Spaces:
Running
Running
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,128 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
import gradio as gr
|
3 |
+
import peft
|
4 |
+
from peft import LoraConfig, PeftModel
|
5 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, CLIPVisionModel, AutoProcessor
|
6 |
+
import torch
|
7 |
+
from PIL import Image
|
8 |
+
import requests
|
9 |
+
import numpy as np
|
10 |
+
import torch.nn as nn
|
11 |
+
import whisperx
|
12 |
+
import ffmpeg, pydub
|
13 |
+
from pydub import AudioSegment
|
14 |
+
|
15 |
+
clip_model_name = "wkcn/TinyCLIP-ViT-61M-32-Text-29M-LAION400M"
|
16 |
+
phi_model_name = "microsoft/phi-2"
|
17 |
+
|
18 |
+
tokenizer = AutoTokenizer.from_pretrained(phi_model_name, trust_remote_code=True)
|
19 |
+
processor = AutoProcessor.from_pretrained(clip_model_name)
|
20 |
+
tokenizer.pad_token = tokenizer.eos_token
|
21 |
+
IMAGE_TOKEN_ID = 23893 # token for word comment
|
22 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
23 |
+
clip_embed = 640
|
24 |
+
phi_embed = 2560
|
25 |
+
compute_type = "float32"
|
26 |
+
audio_batch_size = 1
|
27 |
+
|
28 |
+
import gc
|
29 |
+
|
30 |
+
# models
|
31 |
+
clip_model = CLIPVisionModel.from_pretrained(clip_model_name).to(device)
|
32 |
+
|
33 |
+
projection = torch.nn.Linear(clip_embed, phi_embed).to(device)
|
34 |
+
|
35 |
+
gc.collect()
|
36 |
+
phi_model = AutoModelForCausalLM.from_pretrained(
|
37 |
+
phi_model_name,
|
38 |
+
trust_remote_code=True,
|
39 |
+
)
|
40 |
+
audio_model = whisperx.load_model("small", device, compute_type=compute_type)
|
41 |
+
|
42 |
+
# load weights
|
43 |
+
model_to_merge = PeftModel.from_pretrained(phi_model,'./model_chkpt/')
|
44 |
+
merged_model = model_to_merge.merge_and_unload().to(device)
|
45 |
+
projection.load_state_dict(torch.load('./ft_projection.pth',map_location=torch.device(device)))
|
46 |
+
|
47 |
+
def inference(img=None,img_audio=None,val_q=None):
|
48 |
+
|
49 |
+
max_generate_length = 50
|
50 |
+
val_combined_embeds = []
|
51 |
+
|
52 |
+
with torch.no_grad():
|
53 |
+
|
54 |
+
# image
|
55 |
+
if img is not None:
|
56 |
+
image_processed = processor(images=img, return_tensors="pt").to(device)
|
57 |
+
clip_val_outputs = clip_model(**image_processed).last_hidden_state[:,1:,:]
|
58 |
+
val_image_embeds = projection(clip_val_outputs)
|
59 |
+
|
60 |
+
img_token_tensor = torch.tensor(IMAGE_TOKEN_ID).to(device)
|
61 |
+
img_token_embeds = merged_model.model.embed_tokens(img_token_tensor).unsqueeze(0).unsqueeze(0)
|
62 |
+
|
63 |
+
val_combined_embeds.append(val_image_embeds)
|
64 |
+
val_combined_embeds.append(img_token_embeds)
|
65 |
+
|
66 |
+
# audio
|
67 |
+
if img_audio is not None:
|
68 |
+
|
69 |
+
# accepting only initial few secs speech
|
70 |
+
audio = AudioSegment.from_mp3( img_audio)
|
71 |
+
clipped_audio = audio[:20*1000]
|
72 |
+
clipped_audio.export( 'audio.mp3', format="mp3")
|
73 |
+
result = audio_model.transcribe('audio.mp3')
|
74 |
+
audio_text = ''
|
75 |
+
|
76 |
+
audio_text = result["segments"][0]['text']
|
77 |
+
audio_text = audio_text.strip()
|
78 |
+
audio_tokens = tokenizer(audio_text, return_tensors="pt", return_attention_mask=False)['input_ids'].squeeze(0).to(device)
|
79 |
+
audio_embeds = merged_model.model.embed_tokens(audio_tokens).unsqueeze(0)
|
80 |
+
val_combined_embeds.append(audio_embeds)
|
81 |
+
|
82 |
+
# text question
|
83 |
+
if len(val_q) != 0:
|
84 |
+
val_q_tokenised = tokenizer(val_q, return_tensors="pt", return_attention_mask=False)['input_ids'].squeeze(0).to(device)
|
85 |
+
val_q_embeds = merged_model.model.embed_tokens(val_q_tokenised).unsqueeze(0)
|
86 |
+
val_combined_embeds.append(val_q_embeds)
|
87 |
+
|
88 |
+
# val_combined_emb
|
89 |
+
val_combined_embeds = torch.cat(val_combined_embeds,dim=1)
|
90 |
+
|
91 |
+
predicted_caption = torch.full((1,max_generate_length),50256).to(device)
|
92 |
+
|
93 |
+
for g in range(max_generate_length):
|
94 |
+
phi_output_logits = merged_model(inputs_embeds=val_combined_embeds)['logits']
|
95 |
+
predicted_word_token_logits = phi_output_logits[:, -1, :].unsqueeze(1)
|
96 |
+
predicted_word_token = torch.argmax(predicted_word_token_logits, dim = -1)
|
97 |
+
predicted_caption[:,g] = predicted_word_token.view(1,-1)
|
98 |
+
next_token_embeds = phi_model.model.embed_tokens(predicted_word_token)
|
99 |
+
val_combined_embeds = torch.cat([val_combined_embeds, next_token_embeds], dim=1)
|
100 |
+
|
101 |
+
predicted_captions_decoded = tokenizer.batch_decode(predicted_caption,ignore_index = 50256)[0]
|
102 |
+
|
103 |
+
return predicted_captions_decoded
|
104 |
+
|
105 |
+
with gr.Blocks() as demo:
|
106 |
+
|
107 |
+
gr.Markdown(
|
108 |
+
"""
|
109 |
+
# multi-modalLLM
|
110 |
+
Build using Tiny Clip model and Microsoft's Phi-2 model fine tuned on Instruct 150k.
|
111 |
+
"""
|
112 |
+
)
|
113 |
+
|
114 |
+
# app GUI
|
115 |
+
with gr.Row():
|
116 |
+
with gr.Column():
|
117 |
+
img_input = gr.Image(label='Reference Image',type="pil")
|
118 |
+
img_question = gr.Text(label ='Question related to Image')
|
119 |
+
img_audio = gr.Audio(label="Speak a question", sources=['microphone', 'upload'], type='filepath')
|
120 |
+
with gr.Column():
|
121 |
+
img_answer = gr.Text(label ='Response')
|
122 |
+
|
123 |
+
section_btn = gr.Button("Process")
|
124 |
+
section_btn.click(inference, inputs=[img_input,img_audio,img_question], outputs=[img_answer])
|
125 |
+
|
126 |
+
if __name__ == "__main__":
|
127 |
+
demo.launch(debug=True)
|
128 |
+
|