added kaggle comp solution for facial keypoint

ML/Kaggles/Facial Keypoint Detection Competition/config.py

@@ -0,0 +1,61 @@
+import torch
+import albumentations as A
+from albumentations.pytorch import ToTensorV2
+import cv2
+
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+LEARNING_RATE = 1e-4
+WEIGHT_DECAY = 5e-4
+BATCH_SIZE = 64
+NUM_EPOCHS = 100
+NUM_WORKERS = 4
+CHECKPOINT_FILE = "b0_4.pth.tar"
+PIN_MEMORY = True
+SAVE_MODEL = True
+LOAD_MODEL = True
+
+# Data augmentation for images
+train_transforms = A.Compose(
+    [
+        A.Resize(width=96, height=96),
+        A.Rotate(limit=15, border_mode=cv2.BORDER_CONSTANT, p=0.8),
+        A.IAAAffine(shear=15, scale=1.0, mode="constant", p=0.2),
+        A.RandomBrightnessContrast(contrast_limit=0.5, brightness_limit=0.5, p=0.2),
+        A.OneOf([
+            A.GaussNoise(p=0.8),
+            A.CLAHE(p=0.8),
+            A.ImageCompression(p=0.8),
+            A.RandomGamma(p=0.8),
+            A.Posterize(p=0.8),
+            A.Blur(p=0.8),
+        ], p=1.0),
+        A.OneOf([
+            A.GaussNoise(p=0.8),
+            A.CLAHE(p=0.8),
+            A.ImageCompression(p=0.8),
+            A.RandomGamma(p=0.8),
+            A.Posterize(p=0.8),
+            A.Blur(p=0.8),
+        ], p=1.0),
+        A.ShiftScaleRotate(shift_limit=0.1, scale_limit=0.1, rotate_limit=0, p=0.2, border_mode=cv2.BORDER_CONSTANT),
+        A.Normalize(
+            mean=[0.4897, 0.4897, 0.4897],
+            std=[0.2330, 0.2330, 0.2330],
+            max_pixel_value=255.0,
+        ),
+        ToTensorV2(),
+    ], keypoint_params=A.KeypointParams(format="xy", remove_invisible=False),
+)
+
+
+val_transforms = A.Compose(
+    [
+        A.Resize(height=96, width=96),
+        A.Normalize(
+            mean=[0.4897, 0.4897, 0.4897],
+            std=[0.2330, 0.2330, 0.2330],
+            max_pixel_value=255.0,
+        ),
+        ToTensorV2(),
+    ], keypoint_params=A.KeypointParams(format="xy", remove_invisible=False),
+)

ML/Kaggles/Facial Keypoint Detection Competition/dataset.py

@@ -0,0 +1,50 @@
+import pandas as pd
+import numpy as np
+import config
+import matplotlib.pyplot as plt
+from torch.utils.data import DataLoader, Dataset
+
+
+class FacialKeypointDataset(Dataset):
+    def __init__(self, csv_file, train=True, transform=None):
+        super().__init__()
+        self.data = pd.read_csv(csv_file)
+        self.category_names = ['left_eye_center_x', 'left_eye_center_y', 'right_eye_center_x', 'right_eye_center_y', 'left_eye_inner_corner_x', 'left_eye_inner_corner_y', 'left_eye_outer_corner_x', 'left_eye_outer_corner_y', 'right_eye_inner_corner_x', 'right_eye_inner_corner_y', 'right_eye_outer_corner_x', 'right_eye_outer_corner_y', 'left_eyebrow_inner_end_x', 'left_eyebrow_inner_end_y', 'left_eyebrow_outer_end_x', 'left_eyebrow_outer_end_y', 'right_eyebrow_inner_end_x', 'right_eyebrow_inner_end_y', 'right_eyebrow_outer_end_x', 'right_eyebrow_outer_end_y', 'nose_tip_x', 'nose_tip_y', 'mouth_left_corner_x', 'mouth_left_corner_y', 'mouth_right_corner_x', 'mouth_right_corner_y', 'mouth_center_top_lip_x', 'mouth_center_top_lip_y', 'mouth_center_bottom_lip_x', 'mouth_center_bottom_lip_y']
+        self.transform = transform
+        self.train = train
+
+    def __len__(self):
+        return self.data.shape[0]
+
+    def __getitem__(self, index):
+        if self.train:
+            image = np.array(self.data.iloc[index, 30].split()).astype(np.float32)
+            labels = np.array(self.data.iloc[index, :30].tolist())
+            labels[np.isnan(labels)] = -1
+        else:
+            image = np.array(self.data.iloc[index, 1].split()).astype(np.float32)
+            labels = np.zeros(30)
+
+        ignore_indices = labels == -1
+        labels = labels.reshape(15, 2)
+
+        if self.transform:
+            image = np.repeat(image.reshape(96, 96, 1), 3, 2).astype(np.uint8)
+            augmentations = self.transform(image=image, keypoints=labels)
+            image = augmentations["image"]
+            labels = augmentations["keypoints"]
+
+        labels = np.array(labels).reshape(-1)
+        labels[ignore_indices] = -1
+
+        return image, labels.astype(np.float32)
+
+
+if __name__ == "__main__":
+    ds = FacialKeypointDataset(csv_file="data/train_4.csv", train=True, transform=config.train_transforms)
+    loader = DataLoader(ds, batch_size=1, shuffle=True, num_workers=0)
+
+    for idx, (x, y) in enumerate(loader):
+        plt.imshow(x[0][0].detach().cpu().numpy(), cmap='gray')
+        plt.plot(y[0][0::2].detach().cpu().numpy(), y[0][1::2].detach().cpu().numpy(), "go")
+        plt.show()

ML/Kaggles/Facial Keypoint Detection Competition/extract_images_from_csv.py

@@ -0,0 +1,19 @@
+import numpy as np
+import pandas as pd
+import os
+from PIL import Image
+
+
+def extract_images_from_csv(csv, column, save_folder, resize=(96, 96)):
+    if not os.path.exists(save_folder):
+        os.makedirs(save_folder)
+
+    for idx, image in enumerate(csv[column]):
+        image = np.array(image.split()).astype(np.uint8)
+        image = image.reshape(resize[0], resize[1])
+        img = Image.fromarray(image, 'L')
+        img.save(save_folder+f"img_{idx}.png")
+
+
+csv = pd.read_csv("test.csv")
+extract_images_from_csv(csv, "Image", "data/test/")

ML/Kaggles/Facial Keypoint Detection Competition/test_landmarks.py

@@ -0,0 +1,12 @@
+import face_alignment
+from skimage import io
+import matplotlib.pyplot as plt
+
+fa = face_alignment.FaceAlignment(face_alignment.LandmarksType._2D, flip_input=False)
+input = io.imread('mee.jpg')
+preds = fa.get_landmarks(input)
+print(preds[0].shape)
+plt.imshow(input)
+plt.plot(preds[0][:, 0], preds[0][:, 1], "ro")
+plt.show()
+#plt.savefig('myimage.png')

ML/Kaggles/Facial Keypoint Detection Competition/train.py

@@ -0,0 +1,111 @@
+import torch
+from dataset import FacialKeypointDataset
+from torch import nn, optim
+import os
+import config
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+from efficientnet_pytorch import EfficientNet
+from utils import (
+    load_checkpoint,
+    save_checkpoint,
+    get_rmse,
+    get_submission
+)
+
+
+def train_one_epoch(loader, model, optimizer, loss_fn, scaler, device):
+    losses = []
+    loop = tqdm(loader)
+    num_examples = 0
+    for batch_idx, (data, targets) in enumerate(loop):
+        data = data.to(device=device)
+        targets = targets.to(device=device)
+
+        # forward
+        scores = model(data)
+        scores[targets == -1] = -1
+        loss = loss_fn(scores, targets)
+        num_examples += torch.numel(scores[targets != -1])
+        losses.append(loss.item())
+
+        # backward
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+
+    print(f"Loss average over epoch: {(sum(losses)/num_examples)**0.5}")
+
+
+def main():
+    train_ds = FacialKeypointDataset(
+        csv_file="data/train_4.csv",
+        transform=config.train_transforms,
+    )
+    train_loader = DataLoader(
+        train_ds,
+        batch_size=config.BATCH_SIZE,
+        num_workers=config.NUM_WORKERS,
+        pin_memory=config.PIN_MEMORY,
+        shuffle=True,
+    )
+    val_ds = FacialKeypointDataset(
+        transform=config.val_transforms,
+        csv_file="data/val_4.csv",
+    )
+    val_loader = DataLoader(
+        val_ds,
+        batch_size=config.BATCH_SIZE,
+        num_workers=config.NUM_WORKERS,
+        pin_memory=config.PIN_MEMORY,
+        shuffle=False,
+    )
+
+    test_ds = FacialKeypointDataset(
+        csv_file="data/test.csv",
+        transform=config.val_transforms,
+        train=False,
+    )
+
+    test_loader = DataLoader(
+        test_ds,
+        batch_size=1,
+        num_workers=config.NUM_WORKERS,
+        pin_memory=config.PIN_MEMORY,
+        shuffle=False,
+    )
+    loss_fn = nn.MSELoss(reduction="sum")
+    model = EfficientNet.from_pretrained("efficientnet-b0")
+    model._fc = nn.Linear(1280, 30)
+    model = model.to(config.DEVICE)
+    optimizer = optim.Adam(model.parameters(), lr=config.LEARNING_RATE, weight_decay=config.WEIGHT_DECAY)
+    scaler = torch.cuda.amp.GradScaler()
+
+    model_4 = EfficientNet.from_pretrained("efficientnet-b0")
+    model_4._fc = nn.Linear(1280, 30)
+    model_15 = EfficientNet.from_pretrained("efficientnet-b0")
+    model_15._fc = nn.Linear(1280, 30)
+    model_4 = model_4.to(config.DEVICE)
+    model_15 = model_15.to(config.DEVICE)
+
+    if config.LOAD_MODEL and config.CHECKPOINT_FILE in os.listdir():
+        load_checkpoint(torch.load(config.CHECKPOINT_FILE), model, optimizer, config.LEARNING_RATE)
+        load_checkpoint(torch.load("b0_4.pth.tar"), model_4, optimizer, config.LEARNING_RATE)
+        load_checkpoint(torch.load("b0_15.pth.tar"), model_15, optimizer, config.LEARNING_RATE)
+
+    get_submission(test_loader, test_ds, model_15, model_4)
+
+    for epoch in range(config.NUM_EPOCHS):
+        get_rmse(val_loader, model, loss_fn, config.DEVICE)
+        train_one_epoch(train_loader, model, optimizer, loss_fn, scaler, config.DEVICE)
+
+        # get on validation
+        if config.SAVE_MODEL:
+            checkpoint = {
+                "state_dict": model.state_dict(),
+                "optimizer": optimizer.state_dict(),
+            }
+            save_checkpoint(checkpoint, filename=config.CHECKPOINT_FILE)
+
+if __name__ == "__main__":
+    main()

ML/Kaggles/Facial Keypoint Detection Competition/utils.py

@@ -0,0 +1,71 @@
+import torch
+import numpy as np
+import config
+import pandas as pd
+from tqdm import tqdm
+
+
+def get_submission(loader, dataset, model_15, model_4):
+    """
+    This can be done a lot faster.. but it didn't take
+    too much time to do it in this inefficient way
+    """
+    model_15.eval()
+    model_4.eval()
+    id_lookup = pd.read_csv("data/IdLookupTable.csv")
+    predictions = []
+    image_id = 1
+
+    for image, label in tqdm(loader):
+        image = image.to(config.DEVICE)
+        preds_15 = torch.clip(model_15(image).squeeze(0), 0.0, 96.0)
+        preds_4 = torch.clip(model_4(image).squeeze(0), 0.0, 96.0)
+        feature_names = id_lookup.loc[id_lookup["ImageId"] == image_id]["FeatureName"]
+
+        for feature_name in feature_names:
+            feature_index = dataset.category_names.index(feature_name)
+            if feature_names.shape[0] < 10:
+                predictions.append(preds_4[feature_index].item())
+            else:
+                predictions.append(preds_15[feature_index].item())
+
+        image_id += 1
+
+    df = pd.DataFrame({"RowId": np.arange(1, len(predictions)+1), "Location": predictions})
+    df.to_csv("submission.csv", index=False)
+    model_15.train()
+    model_4.train()
+
+
+def get_rmse(loader, model, loss_fn, device):
+    model.eval()
+    num_examples = 0
+    losses = []
+    for batch_idx, (data, targets) in enumerate(loader):
+        data = data.to(device=device)
+        targets = targets.to(device=device)
+
+        # forward
+        scores = model(data)
+        loss = loss_fn(scores[targets != -1], targets[targets != -1])
+        num_examples += scores[targets != -1].shape[0]
+        losses.append(loss.item())
+
+    model.train()
+    print(f"Loss on val: {(sum(losses)/num_examples)**0.5}")
+
+
+def save_checkpoint(state, filename="my_checkpoint.pth.tar"):
+    print("=> Saving checkpoint")
+    torch.save(state, filename)
+
+
+def load_checkpoint(checkpoint, model, optimizer, lr):
+    print("=> Loading checkpoint")
+    model.load_state_dict(checkpoint["state_dict"])
+    optimizer.load_state_dict(checkpoint["optimizer"])
+
+    # If we don't do this then it will just have learning rate of old checkpoint
+    # and it will lead to many hours of debugging \:
+    for param_group in optimizer.param_groups:
+        param_group["lr"] = lr