File size: 7,558 Bytes
9bc7520
 
 
 
521c83c
9bc7520
 
 
 
 
952b001
ece5c3b
 
 
 
 
 
 
 
 
 
 
 
 
 
9bc7520
ece5c3b
 
 
 
 
 
 
9bc7520
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3f01f45
9bc7520
 
 
 
 
 
 
 
521c83c
d5c337b
 
9bc7520
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f2f8465
9bc7520
 
f2f8465
d5c337b
9bc7520
 
 
 
 
 
 
 
 
 
 
 
 
 
d5c337b
 
 
 
 
 
 
 
 
f2f8465
 
 
 
 
 
d5c337b
 
 
 
 
 
 
 
9bc7520
ece5c3b
9bc7520
d5c337b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f2f8465
d5c337b
 
 
9bc7520
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d5c337b
f2f8465
d5c337b
f2f8465
d5c337b
f2f8465
d5c337b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9bc7520
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
import os
import subprocess
from PIL import Image
import io
import gradio as gr
from transformers import AutoProcessor, TextIteratorStreamer
from transformers import Idefics2ForConditionalGeneration
import torch
from peft import LoraConfig
from transformers import AutoProcessor, BitsAndBytesConfig, IdeficsForVisionText2Text

# Project description
description = """
# Kalbe Farma - Visual Question Answering (VQA) for Medical Imaging

## Overview
The project addresses the challenge of accurate and efficient medical imaging analysis in healthcare, aiming to reduce human error and workload for radiologists. The proposed solution involves developing advanced AI models for Visual Question Answering (VQA) to assist healthcare professionals in analyzing medical images quickly and accurately. These models will be integrated into a user-friendly web application, providing a practical tool for real-world healthcare settings.

## Dataset
The model is trained using the [Hugging face](https://huggingface.co/datasets/flaviagiammarino/vqa-rad/viewer).

Reference: [ScienceDirect](https://www.sciencedirect.com/science/article/abs/pii/S0933365723001252)

## Model Architecture

![Model Architecture](img/Model-Architecture.png)

Reference: [ScienceDirect](https://www.sciencedirect.com/science/article/abs/pii/S0933365723001252)

## Demo
Please select the example below or upload 4 pairs of mammography exam results.
"""

DEVICE = torch.device("cuda")

USE_LORA = False
USE_QLORA = True

if USE_QLORA or USE_LORA:
    lora_config = LoraConfig(
        r=8,
        lora_alpha=8,
        lora_dropout=0.1,
        target_modules='.*(text_model|modality_projection|perceiver_resampler).*(down_proj|gate_proj|up_proj|k_proj|q_proj|v_proj|o_proj).*$',
        use_dora=False if USE_QLORA else True,
        init_lora_weights="gaussian"
    )
    if USE_QLORA:
        bnb_config = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_compute_dtype=torch.float16
        )

    model = Idefics2ForConditionalGeneration.from_pretrained(
        "jihadzakki/idefics2-8b-vqarad-delta",
        torch_dtype=torch.float16,
        quantization_config=bnb_config
    )


processor = AutoProcessor.from_pretrained(
    "HuggingFaceM4/idefics2-8b",
)

def format_answer(image, question, history):
    try:
        messages = [
            {
                "role": "user",
                "content": [
                    {"type": "image"},
                    {"type": "text", "text": question}
                ]
            }
        ]

        text = processor.apply_chat_template(messages, add_generation_prompt=True)
        inputs = processor(text=[text.strip()], images=[image], return_tensors="pt", padding=True)
        inputs = {key: value.to(DEVICE) for key, value in inputs.items()}
        generated_ids = model.generate(**inputs, max_new_tokens=64)
        generated_texts = processor.batch_decode(generated_ids[:, inputs["input_ids"].size(1):], skip_special_tokens=True)[0]

        history.append((image, f"Question: {question} | Answer: {generated_texts}"))

        # Store the predicted answer in a variable before deleting intermediate variables
        predicted_answer = f"Predicted Answer: {generated_texts}"

        # Clear the cache and delete unnecessary variables
        del inputs
        del generated_ids
        del generated_texts
        torch.cuda.empty_cache()

        return predicted_answer, history
    except Exception as e:
        # Clear the cache in case of an error
        torch.cuda.empty_cache()
        return f"Error: {str(e)}", history

def clear_history():
    return "", []

def undo_last(history):
    if history:
        history.pop()
    return "", history

def retry_last(image, question, history):
    if history:
        last_image, last_entry = history[-1]
        return format_answer(last_image, question, history[:-1])
    return "No previous analysis to retry.", history

def switch_theme(mode):
    if mode == "Light Mode":
        return gr.themes.Default()
    else:
        return gr.themes.Soft(primary_hue=gr.themes.colors.blue, secondary_hue=gr.themes.colors.orange)

def save_feedback(feedback):
    return "Thank you for your feedback!"

def display_history(history):
    log_entries = []
    for img, text in history:
        log_entries.append((img, text))
    return log_entries

# Build the Visual QA application using Gradio with improvements
with gr.Blocks(
    theme=gr.themes.Soft(
        font=[gr.themes.GoogleFont("Inconsolata"), "Arial", "sans-serif"],
        primary_hue=gr.themes.colors.blue,
        secondary_hue=gr.themes.colors.red,
    )
) as VisualQAApp:
    gr.Markdown(description, elem_classes="title")  # Display the project description

    gr.Markdown("## Demo")

    with gr.Row():
        with gr.Column():
            image_input = gr.Image(label="Upload image", type="pil")
            question_input = gr.Textbox(show_label=False, placeholder="Enter your question here...")
            submit_button = gr.Button("Submit", variant="primary")

        with gr.Column():
            answer_output = gr.Textbox(label="Result Prediction")

    history_state = gr.State([])  # Initialize the history state

    submit_button.click(
        format_answer,
        inputs=[image_input, question_input, history_state],
        outputs=[answer_output, history_state],
        show_progress=True
    )

    with gr.Row():
        retry_button = gr.Button("Retry")
        undo_button = gr.Button("Undo")
        clear_button = gr.Button("Clear")

        retry_button.click(
            retry_last,
            inputs=[image_input, question_input, history_state],
            outputs=[answer_output, history_state]
        )

        undo_button.click(
            undo_last,
            inputs=[history_state],
            outputs=[answer_output, history_state]
        )

        clear_button.click(
            clear_history,
            inputs=[],
            outputs=[answer_output, history_state]
        )

    with gr.Row():
        history_gallery = gr.Gallery(label="History Log", elem_id="history_log")
        submit_button.click(
            display_history,
            inputs=[history_state],
            outputs=[history_gallery]
        )

    with gr.Accordion("Help", open=False):
        gr.Markdown("**Upload image**: Select the chest X-ray image you want to analyze.")
        gr.Markdown("**Enter your question**: Type the question you have about the image, such as 'Is there any sign of pneumonia?'")
        gr.Markdown("**Submit**: Click the submit button to get the prediction from the model.")

    with gr.Accordion("User Preferences", open=False):
        gr.Markdown("**Mode**: Choose between light and dark mode for your comfort.")
        mode_selector = gr.Radio(choices=["Light Mode", "Dark Mode"], label="Select Mode")
        apply_theme_button = gr.Button("Apply Theme")

        apply_theme_button.click(
            switch_theme,
            inputs=[mode_selector],
            outputs=[],
        )

    with gr.Accordion("Feedback", open=False):
        gr.Markdown("**We value your feedback!** Please provide any feedback you have about this application.")
        feedback_input = gr.Textbox(label="Feedback", lines=4)
        submit_feedback_button = gr.Button("Submit Feedback")

        submit_feedback_button.click(
            save_feedback,
            inputs=[feedback_input],
            outputs=[feedback_input]
        )

VisualQAApp.launch(share=True, debug=True)