app.py

import os 
os.system("git clone https://github.com/mchong6/SOAT.git")
import sys
sys.path.append("SOAT")
import os
import torch
import torchvision
from torch import nn
import numpy as np
import torch.backends.cudnn as cudnn
cudnn.benchmark = True

import math
import matplotlib.pyplot as plt
import torch.nn.functional as F
from model import *
from tqdm import tqdm as tqdm
import pickle
from copy import deepcopy
import warnings
warnings.filterwarnings("ignore", category=UserWarning) # get rid of interpolation warning
import kornia.filters as k
from torchvision.utils import save_image
from util import *
import scipy

import gradio as gr

import PIL

from torchvision import transforms

device = 'cpu' #@param ['cuda', 'cpu']

generator = Generator(256, 512, 8, channel_multiplier=2).eval().to(device)
truncation = 0.7

def display_image(image, size=None, mode='nearest', unnorm=False, title=''):
    # image is [3,h,w] or [1,3,h,w] tensor [0,1]
    if image.is_cuda:
        image = image.cpu()
    if size is not None and image.size(-1) != size:
        image = F.interpolate(image, size=(size,size), mode=mode)
    if image.dim() == 4:
        image = image[0]
    image = ((image.clamp(-1,1)+1)/2).permute(1, 2, 0).detach().numpy()
    return image
    

#mean_latentland = load_model(generator, 'landscape.pt')
#mean_latentface = load_model(generator, 'face.pt')
#mean_latentchurch = load_model(generator, 'church.pt')


def inferece(num, seed):
    mean_latent = load_model(generator, 'landscape.pt')

    num_im =  int(num)
    random_seed =  int(seed)
    
    plt.rcParams['figure.dpi'] = 300
    
    
    # pad determines how much of an image is involve in the blending
    pad = 512//4
    
    all_im = []
    
    random_state = np.random.RandomState(random_seed)
    
    # latent smoothing
    with torch.no_grad():
        z = random_state.randn(num_im, 512).astype(np.float32)
        z = scipy.ndimage.gaussian_filter(z, [.7, 0], mode='wrap')
        z /= np.sqrt(np.mean(np.square(z)))
        z = torch.from_numpy(z).to(device)
    
        source = generator.get_latent(z, truncation=truncation, mean_latent=mean_latent)
            
        # merge images 2 at a time
        for i in range(num_im-1):
            source1 = index_layers(source, i)
            source2 = index_layers(source, i+1)
            all_im.append(generator.merge_extension(source1, source2))
    
    # display intermediate generations    
    # for i in all_im:
    #     display_image(i)
        
    
    b,c,h,w = all_im[0].shape
    panorama_im = torch.zeros(b,c,h,512+(num_im-2)*256)
    
    # We created a series of 2-blended images which we can overlay to form a large panorama
    # add first image
    coord = 256+pad
    panorama_im[..., :coord] = all_im[0][..., :coord]
    
    for im in all_im[1:]:
        panorama_im[..., coord:coord+512-2*pad] = im[..., pad:-pad]
        coord += 512-2*pad
    panorama_im[..., coord:] = all_im[-1][..., 512-pad:]
    
    img = display_image(panorama_im)
    return img

title = "SOAT"
description = "Gradio demo for SOAT Panorama Generaton for landscapes. Generate a panorama using a pretrained stylegan by stitching intermediate activations. To use it, simply add the number of images and random seed number . Read more at the links below."
article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2111.01619' target='_blank'>StyleGAN of All Trades: Image Manipulation with Only Pretrained StyleGAN</a> | <a href='https://github.com/mchong6/SOAT' target='_blank'>Github Repo</a></p>"

gr.Interface(
    inferece, 
    [gr.inputs.Number(default=5, label="Number of Images")
,gr.inputs.Number(default=90, label="Random Seed")
],
    gr.outputs.Image(type="numpy", label="Output"),
    title=title,
    description=description,
    article=article, theme="huggingface",enable_queue=True).launch(debug=True)