from pathlib import Path import gradio as gr import torch import torch.nn.functional as F from fastai.vision.all import load_learner from PIL import Image # cpu-basic has 2 vCPUs; os.cpu_count() reports the host's cores instead torch.set_num_threads(2) learn = load_learner("model.pkl") model = learn.model.eval() categories = list(learn.dls.vocab) # Same as the learner's validation pipeline: squash to 256, center-crop 224, ImageNet normalize MEAN = torch.tensor([0.485, 0.456, 0.406]).view(3, 1, 1) STD = torch.tensor([0.229, 0.224, 0.225]).view(3, 1, 1) def preprocess(img: Image.Image) -> torch.Tensor: # reducing_gap makes downscaling large uploads much cheaper img = img.convert("RGB").resize((256, 256), Image.BILINEAR, reducing_gap=3.0) img = img.crop((16, 16, 240, 240)) x = torch.frombuffer(bytearray(img.tobytes()), dtype=torch.uint8).view(224, 224, 3) return ((x.permute(2, 0, 1).float() / 255 - MEAN) / STD).unsqueeze(0) @torch.inference_mode() def classify_image(img: Image.Image | None) -> dict[str, float]: """Identify the bird species in a photo. Returns the 5 most likely of 526 species with confidences.""" if img is None: raise gr.Error("Upload a bird photo first.") probs = F.softmax(model(preprocess(img))[0], dim=0) top = probs.topk(5) return {categories[i].title(): float(p) for p, i in zip(top.values, top.indices)} classify_image(Image.new("RGB", (256, 256))) # warm-up so the first visitor isn't slow examples = sorted(str(p) for p in Path("examples").glob("*.jpg")) with gr.Blocks(title="Bird Classification") as demo: gr.Markdown( f"# 🐦 Bird Classification\n" f"Upload a photo of a bird to identify it among **{len(categories)} species**. " "Works best with one bird, clearly visible and filling most of the frame." ) with gr.Row(equal_height=True): with gr.Column(): image = gr.Image(type="pil", label="Bird photo", sources=["upload", "webcam", "clipboard"], height=360) with gr.Row(): clear = gr.ClearButton(value="Clear") submit = gr.Button("Identify", variant="primary") label = gr.Label(num_top_classes=5, label="Top 5 predictions") clear.add([image, label]) submit.click(classify_image, image, label, api_name="classify_image") image.upload(classify_image, image, label, api_visibility="private") # after the events, so the public endpoint keeps the name "classify_image" gr.Examples(examples, inputs=image, outputs=label, fn=classify_image, run_on_click=True, cache_examples=False, api_visibility="private") demo.queue(max_size=20).launch(theme=gr.themes.Soft())