endovit_demo.py

import torch
import torchvision.transforms as T
from PIL import Image
from pathlib import Path
from timm.models.vision_transformer import VisionTransformer
from functools import partial
from torch import nn

# requires: pytorch 2.0.1, timm 0.9.16
def process_single_image(image_path, input_size=224, dataset_mean=[0.3464, 0.2280, 0.2228], dataset_std=[0.2520, 0.2128, 0.2093]):
    # Define the transformations
    transform = T.Compose([
        T.Resize((input_size, input_size)),
        T.ToTensor(),
        T.Normalize(mean=dataset_mean, std=dataset_std)
    ])

    # Open the image
    image = Image.open(image_path).convert('RGB')

    # Apply the transformations
    processed_image = transform(image)

    return processed_image


image_paths = sorted(Path('demo_images').glob('*.png'))
images = torch.stack([process_single_image(image_path) for image_path in image_paths])

device = "cuda"
dtype = torch.float16

model_weights = torch.load('endovit_seg.pth')['model']

model = VisionTransformer(patch_size=16, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6)).to(device, dtype).eval()
loading = model.load_state_dict(model_weights, strict=False)
print(loading)
output = model.forward_features(images.to(device, dtype))
print(output.shape)