Made changes for audio
Browse files
app.py
CHANGED
@@ -1,7 +1,7 @@
|
|
1 |
import torch
|
2 |
import torch.nn as nn
|
3 |
import gradio as gr
|
4 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
5 |
|
6 |
|
7 |
class _MLPVectorProjector(nn.Module):
|
@@ -20,10 +20,35 @@ class _MLPVectorProjector(nn.Module):
|
|
20 |
def forward(self, x):
|
21 |
return torch.cat([mlp(x) for mlp in self.mlps], dim=-2)
|
22 |
|
|
|
|
|
23 |
model_name = "microsoft/phi-2"
|
24 |
|
25 |
phi2_text = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)
|
26 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
|
28 |
def example_inference(input_text, count): #, image, img_qn, audio):
|
29 |
pred_text = textMode(input_text, count)
|
@@ -33,8 +58,8 @@ def example_inference(input_text, count): #, image, img_qn, audio):
|
|
33 |
|
34 |
def textMode(text, count):
|
35 |
count = int(count)
|
36 |
-
inputs =
|
37 |
-
prediction =
|
38 |
phi2_text.generate(
|
39 |
**inputs,
|
40 |
max_new_tokens=count,
|
@@ -51,7 +76,8 @@ def imageMode(image, question):
|
|
51 |
return "In progress"
|
52 |
|
53 |
def audioMode(audio):
|
54 |
-
|
|
|
55 |
|
56 |
|
57 |
interface_title = "TSAI-ERA-V1 - Capstone - Multimodal GPT Demo"
|
|
|
1 |
import torch
|
2 |
import torch.nn as nn
|
3 |
import gradio as gr
|
4 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
|
5 |
|
6 |
|
7 |
class _MLPVectorProjector(nn.Module):
|
|
|
20 |
def forward(self, x):
|
21 |
return torch.cat([mlp(x) for mlp in self.mlps], dim=-2)
|
22 |
|
23 |
+
## Text model
|
24 |
+
|
25 |
model_name = "microsoft/phi-2"
|
26 |
|
27 |
phi2_text = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)
|
28 |
+
tokenizer_text = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
|
29 |
+
|
30 |
+
## Audio model
|
31 |
+
model_id_audio = "openai/whisper-large-v3"
|
32 |
+
|
33 |
+
model_audio = AutoModelForSpeechSeq2Seq.from_pretrained(
|
34 |
+
model_id_audio, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True).to("cpu")
|
35 |
+
|
36 |
+
processor_audio = AutoProcessor.from_pretrained(model_id_audio)
|
37 |
+
|
38 |
+
pipe_audio = pipeline(
|
39 |
+
"automatic-speech-recognition",
|
40 |
+
model=model_audio,
|
41 |
+
tokenizer=processor_audio.tokenizer,
|
42 |
+
feature_extractor=processor_audio.feature_extractor,
|
43 |
+
max_new_tokens=128,
|
44 |
+
chunk_length_s=30,
|
45 |
+
batch_size=16,
|
46 |
+
return_timestamps=False,
|
47 |
+
torch_dtype=torch.float32,
|
48 |
+
device="cpu",
|
49 |
+
)
|
50 |
+
|
51 |
+
## image model
|
52 |
|
53 |
def example_inference(input_text, count): #, image, img_qn, audio):
|
54 |
pred_text = textMode(input_text, count)
|
|
|
58 |
|
59 |
def textMode(text, count):
|
60 |
count = int(count)
|
61 |
+
inputs = tokenizer_text(text, return_tensors="pt", return_attention_mask=False)
|
62 |
+
prediction = tokenizer_text.batch_decode(
|
63 |
phi2_text.generate(
|
64 |
**inputs,
|
65 |
max_new_tokens=count,
|
|
|
76 |
return "In progress"
|
77 |
|
78 |
def audioMode(audio):
|
79 |
+
result = pipe_audio(audio)
|
80 |
+
return result["text"]
|
81 |
|
82 |
|
83 |
interface_title = "TSAI-ERA-V1 - Capstone - Multimodal GPT Demo"
|