Update README.md
Browse files
README.md
CHANGED
@@ -8,4 +8,81 @@ tags:
|
|
8 |
- audio
|
9 |
---
|
10 |
|
11 |
-
ONNX-compatible weights for https://huggingface.co/kyutai/mimi
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
- audio
|
9 |
---
|
10 |
|
11 |
+
ONNX-compatible weights for https://huggingface.co/kyutai/mimi
|
12 |
+
|
13 |
+
## Inference sample code
|
14 |
+
```py
|
15 |
+
import onnxruntime as ort
|
16 |
+
|
17 |
+
encoder_session = ort.InferenceSession("encoder_model.onnx")
|
18 |
+
decoder_session = ort.InferenceSession("decoder_model.onnx")
|
19 |
+
|
20 |
+
encoder_inputs = {encoder_session.get_inputs()[0].name: dummy_encoder_inputs.numpy()}
|
21 |
+
encoder_outputs = encoder_session.run(None, encoder_inputs)[0]
|
22 |
+
|
23 |
+
decoder_inputs = {decoder_session.get_inputs()[0].name: encoder_outputs}
|
24 |
+
decoder_outputs = decoder_session.run(None, decoder_inputs)[0]
|
25 |
+
|
26 |
+
# Print the results
|
27 |
+
print("Encoder Output Shape:", encoder_outputs.shape)
|
28 |
+
print("Decoder Output Shape:", decoder_outputs.shape)
|
29 |
+
```
|
30 |
+
|
31 |
+
## Conversion sample code
|
32 |
+
```py
|
33 |
+
import torch
|
34 |
+
import torch.nn as nn
|
35 |
+
from transformers import MimiModel
|
36 |
+
|
37 |
+
class MimiEncoder(nn.Module):
|
38 |
+
def __init__(self, model):
|
39 |
+
super(MimiEncoder, self).__init__()
|
40 |
+
self.model = model
|
41 |
+
|
42 |
+
def forward(self, input_values, padding_mask=None):
|
43 |
+
return self.model.encode(input_values, padding_mask=padding_mask).audio_codes
|
44 |
+
|
45 |
+
class MimiDecoder(nn.Module):
|
46 |
+
def __init__(self, model):
|
47 |
+
super(MimiDecoder, self).__init__()
|
48 |
+
self.model = model
|
49 |
+
|
50 |
+
def forward(self, audio_codes, padding_mask=None):
|
51 |
+
return self.model.decode(audio_codes, padding_mask=padding_mask).audio_values
|
52 |
+
|
53 |
+
model = MimiModel.from_pretrained("kyutai/mimi")
|
54 |
+
encoder = MimiEncoder(model)
|
55 |
+
decoder = MimiDecoder(model)
|
56 |
+
|
57 |
+
dummy_encoder_inputs = torch.randn((5, 1, 82500))
|
58 |
+
torch.onnx.export(
|
59 |
+
encoder,
|
60 |
+
dummy_encoder_inputs,
|
61 |
+
"encoder_model.onnx",
|
62 |
+
export_params=True,
|
63 |
+
opset_version=14,
|
64 |
+
do_constant_folding=True,
|
65 |
+
input_names=['input_values'],
|
66 |
+
output_names=['audio_codes'],
|
67 |
+
dynamic_axes={
|
68 |
+
'input_values': {0: 'batch_size', 1: 'num_channels', 2: 'sequence_length'},
|
69 |
+
'audio_codes': {0: 'batch_size', 2: 'codes_length'},
|
70 |
+
},
|
71 |
+
)
|
72 |
+
|
73 |
+
dummy_decoder_inputs = torch.randint(100, (4, 32, 91))
|
74 |
+
torch.onnx.export(
|
75 |
+
decoder,
|
76 |
+
dummy_decoder_inputs,
|
77 |
+
"decoder_model.onnx",
|
78 |
+
export_params=True,
|
79 |
+
opset_version=14,
|
80 |
+
do_constant_folding=True,
|
81 |
+
input_names=['audio_codes'],
|
82 |
+
output_names=['audio_values'],
|
83 |
+
dynamic_axes={
|
84 |
+
'audio_codes': {0: 'batch_size', 2: 'codes_length'},
|
85 |
+
'audio_values': {0: 'batch_size', 1: 'num_channels', 2: 'sequence_length'},
|
86 |
+
},
|
87 |
+
)
|
88 |
+
```
|