app.py

import torch
import torch.nn as nn
from torchvision.transforms._transforms_video import (CenterCropVideo,NormalizeVideo)
from torchvision.transforms import (Compose,Lambda,RandomCrop,RandomHorizontalFlip,Resize)
from pytorchvideo.transforms import (ApplyTransformToKey,Normalize,RandomShortSideScale,UniformTemporalSubsample,Permute)
import numpy as np
import gradio as gr
import spaces
# torch.save(model.state_dict(), 'model.pth')
video_transform = Compose([
    ApplyTransformToKey(key = 'video',
    transform = Compose([
        UniformTemporalSubsample(20),
        Lambda(lambda x:x/255),
        Normalize((0.45,0.45,0.45),(0.225,0.225,0.225)),
        RandomShortSideScale(min_size = 248, max_size = 256),
        CenterCropVideo(224),
        RandomHorizontalFlip(p=0.5),
    ]),  
                       ),  
])
#============================================================
class mymodel_test(nn.Module):
    def __init__(self):
        super(mymodel_test,self).__init__()
        self.video_model = torch.hub.load('facebookresearch/pytorchvideo','efficient_x3d_xs', pretrained=False)
        self.relu = nn.ReLU()
        
        self.Linear = nn.Linear(400,1)        
        
    def forward(self,x):
        
        x = self.relu(self.video_model(x))
        x = self.Linear(x)
        return x
    
    
#============================================================
    
model_test = mymodel_test()
model_test.load_state_dict(torch.load('model.pth'))
from pytorchvideo.data.encoded_video import EncodedVideo
@spaces.GPU
def interface_video(video_path):
    video = EncodedVideo.from_path(video_path)
    
    video_data = video.get_clip(0,2)
    
    video_data = video_transform(video_data)
    
    video_data['video'].shape
    model = model_test
    
    inputs = video_data['video']
    
    inputs = torch.unsqueeze(inputs, 0 )
    inputs.shape
    preds = model(inputs)
    
    preds = preds.detach().cpu().numpy()
    preds = np.where(preds>0.5,1,0)
    if(preds==0):
        return 'non violence'
    else:
        return 'violence'
    # return preds
demo = gr.Blocks()
with demo:
    with gr.Row():
        with gr.Column():
            with gr.Row():
                input_video = gr.Video(label='Input Video', height=360)
                # input_video = load_video(input_video)
            with gr.Row():
                submit_video_button = gr.Button('Submit')
        with gr.Column():
                label_video = gr.Label()
    submit_video_button.click(fn=iinterface_video, inputs=input_video, outputs=label_video)
demo.launch()