feat: better pasting function

Former-commit-id: 43fedd3f6bafb51fe604e347f59b70cd5b0cc218 [formerly 51bb06c3b98df613710b329d3ade1febaf2b0b23]
Former-commit-id: 46b89acd2b860d272ce8a13cf2c8c955d7545c46

comp.ipynb.REMOVED.git-id

@@ -0,0 +1 @@
+a27ff5fec9fd71b7846a70ebc473984e859912b8

src/evaluate.py

@@ -1,8 +1,8 @@
 import torch
-import torch.nn.functional as F
 from tqdm import tqdm
 
-from src.utils.dice import dice_coeff, multiclass_dice_coeff
+import wandb
+from src.utils.dice import dice_coeff
 
 
 def evaluate(net, dataloader, device):
@@ -27,6 +27,12 @@ def evaluate(net, dataloader, device):
 
             pbar.update(images.shape[0])
 
+    # save some images to wandb
+    table = wandb.Table(columns=["image", "mask", "prediction"])
+    for img, mask, pred in zip(images.to("cpu"), masks_true.to("cpu"), masks_pred.to("cpu")):
+        table.add_data(wandb.Image(img), wandb.Image(mask), wandb.Image(pred))
+    wandb.log({"predictions_table": table}, commit=False)
+
     net.train()
 
     # Fixes a potential division by zero error

src/train.py

@@ -14,7 +14,6 @@ from tqdm import tqdm
 import wandb
 from evaluate import evaluate
 from src.utils.dataset import SphereDataset
-from src.utils.dice import dice_loss
 from unet import UNet
 from utils.paste import RandomPaste
 
@@ -43,7 +42,7 @@ def get_args():
         dest="batch_size",
         metavar="B",
         type=int,
-        default=10,
+        default=16,
         help="Batch size",
     )
     parser.add_argument(
@@ -195,21 +194,21 @@ def main():
 
                     wandb.log(  # log training metrics
                         {
-                            "train/epoch": epoch + step / epoch,
+                            "train/epoch": epoch + step / len(train_loader),
                             "train/train_loss": train_loss,
                         }
                     )
 
                 # Evaluation round
-                val_loss = evaluate(net, val_loader, device)
-                scheduler.step(val_loss)
+                val_score = evaluate(net, val_loader, device)
+                scheduler.step(val_score)
                 wandb.log(  # log validation metrics
                     {
-                        "val/val_loss": val_loss,
+                        "val/val_score": val_score,
                     }
                 )
 
-            print(f"Train Loss: {train_loss:.3f}, Valid Loss: {val_loss:3f}")
+            print(f"Train Loss: {train_loss:.3f}, Valid Score: {val_score:3f}")
 
             # save weights when epoch end
             Path(CHECKPOINT_DIR).mkdir(parents=True, exist_ok=True)

src/utils/paste.py

@@ -2,7 +2,6 @@ import os
 import random as rd
 
 import albumentations as A
-import cv2
 import numpy as np
 from PIL import Image
 
@@ -23,16 +22,16 @@ class RandomPaste(A.DualTransform):
     def __init__(
         self,
         nb,
-        scale_limit,
         path_paste_img_dir,
         path_paste_mask_dir,
+        scale_range=(0.1, 0.2),
         always_apply=True,
         p=1.0,
     ):
         super().__init__(always_apply, p)
         self.path_paste_img_dir = path_paste_img_dir
         self.path_paste_mask_dir = path_paste_mask_dir
-        self.scale_limit = scale_limit
+        self.scale_range = scale_range
         self.nb = nb
 
     @property
@@ -40,76 +39,77 @@ class RandomPaste(A.DualTransform):
         return ["image"]
 
     def apply(self, img, positions, paste_img, paste_mask, **params):
-        img = img.copy()
+        # convert img to Image, needed for `paste` function
+        img = Image.fromarray(img)
 
-        w, h = paste_mask.shape
-        mask_b = paste_mask > 0
-        mask_rgb_b = np.stack([mask_b, mask_b, mask_b], axis=2)
+        for pos in positions:
+            img.paste(paste_img, pos, paste_mask)
 
-        for (x, y) in positions:
-            img[x : x + w, y : y + h] = img[x : x + w, y : y + h] * ~mask_rgb_b + paste_img * mask_rgb_b
-
-        return img
+        return np.asarray(img.convert("RGB"))
 
     def apply_to_mask(self, mask, positions, paste_mask, **params):
-        mask = mask.copy()
+        # convert mask to Image, needed for `paste` function
+        mask = Image.fromarray(mask)
 
-        w, h = paste_mask.shape
-        mask_b = paste_mask > 0
+        for pos in positions:
+            mask.paste(paste_mask, pos, paste_mask)
 
-        for (x, y) in positions:
-            mask[x : x + w, y : y + h] = mask[x : x + w, y : y + h] * ~mask_b + mask_b
-
-        return mask
+        return np.asarray(mask.convert("L"))
 
     def get_params_dependent_on_targets(self, params):
+        # choose a random image inside the image folder
         filename = rd.choice(os.listdir(self.path_paste_img_dir))
 
-        paste_img = np.array(
-            Image.open(
+        # load the "paste" image
+        paste_img = Image.open(
             os.path.join(
                 self.path_paste_img_dir,
                 filename,
             )
-            ).convert("RGB"),
-            dtype=np.uint8,
-        )
+        ).convert("RGBA")
 
-        paste_mask = (
-            np.array(
-                Image.open(
+        # load its respective mask
+        paste_mask = Image.open(
             os.path.join(
                 self.path_paste_mask_dir,
                 filename,
             )
-                ).convert("L"),
-                dtype=np.float32,
-            )
-            / 255
-        )
+        ).convert("LA")
 
+        # load the target image
         target_img = params["image"]
+        target_shape = np.array(target_img.shape[:2], dtype=np.uint)
+        paste_shape = np.array(paste_img.size, dtype=np.uint)
 
-        min_scale = min(
-            target_img.shape[0] / paste_img.shape[0],
-            target_img.shape[1] / paste_img.shape[1],
+        # compute the minimum scaling to fit inside target image
+        min_scale = np.min(target_shape / paste_shape)
+
+        # randomize the relative scaling
+        scale = rd.uniform(*self.scale_range)
+
+        # rotate the image and its mask
+        angle = rd.uniform(0, 360)
+        paste_img = paste_img.rotate(angle, expand=True)
+        paste_mask = paste_mask.rotate(angle, expand=True)
+
+        # scale the "paste" image and its mask
+        paste_img = paste_img.resize(
+            tuple((paste_shape * min_scale * scale).astype(np.uint)),
+            resample=Image.Resampling.LANCZOS,
+        )
+        paste_mask = paste_mask.resize(
+            tuple((paste_shape * min_scale * scale).astype(np.uint)),
+            resample=Image.Resampling.LANCZOS,
         )
 
-        rescale_rotate = A.Compose(
-            [
-                A.Rotate(limit=360, always_apply=True, border_mode=cv2.BORDER_CONSTANT),
-                A.RandomScale(scale_limit=(min_scale * self.scale_limit - 1, -0.99), always_apply=True),
-            ],
-        )
-
-        augmentations = rescale_rotate(image=paste_img, mask=paste_mask)
-        paste_img = augmentations["image"]
-        paste_mask = augmentations["mask"]
+        # update paste_shape after scaling
+        paste_shape = np.array(paste_img.size, dtype=np.uint)
 
+        # generate some positions
         positions = []
         for _ in range(rd.randint(1, self.nb)):
-            x = rd.randint(0, target_img.shape[0] - paste_img.shape[0])
-            y = rd.randint(0, target_img.shape[1] - paste_img.shape[1])
+            x = rd.randint(0, target_shape[0] - paste_shape[0])
+            y = rd.randint(0, target_shape[1] - paste_shape[1])
             positions.append((x, y))
 
         params.update(
@@ -123,4 +123,4 @@ class RandomPaste(A.DualTransform):
         return params
 
     def get_transform_init_args_names(self):
-        return "scale_limit", "path_paste_img_dir", "path_paste_mask_dir"
+        return "scale_range", "path_paste_img_dir", "path_paste_mask_dir"