Spaces:
Runtime error
Runtime error
File size: 16,377 Bytes
ab5fa16 5d2edec | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 | import streamlit as st
import torch
import requests
import os
import json
from transformers import AutoImageProcessor, AutoModelForImageClassification, AutoConfig
from peft import PeftModel, PeftConfig
from PIL import Image
import torch.nn.functional as F
import io
# 1. ํ์ด์ง ๊ธฐ๋ณธ ์ค์ (๊ฐ์ฅ ๋จผ์ ํธ์ถํด์ผ ํจ)
st.set_page_config(page_title="Pokemon Classifier", page_icon="๐พ", layout="wide", initial_sidebar_state="collapsed")
# 2. ์ปค์คํ
CSS ์ ์ฉ (๋ ์๋น์ค์ค๋ฌ์ด ๋๋์ ์ํด)
st.markdown("""
<style>
.main-header {
text-align: center;
padding: 2rem 0 3rem 0;
}
.main-header h1 {
font-size: 3.5rem;
color: #ffcb05;
-webkit-text-stroke: 2px #3c5aa6;
text-shadow: 4px 4px 0px #3c5aa6;
margin-bottom: 0.5rem;
}
.main-header p {
font-size: 1.2rem;
color: #555;
font-weight: 500;
}
/* ์นด๋ ๋๋์ ์ปจํ
์ด๋๋ฅผ ์ํด ์ฌ๋ฐฑ ์กฐ์ */
div[data-testid="stVerticalBlock"] {
gap: 1.2rem;
}
</style>
<div class="main-header">
<h1>Pokemon Classifier</h1>
</div>
""", unsafe_allow_html=True)
# 3. ๋ง๋ฅ ๋ชจ๋ธ ๋ก๋ ํจ์ (์บ์ฑ & ์๋ฌ ๋ฐฉ์ด ์ ์ฉ)
@st.cache_resource(show_spinner=False)
def load_model(model_path):
try:
# ํ๊น
ํ์ด์ค ๋ฆฌํฌ์งํ ๋ฆฌ ๋๋ ๋ก์ปฌ ๊ฒฝ๋ก์์ PEFT ์ค์ ํ์ผ ํ์ธ
try:
config = PeftConfig.from_pretrained(model_path)
is_peft = True
except:
is_peft = False
if is_peft:
base_model_name = config.base_model_name_or_path
try:
processor = AutoImageProcessor.from_pretrained(model_path)
except:
processor = AutoImageProcessor.from_pretrained(base_model_name)
# 1. ํฌ์ผ๋ชฌ ๋ถ๋ฅ๋ ๋ฌด์กฐ๊ฑด 150๊ฐ์ ํด๋์ค์
๋๋ค.
num_labels = 150
# 2. ๋ผ๋ฒจ ๋งคํ ๋์
๋๋ฆฌ ๊ตฌ์ฑ (์ง์ ์์ฑํ์ฌ ์ถฉ๋ ๋ฐฉ์ง)
try:
# ๋ก์ปฌ์ ์ ์ฅ๋ Full FT ๋ชจ๋ธ์ config๊ฐ ์๋ค๋ฉด ๊ฐ์ฅ ์๋ฒฝํ ํฌ์ผ๋ชฌ ์ด๋ฆ ๋์
๋๋ฆฌ ์ฌ์ฉ
config_path = "./saved_model/best_vit_full/config.json"
with open(config_path, "r", encoding="utf-8") as f:
full_config = json.load(f)
id2label = {int(k): v for k, v in full_config["id2label"].items()}
label2id = full_config["label2id"]
except Exception:
try:
# ๋ก์ปฌ ํ์ผ์ด ์์ผ๋ฉด ๋ฐฐํฌ๋ ํ๊น
ํ์ด์ค ์ ์ฅ์์์ ์ฌ๋ฐ๋ฅธ ๋งคํ์ ๊ฐ์ ๋ก ๊ฐ์ ธ์ต๋๋ค.
reference_config = AutoConfig.from_pretrained("gyann/pokemon-vit-full")
id2label = {int(k): v for k, v in reference_config.id2label.items()}
label2id = reference_config.label2id
except Exception:
# ์ตํ์ ์๋จ์ผ๋ก ๋๋ฏธ ์์ฑ
id2label = {i: f"LABEL_{i}" for i in range(num_labels)}
label2id = {f"LABEL_{i}": i for i in range(num_labels)}
# 3. Base ๋ชจ๋ธ์ ๋ก๋ํ ๋ ๋ฐ๋์ num_labels๋ฅผ ๋ช
์ํด์ผ classifier ํค๋ ์ฌ์ด์ฆ๊ฐ 150์ผ๋ก ์ด๊ธฐํ๋ฉ๋๋ค.
base_model = AutoModelForImageClassification.from_pretrained(
base_model_name,
num_labels=num_labels,
id2label=id2label,
label2id=label2id,
ignore_mismatched_sizes=True
)
# 4. 150 ์ฌ์ด์ฆ๋ก ๋ง์ถฐ์ง Base ๋ชจ๋ธ์ LoRA ๊ฐ์ค์น ๊ฒฐํฉ
model = PeftModel.from_pretrained(base_model, model_path)
else:
processor = AutoImageProcessor.from_pretrained(model_path)
model = AutoModelForImageClassification.from_pretrained(model_path)
# ๋น-PEFT ๋ชจ๋ธ๋ config์ ์ด๋ฆ์ด ์๋ ๊ฒฝ์ฐ(Label_15 ๋ฑ)๋ฅผ ๋๋นํด ๋งคํ์ ๋ฎ์ด์์๋๋ค.
if getattr(model.config, "id2label", {}).get(0, "") == "LABEL_0" or getattr(model.config, "id2label", {}).get("0", "") == "LABEL_0":
try:
reference_config = AutoConfig.from_pretrained("gyann/pokemon-vit-full")
model.config.id2label = {int(k): v for k, v in reference_config.id2label.items()}
model.config.label2id = reference_config.label2id
except Exception:
pass
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
model.eval()
return processor, model, device, None
except Exception as e:
return None, None, None, str(e)
# 4. PokeAPI ๋ฐ์ดํฐ ํธ์ถ ํจ์ (์ด๋ฏธ์ง, ํ์
, ํค, ๋ชธ๋ฌด๊ฒ)
@st.cache_data(show_spinner=False)
def get_pokemon_data(pokemon_name):
try:
name_lower = pokemon_name.lower().replace(" ", "-").replace(".", "").replace("'", "")
url = f"https://pokeapi.co/api/v2/pokemon/{name_lower}"
response = requests.get(url)
if response.status_code == 200:
data = response.json()
return {
"artwork_url": data["sprites"]["other"]["official-artwork"]["front_default"],
"types": [t["type"]["name"] for t in data["types"]],
"height": data["height"] / 10.0, # meters
"weight": data["weight"] / 10.0 # kg
}
return None
except:
return None
MODEL_PATHS = {
"ViT Full Fine-tuning": "gyann/pokemon-vit-full",
"ViT + LoRA": "gyann/pokemon-vit-lora",
"ViT + QLoRA (4-bit)": "gyann/pokemon-vit-qlora",
"ResNet50": "gyann/pokemon-resnet50",
"ConvNeXt": "gyann/pokemon-convnext",
"Swin Transformer": "gyann/pokemon-swin"
}
# 5. ์ค์ ๋ฐ ์
๋ ฅ ์์ญ (๊น๋ํ ์ปจํ
์ด๋ UI)
with st.container(border=True):
st.markdown("### โ๏ธ ๋ถ์ ์ค์ ")
setting_col1, setting_col2 = st.columns([1, 2])
with setting_col1:
compare_models = st.toggle("๋ชจ๋ธ ๋น๊ต ๋ชจ๋ ํ์ฑํ", value=False)
st.caption(f"๐ ํ์ฌ ๊ฐ์ ์ฅ์น: **{'GPU (CUDA)' if torch.cuda.is_available() else 'CPU'}**")
with setting_col2:
if compare_models:
col_a, col_b = st.columns(2)
with col_a:
model_name_a = st.selectbox("Model A", list(MODEL_PATHS.keys()), index=0)
with col_b:
model_name_b = st.selectbox("Model B", list(MODEL_PATHS.keys()), index=1)
with st.spinner("๋ชจ๋ธ A ๋ก๋ฉ ์ค..."):
processor_a, model_a, device_a, err_a = load_model(MODEL_PATHS[model_name_a])
with st.spinner("๋ชจ๋ธ B ๋ก๋ฉ ์ค..."):
processor_b, model_b, device_b, err_b = load_model(MODEL_PATHS[model_name_b])
if err_a: st.error(f"Model A ๋ก๋ ์คํจ: {err_a}")
if err_b: st.error(f"Model B ๋ก๋ ์คํจ: {err_b}")
active_models = [
("Model A", model_name_a, processor_a, model_a, device_a),
("Model B", model_name_b, processor_b, model_b, device_b)
]
else:
model_name = st.selectbox("๋ถ์์ ์ฌ์ฉํ ๋ชจ๋ธ", list(MODEL_PATHS.keys()))
with st.spinner("๋ชจ๋ธ ๋ก๋ฉ ์ค..."):
processor, model, device, err = load_model(MODEL_PATHS[model_name])
if err: st.error(f"๋ชจ๋ธ ๋ก๋ ์คํจ: {err}")
active_models = [("Result", model_name, processor, model, device)]
with st.container(border=True):
st.markdown("### ๐ผ๏ธ ์ด๋ฏธ์ง ์ ํ")
EXAMPLE_IMAGES = {
"์ง์ ํ์ผ ์
๋ก๋ํ๊ธฐ": None,
"์์: ํผ์นด์ธ (Pikachu)": "https://raw.githubusercontent.com/PokeAPI/sprites/master/sprites/pokemon/other/official-artwork/25.png",
"์์: ์ด์ํด์จ (Bulbasaur)": "https://raw.githubusercontent.com/PokeAPI/sprites/master/sprites/pokemon/other/official-artwork/1.png",
"์์: ๊ผฌ๋ถ๊ธฐ (Squirtle)": "https://raw.githubusercontent.com/PokeAPI/sprites/master/sprites/pokemon/other/official-artwork/7.png",
"์์: ํ์ด๋ฆฌ (Charmander)": "https://raw.githubusercontent.com/PokeAPI/sprites/master/sprites/pokemon/other/official-artwork/4.png"
}
# ๋ผ๋์ค ๋ฒํผ์ ๊ฐ๋ก๋ก ๋ฐฐ์นํ์ฌ ๋ชจ๋ํ๊ฒ
selected_example = st.radio("ํ
์คํธ ๋ฐฉ์์ ์ ํํ์ธ์", list(EXAMPLE_IMAGES.keys()), horizontal=True, label_visibility="collapsed")
image_to_process = None
if selected_example == "์ง์ ํ์ผ ์
๋ก๋ํ๊ธฐ":
uploaded_file = st.file_uploader("ํฌ์ผ๋ชฌ ์ด๋ฏธ์ง๋ฅผ ๋๋๊ทธ ์ค ๋๋กญํ์ธ์", type=["jpg", "jpeg", "png"], label_visibility="collapsed")
if uploaded_file is not None:
image_to_process = Image.open(uploaded_file).convert("RGB")
else:
example_url = EXAMPLE_IMAGES[selected_example]
try:
response = requests.get(example_url)
if response.status_code == 200:
image_to_process = Image.open(io.BytesIO(response.content)).convert("RGB")
except Exception as e:
st.error(f"์์ ์ด๋ฏธ์ง๋ฅผ ๋ถ๋ฌ์ค๋ ์ค ์ค๋ฅ๊ฐ ๋ฐ์ํ์ต๋๋ค: {e}")
# 6. ์ถ๋ก ๋ฐ ๊ฒฐ๊ณผ ์ถ๋ ฅ ์ปดํฌ๋ํธ
def predict_and_display(name_prefix, model_name, processor, model, device, image):
if model is None:
st.warning(f"๋ชจ๋ธ({model_name})์ด ์ ์์ ์ผ๋ก ๋ก๋๋์ง ์์์ต๋๋ค.")
return
inputs = processor(images=image, return_tensors="pt").to(device)
with torch.no_grad():
outputs = model(**inputs)
logits = outputs.logits
probs = F.softmax(logits, dim=-1)[0]
top5_prob, top5_catid = torch.topk(probs, 5)
cat_id_0 = top5_catid[0].item()
top1_label = model.config.id2label.get(cat_id_0, model.config.id2label.get(str(cat_id_0), "Unknown"))
top1_score = top5_prob[0].item()
poke_data = get_pokemon_data(top1_label)
st.markdown(f"<h5 style='text-align: center; color: #777; margin-bottom: 0;'>{name_prefix}: {model_name}</h5>", unsafe_allow_html=True)
# ๋ฉํธ๋ฆญ
st.metric(label="๐ Top-1 Prediction", value=top1_label.title(), delta=f"{top1_score*100:.1f}% Confidence", delta_color="normal")
tab1, tab2, tab3 = st.tabs(["๐ Overview", "๐ Stats", "๐ Top-5 Details"])
with tab1:
if poke_data and poke_data["artwork_url"]:
st.image(poke_data["artwork_url"], use_container_width=True)
else:
st.info("๊ณต์ ์ผ๋ฌ์คํธ๊ฐ ์์ต๋๋ค.")
with tab2:
if poke_data:
st.markdown(f"**์์ฑ(Types)**: {', '.join(poke_data['types']).title()}")
st.markdown(f"**์ ์ฅ(Height)**: {poke_data['height']} m")
st.markdown(f"**์ฒด์ค(Weight)**: {poke_data['weight']} kg")
else:
st.warning("๋๊ฐ ์ ๋ณด๋ฅผ ๊ฐ์ ธ์ฌ ์ ์์ต๋๋ค.")
with tab3:
for i in range(top5_prob.size(0)):
c_id = top5_catid[i].item()
lbl = model.config.id2label.get(c_id, model.config.id2label.get(str(c_id), "Unknown"))
scr = top5_prob[i].item()
st.caption(f"**{i+1}. {lbl.title()}**")
st.progress(scr, text=f"{scr * 100:.1f}%")
# 7. ๋ฉ์ธ ์คํ ๋ถ
if image_to_process is not None:
st.markdown("---")
with st.container():
if compare_models:
col_img, col_res = st.columns([1, 2.5])
with col_img:
st.markdown("### ๐ท ์
๋ ฅ ์ด๋ฏธ์ง")
st.image(image_to_process, use_container_width=True)
with col_res:
st.markdown("### โจ ๋ถ์ ๊ฒฐ๊ณผ")
col1, col2 = st.columns(2)
with col1:
with st.container(border=True):
m_prefix, m_name, m_proc, m_model, m_dev = active_models[0]
with st.spinner('๋ถ์ ์ค...'):
predict_and_display(m_prefix, m_name, m_proc, m_model, m_dev, image_to_process)
with col2:
with st.container(border=True):
m_prefix, m_name, m_proc, m_model, m_dev = active_models[1]
with st.spinner('๋ถ์ ์ค...'):
predict_and_display(m_prefix, m_name, m_proc, m_model, m_dev, image_to_process)
else:
col1, col2 = st.columns([1, 1.5])
with col1:
st.markdown("### ๐ท ์
๋ ฅ ์ด๋ฏธ์ง")
st.image(image_to_process, use_container_width=True)
with col2:
st.markdown("### โจ ๋ถ์ ๊ฒฐ๊ณผ")
with st.container(border=True):
m_prefix, m_name, m_proc, m_model, m_dev = active_models[0]
with st.spinner('๋ถ์ ์ค...'):
predict_and_display(m_prefix, m_name, m_proc, m_model, m_dev, image_to_process)
# 8. ๋ชจ๋ธ ์ํคํ
์ฒ ์ ๋ณด UI (์ตํ๋จ ๋ฐฐ์น)
ARCHITECTURE_INFO = {
"ViT Full Fine-tuning": {
"title": "Vision Transformer (ViT) - Full FT",
"desc": "์ด๋ฏธ์ง๋ฅผ 16x16 ํฝ์
ํจ์น(Patch)๋ก ๋๋์ด ์ฒ๋ฆฌํฉ๋๋ค. ์ด๋ฏธ์ง ์ ์ฒด์ **์ ์ญ์ ๋ฌธ๋งฅ(Global Context)** ์ ํ์
ํ๋ ๋ฐ ๋ฐ์ด๋๋ฉฐ, ๋ชจ๋ ๊ฐ์ค์น๋ฅผ ์ฌํ์ตํ์ฌ ์ต๊ณ ์ฑ๋ฅ์ ๋์ถํ์ง๋ง ์ฐ์ฐ ๋น์ฉ์ด ๊ฐ์ฅ ํฝ๋๋ค."
},
"ViT + LoRA": {
"title": "ViT with LoRA (Low-Rank Adaptation)",
"desc": "๊ฑฐ๋ํ ์๋ณธ ๋ชจ๋ธ์ ์ผ๋ ค๋๊ณ (Freeze), ํต์ฌ ์ฐ์ฐ์ธต์ ์์ฃผ ์์ **ํ์ต ๊ฐ๋ฅํ ์ฐํ๋ก(์ ๋ญํฌ ํ๋ ฌ)** ๋ฅผ ๋ง๋ถ์
๋๋ค. ๋จ 1%์ ํ๋ผ๋ฏธํฐ๋ง ํ์ตํ์ฌ ๋น์ฉ์ ๊ทน์ ์ผ๋ก ๋ฎ์ถ๋ฉด์๋ Full FT์ ์ ์ฌํ ์ฑ๋ฅ์ ๋
๋๋ค."
},
"ViT + QLoRA (4-bit)": {
"title": "ViT with QLoRA (Quantized LoRA)",
"desc": "LoRA์์ ํ ๋ฐ ๋ ๋์๊ฐ, ์๋ณธ ๋ชจ๋ธ์ **4-bit ์ ๋ฐ๋** ๋ก ์์ถ(Quantization)ํ์ฌ ๋ฉ๋ชจ๋ฆฌ์ ์ ์ฌํฉ๋๋ค. VRAM์ด ๋งค์ฐ ๋ถ์กฑํ ํ๊ฒฝ์์๋ ๋๊ท๋ชจ ๋ชจ๋ธ์ ํ๋ํ ์ ์๊ฒ ํ๋ ์ต์ ํ ๊ธฐ๋ฒ์
๋๋ค."
},
"ResNet50": {
"title": "ResNet50 (Baseline CNN)",
"desc": "ํฉ์ฑ๊ณฑ(Convolution) ํํฐ๋ฅผ ๊ฒน์ณ ์ด๋ฏธ์ง์ **๊ตญ์์ ํจํด(Local Feature)** ์ ์ฐพ์๋ด๋ ์ ํต์ ๊ฐ์์
๋๋ค. ์์ฐจ ์ฐ๊ฒฐ(Residual Connection)๋ก ๊น์ ์ ๊ฒฝ๋ง์ ํ์ต ์์ ์ฑ์ ๋ณด์ฅํฉ๋๋ค."
},
"ConvNeXt": {
"title": "ConvNeXt (Modernized CNN)",
"desc": "ํธ๋์คํฌ๋จธ์ ์ค๊ณ ์ฒ ํ(ํฐ ์ปค๋, LayerNorm, GELU ๋ฑ)์ ์ญ์ผ๋ก CNN์ ๋์
ํ **'๋ชจ๋ CNN'** ์
๋๋ค. CNN์ ์ง์ญ์ ๊ท๋ฉ์ ํธํฅ์ ์ ์งํ๋ฉด์๋ ํธ๋์คํฌ๋จธ๊ธ ์ฑ๋ฅ์ ๋
๋๋ค."
},
"Swin Transformer": {
"title": "Swin Transformer (Hierarchical ViT)",
"desc": "CNN์ฒ๋ผ ์์ ์์ญ(Window)๋ถํฐ ์ ์ ๋์ ์์ญ์ผ๋ก **๊ณ์ธต์ (Hierarchical)** ์ผ๋ก ๋ณํฉํ๋ฉฐ ํ์ตํ๋ ํธ๋์คํฌ๋จธ์
๋๋ค. ViT๊ฐ ๋์น๊ธฐ ์ฌ์ด ๋ฏธ์ธํ ๋ํ
์ผ ๊ตฌ๋ถ์ ๊ฐํฉ๋๋ค."
}
}
st.markdown("---")
st.markdown("#### ๐ง ์ ํ๋ ์ํคํ
์ฒ ์์๋ณด๊ธฐ")
if compare_models:
col_info1, col_info2 = st.columns(2)
with col_info1:
with st.expander(f"{model_name_a} ๊ตฌ์กฐ", expanded=False):
info = ARCHITECTURE_INFO.get(model_name_a, {"title": model_name_a, "desc": "์ค๋ช
์ด ์ค๋น๋์ง ์์์ต๋๋ค."})
st.markdown(f"**{info['title']}**\n\n{info['desc']}")
with col_info2:
with st.expander(f"{model_name_b} ๊ตฌ์กฐ", expanded=False):
info = ARCHITECTURE_INFO.get(model_name_b, {"title": model_name_b, "desc": "์ค๋ช
์ด ์ค๋น๋์ง ์์์ต๋๋ค."})
st.markdown(f"**{info['title']}**\n\n{info['desc']}")
else:
with st.expander(f"{model_name} ๊ตฌ์กฐ", expanded=False):
info = ARCHITECTURE_INFO.get(model_name, {"title": model_name, "desc": "์ค๋ช
์ด ์ค๋น๋์ง ์์์ต๋๋ค."})
st.markdown(f"**{info['title']}**\n\n{info['desc']}") |