| """Standalone inference for the Planktoscope phytoplankton classifier. |
| Requires: torch, timm, safetensors, torchvision, pillow, huggingface_hub. |
| """ |
| import json |
| import torch |
| import timm |
| from huggingface_hub import hf_hub_download |
| from safetensors.torch import load_file |
| from PIL import Image, ImageOps |
| import torchvision.transforms as T |
|
|
| REPO = "patcdaniel/planktoscope-phytoplankton-classifier" |
| cfg = json.load(open(hf_hub_download(REPO, "config.json"))) |
|
|
| |
| backbone = timm.create_model("vit_small_patch14_dinov2.lvd142m", |
| pretrained=False, num_classes=0, |
| img_size=cfg["image_size"]) |
| head = torch.nn.Linear(cfg["embedding_dim"], cfg["num_labels"]) |
| sd = load_file(hf_hub_download(REPO, "model.safetensors")) |
| backbone.load_state_dict({k[9:]: v for k, v in sd.items() if k.startswith("backbone.")}) |
| head.load_state_dict({k[5:]: v for k, v in sd.items() if k.startswith("head.")}) |
| backbone.eval(); head.eval() |
|
|
| |
| fill = tuple(cfg["preprocessing"]["pad_to_square_fill"]) |
| size = cfg["image_size"] |
| def preprocess(img: Image.Image): |
| w, h = img.size; s = max(w, h) |
| canvas = Image.new("RGB", (s, s), fill) |
| canvas.paste(img.convert("RGB"), ((s - w) // 2, (s - h) // 2)) |
| tf = T.Compose([T.Resize((size, size)), T.ToTensor(), |
| T.Normalize(cfg["preprocessing"]["normalize_mean"], |
| cfg["preprocessing"]["normalize_std"])]) |
| return tf(canvas) |
|
|
| @torch.no_grad() |
| def predict(path): |
| x = preprocess(Image.open(path)).unsqueeze(0) |
| probs = head(backbone(x)).softmax(1)[0] |
| conf, idx = probs.max(0) |
| label = cfg["id2label"][str(int(idx))] |
| |
| if conf < cfg["recommended_min_confidence"]: |
| return "Unassigned", float(conf) |
| return label, float(conf) |
|
|
| print(predict("example.jpg")) |
|
|
