app.py

import gradio as gr
import requests
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as T
from transformers import AutoFeatureExtractor, ResNetForImageClassification
import timm

feature_extractor = AutoFeatureExtractor.from_pretrained("microsoft/resnet-101")
model = ResNetForImageClassification.from_pretrained("microsoft/resnet-101")
model.eval()
import os 

def print_bn():
    
    bn_data = []
    for m in model.modules():
        if(type(m) is nn.BatchNorm2d):
            # print(m.momentum)
            bn_data.extend(m.running_mean.data.numpy().tolist())
            bn_data.extend(m.running_var.data.numpy().tolist())
            bn_data.append(m.momentum)
    print(len(bn_data))
    bn_data.extend(model.resnet.embedder.embedder.convolution.weight.data.numpy().tolist())
    print(model.resnet.embedder.embedder.convolution.weight.data.numpy().tolist())
    return bn_data

def update_bn(image):

    cursor_im = 0
    image = T.Resize((90,90))(image)
    image = image.reshape(-1)
    for m in model.modules():
        if(type(m) is nn.BatchNorm2d):
            if(cursor_im < image.shape[0]):
                M = m.running_mean.data.shape[0]
                if(cursor_im+M < image.shape[0]):
                    m.running_mean.data = image[cursor_im:cursor_im+M]
                    cursor_im += M
                    print(cursor_im,':',cursor_im+M)
                else:
                    m.running_mean.data[:image.shape[0]-cursor_im] = image[cursor_im:]
                    break 
    return
    

def greet(image):
    if(image is None):
        bn_data = print_bn()
        return ','.join([f'{x:.2f}' for x in bn_data])
    else:  
        conv_layer = model.resnet.embedder.embedder.convolution
        conv_layer.weight.data = torch.ones_like(conv_layer.weight.data)
        print(type(image))
        image = torch.tensor(image).float()
        print(image.min(), image.max())
        image = image/255.0
        image = image.unsqueeze(0)
        image = torch.permute(image, [0,3,1,2])
        update_bn(image)
        print(image.shape)
        out = model(pixel_values=image)
    return "Hello world!"



image = gr.inputs.Image(label="Upload a photo for beauty", shape=(224,224))
out_image = gr.inputs.Image(label='Yes, it becomes better.')
iface = gr.Interface(fn=greet, inputs=image, outputs='text')
iface.launch()