event-clip

Example to use pretrained checkpoints.

from huggingface_hub import hf_hub_download
import torch
import clip
import torch.nn.functional as F
import numpy as np
import cv2
import torchvision.transforms as transforms


def generate_event_image(frames, threshold=10):
    frames = np.array(frames)  
    num_frames, height, width, _ = frames.shape
    event_images = []
    
    for i in range(1, num_frames):
        diff = cv2.absdiff(frames[i], frames[i-1])
        gray_diff = cv2.cvtColor(diff, cv2.COLOR_RGB2GRAY)
        _, event_image = cv2.threshold(gray_diff, threshold, 255, cv2.THRESH_BINARY)
        event_images.append(event_image)

    return torch.tensor(event_images).sum(dim=0)


ckpt_path = hf_hub_download(
    repo_id="Eavn/event-clip",      
    filename="vitb.pt", # or vitl.pt for pretraining checkpoints
    repo_type="model"               
)

model, preprocess = clip.load("ViT-B/32")
# model, preprocess = clip.load("ViT-L/14")

state_dict = torch.load(ckpt_path)["checkpoint"]
new_state_dict = {}
for key in state_dict.keys():
    if 'encoder_k' in key:
        new_state_dict[key.replace('encoder_k.', '')] = state_dict[key]
model.load_state_dict(new_state_dict)

transform = transforms.Compose([
    transforms.Resize((224, 224)),
])

stack_size = 16
threshold = 10
clamp = 10
text = 'Put the Text Here'
text = clip.tokenize([text]).cuda()

images = (np.random.rand(32, 224, 224, 3) * 255).astype(np.uint8)
event = generate_event_image(
    images[:stack_size], 
    threshold=threshold
)
if clamp > 0:
    event = torch.clamp(event, min=0, max=clamp)
event = event / event.max()
event = torch.stack([event, event, event])
event = transform(event)
event = event.cuda().unsqueeze(0)

logits_per_event, _ = model(event, text)

arxiv.org/abs/2505.02393 arxiv.org/abs/2412.03093