File size: 4,147 Bytes
337a98d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 | import os
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt
from sklearn.manifold import TSNE
import torch
import torchvision.transforms as transforms
from mpl_toolkits.mplot3d import Axes3D
from tqdm import tqdm
import clip # For CLIP model
def load_models(device):
"""Load both DINOv2 and CLIP models"""
# DINOv2
dinov2 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitb14').to(device).eval()
# CLIP
clip_model, clip_preprocess = clip.load("ViT-B/32", device=device)
return {
'dinov2': dinov2,
'clip': clip_model,
'clip_preprocess': clip_preprocess
}
def load_frames(folder_path, frame_ext='.png'):
frame_files = sorted(
[f for f in os.listdir(folder_path) if f.endswith(frame_ext)],
key=lambda x: int(x.split('.')[0]))
return [Image.open(os.path.join(folder_path, f)).convert('RGB') for f in tqdm(frame_files, desc="Loading frames")]
def extract_embeddings(frames, model_dict, model_type, batch_size=8, device='cuda'):
embeddings = []
with torch.no_grad():
if model_type == 'dinov2':
transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
for i in tqdm(range(0, len(frames), batch_size), desc="Extracting DINOv2 embeddings"):
batch = torch.stack([transform(frames[j]) for j in range(i, min(i+batch_size, len(frames)))]).to(device)
outputs = model_dict['dinov2'](batch)
embeddings.append(outputs.cpu().numpy())
elif model_type == 'clip':
for i in tqdm(range(0, len(frames), batch_size), desc="Extracting CLIP embeddings"):
batch = [model_dict['clip_preprocess'](frames[j]) for j in range(i, min(i+batch_size, len(frames)))]
batch = torch.stack(batch).to(device)
outputs = model_dict['clip'].encode_image(batch)
embeddings.append(outputs.cpu().numpy())
return np.concatenate(embeddings)
def visualize_tsne(embeddings, title_suffix="", perplexity=15):
plt.figure(figsize=(20, 8))
# 2D t-SNE
plt.subplot(1, 2, 1)
tsne_2d = TSNE(n_components=2, perplexity=perplexity, random_state=42)
emb_2d = tsne_2d.fit_transform(embeddings)
plt.scatter(emb_2d[:, 0], emb_2d[:, 1], c=range(len(embeddings)), cmap='viridis', alpha=0.7)
plt.colorbar(label='Frame Number')
plt.title(f'2D t-SNE {title_suffix}')
# 3D t-SNE
ax = plt.subplot(1, 2, 2, projection='3d')
tsne_3d = TSNE(n_components=3, perplexity=perplexity, random_state=42)
emb_3d = tsne_3d.fit_transform(embeddings)
sc = ax.scatter(emb_3d[:, 0], emb_3d[:, 1], emb_3d[:, 2],
c=range(len(embeddings)), cmap='viridis', alpha=0.7)
plt.colorbar(sc, label='Frame Number')
ax.set_title(f'3D t-SNE {title_suffix}')
plt.tight_layout()
plt.show()
def main():
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('folder_path', type=str)
parser.add_argument('--ext', type=str, default='.png')
parser.add_argument('--batch_size', type=int, default=8)
parser.add_argument('--perplexity', type=int, default=15)
args = parser.parse_args()
device = 'cuda' if torch.cuda.is_available() else 'cpu'
print(f"Using device: {device}")
# Load models
print("Loading models...")
models = load_models(device)
frames = load_frames(args.folder_path, args.ext)
# DINOv2 Visualization
print("\nProcessing DINOv2 embeddings...")
dinov2_emb = extract_embeddings(frames, models, 'dinov2', args.batch_size, device)
visualize_tsne(dinov2_emb, "(DINOv2)", args.perplexity)
# CLIP Visualization
print("\nProcessing CLIP embeddings...")
clip_emb = extract_embeddings(frames, models, 'clip', args.batch_size, device)
visualize_tsne(clip_emb, "(CLIP)", args.perplexity)
if __name__ == '__main__':
main()
|