |
+
self.conv_block6 = ConvBlock(in_channels=1024, out_channels=2048)
|
| 1253 |
+
|
| 1254 |
+
self.downsample_ratio = 32
|
| 1255 |
+
|
| 1256 |
+
self.fc1 = nn.Linear(2048, 2048, bias=True)
|
| 1257 |
+
self.fc_emb_size = 2048
|
| 1258 |
+
|
| 1259 |
+
def forward(self, input_dict):
|
| 1260 |
+
lms = input_dict["lms"]
|
| 1261 |
+
wave_length = input_dict["wav_len"]
|
| 1262 |
+
|
| 1263 |
+
x = lms # (batch_size, mel_bins, time_steps)
|
| 1264 |
+
x = x.transpose(1, 2)
|
| 1265 |
+
x = x.unsqueeze(1) # (batch_size, 1, time_steps, mel_bins)
|
| 1266 |
+
|
| 1267 |
+
x = x.transpose(1, 3)
|
| 1268 |
+
x = self.bn0(x)
|
| 1269 |
+
x = x.transpose(1, 3)
|
| 1270 |
+
|
| 1271 |
+
x = self.conv_block1(x, pool_size=(2, 2), pool_type='avg')
|
| 1272 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
| 1273 |
+
x = self.conv_block2(x, pool_size=(2, 2), pool_type='avg')
|
| 1274 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
| 1275 |
+
x = self.conv_block3(x, pool_size=(2, 2), pool_type='avg')
|
| 1276 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
| 1277 |
+
x = self.conv_block4(x, pool_size=(2, 2), pool_type='avg')
|
| 1278 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
| 1279 |
+
x = self.conv_block5(x, pool_size=(2, 2), pool_type='avg')
|
| 1280 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
| 1281 |
+
x = self.conv_block6(x, pool_size=(1, 1), pool_type='avg')
|
| 1282 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
| 1283 |
+
x = torch.mean(x, dim=3)
|
| 1284 |
+
attn_emb = x.transpose(1, 2)
|
| 1285 |
+
|
| 1286 |
+
wave_length = torch.as_tensor(wave_length)
|
| 1287 |
+
feat_length = torch.div(wave_length, self.hop_length,
|
| 1288 |
+
rounding_mode="floor") + 1
|
| 1289 |
+
feat_length = torch.div(feat_length, self.downsample_ratio,
|
| 1290 |
+
rounding_mode="floor")
|
| 1291 |
+
x_max = max_with_lens(attn_emb, feat_length)
|
| 1292 |
+
x_mean = mean_with_lens(attn_emb, feat_length)
|
| 1293 |
+
x = x_max + x_mean
|
| 1294 |
+
x = F.dropout(x, p=0.5, training=self.training)
|
| 1295 |
+
x = F.relu_(self.fc1(x))
|
| 1296 |
+
fc_emb = F.dropout(x, p=0.5, training=self.training)
|
| 1297 |
+
|
| 1298 |
+
output_dict = {
|
| 1299 |
+
'fc_emb': fc_emb,
|
| 1300 |
+
'attn_emb': attn_emb,
|
| 1301 |
+
'attn_emb_len': feat_length
|
| 1302 |
+
}
|
| 1303 |
+
|
| 1304 |
+
return output_dict
|
| 1305 |
+
|
| 1306 |
+
|
| 1307 |
+
class RnnEncoder(nn.Module):
|
| 1308 |
+
|
| 1309 |
+
def __init__(self,
|
| 1310 |
+
attn_feat_dim,
|
| 1311 |
+
pooling="mean",
|
| 1312 |
+
**kwargs):
|
| 1313 |
+
super().__init__()
|
| 1314 |
+
self.pooling = pooling
|
| 1315 |
+
self.hidden_size = kwargs.get('hidden_size', 512)
|
| 1316 |
+
self.bidirectional = kwargs.get('bidirectional', False)
|
| 1317 |
+
self.num_layers = kwargs.get('num_layers', 1)
|
| 1318 |
+
self.dropout = kwargs.get('dropout', 0.2)
|
| 1319 |
+
self.rnn_type = kwargs.get('rnn_type', "GRU")
|
| 1320 |
+
self.in_bn = kwargs.get('in_bn', False)
|
| 1321 |
+
self.embed_dim = self.hidden_size * (self.bidirectional + 1)
|
| 1322 |
+
self.network = getattr(nn, self.rnn_type)(
|
| 1323 |
+
attn_feat_dim,
|
| 1324 |
+
self.hidden_size,
|
| 1325 |
+
num_layers=self.num_layers,
|
| 1326 |
+
bidirectional=self.bidirectional,
|
| 1327 |
+
dropout=self.dropout,
|
| 1328 |
+
batch_first=True)
|
| 1329 |
+
if self.in_bn:
|
| 1330 |
+
self.bn = nn.BatchNorm1d(self.embed_dim)
|
| 1331 |
+
|
| 1332 |
+
def forward(self, input_dict):
|
| 1333 |
+
x = input_dict["attn"]
|
| 1334 |
+
lens = input_dict["attn_len"]
|
| 1335 |
+
lens = torch.as_tensor(lens)
|
| 1336 |
+
# x: [N, T, E]
|
| 1337 |
+
if self.in_bn:
|
| 1338 |
+
x = pack_wrapper(self.bn, x, lens)
|
| 1339 |
+
out = pack_wrapper(self.network, x, lens)
|
| 1340 |
+
# out: [N, T, hidden]
|
| 1341 |
+
attn_emb = out
|
| 1342 |
+
fc_emb = embedding_pooling(out, lens, self.pooling)
|
| 1343 |
+
return {
|
| 1344 |
+
"attn_emb": attn_emb,
|
| 1345 |
+
"fc_emb": fc_emb,
|
| 1346 |
+
"attn_emb_len": lens
|
| 1347 |
+
}
|
| 1348 |
+
|
| 1349 |
+
|
| 1350 |
+
class Cnn14RnnEncoder(nn.Module):
|
| 1351 |
+
|
| 1352 |
+
def __init__(self,
|
| 1353 |
+
sample_rate,
|
| 1354 |
+
rnn_bidirectional,
|
| 1355 |
+
rnn_hidden_size,
|
| 1356 |
+
rnn_dropout,
|
| 1357 |
+
rnn_num_layers):
|
| 1358 |
+
super().__init__()
|
| 1359 |
+
self.cnn = Cnn14Encoder(sample_rate=sample_rate)
|
| 1360 |
+
self.rnn = RnnEncoder(
|
| 1361 |
+
2048,
|
| 1362 |
+
bidirectional=rnn_bidirectional,
|
| 1363 |
+
hidden_size=rnn_hidden_size,
|
| 1364 |
+
dropout=rnn_dropout,
|
| 1365 |
+
num_layers=rnn_num_layers,
|
| 1366 |
+
)
|
| 1367 |
+
|
| 1368 |
+
def forward(self, input_dict):
|
| 1369 |
+
output_dict = self.cnn(input_dict)
|
| 1370 |
+
output_dict["attn"] = output_dict["attn_emb"]
|
| 1371 |
+
output_dict["attn_len"] = output_dict["attn_emb_len"]
|
| 1372 |
+
del output_dict["attn_emb"], output_dict["attn_emb_len"]
|
| 1373 |
+
output_dict = self.rnn(output_dict)
|
| 1374 |
+
return output_dict
|
| 1375 |
+
|
| 1376 |
+
|
| 1377 |
+
class Seq2SeqAttention(nn.Module):
|
| 1378 |
+
|
| 1379 |
+
def __init__(self, hs_enc, hs_dec, attn_size):
|
| 1380 |
+
"""
|
| 1381 |
+
Args:
|
| 1382 |
+
hs_enc: encoder hidden size
|
| 1383 |
+
hs_dec: decoder hidden size
|
| 1384 |
+
attn_size: attention vector size
|
| 1385 |
+
"""
|
| 1386 |
+
super(Seq2SeqAttention, self).__init__()
|
| 1387 |
+
self.h2attn = nn.Linear(hs_enc + hs_dec, attn_size)
|
| 1388 |
+
self.v = nn.Parameter(torch.randn(attn_size))
|
| 1389 |
+
|
| 1390 |
+
def forward(self, h_dec, h_enc, src_lens):
|
| 1391 |
+
"""
|
| 1392 |
+
Args:
|
| 1393 |
+
h_dec: decoder hidden (query), [N, hs_dec]
|
| 1394 |
+
h_enc: encoder memory (key/value), [N, src_max_len, hs_enc]
|
| 1395 |
+
src_lens: source (encoder memory) lengths, [N, ]
|
| 1396 |
+
"""
|
| 1397 |
+
N = h_enc.size(0)
|
| 1398 |
+
src_max_len = h_enc.size(1)
|
| 1399 |
+
h_dec = h_dec.unsqueeze(1).repeat(1, src_max_len, 1) # [N, src_max_len, hs_dec]
|
| 1400 |
+
|
| 1401 |
+
attn_input = torch.cat((h_dec, h_enc), dim=-1)
|
| 1402 |
+
attn_out = torch.tanh(self.h2attn(attn_input)) # [N, src_max_len, attn_size]
|
| 1403 |
+
|
| 1404 |
+
v = self.v.repeat(N, 1).unsqueeze(1) # [N, 1, attn_size]
|
| 1405 |
+
score = torch.bmm(v, attn_out.transpose(1, 2)).squeeze(1) # [N, src_max_len]
|
| 1406 |
+
|
| 1407 |
+
idxs = torch.arange(src_max_len).repeat(N).view(N, src_max_len)
|
| 1408 |
+
mask = (idxs < src_lens.view(-1, 1)).to(h_dec.device)
|
| 1409 |
+
|
| 1410 |
+
score = score.masked_fill(mask == 0, -1e10)
|
| 1411 |
+
weights = torch.softmax(score, dim=-1) # [N, src_max_len]
|
| 1412 |
+
ctx = torch.bmm(weights.unsqueeze(1), h_enc).squeeze(1) # [N, hs_enc]
|
| 1413 |
+
|
| 1414 |
+
return ctx, weights
|
| 1415 |
+
|
| 1416 |
+
|
| 1417 |
+
class RnnDecoder(BaseDecoder):
|
| 1418 |
+
|
| 1419 |
+
def __init__(self, emb_dim, vocab_size, fc_emb_dim, attn_emb_dim,
|
| 1420 |
+
dropout, d_model, **kwargs):
|
| 1421 |
+
super().__init__(emb_dim, vocab_size, fc_emb_dim, attn_emb_dim,
|
| 1422 |
+
dropout,)
|
| 1423 |
+
self.d_model = d_model
|
| 1424 |
+
self.num_layers = kwargs.get('num_layers', 1)
|
| 1425 |
+
self.bidirectional = kwargs.get('bidirectional', False)
|
| 1426 |
+
self.rnn_type = kwargs.get('rnn_type', "GRU")
|
| 1427 |
+
self.classifier = nn.Linear(
|
| 1428 |
+
self.d_model * (self.bidirectional + 1), vocab_size)
|
| 1429 |
+
|
| 1430 |
+
def forward(self, x):
|
| 1431 |
+
raise NotImplementedError
|
| 1432 |
+
|
| 1433 |
+
def init_hidden(self, bs, device):
|
| 1434 |
+
num_dire = self.bidirectional + 1
|
| 1435 |
+
n_layer = self.num_layers
|
| 1436 |
+
hid_dim = self.d_model
|
| 1437 |
+
if self.rnn_type == "LSTM":
|
| 1438 |
+
return (torch.zeros(num_dire * n_layer, bs, hid_dim).to(device),
|
| 1439 |
+
torch.zeros(num_dire * n_layer, bs, hid_dim).to(device))
|
| 1440 |
+
else:
|
| 1441 |
+
return torch.zeros(num_dire * n_layer, bs, hid_dim).to(device)
|
| 1442 |
+
|
| 1443 |
+
|
| 1444 |
+
class BahAttnCatFcDecoder(RnnDecoder):
|
| 1445 |
+
|
| 1446 |
+
def __init__(self, emb_dim, vocab_size, fc_emb_dim, attn_emb_dim,
|
| 1447 |
+
dropout, d_model, **kwargs):
|
| 1448 |
+
"""
|
| 1449 |
+
concatenate fc, attn, word to feed to the rnn
|
| 1450 |
+
"""
|
| 1451 |
+
super().__init__(emb_dim, vocab_size, fc_emb_dim, attn_emb_dim,
|
| 1452 |
+
dropout, d_model, **kwargs)
|
| 1453 |
+
attn_size = kwargs.get("attn_size", self.d_model)
|
| 1454 |
+
self.model = getattr(nn, self.rnn_type)(
|
| 1455 |
+
input_size=self.emb_dim * 3,
|
| 1456 |
+
hidden_size=self.d_model,
|
| 1457 |
+
batch_first=True,
|
| 1458 |
+
num_layers=self.num_layers,
|
| 1459 |
+
bidirectional=self.bidirectional)
|
| 1460 |
+
self.attn = Seq2SeqAttention(self.attn_emb_dim,
|
| 1461 |
+
self.d_model * (self.bidirectional + 1) * \
|
| 1462 |
+
self.num_layers,
|
| 1463 |
+
attn_size)
|
| 1464 |
+
self.fc_proj = nn.Linear(self.fc_emb_dim, self.emb_dim)
|
| 1465 |
+
self.ctx_proj = nn.Linear(self.attn_emb_dim, self.emb_dim)
|
| 1466 |
+
|
| 1467 |
+
def forward(self, input_dict):
|
| 1468 |
+
word = input_dict["word"]
|
| 1469 |
+
state = input_dict.get("state", None) # [n_layer * n_dire, bs, d_model]
|
| 1470 |
+
fc_emb = input_dict["fc_emb"]
|
| 1471 |
+
attn_emb = input_dict["attn_emb"]
|
| 1472 |
+
attn_emb_len = input_dict["attn_emb_len"]
|
| 1473 |
+
|
| 1474 |
+
word = word.to(fc_emb.device)
|
| 1475 |
+
embed = self.in_dropout(self.word_embedding(word))
|
| 1476 |
+
|
| 1477 |
+
# embed: [N, 1, embed_size]
|
| 1478 |
+
if state is None:
|
| 1479 |
+
state = self.init_hidden(word.size(0), fc_emb.device)
|
| 1480 |
+
if self.rnn_type == "LSTM":
|
| 1481 |
+
query = state[0].transpose(0, 1).flatten(1)
|
| 1482 |
+
else:
|
| 1483 |
+
query = state.transpose(0, 1).flatten(1)
|
| 1484 |
+
c, attn_weight = self.attn(query, attn_emb, attn_emb_len)
|
| 1485 |
+
|
| 1486 |
+
p_fc_emb = self.fc_proj(fc_emb)
|
| 1487 |
+
p_ctx = self.ctx_proj(c)
|
| 1488 |
+
rnn_input = torch.cat((embed, p_ctx.unsqueeze(1), p_fc_emb.unsqueeze(1)),
|
| 1489 |
+
dim=-1)
|
| 1490 |
+
|
| 1491 |
+
out, state = self.model(rnn_input, state)
|
| 1492 |
+
|
| 1493 |
+
output = {
|
| 1494 |
+
"state": state,
|
| 1495 |
+
"embed": out,
|
| 1496 |
+
"logit": self.classifier(out),
|
| 1497 |
+
"attn_weight": attn_weight
|
| 1498 |
+
}
|
| 1499 |
+
return output
|
| 1500 |
+
|
| 1501 |
+
|
| 1502 |
+
class TemporalBahAttnDecoder(BahAttnCatFcDecoder):
|
| 1503 |
+
|
| 1504 |
+
def __init__(self, emb_dim, vocab_size, fc_emb_dim, attn_emb_dim,
|
| 1505 |
+
dropout, d_model, **kwargs):
|
| 1506 |
+
"""
|
| 1507 |
+
concatenate fc, attn, word to feed to the rnn
|
| 1508 |
+
"""
|
| 1509 |
+
super().__init__(emb_dim, vocab_size, fc_emb_dim, attn_emb_dim,
|
| 1510 |
+
dropout, d_model, **kwargs)
|
| 1511 |
+
self.temporal_embedding = nn.Embedding(4, emb_dim)
|
| 1512 |
+
|
| 1513 |
+
def forward(self, input_dict):
|
| 1514 |
+
word = input_dict["word"]
|
| 1515 |
+
state = input_dict.get("state", None) # [n_layer * n_dire, bs, d_model]
|
| 1516 |
+
fc_embs = input_dict["fc_emb"]
|
| 1517 |
+
attn_embs = input_dict["attn_emb"]
|
| 1518 |
+
attn_emb_lens = input_dict["attn_emb_len"]
|
| 1519 |
+
temporal_tag = input_dict["temporal_tag"]
|
| 1520 |
+
|
| 1521 |
+
if input_dict["t"] == 0:
|
| 1522 |
+
embed = self.in_dropout(
|
| 1523 |
+
self.temporal_embedding(temporal_tag)).unsqueeze(1)
|
| 1524 |
+
elif word.size(-1) == self.fc_emb_dim: # fc_embs
|
| 1525 |
+
embed = word.unsqueeze(1)
|
| 1526 |
+
elif word.size(-1) == 1: # word
|
| 1527 |
+
word = word.to(fc_embs.device)
|
| 1528 |
+
embed = self.in_dropout(self.word_embedding(word))
|
| 1529 |
+
else:
|
| 1530 |
+
raise Exception(f"problem with word input size {word.size()}")
|
| 1531 |
+
|
| 1532 |
+
# embed: [N, 1, embed_size]
|
| 1533 |
+
if state is None:
|
| 1534 |
+
state = self.init_hidden(word.size(0), fc_embs.device)
|
| 1535 |
+
if self.rnn_type == "LSTM":
|
| 1536 |
+
query = state[0].transpose(0, 1).flatten(1)
|
| 1537 |
+
else:
|
| 1538 |
+
query = state.transpose(0, 1).flatten(1)
|
| 1539 |
+
c, attn_weight = self.attn(query, attn_embs, attn_emb_lens)
|
| 1540 |
+
|
| 1541 |
+
p_ctx = self.ctx_proj(c)
|
| 1542 |
+
p_fc_embs = self.fc_proj(fc_embs)
|
| 1543 |
+
p_ctx = self.ctx_proj(c)
|
| 1544 |
+
rnn_input = torch.cat((embed, p_ctx.unsqueeze(1), p_fc_embs.unsqueeze(1)), dim=-1)
|
| 1545 |
+
|
| 1546 |
+
out, state = self.model(rnn_input, state)
|
| 1547 |
+
|
| 1548 |
+
output = {
|
| 1549 |
+
"state": state,
|
| 1550 |
+
"embed": out,
|
| 1551 |
+
"logit": self.classifier(out),
|
| 1552 |
+
"attn_weight": attn_weight
|
| 1553 |
+
}
|
| 1554 |
+
return output
|
| 1555 |
+
|
| 1556 |
+
|
| 1557 |
+
class Seq2SeqAttnModel(CaptionModel):
|
| 1558 |
+
|
| 1559 |
+
def __init__(self, encoder, decoder, **kwargs):
|
| 1560 |
+
if not hasattr(self, "compatible_decoders"):
|
| 1561 |
+
self.compatible_decoders = (
|
| 1562 |
+
BahAttnCatFcDecoder,
|
| 1563 |
+
)
|
| 1564 |
+
super().__init__(encoder, decoder, **kwargs)
|
| 1565 |
+
|
| 1566 |
+
|
| 1567 |
+
def seq_forward(self, input_dict):
|
| 1568 |
+
# Bahdanau attention only supports step-by-step implementation, so we implement forward in
|
| 1569 |
+
# step-by-step manner whether in training or evaluation
|
| 1570 |
+
return self.stepwise_forward(input_dict)
|
| 1571 |
+
|
| 1572 |
+
def prepare_output(self, input_dict):
|
| 1573 |
+
output = super().prepare_output(input_dict)
|
| 1574 |
+
attn_weight = torch.empty(output["seq"].size(0),
|
| 1575 |
+
input_dict["attn_emb"].size(1), output["seq"].size(1))
|
| 1576 |
+
output["attn_weight"] = attn_weight
|
| 1577 |
+
return output
|
| 1578 |
+
|
| 1579 |
+
def prepare_decoder_input(self, input_dict, output):
|
| 1580 |
+
decoder_input = {
|
| 1581 |
+
"fc_emb": input_dict["fc_emb"],
|
| 1582 |
+
"attn_emb": input_dict["attn_emb"],
|
| 1583 |
+
"attn_emb_len": input_dict["attn_emb_len"]
|
| 1584 |
+
}
|
| 1585 |
+
t = input_dict["t"]
|
| 1586 |
+
###############
|
| 1587 |
+
# determine input word
|
| 1588 |
+
################
|
| 1589 |
+
if input_dict["mode"] == "train" and random.random() < input_dict["ss_ratio"]: # training, scheduled sampling
|
| 1590 |
+
word = input_dict["cap"][:, t]
|
| 1591 |
+
else:
|
| 1592 |
+
if t == 0:
|
| 1593 |
+
word = torch.tensor([self.start_idx,] * input_dict["fc_emb"].size(0)).long()
|
| 1594 |
+
else:
|
| 1595 |
+
word = output["seq"][:, t-1]
|
| 1596 |
+
# word: [N,]
|
| 1597 |
+
decoder_input["word"] = word.unsqueeze(1)
|
| 1598 |
+
|
| 1599 |
+
################
|
| 1600 |
+
# prepare rnn state
|
| 1601 |
+
################
|
| 1602 |
+
if t > 0:
|
| 1603 |
+
decoder_input["state"] = output["state"]
|
| 1604 |
+
return decoder_input
|
| 1605 |
+
|
| 1606 |
+
def stepwise_process_step(self, output, output_t):
|
| 1607 |
+
super().stepwise_process_step(output, output_t)
|
| 1608 |
+
output["state"] = output_t["state"]
|
| 1609 |
+
t = output_t["t"]
|
| 1610 |
+
output["attn_weight"][:, :, t] = output_t["attn_weight"]
|
| 1611 |
+
|
| 1612 |
+
def prepare_beamsearch_output(self, input_dict):
|
| 1613 |
+
output = super().prepare_beamsearch_output(input_dict)
|
| 1614 |
+
beam_size = input_dict["beam_size"]
|
| 1615 |
+
max_length = input_dict["max_length"]
|
| 1616 |
+
output["attn_weight"] = torch.empty(beam_size,
|
| 1617 |
+
max(input_dict["attn_emb_len"]), max_length)
|
| 1618 |
+
return output
|
| 1619 |
+
|
| 1620 |
+
def prepare_beamsearch_decoder_input(self, input_dict, output_i):
|
| 1621 |
+
decoder_input = {}
|
| 1622 |
+
t = input_dict["t"]
|
| 1623 |
+
i = input_dict["sample_idx"]
|
| 1624 |
+
beam_size = input_dict["beam_size"]
|
| 1625 |
+
###############
|
| 1626 |
+
# prepare fc embeds
|
| 1627 |
+
################
|
| 1628 |
+
if t == 0:
|
| 1629 |
+
fc_emb = repeat_tensor(input_dict["fc_emb"][i], beam_size)
|
| 1630 |
+
output_i["fc_emb"] = fc_emb
|
| 1631 |
+
decoder_input["fc_emb"] = output_i["fc_emb"]
|
| 1632 |
+
|
| 1633 |
+
###############
|
| 1634 |
+
# prepare attn embeds
|
| 1635 |
+
################
|
| 1636 |
+
if t == 0:
|
| 1637 |
+
attn_emb = repeat_tensor(input_dict["attn_emb"][i], beam_size)
|
| 1638 |
+
attn_emb_len = repeat_tensor(input_dict["attn_emb_len"][i], beam_size)
|
| 1639 |
+
output_i["attn_emb"] = attn_emb
|
| 1640 |
+
output_i["attn_emb_len"] = attn_emb_len
|
| 1641 |
+
decoder_input["attn_emb"] = output_i["attn_emb"]
|
| 1642 |
+
decoder_input["attn_emb_len"] = output_i["attn_emb_len"]
|
| 1643 |
+
|
| 1644 |
+
###############
|
| 1645 |
+
# determine input word
|
| 1646 |
+
################
|
| 1647 |
+
if t == 0:
|
| 1648 |
+
word = torch.tensor([self.start_idx,] * beam_size).long()
|
| 1649 |
+
else:
|
| 1650 |
+
word = output_i["next_word"]
|
| 1651 |
+
decoder_input["word"] = word.unsqueeze(1)
|
| 1652 |
+
|
| 1653 |
+
################
|
| 1654 |
+
# prepare rnn state
|
| 1655 |
+
################
|
| 1656 |
+
if t > 0:
|
| 1657 |
+
if self.decoder.rnn_type == "LSTM":
|
| 1658 |
+
decoder_input["state"] = (output_i["state"][0][:, output_i["prev_words_beam"], :].contiguous(),
|
| 1659 |
+
output_i["state"][1][:, output_i["prev_words_beam"], :].contiguous())
|
| 1660 |
+
else:
|
| 1661 |
+
decoder_input["state"] = output_i["state"][:, output_i["prev_words_beam"], :].contiguous()
|
| 1662 |
+
|
| 1663 |
+
return decoder_input
|
| 1664 |
+
|
| 1665 |
+
def beamsearch_process_step(self, output_i, output_t):
|
| 1666 |
+
t = output_t["t"]
|
| 1667 |
+
output_i["state"] = output_t["state"]
|
| 1668 |
+
output_i["attn_weight"][..., t] = output_t["attn_weight"]
|
| 1669 |
+
output_i["attn_weight"] = output_i["attn_weight"][output_i["prev_words_beam"], ...]
|
| 1670 |
+
|
| 1671 |
+
def beamsearch_process(self, output, output_i, input_dict):
|
| 1672 |
+
super().beamsearch_process(output, output_i, input_dict)
|
| 1673 |
+
i = input_dict["sample_idx"]
|
| 1674 |
+
output["attn_weight"][i] = output_i["attn_weight"][0]
|
| 1675 |
+
|
| 1676 |
+
def prepare_dbs_decoder_input(self, input_dict, output_i):
|
| 1677 |
+
decoder_input = {}
|
| 1678 |
+
t = input_dict["t"]
|
| 1679 |
+
i = input_dict["sample_idx"]
|
| 1680 |
+
bdash = input_dict["bdash"]
|
| 1681 |
+
divm = input_dict["divm"]
|
| 1682 |
+
|
| 1683 |
+
local_time = t - divm
|
| 1684 |
+
###############
|
| 1685 |
+
# prepare fc embeds
|
| 1686 |
+
################
|
| 1687 |
+
# repeat only at the first timestep to save consumption
|
| 1688 |
+
if t == 0:
|
| 1689 |
+
fc_emb = repeat_tensor(input_dict["fc_emb"][i], bdash).unsqueeze(1)
|
| 1690 |
+
output_i["fc_emb"] = fc_emb
|
| 1691 |
+
decoder_input["fc_emb"] = output_i["fc_emb"]
|
| 1692 |
+
|
| 1693 |
+
###############
|
| 1694 |
+
# prepare attn embeds
|
| 1695 |
+
################
|
| 1696 |
+
if t == 0:
|
| 1697 |
+
attn_emb = repeat_tensor(input_dict["attn_emb"][i], bdash)
|
| 1698 |
+
attn_emb_len = repeat_tensor(input_dict["attn_emb_len"][i], bdash)
|
| 1699 |
+
output_i["attn_emb"] = attn_emb
|
| 1700 |
+
output_i["attn_emb_len"] = attn_emb_len
|
| 1701 |
+
decoder_input["attn_emb"] = output_i["attn_emb"]
|
| 1702 |
+
decoder_input["attn_emb_len"] = output_i["attn_emb_len"]
|
| 1703 |
+
|
| 1704 |
+
###############
|
| 1705 |
+
# determine input word
|
| 1706 |
+
################
|
| 1707 |
+
if local_time == 0:
|
| 1708 |
+
word = torch.tensor([self.start_idx,] * bdash).long()
|
| 1709 |
+
else:
|
| 1710 |
+
word = output_i["next_word"][divm]
|
| 1711 |
+
decoder_input["word"] = word.unsqueeze(1)
|
| 1712 |
+
|
| 1713 |
+
################
|
| 1714 |
+
# prepare rnn state
|
| 1715 |
+
################
|
| 1716 |
+
if local_time > 0:
|
| 1717 |
+
if self.decoder.rnn_type == "LSTM":
|
| 1718 |
+
decoder_input["state"] = (
|
| 1719 |
+
output_i["state"][0][divm][
|
| 1720 |
+
:, output_i["prev_words_beam"][divm], :].contiguous(),
|
| 1721 |
+
output_i["state"][1][divm][
|
| 1722 |
+
:, output_i["prev_words_beam"][divm], :].contiguous()
|
| 1723 |
+
)
|
| 1724 |
+
else:
|
| 1725 |
+
decoder_input["state"] = output_i["state"][divm][
|
| 1726 |
+
:, output_i["prev_words_beam"][divm], :].contiguous()
|
| 1727 |
+
|
| 1728 |
+
return decoder_input
|
| 1729 |
+
|
| 1730 |
+
def dbs_process_step(self, output_i, output_t):
|
| 1731 |
+
divm = output_t["divm"]
|
| 1732 |
+
output_i["state"][divm] = output_t["state"]
|
| 1733 |
+
# TODO attention weight
|
| 1734 |
+
|
| 1735 |
+
|
| 1736 |
+
class TemporalSeq2SeqAttnModel(Seq2SeqAttnModel):
|
| 1737 |
+
|
| 1738 |
+
def __init__(self, encoder, decoder, **kwargs):
|
| 1739 |
+
if not hasattr(self, "compatible_decoders"):
|
| 1740 |
+
self.compatible_decoders = (
|
| 1741 |
+
TemporalBahAttnDecoder,
|
| 1742 |
+
)
|
| 1743 |
+
super().__init__(encoder, decoder, **kwargs)
|
| 1744 |
+
self.train_forward_keys = ["cap", "cap_len", "ss_ratio", "temporal_tag"]
|
| 1745 |
+
self.inference_forward_keys = ["sample_method", "max_length", "temp", "temporal_tag"]
|
| 1746 |
+
|
| 1747 |
+
|
| 1748 |
+
def prepare_decoder_input(self, input_dict, output):
|
| 1749 |
+
decoder_input = super().prepare_decoder_input(input_dict, output)
|
| 1750 |
+
decoder_input["temporal_tag"] = input_dict["temporal_tag"]
|
| 1751 |
+
decoder_input["t"] = input_dict["t"]
|
| 1752 |
+
|
| 1753 |
+
return decoder_input
|
| 1754 |
+
|
| 1755 |
+
|
| 1756 |
+
def prepare_beamsearch_decoder_input(self, input_dict, output_i):
|
| 1757 |
+
decoder_input = super().prepare_beamsearch_decoder_input(input_dict, output_i)
|
| 1758 |
+
t = input_dict["t"]
|
| 1759 |
+
i = input_dict["sample_idx"]
|
| 1760 |
+
beam_size = input_dict["beam_size"]
|
| 1761 |
+
###############
|
| 1762 |
+
# prepare temporal_tag
|
| 1763 |
+
################
|
| 1764 |
+
if t == 0:
|
| 1765 |
+
temporal_tag = repeat_tensor(input_dict["temporal_tag"][i], beam_size)
|
| 1766 |
+
output_i["temporal_tag"] = temporal_tag
|
| 1767 |
+
decoder_input["temporal_tag"] = output_i["temporal_tag"]
|
| 1768 |
+
decoder_input["t"] = input_dict["t"]
|
| 1769 |
+
|
| 1770 |
+
return decoder_input
|
| 1771 |
+
|
| 1772 |
+
def prepare_dbs_decoder_input(self, input_dict, output_i):
|
| 1773 |
+
decoder_input = super.prepare_dbs_decoder_input(input_dict, output_i)
|
| 1774 |
+
t = input_dict["t"]
|
| 1775 |
+
i = input_dict["sample_idx"]
|
| 1776 |
+
bdash = input_dict["bdash"]
|
| 1777 |
+
|
| 1778 |
+
###############
|
| 1779 |
+
# prepare temporal tag
|
| 1780 |
+
################
|
| 1781 |
+
# repeat only at the first timestep to save consumption
|
| 1782 |
+
if t == 0:
|
| 1783 |
+
temporal_tag = repeat_tensor(input_dict["temporal_tag"][i], bdash)
|
| 1784 |
+
output_i["temporal_tag"] = temporal_tag
|
| 1785 |
+
decoder_input["temporal_tag"] = output_i["temporal_tag"]
|
| 1786 |
+
decoder_input["t"] = input_dict["t"]
|
| 1787 |
+
|
| 1788 |
+
return decoder_input
|
| 1789 |
+
|
| 1790 |
+
|
| 1791 |
+
class Cnn8rnnSedModel(nn.Module):
|
| 1792 |
+
def __init__(self, classes_num):
|
| 1793 |
+
|
| 1794 |
+
super().__init__()
|
| 1795 |
+
|
| 1796 |
+
self.time_resolution = 0.01
|
| 1797 |
+
self.interpolate_ratio = 4 # Downsampled ratio
|
| 1798 |
+
|
| 1799 |
+
self.bn0 = nn.BatchNorm2d(64)
|
| 1800 |
+
|
| 1801 |
+
self.conv_block1 = ConvBlock(in_channels=1, out_channels=64)
|
| 1802 |
+
self.conv_block2 = ConvBlock(in_channels=64, out_channels=128)
|
| 1803 |
+
self.conv_block3 = ConvBlock(in_channels=128, out_channels=256)
|
| 1804 |
+
self.conv_block4 = ConvBlock(in_channels=256, out_channels=512)
|
| 1805 |
+
|
| 1806 |
+
self.fc1 = nn.Linear(512, 512, bias=True)
|
| 1807 |
+
self.rnn = nn.GRU(512, 256, bidirectional=True, batch_first=True)
|
| 1808 |
+
self.fc_audioset = nn.Linear(512, classes_num, bias=True)
|
| 1809 |
+
|
| 1810 |
+
def forward(self, lms):
|
| 1811 |
+
output = self.forward_prob(lms)
|
| 1812 |
+
framewise_output = output["framewise_output"].cpu().numpy()
|
| 1813 |
+
thresholded_predictions = double_threshold(
|
| 1814 |
+
framewise_output, 0.75, 0.25)
|
| 1815 |
+
decoded_tags = decode_with_timestamps(
|
| 1816 |
+
thresholded_predictions, self.time_resolution
|
| 1817 |
+
)
|
| 1818 |
+
return decoded_tags
|
| 1819 |
+
|
| 1820 |
+
def forward_prob(self, lms):
|
| 1821 |
+
"""
|
| 1822 |
+
lms: (batch_size, mel_bins, time_steps)"""
|
| 1823 |
+
|
| 1824 |
+
x = lms
|
| 1825 |
+
x = x.transpose(1, 2)
|
| 1826 |
+
x = x.unsqueeze(1)
|
| 1827 |
+
|
| 1828 |
+
frames_num = x.shape[2]
|
| 1829 |
+
|
| 1830 |
+
x = x.transpose(1, 3)
|
| 1831 |
+
x = self.bn0(x)
|
| 1832 |
+
x = x.transpose(1, 3)
|
| 1833 |
+
|
| 1834 |
+
x = self.conv_block1(x, pool_size=(2, 2), pool_type='avg+max')
|
| 1835 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
| 1836 |
+
x = self.conv_block2(x, pool_size=(2, 2), pool_type='avg+max')
|
| 1837 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
| 1838 |
+
x = self.conv_block3(x, pool_size=(1, 2), pool_type='avg+max')
|
| 1839 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
| 1840 |
+
x = self.conv_block4(x, pool_size=(1, 2), pool_type='avg+max')
|
| 1841 |
+
x = F.dropout(x, p=0.2, training=self.training) # (batch_size, 256, time_steps / 4, mel_bins / 16)
|
| 1842 |
+
x = torch.mean(x, dim=3)
|
| 1843 |
+
|
| 1844 |
+
x = x.transpose(1, 2)
|
| 1845 |
+
x = F.dropout(x, p=0.5, training=self.training)
|
| 1846 |
+
x = F.relu_(self.fc1(x))
|
| 1847 |
+
x, _ = self.rnn(x)
|
| 1848 |
+
segmentwise_output = torch.sigmoid(self.fc_audioset(x)).clamp(1e-7, 1.)
|
| 1849 |
+
|
| 1850 |
+
framewise_output = interpolate(segmentwise_output,
|
| 1851 |
+
self.interpolate_ratio)
|
| 1852 |
+
framewise_output = pad_framewise_output(framewise_output, frames_num)
|
| 1853 |
+
|
| 1854 |
+
output_dict = {
|
| 1855 |
+
"segmentwise_output": segmentwise_output,
|
| 1856 |
+
'framewise_output': framewise_output,
|
| 1857 |
+
}
|
| 1858 |
+
|
| 1859 |
+
return output_dict
|
| 1860 |
+
|
| 1861 |
+
|
| 1862 |
+
class Cnn14RnnTempAttnGruConfig(PretrainedConfig):
|
| 1863 |
+
|
| 1864 |
+
def __init__(
|
| 1865 |
+
self,
|
| 1866 |
+
sample_rate: int = 32000,
|
| 1867 |
+
encoder_rnn_bidirectional: bool = True,
|
| 1868 |
+
encoder_rnn_hidden_size: int = 256,
|
| 1869 |
+
encoder_rnn_dropout: float = 0.5,
|
| 1870 |
+
encoder_rnn_num_layers: int = 3,
|
| 1871 |
+
decoder_emb_dim: int = 512,
|
| 1872 |
+
vocab_size: int = 4981,
|
| 1873 |
+
fc_emb_dim: int = 512,
|
| 1874 |
+
attn_emb_dim: int = 512,
|
| 1875 |
+
decoder_rnn_type: str = "GRU",
|
| 1876 |
+
decoder_num_layers: int = 1,
|
| 1877 |
+
decoder_d_model: int = 512,
|
| 1878 |
+
decoder_dropout: float = 0.5,
|
| 1879 |
+
**kwargs
|
| 1880 |
+
):
|
| 1881 |
+
self.sample_rate = sample_rate
|
| 1882 |
+
self.encoder_rnn_bidirectional = encoder_rnn_bidirectional
|
| 1883 |
+
self.encoder_rnn_hidden_size = encoder_rnn_hidden_size
|
| 1884 |
+
self.encoder_rnn_dropout = encoder_rnn_dropout
|
| 1885 |
+
self.encoder_rnn_num_layers = encoder_rnn_num_layers
|
| 1886 |
+
self.decoder_emb_dim = decoder_emb_dim
|
| 1887 |
+
self.vocab_size = vocab_size
|
| 1888 |
+
self.fc_emb_dim = fc_emb_dim
|
| 1889 |
+
self.attn_emb_dim = attn_emb_dim
|
| 1890 |
+
self.decoder_rnn_type = decoder_rnn_type
|
| 1891 |
+
self.decoder_num_layers = decoder_num_layers
|
| 1892 |
+
self.decoder_d_model = decoder_d_model
|
| 1893 |
+
self.decoder_dropout = decoder_dropout
|
| 1894 |
+
super().__init__(**kwargs)
|
| 1895 |
+
|
| 1896 |
+
|
| 1897 |
+
class Cnn14RnnTempAttnGruModel(PreTrainedModel):
|
| 1898 |
+
config_class = Cnn14RnnTempAttnGruConfig
|
| 1899 |
+
|
| 1900 |
+
def __init__(self, config):
|
| 1901 |
+
super().__init__(config)
|
| 1902 |
+
sample_rate = config.sample_rate
|
| 1903 |
+
sr_to_fmax = {
|
| 1904 |
+
32000: 14000,
|
| 1905 |
+
16000: 8000
|
| 1906 |
+
}
|
| 1907 |
+
self.melspec_extractor = transforms.MelSpectrogram(
|
| 1908 |
+
sample_rate=sample_rate,
|
| 1909 |
+
n_fft=32 * sample_rate // 1000,
|
| 1910 |
+
win_length=32 * sample_rate // 1000,
|
| 1911 |
+
hop_length=10 * sample_rate // 1000,
|
| 1912 |
+
f_min=50,
|
| 1913 |
+
f_max=sr_to_fmax[sample_rate],
|
| 1914 |
+
n_mels=64,
|
| 1915 |
+
norm="slaney",
|
| 1916 |
+
mel_scale="slaney"
|
| 1917 |
+
)
|
| 1918 |
+
self.db_transform = transforms.AmplitudeToDB()
|
| 1919 |
+
|
| 1920 |
+
encoder = Cnn14RnnEncoder(
|
| 1921 |
+
sample_rate=config.sample_rate,
|
| 1922 |
+
rnn_bidirectional=config.encoder_rnn_bidirectional,
|
| 1923 |
+
rnn_hidden_size=config.encoder_rnn_hidden_size,
|
| 1924 |
+
rnn_dropout=config.encoder_rnn_dropout,
|
| 1925 |
+
rnn_num_layers=config.encoder_rnn_num_layers
|
| 1926 |
+
)
|
| 1927 |
+
decoder = TemporalBahAttnDecoder(
|
| 1928 |
+
emb_dim=config.decoder_emb_dim,
|
| 1929 |
+
vocab_size=config.vocab_size,
|
| 1930 |
+
fc_emb_dim=config.fc_emb_dim,
|
| 1931 |
+
attn_emb_dim=config.attn_emb_dim,
|
| 1932 |
+
rnn_type=config.decoder_rnn_type,
|
| 1933 |
+
num_layers=config.decoder_num_layers,
|
| 1934 |
+
d_model=config.decoder_d_model,
|
| 1935 |
+
dropout=config.decoder_dropout,
|
| 1936 |
+
)
|
| 1937 |
+
cap_model = TemporalSeq2SeqAttnModel(encoder, decoder)
|
| 1938 |
+
sed_model = Cnn8rnnSedModel(classes_num=447)
|
| 1939 |
+
self.cap_model = cap_model
|
| 1940 |
+
self.sed_model = sed_model
|
| 1941 |
+
|
| 1942 |
+
def forward(self,
|
| 1943 |
+
audio: torch.Tensor,
|
| 1944 |
+
audio_length: Union[List, np.ndarray, torch.Tensor],
|
| 1945 |
+
temporal_tag: Union[List, np.ndarray, torch.Tensor] = None,
|
| 1946 |
+
sample_method: str = "beam",
|
| 1947 |
+
beam_size: int = 3,
|
| 1948 |
+
max_length: int = 20,
|
| 1949 |
+
temp: float = 1.0,):
|
| 1950 |
+
device = self.device
|
| 1951 |
+
mel_spec = self.melspec_extractor(audio.to(device))
|
| 1952 |
+
log_mel_spec = self.db_transform(mel_spec)
|
| 1953 |
+
|
| 1954 |
+
sed_tag = self.sed_model(log_mel_spec)
|
| 1955 |
+
sed_tag = torch.as_tensor(sed_tag).to(device)
|
| 1956 |
+
if temporal_tag is not None:
|
| 1957 |
+
temporal_tag = torch.as_tensor(temporal_tag).to(device)
|
| 1958 |
+
temporal_tag = torch.stack([temporal_tag, sed_tag], dim=0)
|
| 1959 |
+
temporal_tag = torch.min(temporal_tag, dim=0).values
|
| 1960 |
+
else:
|
| 1961 |
+
temporal_tag = sed_tag
|
| 1962 |
+
|
| 1963 |
+
input_dict = {
|
| 1964 |
+
"lms": log_mel_spec,
|
| 1965 |
+
"wav_len": audio_length,
|
| 1966 |
+
"temporal_tag": temporal_tag,
|
| 1967 |
+
"mode": "inference",
|
| 1968 |
+
"sample_method": sample_method,
|
| 1969 |
+
"max_length": max_length,
|
| 1970 |
+
"temp": temp,
|
| 1971 |
+
}
|
| 1972 |
+
if sample_method == "beam":
|
| 1973 |
+
input_dict["beam_size"] = beam_size
|
| 1974 |
+
return self.cap_model(input_dict)["seq"].cpu()
|
pytorch_model.bin
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:49356c485fbafb022c03b351009a4634d91bd89eee01baf3c31e82e836b0fa3a
|
| 3 |
+
size 54696313
|