Create preprocess.py

preprocess.py

@@ -0,0 +1,72 @@
+import torch
+from torchvision import transforms
+from dcae.dcae import DCAE
+from datasets import load_dataset, Features, Array3D, Value, Image
+
+# Constants
+COMPRESSION_FACTOR = "f128"  # Options: "f32", "f64", "f128"
+OG_DATASET = "pravsels/FFHQ_1024"
+UPLOAD_DATASET = f"SwayStar123/FFHQ_1024_DC-AE_{COMPRESSION_FACTOR}"
+MODEL_PATHS = {
+    "f32": "dc-ae-f32c32-mix-1.0",
+    "f64": "dc-ae-f64c128-mix-1.0",
+    "f128": "dc-ae-f128c512-mix-1.0"
+}
+CACHE_DIR = "models/dc_ae"
+DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+DTYPE = torch.bfloat16
+BATCH_SIZE = 30
+
+# Define transforms
+transform = transforms.Compose([
+    transforms.ToTensor(),
+    transforms.Normalize((0.5, 0.5, 0.5,), (0.5, 0.5, 0.5,))
+])
+
+def denorm(x):
+    return (x * 0.5 + 0.5).clamp(0, 1)
+
+def main():
+    model_name = MODEL_PATHS[COMPRESSION_FACTOR]
+    dc_ae = DCAE(model_name, device=DEVICE, dtype=DTYPE, cache_dir=CACHE_DIR).eval()
+
+    # Get the shape of the latent representations
+    dummy_input = torch.randn(1, 3, 1024, 1024).to(DTYPE).to(DEVICE)
+    with torch.no_grad():
+        dummy_latent = dc_ae.encode(dummy_input).cpu()
+    latent_shape = dummy_latent.shape[1:]
+    print(f"Latent shape: {latent_shape}")
+    
+    features = Features({
+        'label': Value('int64'),
+        'latent': Array3D(dtype='float16', shape=latent_shape)
+    })
+    
+    dataset = load_dataset(OG_DATASET, split="train")
+
+    def process_batch(batch):
+        images = [img.convert("RGB") for img in batch["image"]]
+        img_tensors = torch.stack([transform(img) for img in images]).to(DTYPE).to(DEVICE)
+        with torch.no_grad():
+            latents = dc_ae.encode(img_tensors).cpu().to(torch.float16).numpy()
+        batch["latent"] = latents
+        return batch
+
+    processed_dataset = dataset.map(
+        process_batch,
+        batched=True,
+        batch_size=BATCH_SIZE,
+    )
+    
+    # Drop the image column
+    processed_dataset = processed_dataset.remove_columns(["image"])
+
+    # Push the dataset to the hub
+    processed_dataset.push_to_hub(
+        repo_id=UPLOAD_DATASET
+    )
+    
+    print(f"Dataset uploaded to Hugging Face Hub: {UPLOAD_DATASET}")
+
+if __name__ == "__main__":
+    main()