Spaces:
Build error
Build error
Update networks.py
Browse files- networks.py +5 -73
networks.py
CHANGED
@@ -1,79 +1,11 @@
|
|
1 |
-
import peft
|
2 |
-
import torch
|
3 |
-
import whisperx
|
4 |
-
import torch.nn as nn
|
5 |
from config import Config
|
6 |
-
from transformers import CLIPVisionModel, AutoModelForCausalLM
|
7 |
|
|
|
8 |
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
torch_dtype=torch.float16,
|
13 |
-
#device_map="cuda",
|
14 |
-
low_cpu_mem_usage=True,
|
15 |
-
return_dict=True,
|
16 |
-
trust_remote_code=True)
|
17 |
-
|
18 |
-
peft_model = peft.PeftModel.from_pretrained(text_model, 'models/29000')
|
19 |
-
projection = load_projection_model("models/MModalGPT-FINETUNE-step=29000-loss=3.45.ckpt", 768, 2560)
|
20 |
-
|
21 |
-
clip_model = CLIPVisionModel.from_pretrained(model_name)
|
22 |
-
audio_model = whisperx.load_model("small", device.type, compute_type="float16")
|
23 |
-
|
24 |
|
25 |
projection = projection.to(device)
|
26 |
peft_model = peft_model.to(device)
|
27 |
-
clip_model = clip_model.to(device)
|
28 |
-
|
29 |
-
|
30 |
-
def load_projection_model(path, clip_embed, phi_embed):
|
31 |
-
"""Loads a Projections model instance from a checkpoint and returns it with weights loaded.
|
32 |
-
|
33 |
-
Args:
|
34 |
-
path (str): Path to the checkpoint file.
|
35 |
-
|
36 |
-
Returns:
|
37 |
-
torch.nn.Module: The loaded Projections model instance.
|
38 |
-
"""
|
39 |
-
|
40 |
-
state_dict = torch.load(path)['state_dict']
|
41 |
-
new_state_dict = {k.replace('projection.', ''): v for k, v in state_dict.items()}
|
42 |
-
|
43 |
-
model = Projections(clip_embed, phi_embed)
|
44 |
-
model.load_state_dict(new_state_dict)
|
45 |
-
|
46 |
-
return model
|
47 |
-
|
48 |
-
|
49 |
-
class Projections(nn.Module):
|
50 |
-
def __init__(
|
51 |
-
self,
|
52 |
-
clip_embed,
|
53 |
-
phi_embed,
|
54 |
-
num_projection_layers=6,
|
55 |
-
):
|
56 |
-
super().__init__()
|
57 |
-
|
58 |
-
self.norm = nn.LayerNorm(phi_embed)
|
59 |
-
self.output = nn.Linear(clip_embed, phi_embed)
|
60 |
-
self.projection_layers = nn.ModuleList(
|
61 |
-
[
|
62 |
-
nn.Sequential(
|
63 |
-
nn.Linear(phi_embed, phi_embed),
|
64 |
-
nn.GELU(),
|
65 |
-
nn.Linear(phi_embed, phi_embed),
|
66 |
-
)
|
67 |
-
for _ in range(num_projection_layers)
|
68 |
-
]
|
69 |
-
)
|
70 |
-
|
71 |
-
def forward(self, x):
|
72 |
-
x = self.output(x)
|
73 |
-
self.norm(x)
|
74 |
-
for layer in self.projection_layers:
|
75 |
-
residual = x
|
76 |
-
x = layer(x) + residual
|
77 |
-
|
78 |
-
return x
|
79 |
-
|
|
|
|
|
|
|
|
|
|
|
1 |
from config import Config
|
|
|
2 |
|
3 |
+
device = Config.device
|
4 |
|
5 |
+
audio_model = Config.audio_model
|
6 |
+
text_model, peft_model = Config.text_model, Config.peft_model
|
7 |
+
projection, clip_model = Config.projection, Config.clip_model
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
|
9 |
projection = projection.to(device)
|
10 |
peft_model = peft_model.to(device)
|
11 |
+
clip_model = clip_model.to(device)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|