app.py

import gradio as gr
import torch
import math
import cv2
import os
import sys
import FFV1MT_MS
import flow_tools

model = FFV1MT_MS.FFV1DNN()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

print('Number fo parameters: {}'.format(model.num_parameters()))
model.to(device)
model_dict = torch.load('Model_example.pth.tar')['state_dict']
# save model
model.load_state_dict(model_dict, strict=True)
model.eval()


def process_images(videos, x, y):
    # read video file
    cap = cv2.VideoCapture(videos)
    # transform images to a list of images ndarray
    images = []
    while True:
        ret, frame = cap.read()
        if ret:
            images.append(frame)
        else:
            break
    if len(images) < 11:
        print('video is too short')
        return
    # only use the first 11 frames
    images = images[:11]
    # transform images to a list of images tensor
    images = [torch.from_numpy(img).permute(2, 0, 1).float().to(device).unsqueeze(0) / 255.0 for img in images]
    # if the max size of the image is larger than 1024, resize the image to 1024 with same ratio
    max_size = max(images[0].shape[2], images[0].shape[3])
    if max_size > 832:
        ratio = 832 / max_size
        images = [torch.nn.functional.interpolate(img, scale_factor=ratio, mode='bicubic', align_corners=True) for img
                  in images]
    # transform color image to gray image

    result = model.forward_viz(images, layer=7, x=x, y=y)
    flow = result['flow']
    attention = result['attention']
    activation = result['activation']

    return [flow, activation, attention]


title = "Modelling Human Visual Motion Processing with Trainable Motion Energy Sensing and a Self-attention Network "
description = "## Introduction(^_^)\n" \
              " The intersection of cognitive neuroscience and computer vision offers exciting advancements in " \
              "how machines perceive motion. Our research bridges the gap between these fields by proposing a novel " \
              "image-computable model that aligns with human motion perception mechanisms. By integrating trainable" \
              " motion energy sensing with recurrent self-attention networks, we can simulate the complex motion " \
              "processing of the human visual cortex, particularly the V1-MT pathway. Our model not only parallels" \
              " physiological responses in V1 and MT neurons but also replicates human psychophysical responses " \
              "to dynamic stimuli. \n\n\n" \
              "![](https://drive.google.com/uc?id=10PcKzQ9X1nsXKUi8OPR0jN_ZsjlCAV47) \n" \
              "## Environment Configuration \n" \
              "To run our model, the basic environment configuration is required:\n" \
              '- Python 3.8 or higher \n' \
              '- Pyotrch 2.0 \n' \
              '- CUDA Toolkit 11.x (for GPU acceleration)\n' \
              '- opencv-python \n' \
              '- Imageio \n' \
              '- Matplotlib \n\n' \
              "## Preprint Paper \n" \
              "The paper is available at [arXiv](https://arxiv.org/abs/2305.09156) \n" \
              "## Video Presentation \n" \
              "The video presentation is available at [Video Record](https://recorder-v3.slideslive.com/?share=85662&s=6afe157c-e764-4e3c-9302-2c6dd6887db1/). \n" \
              "## Conference Website \n" \
              "The project is presented at [NeurIPS 2023](https://neurips.cc/virtual/2023/poster/70202). \n" \
              "## Below is the interactive demo of our model.  You can select the videos examples below or upload your own videos.  The model outputs the motion flow field, the activation of the first stage, and the attention map of the second stage." \
              "We also provide two sliders to adjust the location of the attention visualizer. \n" \
              " **Note**: The demo is running on CPU, so it may take a while to process the video.  \n"

examples = [["example_1.mp4", 62, 56], ["example_2.mp4", 59, 55], ["example_3.mp4", 50, 50], ["example_4.mp4", 50, 50],
            ["example_5.mp4", 39, 72]]

md = "![](https://drive.google.com/uc?id=1WBqYsKRwn_78A72MJBrk643l3-gfAssP) \n"  \
     "## Author \n" \
     "This project page is developed by Zitang Sun (zitangsun96 @ gmail.com)\n" \
     "## LICENSE \n" \
     "This project is licensed under the terms of the MIT license. \n"
iface = gr.Interface(fn=process_images,
                     inputs=[gr.Video(label="Upload video or use the example images below"),
                             gr.Slider(0, 100, label='X location of attention visualizer'),
                             gr.Slider(0, 100, label='Y location of attention visualizer')],
                     # out put is three images
                     outputs=[gr.Image(type="numpy", label="Motion flow field"),
                              gr.Image(type="numpy", label="Activation of Stage I"),
                              gr.Image(type="numpy", label="Attention map of Stage II")],
                     title=title,
                     description=description,
                     article=md,
                     examples=examples)

iface.launch(debug=True, share=True)