Minimal Space: requirements, model, app, placeholder classes

app.py

@@ -0,0 +1,65 @@
+from __future__ import annotations
+from pathlib import Path
+import gradio as gr
+import torch, torch.nn.functional as F
+from PIL import Image
+from torchvision import transforms
+from model import build_model
+
+ROOT = Path(__file__).resolve().parent
+CKPT = ROOT / "ckpt_final320" / "best.pt"
+CLASSES_TXT = ROOT / "classes.txt"
+
+IMAGENET_MEAN = (0.485, 0.456, 0.406)
+IMAGENET_STD  = (0.229, 0.224, 0.225)
+RES = 320
+
+def _load_classes(p: Path):
+    return [ln.strip() for ln in p.read_text().splitlines() if ln.strip()] if p.exists() else []
+
+CLASSES = _load_classes(CLASSES_TXT)
+MODEL = None
+
+def _maybe_load_model():
+    global MODEL, CLASSES
+    if MODEL is not None:
+        return
+    CLASSES = _load_classes(CLASSES_TXT) or ["class_0","class_1"]
+    m = build_model(len(CLASSES), pretrained=False)
+    if CKPT.exists():
+        sd = torch.load(CKPT, map_location="cpu")
+        sd = sd["model"] if isinstance(sd, dict) and "model" in sd else sd
+        m.load_state_dict(sd, strict=False)
+    m.eval()
+    MODEL = m
+
+TFM = transforms.Compose([
+    transforms.Resize(int(RES*256/224)),
+    transforms.CenterCrop(RES),
+    transforms.ToTensor(),
+    transforms.Normalize(IMAGENET_MEAN, IMAGENET_STD),
+])
+
+def predict(img: Image.Image):
+    _maybe_load_model()
+    if img is None:
+        return "Please upload an image."
+    x = TFM(img.convert("RGB")).unsqueeze(0)
+    with torch.inference_mode():
+        logits = MODEL(x)
+        probs = F.softmax(logits, dim=1)[0]
+    i = int(torch.argmax(probs).item())
+    p = float(probs[i].item())
+    name = CLASSES[i] if i < len(CLASSES) else f"class_{i}"
+    note = "" if CKPT.exists() else "_Running without checkpoint (demo mode). Upload ckpt_final320/best.pt to get real predictions._"
+    return f"**Prediction:** {name} — **{p*100:.2f}%**\n\n{note}"
+
+with gr.Blocks(title="Bird Classifier (Minimal)") as demo:
+    gr.Markdown("### Bird Classifier — Minimal Space\nUpload an image and predict. (Works even without checkpoint.)")
+    img = gr.Image(type="pil", label="Image")
+    btn = gr.Button("Predict", variant="primary")
+    out = gr.Markdown()
+    btn.click(predict, inputs=[img], outputs=[out], show_progress="full")
+
+if __name__ == "__main__":
+    demo.launch()

classes.txt

@@ -0,0 +1,2 @@
+bird_a
+bird_b

model.py

@@ -0,0 +1,9 @@
+from __future__ import annotations
+import torch.nn as nn
+from torchvision.models import resnet50
+
+def build_model(num_classes: int, pretrained: bool = False) -> nn.Module:
+    m = resnet50(pretrained=pretrained)
+    in_features = m.fc.in_features
+    m.fc = nn.Linear(in_features, num_classes)
+    return m

requirements.txt

@@ -0,0 +1,7 @@
+gradio==4.44.0
+torch
+torchvision
+numpy
+pillow
+matplotlib
+scikit-learn