Spaces:
Running
on
Zero
Running
on
Zero
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,223 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import gradio as gr
|
3 |
+
import torch
|
4 |
+
from transformers import AutoProcessor, Gemma3nForConditionalGeneration, TextIteratorStreamer
|
5 |
+
from PIL import Image
|
6 |
+
import threading
|
7 |
+
import traceback
|
8 |
+
import spaces
|
9 |
+
|
10 |
+
# -----------------------------
|
11 |
+
# Config
|
12 |
+
# -----------------------------
|
13 |
+
MODEL_ID = "yasserrmd/GemmaECG-Vision"
|
14 |
+
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
15 |
+
DTYPE = torch.bfloat16 if DEVICE == "cuda" else torch.float32 # safe CPU dtype
|
16 |
+
|
17 |
+
# Generation defaults
|
18 |
+
GEN_KW = dict(
|
19 |
+
max_new_tokens=768,
|
20 |
+
do_sample=True,
|
21 |
+
temperature=1.0,
|
22 |
+
top_p=0.95,
|
23 |
+
top_k=64,
|
24 |
+
use_cache=True,
|
25 |
+
)
|
26 |
+
|
27 |
+
# Clinical prompt
|
28 |
+
CLINICAL_PROMPT = """You are a clinical assistant specialized in ECG interpretation. Given an ECG image, generate a concise, structured, and medically accurate report.
|
29 |
+
|
30 |
+
Use this exact format:
|
31 |
+
|
32 |
+
Rhythm:
|
33 |
+
PR Interval:
|
34 |
+
QRS Duration:
|
35 |
+
Axis:
|
36 |
+
Bundle Branch Blocks:
|
37 |
+
Atrial Abnormalities:
|
38 |
+
Ventricular Hypertrophy:
|
39 |
+
Q Wave or QS Complexes:
|
40 |
+
T Wave Abnormalities:
|
41 |
+
ST Segment Changes:
|
42 |
+
Final Impression:
|
43 |
+
|
44 |
+
Guidance:
|
45 |
+
- Confirm sinus rhythm only if consistent P waves precede each QRS.
|
46 |
+
- Describe PACs only if early, ectopic P waves are visible.
|
47 |
+
- Do not diagnose myocardial infarction solely based on QS complexes unless accompanied by other signs (e.g., ST elevation, reciprocal changes, poor R wave progression).
|
48 |
+
- Only mention axis deviation if QRS axis is clearly rightward (RAD) or leftward (LAD).
|
49 |
+
- Use terms like "suggestive of" or "possible" for uncertain findings.
|
50 |
+
- Avoid repetition and keep the report clinically focused.
|
51 |
+
- Do not include external references or source citations.
|
52 |
+
- Do not diagnose left bundle branch block unless QRS duration is ≥120 ms with typical morphology in leads I, V5, V6.
|
53 |
+
- Mark T wave changes in inferior leads as “nonspecific” unless clear ST elevation or reciprocal depression is present.
|
54 |
+
|
55 |
+
Your goal is to provide a structured ECG summary useful for a cardiologist or internal medicine physician.
|
56 |
+
"""
|
57 |
+
|
58 |
+
# -----------------------------
|
59 |
+
# Load model & processor
|
60 |
+
# -----------------------------
|
61 |
+
model = Gemma3nForConditionalGeneration.from_pretrained(
|
62 |
+
MODEL_ID, torch_dtype=DTYPE
|
63 |
+
).to(DEVICE).eval()
|
64 |
+
processor = AutoProcessor.from_pretrained(MODEL_ID)
|
65 |
+
|
66 |
+
# -----------------------------
|
67 |
+
# Inference (streaming) function
|
68 |
+
# -----------------------------
|
69 |
+
@spaces.GPU
|
70 |
+
def analyze_ecg_stream(image: Image.Image):
|
71 |
+
"""
|
72 |
+
Streams model output into the Gradio textbox.
|
73 |
+
Yields incremental text chunks.
|
74 |
+
"""
|
75 |
+
if image is None:
|
76 |
+
yield "Please upload an ECG image."
|
77 |
+
return
|
78 |
+
|
79 |
+
# Build a multimodal chat-style message; rely on the model's chat template to inject image tokens.
|
80 |
+
messages = [
|
81 |
+
{
|
82 |
+
"role": "user",
|
83 |
+
"content": [
|
84 |
+
{"type": "text", "text": CLINICAL_PROMPT},
|
85 |
+
{"type": "image"},
|
86 |
+
],
|
87 |
+
}
|
88 |
+
]
|
89 |
+
|
90 |
+
try:
|
91 |
+
# Try with chat template first (recommended for chat-tuned models)
|
92 |
+
chat_text = processor.apply_chat_template(messages, add_generation_prompt=True)
|
93 |
+
|
94 |
+
model_inputs = processor(
|
95 |
+
text=chat_text,
|
96 |
+
images=image,
|
97 |
+
return_tensors="pt",
|
98 |
+
)
|
99 |
+
model_inputs = {k: v.to(DEVICE) for k, v in model_inputs.items()}
|
100 |
+
|
101 |
+
except Exception as e:
|
102 |
+
# If the template or image-token count fails, fallback to a simple text+image pack.
|
103 |
+
# This handles errors like:
|
104 |
+
# "Number of images does not match number of special image tokens..."
|
105 |
+
fallback_note = (
|
106 |
+
"\n[Note] Falling back to a simpler prompt packing due to template/image token mismatch."
|
107 |
+
)
|
108 |
+
try:
|
109 |
+
model_inputs = processor(
|
110 |
+
text=CLINICAL_PROMPT,
|
111 |
+
images=image,
|
112 |
+
return_tensors="pt",
|
113 |
+
)
|
114 |
+
model_inputs = {k: v.to(DEVICE) for k, v in model_inputs.items()}
|
115 |
+
# Surface a short note at the start of the stream so user knows why
|
116 |
+
yield fallback_note + "\n"
|
117 |
+
except Exception as inner_e:
|
118 |
+
err_msg = f"Input preparation failed:\n{repr(e)}\n{repr(inner_e)}"
|
119 |
+
yield err_msg
|
120 |
+
return
|
121 |
+
|
122 |
+
# Prepare streamer
|
123 |
+
streamer = TextIteratorStreamer(
|
124 |
+
processor.tokenizer,
|
125 |
+
skip_prompt=True,
|
126 |
+
skip_special_tokens=True,
|
127 |
+
)
|
128 |
+
|
129 |
+
# Launch generation in a background thread
|
130 |
+
generated_text = []
|
131 |
+
def _generate():
|
132 |
+
try:
|
133 |
+
model.generate(
|
134 |
+
**model_inputs,
|
135 |
+
streamer=streamer,
|
136 |
+
**GEN_KW
|
137 |
+
)
|
138 |
+
except Exception as gen_e:
|
139 |
+
# Put traceback into the stream so the user sees it (useful during debugging)
|
140 |
+
tb = traceback.format_exc()
|
141 |
+
streamer.put("\n\n[Generation Error]\n" + str(gen_e) + "\n" + tb)
|
142 |
+
finally:
|
143 |
+
streamer.end()
|
144 |
+
|
145 |
+
thread = threading.Thread(target=_generate)
|
146 |
+
thread.start()
|
147 |
+
|
148 |
+
# Collect incremental tokens and yield buffer
|
149 |
+
buffer = ""
|
150 |
+
for token in streamer:
|
151 |
+
buffer += token
|
152 |
+
# Stream into Gradio textbox
|
153 |
+
yield buffer
|
154 |
+
|
155 |
+
def reset():
|
156 |
+
return None, ""
|
157 |
+
|
158 |
+
# -----------------------------
|
159 |
+
# Gradio UI
|
160 |
+
# -----------------------------
|
161 |
+
with gr.Blocks(css="""
|
162 |
+
.disclaimer {
|
163 |
+
padding: 12px 16px;
|
164 |
+
border: 1px solid #b91c1c;
|
165 |
+
background: #fef2f2;
|
166 |
+
color: #7f1d1d;
|
167 |
+
border-radius: 8px;
|
168 |
+
font-weight: 600;
|
169 |
+
}
|
170 |
+
.footer-note {
|
171 |
+
font-size: 12px;
|
172 |
+
color: #374151;
|
173 |
+
}
|
174 |
+
.gr-button { background-color: #1e3a8a; color: #ffffff; }
|
175 |
+
""") as demo:
|
176 |
+
gr.Markdown("## 🩺 ECG Interpretation Assistant — Gemma-ECG-Vision")
|
177 |
+
gr.HTML("""
|
178 |
+
<div class="disclaimer">
|
179 |
+
⚠️ <strong>Important Medical Disclaimer:</strong> This tool is for <u>education and research</u> purposes only.
|
180 |
+
It is <u>not</u> a medical device and must not be used for diagnosis or treatment.
|
181 |
+
Always consult a licensed clinician for interpretation and clinical decisions.
|
182 |
+
</div>
|
183 |
+
""")
|
184 |
+
|
185 |
+
with gr.Row():
|
186 |
+
image_input = gr.Image(type="pil", label="Upload ECG Image", height=320)
|
187 |
+
output_box = gr.Textbox(
|
188 |
+
label="Generated ECG Report (Streaming)",
|
189 |
+
lines=24,
|
190 |
+
show_copy_button=True,
|
191 |
+
autoscroll=True,
|
192 |
+
)
|
193 |
+
|
194 |
+
with gr.Row():
|
195 |
+
with gr.Column():
|
196 |
+
submit_btn = gr.Button("Generate Report", variant="primary")
|
197 |
+
with gr.Column():
|
198 |
+
reset_btn = gr.Button("Reset")
|
199 |
+
|
200 |
+
# Wire actions: analyze_ecg_stream yields partial strings for streaming
|
201 |
+
submit_btn.click(
|
202 |
+
fn=analyze_ecg_stream,
|
203 |
+
inputs=image_input,
|
204 |
+
outputs=output_box,
|
205 |
+
queue=True,
|
206 |
+
api_name="analyze_ecg",
|
207 |
+
)
|
208 |
+
reset_btn.click(fn=reset, outputs=[image_input, output_box])
|
209 |
+
|
210 |
+
gr.Markdown(
|
211 |
+
"""
|
212 |
+
<div class="footer-note">
|
213 |
+
Model: <code>{model_id}</code> | Device: <code>{device}</code><br>
|
214 |
+
Tip: Larger images can improve recognition of fine waveform details (P waves, ST segments).
|
215 |
+
Ensure lead labels are visible when possible.
|
216 |
+
</div>
|
217 |
+
""".format(model_id=MODEL_ID, device=DEVICE)
|
218 |
+
)
|
219 |
+
|
220 |
+
# Enable queuing for proper streaming under concurrency
|
221 |
+
demo.queue(concurrency_count=2, max_size=16)
|
222 |
+
# In hosted notebooks, you can set share=True if needed
|
223 |
+
demo.launch(share=False, debug=True)
|