Upload 4 files
Browse files- bsq2048/config.yml +40 -0
- bsq2048/pytorch_model.bin +3 -0
- bsq32/config.yml +40 -0
- bsq32/pytorch_model.bin +3 -0
bsq2048/config.yml
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
_target_: default_model.AstralQuantizer
|
| 2 |
+
tokenizer_name: "openai/whisper-small"
|
| 3 |
+
ssl_model_name: "facebook/hubert-large-ll60k"
|
| 4 |
+
ssl_output_layer: 18
|
| 5 |
+
encoder:
|
| 6 |
+
_target_: modules.convnext.ConvNeXtV2Stage
|
| 7 |
+
dim: 512
|
| 8 |
+
num_blocks: 12
|
| 9 |
+
intermediate_dim: 1536
|
| 10 |
+
dilation: 1
|
| 11 |
+
input_dim: 1024
|
| 12 |
+
quantizer:
|
| 13 |
+
_target_: modules.bsq.BinarySphericalQuantize
|
| 14 |
+
codebook_size: 2048 # codebook size, must be a power of 2
|
| 15 |
+
dim: 512
|
| 16 |
+
entropy_loss_weight: 0.1
|
| 17 |
+
diversity_gamma: 1.0
|
| 18 |
+
spherical: True
|
| 19 |
+
enable_entropy_loss: True
|
| 20 |
+
soft_entropy_loss: True
|
| 21 |
+
decoder:
|
| 22 |
+
_target_: modules.convnext.ConvNeXtV2Stage
|
| 23 |
+
dim: 512
|
| 24 |
+
num_blocks: 12
|
| 25 |
+
intermediate_dim: 1536
|
| 26 |
+
dilation: 1
|
| 27 |
+
output_dim: 1024
|
| 28 |
+
gin_channels: 192
|
| 29 |
+
asr_decoder:
|
| 30 |
+
_target_: modules.asr_decoder.ASRDecoder
|
| 31 |
+
hidden_dim: 768
|
| 32 |
+
num_heads: 12
|
| 33 |
+
depth: 12
|
| 34 |
+
block_size: 4096
|
| 35 |
+
in_channels: 512
|
| 36 |
+
n_vocab: 51866
|
| 37 |
+
bos_id: 50528
|
| 38 |
+
eos_id: 50527
|
| 39 |
+
dropout_rate: 0.0
|
| 40 |
+
attn_dropout_rate: 0.0
|
bsq2048/pytorch_model.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:198cc8b2bbde6c1818cf44f748d146f80b08438532b6f0170ddbef78f7e98c2f
|
| 3 |
+
size 1912057564
|
bsq32/config.yml
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
_target_: default_model.AstralQuantizer
|
| 2 |
+
tokenizer_name: "openai/whisper-small"
|
| 3 |
+
ssl_model_name: "facebook/hubert-large-ll60k"
|
| 4 |
+
ssl_output_layer: 18
|
| 5 |
+
encoder:
|
| 6 |
+
_target_: modules.convnext.ConvNeXtV2Stage
|
| 7 |
+
dim: 512
|
| 8 |
+
num_blocks: 12
|
| 9 |
+
intermediate_dim: 1536
|
| 10 |
+
dilation: 1
|
| 11 |
+
input_dim: 1024
|
| 12 |
+
quantizer:
|
| 13 |
+
_target_: modules.bsq.BinarySphericalQuantize
|
| 14 |
+
codebook_size: 32 # codebook size, must be a power of 2
|
| 15 |
+
dim: 512
|
| 16 |
+
entropy_loss_weight: 0.1
|
| 17 |
+
diversity_gamma: 1.0
|
| 18 |
+
spherical: True
|
| 19 |
+
enable_entropy_loss: True
|
| 20 |
+
soft_entropy_loss: True
|
| 21 |
+
decoder:
|
| 22 |
+
_target_: modules.convnext.ConvNeXtV2Stage
|
| 23 |
+
dim: 512
|
| 24 |
+
num_blocks: 12
|
| 25 |
+
intermediate_dim: 1536
|
| 26 |
+
dilation: 1
|
| 27 |
+
output_dim: 1024
|
| 28 |
+
gin_channels: 192
|
| 29 |
+
asr_decoder:
|
| 30 |
+
_target_: modules.asr_decoder.ASRDecoder
|
| 31 |
+
hidden_dim: 768
|
| 32 |
+
num_heads: 12
|
| 33 |
+
depth: 12
|
| 34 |
+
block_size: 4096
|
| 35 |
+
in_channels: 512
|
| 36 |
+
n_vocab: 51866
|
| 37 |
+
bos_id: 50528
|
| 38 |
+
eos_id: 50527
|
| 39 |
+
dropout_rate: 0.0
|
| 40 |
+
attn_dropout_rate: 0.0
|
bsq32/pytorch_model.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fc023addd00827f1c462d096c9da37840c42421f64b36665a3207f6b8358ea5a
|
| 3 |
+
size 1912031480
|