app.py

import gradio as gr
import pandas as pd
from torchvision.io import read_video
import torch.nn.functional as F
import torch, hiera
df=pd.read_csv('Kinetic400.csv')
model = hiera.hiera_base_16x224(pretrained=True, checkpoint="mae_k400_ft_k400")
def recognize(vid):
  frames, audio, info = read_video(vid, pts_unit='sec', output_format='THWC')
  frames = frames.float() / 255  # Convert from byte to float
  frames = torch.stack([frames[:64], frames[64:128]], dim=0)
  frames = frames[:, ::4]  # Sample every 4 frames
  frames = frames.permute(0, 4, 1, 2, 3).contiguous()
  frames = F.interpolate(frames, size=(16, 224, 224), mode="trilinear")
  torch.Size([2, 3, 16, 224, 224])
  frames = frames - torch.tensor([0.45, 0.45, 0.45]).view(1, -1, 1, 1, 1)
  frames = frames / torch.tensor([0.225, 0.225, 0.255]).view(1, -1, 1, 1, 1)
  out = model(frames)
  out = out.mean(0)
  out1=out.argmax(dim=-1).item()
  out2=df.iloc[out1,1]
  return out2
demo = gr.Interface(fn=recognize, inputs=gr.Video(type="file"),outputs='text',examples= [['dog.mp4']])
demo.launch()