app.py

import os

os.environ['MKL_SERVICE_FORCE_INTEL'] = '1'

from loss_functions import ContentLoss, StyleLoss
import torchvision.models as models

from torch import optim
from pathlib import Path
import gradio as gr
import cv2
from PIL import Image
import cv2
import random, os
import numpy as np
import torch
import torch.nn as nn
import torchvision.transforms as transforms


seed = 2023
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = True
    
# model = create_vgg_model()
device = 'cuda' if torch.cuda.is_available() else 'cpu'
cnn = models.vgg16(pretrained=True).features.to(device).eval()


duration = 5

content_layers = ['conv_4']
style_layers = ['conv_1', 'conv_2', 'conv_3', 'conv_4', 'conv_5']



def predict(content_img, style_img, style, content, lr, epoch):
    # progress(0, desc="Starting...")
    i = 0
    content_losses = []
    style_losses = []
    model = nn.Sequential().to(device)

    imsize = tuple(content_img.shape[:-1])

    loader = transforms.Compose([
    transforms.ToTensor()])

    style_img = cv2.resize(style_img, imsize)

    content_img = loader(content_img).to(device).unsqueeze(0)
    style_img = loader(style_img).to(device).unsqueeze(0)

    print(content_img.shape, style_img.shape)

    for layer in cnn.children():
        if isinstance(layer, nn.Conv2d):
            i += 1
            name = 'conv_{}'.format(i)

        elif isinstance(layer, nn.ReLU):
            name = 'relu_{}'.format(i)
            layer = nn.ReLU(inplace=False)

        elif isinstance(layer, nn.MaxPool2d):
            name = 'pool_{}'.format(i)

        elif isinstance(layer, nn.BatchNorm2d):
            name = 'bn_{}'.format(i)

        else:
            raise RuntimeError('Unrecognized layer: {}'.format(layer.__class__.__name__))

        model.add_module(name, layer)

        if name in content_layers:
            target = model(content_img).detach()
            content_loss = ContentLoss(target)
            model.add_module("content_loss_{}".format(i), content_loss)
            content_losses.append(content_loss)

        if name in style_layers:
            target_feature = model(style_img).detach()
            style_loss = StyleLoss(target_feature)
            model.add_module("style_loss_{}".format(i), style_loss)
            style_losses.append(style_loss)

    for i in range(len(model) - 1, -1, -1):
        if isinstance(model[i], ContentLoss) or isinstance(model[i], StyleLoss):
            break

    model = model[:(i + 1)]
    # model = torch.compile(model)
    input_img = torch.randn(content_img.data.size(), device=device) 
    input_img.requires_grad_(True)
    model.requires_grad_(False)
    optimizer = optim.Adam([input_img], lr) #We are using input_img instead of model.parameters bcos input_img is modified


    i_=0
    img_history = []
    for i_ in range(epoch):
        with torch.no_grad():
            input_img.clamp_(0, 1)

        optimizer.zero_grad()
        model(input_img)

        temp_style_loss = 0
        temp_content_loss = 0

        for i in style_losses:
            temp_style_loss = temp_style_loss + i.loss

        for i in content_losses:
            temp_content_loss = temp_content_loss + i.loss

        loss = temp_style_loss*style + temp_content_loss*content
        loss.backward()
        optimizer.step()

        img_history.append(np.uint8(torch.permute(input_img[0].clone().cpu().detach(), (1, 2, 0)).numpy()*255.0))

        i_+=1

    with torch.no_grad():
        input_img.clamp_(0, 1)

    print(input_img.shape)


    fps = len(img_history) //duration
    fourcc = cv2.VideoWriter_fourcc(*'mp4v')
    pic = img_history[-1]
    out = cv2.VideoWriter('output.mp4', fourcc, int(fps), (pic.shape[1], pic.shape[0]))

    for img in img_history:
        out.write(img[::,::,::-1])
    out.release()

    return Image.fromarray(np.uint8(torch.permute(input_img[0], (1, 2, 0)).cpu().detach().numpy()*255.0)), 'output.mp4'


example_list =  [['content/content2.jpg',
                  'style/style2.jpg', 100000, 0.6, 0.3, 400],

                 ['content/content2.jpg',
                  'style/curvy.jpeg', 100000, 1, 0.3, 400],


                ['content/content2.jpg',
                  'style/water_color.jpg', 30000, 1, 0.1, 300],


                ['content/content2.jpg',
                  'style/rgb.png', 50000, 1, 0.1, 400],

                ['style/water_color.jpg',
                  'style/rgb.png', 70000, 1, 0.1, 400],

]

title = "Neural Style Transfer 🎨"
description = "You can run the code on [Kaggle](https://www.kaggle.com/frozenwolf/neural-style-transfer). See the code on [GitHub](https://github.com/FrozenWolf-Cyber/Neural-Style-Transfer) for Neural Style Transfer comparison between VGG16 and AlexNet"
article = ""


content_input = gr.inputs.Image(label="Upload an image to which you want the style to be applied.",shape= (256,256))
style_input = gr.inputs.Image( label="Upload Style Image ",shape= (256,256), )


style_slider = gr.inputs.Slider(1,100000,label="Adjust Style Density" ,default=100000,)
content_slider = gr.inputs.Slider(1/100000,1,label="Content Sharpness" ,default=1,)
lr_slider = gr.inputs.Slider(0.001,1,label="Learning Rate" ,default=0.1,)
epoch_slider = gr.inputs.Slider(50,500,label="Epoch Slider" ,default=100,)

demo = gr.Interface(fn=predict,
                    inputs=[content_input,
                                style_input,
                                style_slider ,
                                content_slider,
                                lr_slider,
                                epoch_slider
                                # style_checkbox
                                ],
                    outputs=[gr.Image(shape= (256,256),),
                             gr.Video(shape= (256,256),)],
                    examples=example_list,
                    title=title,
                    description=description,
                    article=article)


demo.launch(debug=False,
            share=False)