File size: 4,431 Bytes
d25fd03
1ba239b
 
 
d25fd03
1ba239b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d25fd03
 
 
 
 
 
1ba239b
 
 
 
 
 
d25fd03
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
import gradio as gr
import cv2
import gradio as gr
import os
from PIL import Image
import numpy as np
import torch
from torch.autograd import Variable
from torchvision import transforms
import torch.nn.functional as F
import matplotlib.pyplot as plt
import warnings
warnings.filterwarnings("ignore")


# project imports
from data_loader_cache import normalize, im_reader, im_preprocess 
from models import *

#Helpers
device = 'cuda' if torch.cuda.is_available() else 'cpu'

    
class GOSNormalize(object):
    '''
    Normalize the Image using torch.transforms
    '''
    def __init__(self, mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225]):
        self.mean = mean
        self.std = std

    def __call__(self,image):
        image = normalize(image,self.mean,self.std)
        return image


transform =  transforms.Compose([GOSNormalize([0.5,0.5,0.5],[1.0,1.0,1.0])])

def load_image(im_path, hypar):
    im = im_reader(im_path)
    im, im_shp = im_preprocess(im, hypar["cache_size"])
    im = torch.divide(im,255.0)
    shape = torch.from_numpy(np.array(im_shp))
    return transform(im).unsqueeze(0), shape.unsqueeze(0) # make a batch of image, shape


def build_model(hypar,device):
    net = hypar["model"]#GOSNETINC(3,1)

    # convert to half precision
    if(hypar["model_digit"]=="half"):
        net.half()
        for layer in net.modules():
            if isinstance(layer, nn.BatchNorm2d):
                layer.float()

    net.to(device)

    if(hypar["restore_model"]!=""):
        net.load_state_dict(torch.load(hypar["model_path"]+"/"+hypar["restore_model"], map_location=device))
        net.to(device)
    net.eval()  
    return net

    
def predict(net,  inputs_val, shapes_val, hypar, device):
    '''
    Given an Image, predict the mask
    '''
    net.eval()

    if(hypar["model_digit"]=="full"):
        inputs_val = inputs_val.type(torch.FloatTensor)
    else:
        inputs_val = inputs_val.type(torch.HalfTensor)

  
    inputs_val_v = Variable(inputs_val, requires_grad=False).to(device) # wrap inputs in Variable
   
    ds_val = net(inputs_val_v)[0] # list of 6 results

    pred_val = ds_val[0][0,:,:,:] # B x 1 x H x W    # we want the first one which is the most accurate prediction

    ## recover the prediction spatial size to the orignal image size
    pred_val = torch.squeeze(F.upsample(torch.unsqueeze(pred_val,0),(shapes_val[0][0],shapes_val[0][1]),mode='bilinear'))

    ma = torch.max(pred_val)
    mi = torch.min(pred_val)
    pred_val = (pred_val-mi)/(ma-mi) # max = 1

    if device == 'cuda': torch.cuda.empty_cache()
    return (pred_val.detach().cpu().numpy()*255).astype(np.uint8) # it is the mask we need
    
# Set Parameters
hypar = {} # paramters for inferencing


hypar["model_path"] ="./saved_models" ## load trained weights from this path
hypar["restore_model"] = "isnet.pth" ## name of the to-be-loaded weights
hypar["interm_sup"] = False ## indicate if activate intermediate feature supervision

##  choose floating point accuracy --
hypar["model_digit"] = "full" ## indicates "half" or "full" accuracy of float number
hypar["seed"] = 0

hypar["cache_size"] = [1024, 1024] ## cached input spatial resolution, can be configured into different size

## data augmentation parameters ---
hypar["input_size"] = [1024, 1024] ## mdoel input spatial size, usually use the same value hypar["cache_size"], which means we don't further resize the images
hypar["crop_size"] = [1024, 1024] ## random crop size from the input, it is usually set as smaller than hypar["cache_size"], e.g., [920,920] for data augmentation

hypar["model"] = ISNetDIS()

 # Build Model
net = build_model(hypar, device)


def inference(image: Image):
  image_path = image
  
  image_tensor, orig_size = load_image(image_path, hypar) 
  mask = predict(net, image_tensor, orig_size, hypar, device)
  
  pil_mask = Image.fromarray(mask).convert('L')
  im_rgb = Image.open(image).convert("RGB")
  
  im_rgba = im_rgb.copy()
  im_rgba.putalpha(pil_mask)

  return im_rgba

def bw(image_file:Image):
    img = Image.open(image_file)
    img = img.convert("L")
    return img

iface = gr.Interface(fn=inference, 
                    inputs=gr.Image(type='filepath'), 
                    outputs=["image"],
                    title="Remove Background",
                    description="Uses <a href='https://github.com/xuebinqin/DIS'>DIS</a> to remove background"
                    )
iface.launch()