Spaces:
Running on Zero
Running on Zero
| """ | |
| The crop contract and the model's input transform, shared by training (src/epills/training) and the API. | |
| Crop contract: a pill is cropped as a square window WINDOW x the long side of its *tight* box (the pill only, as a | |
| detector gives it), centred on the box, from the full-resolution photo. Outside the photo the edge pixels are | |
| repeated. A detector (Stage 1) sends boxes; the API crops with crop_window. | |
| Each model's window is chosen on validation by scripts/calibrate_stage2.py --windows and travels with the model | |
| (Embedder.window); run 3: 1.4 (validation: 1.3-1.7 all within 0.6 points). WINDOW = 1.7 is only the default for a | |
| model without one: it reproduces the training crops, where the pill alone fills a median 59% of the crop side | |
| (measured 2026-10-09 on 581 training crops; 10-90%: 41-69%; the C3PI cropper's 1.4 applies to its own boxes, | |
| which include the shadow). | |
| """ | |
| import numpy as np | |
| import torch | |
| from PIL import Image | |
| from torchvision.transforms import v2 as T | |
| IMAGE_SIZE = 224 # model input | |
| CROP_SIZE = 384 # stored crops | |
| WINDOW = 1.7 # crop window / long side of a tight pill box | |
| def eval_transform(size=IMAGE_SIZE): | |
| """PIL crop -> float tensor (3, size, size) in [0, 1]; what the model sees at evaluation and in the API.""" | |
| return T.Compose([T.ToImage(), T.Resize(size, antialias=True), T.CenterCrop(size), | |
| T.ToDtype(torch.float32, scale=True)]) | |
| def crop_window(image, box, window=WINDOW, out_size=CROP_SIZE): | |
| """ | |
| The contract crop: a PIL image of the square window `window` x the long side of box = (x0, y0, x1, y1), in | |
| the photo's pixels, resized to out_size. Parts outside the photo repeat the edge pixels. | |
| """ | |
| x0, y0, x1, y1 = (float(v) for v in box) | |
| if x1 <= x0 or y1 <= y0: | |
| raise ValueError(f"empty box {box}") | |
| side = window * max(x1 - x0, y1 - y0) | |
| cx, cy = (x0 + x1) / 2, (y0 + y1) / 2 | |
| left, top = int(round(cx - side / 2)), int(round(cy - side / 2)) | |
| s = max(1, int(round(side))) | |
| arr = np.asarray(image.convert("RGB")) | |
| h, w = arr.shape[:2] | |
| pad = ((max(0, -top), max(0, top + s - h)), (max(0, -left), max(0, left + s - w)), (0, 0)) | |
| if any(p for pair in pad[:2] for p in pair): | |
| arr = np.pad(arr, pad, mode="edge") | |
| left, top = left + pad[1][0], top + pad[0][0] | |
| crop = Image.fromarray(arr[top:top + s, left:left + s]) | |
| return crop.resize((out_size, out_size), Image.Resampling.BICUBIC) | |