app.py

import numpy as np

import gradio as gr
import os
#os.system("sudo apt-get install nvIDia-cuda-toolkit")
os.system("pip3 install torch")
os.system("/usr/local/bin/python -m pip install --upgrade pip")
os.system("pip3 install collections")
os.system("pip3 install torchvision")
os.system("pip3 install einops")
os.system("pip3 install git+https://github.com/lucasb-eyer/pydensecrf.git")
#os.system("pip install argparse")
import pydensecrf.densecrf as dcrf
from PIL import Image
import torch
import torch.nn.functional as F
from torchvision import transforms
from model_video import build_model
import numpy as np
import collections

def crf_refine(img, annos):
    def _sigmoid(x):
        return 1 / (1 + np.exp(-x))

    assert img.dtype == np.uint8
    assert annos.dtype == np.uint8
    assert img.shape[:2] == annos.shape

    # img and annos should be np array with data type uint8

    EPSILON = 1e-8

    M = 2  # salient or not
    tau = 1.05
    # Setup the CRF model
    d = dcrf.DenseCRF2D(img.shape[1], img.shape[0], M)

    anno_norm = annos / 255.

    n_energy = -np.log((1.0 - anno_norm + EPSILON)) / (tau * _sigmoid(1 - anno_norm))
    p_energy = -np.log(anno_norm + EPSILON) / (tau * _sigmoid(anno_norm))

    U = np.zeros((M, img.shape[0] * img.shape[1]), dtype='float32')
    U[0, :] = n_energy.flatten()
    U[1, :] = p_energy.flatten()

    d.setUnaryEnergy(U)

    d.addPairwiseGaussian(sxy=3, compat=3)
    d.addPairwiseBilateral(sxy=60, srgb=5, rgbim=img, compat=5)

    # Do the inference
    infer = np.array(d.inference(1)).astype('float32')
    res = infer[1, :]

    res = res * 255
    res = res.reshape(img.shape[:2])
    return res.astype('uint8')

#import argparse
device='cpu'
net = build_model(device).to(device)
#net=torch.nn.DataParallel(net)
model_path = 'image_best.pth'
print(model_path)
weight=torch.load(model_path,map_location=torch.device(device))
#print(type(weight))
new_dict=collections.OrderedDict()
for k in weight.keys():
  new_dict[k[len('module.'):]]=weight[k]
net.load_state_dict(new_dict)
net.eval()
net = net.to(device)
def test(gpu_id, net, img_list, group_size, img_size):
    print('test')
    #device=device
    hl,wl=[_.shape[0] for _ in img_list],[_.shape[1] for _ in img_list]
    img_transform = transforms.Compose([transforms.Resize((img_size, img_size)), transforms.ToTensor(),
                    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
    img_transform_gray = transforms.Compose([transforms.Resize((img_size, img_size)), transforms.ToTensor(),
                                             transforms.Normalize(mean=[0.449], std=[0.226])])
    with torch.no_grad():
        
        group_img=torch.rand(5,3,224,224)
        for i in range(5):
          group_img[i]=img_transform(Image.fromarray(img_list[i]))
        _,pred_mask=net(group_img*1)
        pred_mask=(pred_mask.detach().squeeze()*255)#.numpy().astype(np.uint8)
        pred_mask=[F.interpolate(pred_mask[i].reshape(1,1,pred_mask[i].shape[-2],pred_mask[i].shape[-1]),size=(wl[i],hl[i]),mode='bilinear').numpy().astype(np.uint8) for i in range(5)]
        #pred_mask=[crf_refine(((group_img[i]-group_img[i].min())/(group_img[i].max()-group_img[i].min())*255).permute(1,2,0).contiguous().numpy().astype(np.uint8),pred_mask[i]) for i in range(5)]
        pred_mask=[crf_refine(img_list[i],pred_mask[i]) for i in range(5)]
        print(pred_mask[0].shape)
        result = [Image.fromarray((torch.from_numpy(pred_mask[i]).unsqueeze(2).repeat(1,1,3)).numpy()) for i in range(5)]
        #w, h = 224,224#Image.open(image_list[i][j]).size
        #result = result.resize((w, h), Image.BILINEAR)
        #result.convert('L').save('0.png')
        print('done')
        return result
test('cpu',net,[(torch.rand(224,224,3)*255).numpy().astype(np.uint8) for i in range(5)],5,224)
def sepia(img1,img2,img3,img4,img5):
  print('sepia')
  '''ans=[]
  print(len(input_imgs))
  for input_img in input_imgs:
    sepia_filter = np.array(
        [[0.393, 0.769, 0.189], [0.349, 0.686, 0.168], [0.272, 0.534, 0.131]]
    )
    sepia_img = input_img.dot(sepia_filter.T)
    sepia_img /= sepia_img.max()
    ans.append(input_img)'''
  img_list=[img1,img2,img3,img4,img5]
  h_list,w_list=[_.shape[0] for _ in img_list],[_.shape[1] for _ in img_list]
  #print(type(img1))
  #print(img1.shape)
  result_list=test(device,net,img_list,5,224)
  #result_list=[result_list[i].resize((w_list[i], h_list[i]), Image.BILINEAR) for i in range(5)]
  img1,img2,img3,img4,img5=result_list#test('cpu',net,img_list,5,224)
  return img1,img2,img3,img4,img5

#gr.Image(shape=(224, 2))
demo = gr.Interface(sepia, inputs=["image","image","image","image","image"], outputs=["image","image","image","image","image"])#gr.Interface(sepia, gr.Image(shape=(200, 200)), "image")

demo.launch(debug=True)