app.py

import torch
import torch.nn as nn
import torch.optim as optim
import torchvision.transforms as transforms
import torchvision.models as models
from torchvision.utils import save_image
import numpy as np
import gradio as gr

IMAGE_SIZE = 244 # VGG image input size - we use VGG 19 as our pretrained CNN

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

cnn = models.vgg19(weights=None)
state_dict = torch.load("vgg19-dcbb9e9d.pth")
cnn.load_state_dict(state_dict)

class VGG(nn.Module):
  def __init__(self):
    super(VGG, self).__init__()
    self.layers = ['0', '5', '10', '19', '28'] # layers we use as representations
    self.model = cnn.features[:29] # we don't care about later layers

  def forward(self, x):
    features = []

    for layer_num, layer in enumerate(self.model):
      x = layer(x)

      # we don't care about the model output - we care about the output of individual layers
      if str(layer_num) in self.layers:
        features.append(x)

    return features
  

gradio_transforms = transforms.Compose([
        transforms.ToTensor(),
        transforms.Resize([IMAGE_SIZE, IMAGE_SIZE])
])


def sanitize_inputs(epochs, lr, cl, sl):
  if epochs < 1:
    return ["Epochs must be positive", None]
  if not isinstance(epochs, int):
    return ["Epochs must be an integer", None]
  if lr < 0:
    return ["Learning rate must be positive", None]
  if lr > 1:
    return ["Learning rate must be less than one", None]
  if cl < 0 or cl > 1:
    return ["Content loss weight must be between 0 and 1", None]
  if sl < 0 or sl > 1:
    return ["Style loss weight must be between 0 and 1", None]

  return None

def train(Epochs, Learning_Rate, Content_Loss, Style_Loss, Content_Image, Style_Image):
  errors = sanitize_inputs(Epochs, Learning_Rate, Content_Loss, Style_Loss)
  if errors is not None:
    return errors

  test = Content_Image

  content = gradio_transforms(Content_Image).unsqueeze(0).to(device)
  style = gradio_transforms(Style_Image).unsqueeze(0).to(device)
  generated = content.clone().requires_grad_(True).to(device)

  model = VGG().to(device).eval()
  optimizer = optim.Adam([generated], lr=Learning_Rate)

  for epoch in range(Epochs):
    generatedFeatures = model(generated)
    contentFeatures = model(content)
    styleFeatures = model(style)

    styleLoss = 0
    contentLoss = 0

    for genFeat, contFeat, styleFeat in zip(generatedFeatures, contentFeatures, styleFeatures):
      batch_size, channel, height, width = genFeat.shape

      contentLoss += torch.mean((genFeat - contFeat) ** 2)

      G = genFeat.view(channel, height * width).mm(genFeat.view(channel, height * width).t())
      A = styleFeat.view(channel, height * width).mm(styleFeat.view(channel, height * width).t())

      styleLoss += torch.mean((G - A) ** 2)

    total_loss = Content_Loss * contentLoss + Style_Loss * styleLoss
    optimizer.zero_grad()
    total_loss.backward()
    optimizer.step()


  save_image(generated, "generated_gradio.png")

  return ["No errors! Enjoy your new image!", "generated_gradio.png"]

demo = gr.Interface(
    fn=train,
    inputs=["number", "number", "number", "number", "image", "image"],
    outputs=[
        gr.Label(label="Error Messages"),
        gr.Image(label="Generated Image"),
    ],
    title="Neural Style Transfer",
    description="Perform neural style transfer on images of your choice!  Provide a content image that contains the content you want to transform and a style image that contains the style you want to emulate.\n\nNote: Huggingface requires users to pay to gain access to GPUs, so this model is hosted on a cpu. Training for many epochs will take a VERY long time.  Using a larger learning rate (e.g., 0.01) can help reduce the number of epochs you need.",
    theme=gr.themes.Soft()
)

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