File size: 8,999 Bytes
0e9f127
ebe0135
0e9f127
 
45c7be0
9e25965
45c7be0
0e9f127
701a46b
9e25965
0e9f127
45c7be0
0e9f127
660df5d
9e25965
660df5d
 
9e25965
 
 
660df5d
9e25965
 
 
660df5d
 
 
9e25965
 
660df5d
45c7be0
9e25965
660df5d
45c7be0
660df5d
0e9f127
45c7be0
9e25965
660df5d
9e25965
45c7be0
9e25965
0e9f127
9e25965
 
660df5d
9e25965
 
660df5d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
701a46b
 
9e25965
701a46b
 
45c7be0
701a46b
45c7be0
701a46b
660df5d
 
 
 
701a46b
 
 
 
9e25965
 
 
 
 
 
 
45c7be0
660df5d
 
45c7be0
9e25965
 
 
 
701a46b
660df5d
45c7be0
660df5d
 
9e25965
 
 
660df5d
 
 
 
 
9e25965
 
 
660df5d
 
45c7be0
9e25965
 
 
 
660df5d
 
9e25965
 
 
660df5d
9e25965
 
 
660df5d
9e25965
 
 
 
 
660df5d
9e25965
 
 
 
 
 
 
 
660df5d
9e25965
 
660df5d
9e25965
45c7be0
 
 
 
 
 
 
660df5d
45c7be0
9e25965
660df5d
9e25965
45c7be0
 
9e25965
45c7be0
 
660df5d
9e25965
45c7be0
9e25965
 
 
 
660df5d
 
 
 
9e25965
660df5d
0e9f127
93307ce
 
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

import os
import gradio as gr
import torch
import gc
from PIL import Image
from transformers import AutoModelForImageTextToText, AutoProcessor
import json
import re
from typing import Dict, List, Any, Optional

HF_TOKEN = os.environ.get("HF_TOKEN")

# ── Model Cache ──────────────────────────────────────────────
_model_cache = {}
MAX_CACHED_MODELS = 2
QWEN_VL_IMG_TOKEN = "<|vision_start|><|image_pad|><|vision_end|>"

def load_model(model_id: str):
    if model_id in _model_cache:
        print(f"⚑ Cache Hit: {model_id}")
        return _model_cache[model_id]

    if len(_model_cache) >= MAX_CACHED_MODELS:
        oldest = list(_model_cache.keys())[0]
        print(f"🧹 Unloading: {oldest}")
        del _model_cache[oldest]
        gc.collect()

    print(f"⏳ Loading: {model_id}")
    try:
        processor = AutoProcessor.from_pretrained(model_id, token=HF_TOKEN)
        device_map = "auto" if torch.cuda.is_available() else "cpu"
        model = AutoModelForImageTextToText.from_pretrained(
            model_id, device_map=device_map, low_cpu_mem_usage=True, token=HF_TOKEN
        )
        model.eval()
        _model_cache[model_id] = (processor, model)
        print(f"βœ… Loaded: {model_id}")
        return processor, model
    except Exception as e:
        return None, None

def ui_model_change(model_id):
    processor, model = load_model(model_id)
    if model: return f"βœ… Model Active: {model_id}"
    return f"❌ Failed to load {model_id}"

# ── THE FIX: prepare_inputs (from your reference app.py) ──────
# Yeh function mixed content (string + list) ko flat format me
# convert karke processor ko safe tarike se deta hai
def prepare_inputs(processor, model, messages: List[Dict]) -> Dict:
    pil_images = []
    flat_messages = []

    for msg in messages:
        role    = msg.get("role", "user")
        content = msg.get("content", "")

        if isinstance(content, list):
            parts = []
            for item in content:
                if not isinstance(item, dict):
                    parts.append(str(item))
                    continue
                t = item.get("type", "")
                if t == "text":
                    parts.append(item.get("text", ""))
                elif t == "image":
                    img = item.get("image")
                    if img is not None and isinstance(img, Image.Image):
                        pil_images.append(img)
                    parts.append(QWEN_VL_IMG_TOKEN)
            flat_messages.append({"role": role, "content": "".join(parts)})
        else:
            # History string messages directly add kar do
            flat_messages.append({"role": role, "content": str(content)})

    text = processor.apply_chat_template(flat_messages, tokenize=False, add_generation_prompt=True)

    if pil_images and hasattr(processor, "image_processor"):
        inputs = processor(text=[text], images=pil_images, padding=True, return_tensors="pt")
    else:
        inputs = processor(text=[text], padding=True, return_tensors="pt")

    return {k: v.to(model.device) if torch.is_tensor(v) else v for k, v in inputs.items()}

# ── Enterprise OCR ────────────────────────────────────────────
def extract_tag(tag, text):
    match = re.search(f"<(?:{tag})?>(.*?)</(?:{tag})?", text, re.IGNORECASE)
    if not match: match = re.search(f"<{tag}>(.*?)</{tag}>", text, re.IGNORECASE)
    return match.group(1).strip() if match else "UNKNOWN"

def build_enterprise_json(raw_text):
    result_json = {
      "DocumentMetadata": {"document_type": "Resident Card", "has_mrz": True},
      "StructuredData": {
        "civil_number": extract_tag("ID", raw_text),
        "full_name": extract_tag("NAME", raw_text),
        "date_of_birth": extract_tag("DOB", raw_text),
        "nationality": extract_tag("NAT", raw_text)
      }
    }
    return json.dumps(result_json, indent=2, ensure_ascii=False)

def run_document_scan(front_img, model_name):
    if front_img is None: return "Error: Please upload document image."
    processor, model = load_model(model_name)
    if not model: return "Error: Model not loaded."

    prompt = "Extract details inside these XML tags ONLY:\n<ID></ID>\n<NAME></NAME>\n<DOB></DOB>\n<NAT></NAT>"
    messages = [{"role": "user", "content": [{"type": "image", "image": front_img}, {"type": "text", "text": prompt}]}]

    try:
        inputs = prepare_inputs(processor, model, messages)
        with torch.no_grad():
            generated_ids = model.generate(**inputs, max_new_tokens=150, temperature=0.1)
        trimmed = [out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs["input_ids"], generated_ids)]
        raw_output = processor.batch_decode(trimmed, skip_special_tokens=True)[0]
        return build_enterprise_json(raw_output)
    except Exception as e:
        return f"Extraction Failed: {str(e)}"

# ── Chat ──────────────────────────────────────────────────────
def process_chat(text: str, image: Optional[Image.Image], history: List[Dict], model_name: str) -> str:
    processor, model = load_model(model_name)
    if not model: return "Error: Model not loaded."

    # Build history messages first
    messages = [{"role": m["role"], "content": m["content"]}
                for m in history if m.get("role") in ("user", "assistant")]

    # Current message with optional image (as list)
    content = []
    if image is not None:
        content.append({"type": "image", "image": image})
    if text:
        content.append({"type": "text", "text": text})

    if content:
        messages.append({"role": "user", "content": content})

    try:
        # prepare_inputs now handles mixed string/list content safely
        inputs = prepare_inputs(processor, model, messages)
        with torch.no_grad():
            generated_ids = model.generate(**inputs, max_new_tokens=512, temperature=0.7, top_p=0.9)
        trimmed = [o[len(i):] for i, o in zip(inputs['input_ids'], generated_ids)]
        return processor.batch_decode(trimmed, skip_special_tokens=True)[0]
    except Exception as e:
        return f"❌ Error: {str(e)}"

def chat_fn(message: Dict[str, Any], history: List[Dict], model_name: str):
    text  = message.get("text", "")
    files = message.get("files", [])
    image = None
    if files:
        try: image = Image.open(files[0]).convert("RGB")
        except Exception as e: print(f"Image error: {e}")

    response = process_chat(text, image, history, model_name)

    display_text = f"{text}\nπŸ“Ž [Image attached]" if image else text
    history.append({"role": "user", "content": display_text})
    history.append({"role": "assistant", "content": response})
    return gr.update(value={"text": "", "files": []}), history

# ── Gradio UI ─────────────────────────────────────────────────
with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("# πŸͺͺ CSM Smart Document Engine")
    gr.Markdown("_On-Demand Caching β€’ Document Scanner β€’ Intelligent Multi-Turn Chat_")

    with gr.Row(variant="panel"):
        model_dropdown = gr.Dropdown(
            choices=[
                "Chhagan005/CSM-KIE-Universal",
                "Chhagan005/CSM-DocExtract-8N",
                "Chhagan005/CSM-DocExtract-4N",
            ],
            label="πŸ€– Select Model", value="Chhagan005/CSM-KIE-Universal", interactive=True
        )
        status_bar = gr.Textbox(label="Memory Status", value="Select a model to load into memory", interactive=False)

    model_dropdown.change(fn=ui_model_change, inputs=[model_dropdown], outputs=[status_bar])

    with gr.Tabs():
        with gr.TabItem("πŸ“„ Document Scanner"):
            with gr.Row():
                with gr.Column():
                    doc_img  = gr.Image(type="pil", label="Upload ID Card")
                    scan_btn = gr.Button("πŸ” Extract JSON", variant="primary")
                with gr.Column():
                    json_output = gr.Code(language="json", label="Enterprise Result")
            scan_btn.click(fn=run_document_scan, inputs=[doc_img, model_dropdown], outputs=[json_output])

        with gr.TabItem("πŸ’¬ Intelligent Chat"):
            chatbot  = gr.Chatbot(label="Chat History", height=450, value=[])
            chat_msg = gr.MultimodalTextbox(
                label="Message", placeholder="Type a message or click πŸ“Ž to attach an image...",
                file_types=["image"], submit_btn=True
            )
            chat_msg.submit(fn=chat_fn, inputs=[chat_msg, chatbot, model_dropdown], outputs=[chat_msg, chatbot])

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860)