Spaces:
Sleeping
Sleeping
Update app.py
Browse filesUI design fixes done
app.py
CHANGED
|
@@ -3,26 +3,24 @@ from PIL import Image
|
|
| 3 |
import torch
|
| 4 |
import os
|
| 5 |
|
| 6 |
-
|
| 7 |
-
# NOTE: inference.py must be present and contain these functions for VQA to work.
|
| 8 |
from inference import load_for_inference, predict
|
| 9 |
|
| 10 |
-
# -----------------------
|
| 11 |
# Load TAMGA VQA model
|
| 12 |
-
|
| 13 |
TAMGA_REPO = "Mueris/TurkishVLMTAMGA"
|
| 14 |
-
|
| 15 |
if 'load_for_inference' in globals():
|
| 16 |
tamga_model, tamga_tokenizer, tamga_device = load_for_inference(TAMGA_REPO)
|
| 17 |
else:
|
| 18 |
-
|
| 19 |
print("Warning: inference.py functions not loaded. Using placeholder values.")
|
| 20 |
tamga_model, tamga_tokenizer, tamga_device = None, None, 'cpu'
|
| 21 |
|
| 22 |
|
| 23 |
-
|
| 24 |
# Load BLIP Caption Model
|
| 25 |
-
|
| 26 |
from transformers import BlipProcessor, BlipForConditionalGeneration
|
| 27 |
|
| 28 |
CAPTION_REPO = "Mueris/TurkishVLMTAMGA-CaptioningModel"
|
|
@@ -32,17 +30,15 @@ caption_model.to("cuda" if torch.cuda.is_available() else "cpu")
|
|
| 32 |
caption_device = caption_model.device
|
| 33 |
|
| 34 |
|
| 35 |
-
|
| 36 |
# Utility Functions
|
| 37 |
-
|
| 38 |
|
| 39 |
def toggle_question_input(model_choice):
|
| 40 |
|
| 41 |
if model_choice == "TAMGA VQA":
|
| 42 |
-
# VQA seçildiyse: Grubu GÖSTER, Metin kutusuna DOKUNMA (değişiklik yok)
|
| 43 |
return gr.update(visible=True), gr.update()
|
| 44 |
else:
|
| 45 |
-
# BLIP seçildiyse: Grubu GİZLE, Metin kutusunu TEMİZLE
|
| 46 |
return gr.update(visible=False), gr.update(value="")
|
| 47 |
|
| 48 |
def select_quick_question(quick_question):
|
|
@@ -52,9 +48,9 @@ def select_quick_question(quick_question):
|
|
| 52 |
return gr.update(value=quick_question), gr.update(value=None)
|
| 53 |
|
| 54 |
|
| 55 |
-
|
| 56 |
# Main Inference Function
|
| 57 |
-
|
| 58 |
def answer(model_choice, image, question):
|
| 59 |
|
| 60 |
if image is None:
|
|
@@ -70,8 +66,7 @@ def answer(model_choice, image, question):
|
|
| 70 |
return "**Hata: TAMGA VQA modeli yüklenemedi. 'inference.py' dosyasını ve bağımlılıkları kontrol edin.**"
|
| 71 |
|
| 72 |
pil_image = Image.fromarray(image)
|
| 73 |
-
|
| 74 |
-
# Assuming predict function is correctly implemented in inference.py
|
| 75 |
response = predict(tamga_model, tamga_tokenizer, tamga_device, pil_image, question)
|
| 76 |
return f"**Cevap:** {response}"
|
| 77 |
|
|
@@ -79,9 +74,7 @@ def answer(model_choice, image, question):
|
|
| 79 |
elif model_choice == "BLIP Caption (Fine-Tuned)":
|
| 80 |
|
| 81 |
pil_image = Image.fromarray(image)
|
| 82 |
-
# Ensure device is correctly set for inputs
|
| 83 |
inputs = caption_processor(images=pil_image, return_tensors="pt").to(caption_device)
|
| 84 |
-
# Generate caption
|
| 85 |
output = caption_model.generate(**inputs, max_new_tokens=64)
|
| 86 |
caption = caption_processor.decode(output[0], skip_special_tokens=True)
|
| 87 |
return f"**Açıklama:** {caption}"
|
|
@@ -89,9 +82,9 @@ def answer(model_choice, image, question):
|
|
| 89 |
return "**Model bulunamadı.**"
|
| 90 |
|
| 91 |
|
| 92 |
-
|
| 93 |
-
# CSS
|
| 94 |
-
|
| 95 |
css = """
|
| 96 |
#col-container {
|
| 97 |
max-width: 1100px;
|
|
@@ -224,19 +217,19 @@ button[variant="primary"]:hover {
|
|
| 224 |
# -----------------------
|
| 225 |
VQA_QUESTION_CHOICES = [
|
| 226 |
"Bu görselde kaç tane insan figürü var?",
|
| 227 |
-
"
|
| 228 |
"Fotoğrafta ne tür bir araç görülüyor?",
|
| 229 |
-
"Bu
|
| 230 |
]
|
| 231 |
|
| 232 |
|
| 233 |
-
|
| 234 |
# UI Layout
|
| 235 |
-
|
| 236 |
with gr.Blocks(css=css) as demo:
|
| 237 |
|
| 238 |
-
gr.HTML("<div id='title'>🇹🇷 TAMGA — Çok Modelli Görsel Dil
|
| 239 |
-
gr.HTML("<div id='subtitle'>VQA veya
|
| 240 |
|
| 241 |
with gr.Row(elem_id="col-container"):
|
| 242 |
|
|
@@ -260,7 +253,7 @@ with gr.Blocks(css=css) as demo:
|
|
| 260 |
)
|
| 261 |
# ----------------------
|
| 262 |
|
| 263 |
-
|
| 264 |
with gr.Group(visible=True) as vqa_inputs_group:
|
| 265 |
question = gr.Textbox(
|
| 266 |
label="Soru (Sadece VQA Modeli İçin)",
|
|
@@ -273,9 +266,9 @@ with gr.Blocks(css=css) as demo:
|
|
| 273 |
quick_question_radio = gr.Radio(
|
| 274 |
choices=VQA_QUESTION_CHOICES,
|
| 275 |
label="Hızlı Sorular",
|
| 276 |
-
value=None,
|
| 277 |
elem_id="quick-questions",
|
| 278 |
-
container=False
|
| 279 |
)
|
| 280 |
# --------------------------------------
|
| 281 |
|
|
@@ -286,7 +279,7 @@ with gr.Blocks(css=css) as demo:
|
|
| 286 |
|
| 287 |
output = gr.Markdown(elem_classes="output-box")
|
| 288 |
|
| 289 |
-
# Button click
|
| 290 |
submit_btn.click(
|
| 291 |
fn=answer,
|
| 292 |
inputs=[model_choice, image, question],
|
|
@@ -301,7 +294,7 @@ with gr.Blocks(css=css) as demo:
|
|
| 301 |
outputs=[vqa_inputs_group, question],
|
| 302 |
queue=False
|
| 303 |
)
|
| 304 |
-
|
| 305 |
|
| 306 |
|
| 307 |
# --- Quick Question Selection Logic ---
|
|
|
|
| 3 |
import torch
|
| 4 |
import os
|
| 5 |
|
| 6 |
+
|
|
|
|
| 7 |
from inference import load_for_inference, predict
|
| 8 |
|
|
|
|
| 9 |
# Load TAMGA VQA model
|
| 10 |
+
|
| 11 |
TAMGA_REPO = "Mueris/TurkishVLMTAMGA"
|
| 12 |
+
|
| 13 |
if 'load_for_inference' in globals():
|
| 14 |
tamga_model, tamga_tokenizer, tamga_device = load_for_inference(TAMGA_REPO)
|
| 15 |
else:
|
| 16 |
+
|
| 17 |
print("Warning: inference.py functions not loaded. Using placeholder values.")
|
| 18 |
tamga_model, tamga_tokenizer, tamga_device = None, None, 'cpu'
|
| 19 |
|
| 20 |
|
| 21 |
+
|
| 22 |
# Load BLIP Caption Model
|
| 23 |
+
|
| 24 |
from transformers import BlipProcessor, BlipForConditionalGeneration
|
| 25 |
|
| 26 |
CAPTION_REPO = "Mueris/TurkishVLMTAMGA-CaptioningModel"
|
|
|
|
| 30 |
caption_device = caption_model.device
|
| 31 |
|
| 32 |
|
| 33 |
+
|
| 34 |
# Utility Functions
|
| 35 |
+
|
| 36 |
|
| 37 |
def toggle_question_input(model_choice):
|
| 38 |
|
| 39 |
if model_choice == "TAMGA VQA":
|
|
|
|
| 40 |
return gr.update(visible=True), gr.update()
|
| 41 |
else:
|
|
|
|
| 42 |
return gr.update(visible=False), gr.update(value="")
|
| 43 |
|
| 44 |
def select_quick_question(quick_question):
|
|
|
|
| 48 |
return gr.update(value=quick_question), gr.update(value=None)
|
| 49 |
|
| 50 |
|
| 51 |
+
|
| 52 |
# Main Inference Function
|
| 53 |
+
|
| 54 |
def answer(model_choice, image, question):
|
| 55 |
|
| 56 |
if image is None:
|
|
|
|
| 66 |
return "**Hata: TAMGA VQA modeli yüklenemedi. 'inference.py' dosyasını ve bağımlılıkları kontrol edin.**"
|
| 67 |
|
| 68 |
pil_image = Image.fromarray(image)
|
| 69 |
+
|
|
|
|
| 70 |
response = predict(tamga_model, tamga_tokenizer, tamga_device, pil_image, question)
|
| 71 |
return f"**Cevap:** {response}"
|
| 72 |
|
|
|
|
| 74 |
elif model_choice == "BLIP Caption (Fine-Tuned)":
|
| 75 |
|
| 76 |
pil_image = Image.fromarray(image)
|
|
|
|
| 77 |
inputs = caption_processor(images=pil_image, return_tensors="pt").to(caption_device)
|
|
|
|
| 78 |
output = caption_model.generate(**inputs, max_new_tokens=64)
|
| 79 |
caption = caption_processor.decode(output[0], skip_special_tokens=True)
|
| 80 |
return f"**Açıklama:** {caption}"
|
|
|
|
| 82 |
return "**Model bulunamadı.**"
|
| 83 |
|
| 84 |
|
| 85 |
+
|
| 86 |
+
# CSS
|
| 87 |
+
|
| 88 |
css = """
|
| 89 |
#col-container {
|
| 90 |
max-width: 1100px;
|
|
|
|
| 217 |
# -----------------------
|
| 218 |
VQA_QUESTION_CHOICES = [
|
| 219 |
"Bu görselde kaç tane insan figürü var?",
|
| 220 |
+
"Görselde ne görüyorsun?",
|
| 221 |
"Fotoğrafta ne tür bir araç görülüyor?",
|
| 222 |
+
"Bu görselde hava aracı var mı?"
|
| 223 |
]
|
| 224 |
|
| 225 |
|
| 226 |
+
|
| 227 |
# UI Layout
|
| 228 |
+
|
| 229 |
with gr.Blocks(css=css) as demo:
|
| 230 |
|
| 231 |
+
gr.HTML("<div id='title'>🇹🇷 TAMGA — Çok Modelli Türkçe Görsel Dil Modeli</div>")
|
| 232 |
+
gr.HTML("<div id='subtitle'>TAMGA VQA (Soru Cevap) veya TAMGA Görsel Açıklama modellerinden birini seçin.</div>")
|
| 233 |
|
| 234 |
with gr.Row(elem_id="col-container"):
|
| 235 |
|
|
|
|
| 253 |
)
|
| 254 |
# ----------------------
|
| 255 |
|
| 256 |
+
|
| 257 |
with gr.Group(visible=True) as vqa_inputs_group:
|
| 258 |
question = gr.Textbox(
|
| 259 |
label="Soru (Sadece VQA Modeli İçin)",
|
|
|
|
| 266 |
quick_question_radio = gr.Radio(
|
| 267 |
choices=VQA_QUESTION_CHOICES,
|
| 268 |
label="Hızlı Sorular",
|
| 269 |
+
value=None,
|
| 270 |
elem_id="quick-questions",
|
| 271 |
+
container=False
|
| 272 |
)
|
| 273 |
# --------------------------------------
|
| 274 |
|
|
|
|
| 279 |
|
| 280 |
output = gr.Markdown(elem_classes="output-box")
|
| 281 |
|
| 282 |
+
# Button click run model
|
| 283 |
submit_btn.click(
|
| 284 |
fn=answer,
|
| 285 |
inputs=[model_choice, image, question],
|
|
|
|
| 294 |
outputs=[vqa_inputs_group, question],
|
| 295 |
queue=False
|
| 296 |
)
|
| 297 |
+
|
| 298 |
|
| 299 |
|
| 300 |
# --- Quick Question Selection Logic ---
|