Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,5 +1,5 @@
|
|
1 |
import torch
|
2 |
-
from transformers import CLIPProcessor, CLIPModel, WhisperProcessor, WhisperForConditionalGeneration, AutoTokenizer,
|
3 |
import gradio as gr
|
4 |
import soundfile as sf
|
5 |
|
@@ -15,15 +15,16 @@ print("Loading Whisper model...")
|
|
15 |
whisper_model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")
|
16 |
whisper_processor = WhisperProcessor.from_pretrained("openai/whisper-small")
|
17 |
|
18 |
-
print("Loading
|
19 |
-
tokenizer = AutoTokenizer.from_pretrained("
|
20 |
-
text_model =
|
21 |
|
22 |
# ------------------------------
|
23 |
# Define Projection Layers
|
24 |
# ------------------------------
|
25 |
print("Initializing image projection layer...")
|
26 |
-
#
|
|
|
27 |
image_projection = torch.nn.Linear(512, 768)
|
28 |
|
29 |
# ------------------------------
|
@@ -33,11 +34,11 @@ image_projection = torch.nn.Linear(512, 768)
|
|
33 |
def multimodal_inference(text_input, image_input, audio_input):
|
34 |
"""
|
35 |
Processes text, image, and audio inputs:
|
36 |
-
- Text: used directly.
|
37 |
-
- Image: processed via CLIP
|
38 |
-
- Audio: transcribed using Whisper.
|
39 |
|
40 |
-
The
|
41 |
"""
|
42 |
prompt = ""
|
43 |
|
@@ -54,7 +55,7 @@ def multimodal_inference(text_input, image_input, audio_input):
|
|
54 |
# Normalize and project image features
|
55 |
image_features = image_features / image_features.norm(p=2, dim=-1, keepdim=True)
|
56 |
projected_image = image_projection(image_features)
|
57 |
-
# For demo
|
58 |
prompt += " [IMAGE_EMBEDDING]"
|
59 |
except Exception as e:
|
60 |
print("Error processing image:", e)
|
@@ -79,16 +80,16 @@ def multimodal_inference(text_input, image_input, audio_input):
|
|
79 |
|
80 |
print("Final fused prompt:", prompt)
|
81 |
|
82 |
-
#
|
83 |
inputs = tokenizer(prompt, return_tensors="pt")
|
84 |
with torch.no_grad():
|
85 |
generated_ids = text_model.generate(
|
86 |
**inputs,
|
87 |
max_length=200,
|
88 |
-
temperature=0.7, #
|
89 |
-
top_p=0.9, #
|
90 |
-
repetition_penalty=1.2,#
|
91 |
-
do_sample=True #
|
92 |
)
|
93 |
generated_text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
|
94 |
|
@@ -106,8 +107,8 @@ iface = gr.Interface(
|
|
106 |
gr.Audio(type="filepath", label="Audio Input (Optional)")
|
107 |
],
|
108 |
outputs="text",
|
109 |
-
title="Multi-Modal LLM Demo",
|
110 |
-
description="This demo accepts text, image, and audio inputs, processes each modality, and produces a text response."
|
111 |
)
|
112 |
|
113 |
if __name__ == "__main__":
|
|
|
1 |
import torch
|
2 |
+
from transformers import CLIPProcessor, CLIPModel, WhisperProcessor, WhisperForConditionalGeneration, AutoTokenizer, AutoModelForSeq2SeqLM
|
3 |
import gradio as gr
|
4 |
import soundfile as sf
|
5 |
|
|
|
15 |
whisper_model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")
|
16 |
whisper_processor = WhisperProcessor.from_pretrained("openai/whisper-small")
|
17 |
|
18 |
+
print("Loading Flan-T5 model (instruction-tuned for better responses)...")
|
19 |
+
tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-large")
|
20 |
+
text_model = AutoModelForSeq2SeqLM.from_pretrained("google/flan-t5-large")
|
21 |
|
22 |
# ------------------------------
|
23 |
# Define Projection Layers
|
24 |
# ------------------------------
|
25 |
print("Initializing image projection layer...")
|
26 |
+
# This linear layer projects CLIP's 512-dimensional image embeddings to Flan-T5's expected dimension.
|
27 |
+
# (For a real system, you would fine-tune this layer.)
|
28 |
image_projection = torch.nn.Linear(512, 768)
|
29 |
|
30 |
# ------------------------------
|
|
|
34 |
def multimodal_inference(text_input, image_input, audio_input):
|
35 |
"""
|
36 |
Processes text, image, and audio inputs:
|
37 |
+
- Text: is used directly.
|
38 |
+
- Image: is processed via CLIP; its embedding is projected and a placeholder is appended.
|
39 |
+
- Audio: is transcribed using Whisper.
|
40 |
|
41 |
+
The combined prompt is then fed into Flan-T5 to generate a text response.
|
42 |
"""
|
43 |
prompt = ""
|
44 |
|
|
|
55 |
# Normalize and project image features
|
56 |
image_features = image_features / image_features.norm(p=2, dim=-1, keepdim=True)
|
57 |
projected_image = image_projection(image_features)
|
58 |
+
# For this demo, we append a placeholder tag to indicate image information.
|
59 |
prompt += " [IMAGE_EMBEDDING]"
|
60 |
except Exception as e:
|
61 |
print("Error processing image:", e)
|
|
|
80 |
|
81 |
print("Final fused prompt:", prompt)
|
82 |
|
83 |
+
# Tokenize and generate text using Flan-T5
|
84 |
inputs = tokenizer(prompt, return_tensors="pt")
|
85 |
with torch.no_grad():
|
86 |
generated_ids = text_model.generate(
|
87 |
**inputs,
|
88 |
max_length=200,
|
89 |
+
temperature=0.7, # Moderate randomness
|
90 |
+
top_p=0.9, # Nucleus sampling to limit token choices
|
91 |
+
repetition_penalty=1.2,# Penalize repeated tokens
|
92 |
+
do_sample=True # Enable sampling for more varied responses
|
93 |
)
|
94 |
generated_text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
|
95 |
|
|
|
107 |
gr.Audio(type="filepath", label="Audio Input (Optional)")
|
108 |
],
|
109 |
outputs="text",
|
110 |
+
title="Multi-Modal LLM Demo with Flan-T5",
|
111 |
+
description="This demo accepts text, image, and audio inputs, processes each modality, and produces a text response using an instruction-tuned model."
|
112 |
)
|
113 |
|
114 |
if __name__ == "__main__":
|