Spaces:
Running
Running
Jason Adrian
commited on
Commit
•
d360108
1
Parent(s):
eff3c10
bodypartxr classifier
Browse files- app.py +82 -4
- resnet18.py +129 -0
app.py
CHANGED
@@ -1,7 +1,85 @@
|
|
1 |
import gradio as gr
|
|
|
|
|
|
|
2 |
|
3 |
-
|
4 |
-
return "Hello " + name + "!!"
|
5 |
|
6 |
-
|
7 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
+
import torch
|
3 |
+
from torchvision.transforms import transforms
|
4 |
+
import numpy as np
|
5 |
|
6 |
+
from resnet18 import ResNet18
|
|
|
7 |
|
8 |
+
model = ResNet18(1, 5)
|
9 |
+
|
10 |
+
checkpoint = torch.load('C:\jason\semester 8\Magang\Hugging-face-bodypartxr\bodypartxr\acc=0.94.ckpt')
|
11 |
+
|
12 |
+
# The state dict will contains net.layer_name
|
13 |
+
# Our model doesn't contains `net.` so we have to rename it
|
14 |
+
state_dict = checkpoint['state_dict']
|
15 |
+
for key in list(state_dict.keys()):
|
16 |
+
if 'net.' in key:
|
17 |
+
state_dict[key.replace('net.', '')] = state_dict[key]
|
18 |
+
del state_dict[key]
|
19 |
+
|
20 |
+
model.load_state_dict(state_dict)
|
21 |
+
model.eval()
|
22 |
+
|
23 |
+
class_names = ['abdominal', 'adult', 'others', 'pediatric', 'spine']
|
24 |
+
class_names.sort()
|
25 |
+
|
26 |
+
transformation_pipeline = transforms.Compose([
|
27 |
+
transforms.ToPILImage(),
|
28 |
+
transforms.Grayscale(num_output_channels=1),
|
29 |
+
transforms.CenterCrop((384, 384)),
|
30 |
+
transforms.ToTensor(),
|
31 |
+
transforms.Normalize(mean=[0.50807575], std=[0.20823])
|
32 |
+
])
|
33 |
+
|
34 |
+
|
35 |
+
def preprocess_image(image: np.ndarray):
|
36 |
+
"""Preprocess the input image.
|
37 |
+
|
38 |
+
Note that the input image is in RGB mode.
|
39 |
+
|
40 |
+
Parameters
|
41 |
+
----------
|
42 |
+
image: np.ndarray
|
43 |
+
Input image from callback.
|
44 |
+
"""
|
45 |
+
|
46 |
+
image = transformation_pipeline(image)
|
47 |
+
image = torch.unsqueeze(image, 0)
|
48 |
+
|
49 |
+
return image
|
50 |
+
|
51 |
+
|
52 |
+
def image_classifier(inp):
|
53 |
+
"""Image Classifier Function.
|
54 |
+
|
55 |
+
Parameters
|
56 |
+
----------
|
57 |
+
inp: Optional[np.ndarray] = None
|
58 |
+
Input image from callback
|
59 |
+
|
60 |
+
Returns
|
61 |
+
-------
|
62 |
+
Dict
|
63 |
+
A dictionary class names and its probability
|
64 |
+
"""
|
65 |
+
|
66 |
+
# If input not valid, return dummy data or raise error
|
67 |
+
if inp is None:
|
68 |
+
return {'cat': 0.3, 'dog': 0.7}
|
69 |
+
|
70 |
+
# preprocess
|
71 |
+
image = preprocess_image(inp)
|
72 |
+
image = image.to(dtype=torch.float32)
|
73 |
+
|
74 |
+
# inference
|
75 |
+
result = model(image)
|
76 |
+
|
77 |
+
# postprocess
|
78 |
+
result = torch.nn.functional.softmax(result, dim=1) # apply softmax
|
79 |
+
result = result[0].detach().numpy().tolist() # take the first batch
|
80 |
+
labeled_result = {name:score for name, score in zip(class_names, result)}
|
81 |
+
|
82 |
+
return labeled_result
|
83 |
+
|
84 |
+
demo = gr.Interface(fn=image_classifier, inputs="image", outputs="label")
|
85 |
+
demo.launch()
|
resnet18.py
ADDED
@@ -0,0 +1,129 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional
|
2 |
+
|
3 |
+
import torch.nn as nn
|
4 |
+
import torch
|
5 |
+
|
6 |
+
class BasicBlock(nn.Module):
|
7 |
+
"""ResNet Basic Block.
|
8 |
+
|
9 |
+
Parameters
|
10 |
+
----------
|
11 |
+
in_channels : int
|
12 |
+
Number of input channels
|
13 |
+
out_channels : int
|
14 |
+
Number of output channels
|
15 |
+
stride : int, optional
|
16 |
+
Convolution stride size, by default 1
|
17 |
+
identity_downsample : Optional[torch.nn.Module], optional
|
18 |
+
Downsampling layer, by default None
|
19 |
+
"""
|
20 |
+
|
21 |
+
def __init__(self,
|
22 |
+
in_channels: int,
|
23 |
+
out_channels: int,
|
24 |
+
stride: int = 1,
|
25 |
+
identity_downsample: Optional[torch.nn.Module] = None):
|
26 |
+
super(BasicBlock, self).__init__()
|
27 |
+
self.conv1 = nn.Conv2d(in_channels,
|
28 |
+
out_channels,
|
29 |
+
kernel_size = 3,
|
30 |
+
stride = stride,
|
31 |
+
padding = 1)
|
32 |
+
self.bn1 = nn.BatchNorm2d(out_channels)
|
33 |
+
self.relu = nn.ReLU()
|
34 |
+
self.conv2 = nn.Conv2d(out_channels,
|
35 |
+
out_channels,
|
36 |
+
kernel_size = 3,
|
37 |
+
stride = 1,
|
38 |
+
padding = 1)
|
39 |
+
self.bn2 = nn.BatchNorm2d(out_channels)
|
40 |
+
self.identity_downsample = identity_downsample
|
41 |
+
|
42 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
43 |
+
"""Apply forward computation."""
|
44 |
+
identity = x
|
45 |
+
x = self.conv1(x)
|
46 |
+
x = self.bn1(x)
|
47 |
+
x = self.relu(x)
|
48 |
+
x = self.conv2(x)
|
49 |
+
x = self.bn2(x)
|
50 |
+
|
51 |
+
# Apply an operation to the identity output.
|
52 |
+
# Useful to reduce the layer size and match from conv2 output
|
53 |
+
if self.identity_downsample is not None:
|
54 |
+
identity = self.identity_downsample(identity)
|
55 |
+
x += identity
|
56 |
+
x = self.relu(x)
|
57 |
+
return x
|
58 |
+
|
59 |
+
class ResNet18(nn.Module):
|
60 |
+
"""Construct ResNet-18 Model.
|
61 |
+
|
62 |
+
Parameters
|
63 |
+
----------
|
64 |
+
input_channels : int
|
65 |
+
Number of input channels
|
66 |
+
num_classes : int
|
67 |
+
Number of class outputs
|
68 |
+
"""
|
69 |
+
|
70 |
+
def __init__(self, input_channels, num_classes):
|
71 |
+
|
72 |
+
super(ResNet18, self).__init__()
|
73 |
+
self.conv1 = nn.Conv2d(input_channels,
|
74 |
+
64, kernel_size = 7,
|
75 |
+
stride = 2, padding=3)
|
76 |
+
self.bn1 = nn.BatchNorm2d(64)
|
77 |
+
self.relu = nn.ReLU()
|
78 |
+
self.maxpool = nn.MaxPool2d(kernel_size = 3,
|
79 |
+
stride = 2,
|
80 |
+
padding = 1)
|
81 |
+
|
82 |
+
self.layer1 = self._make_layer(64, 64, stride = 1)
|
83 |
+
self.layer2 = self._make_layer(64, 128, stride = 2)
|
84 |
+
self.layer3 = self._make_layer(128, 256, stride = 2)
|
85 |
+
self.layer4 = self._make_layer(256, 512, stride = 2)
|
86 |
+
|
87 |
+
# Last layers
|
88 |
+
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
|
89 |
+
self.fc = nn.Linear(512, num_classes)
|
90 |
+
|
91 |
+
def identity_downsample(self, in_channels: int, out_channels: int) -> nn.Module:
|
92 |
+
"""Downsampling block to reduce the feature sizes."""
|
93 |
+
return nn.Sequential(
|
94 |
+
nn.Conv2d(in_channels,
|
95 |
+
out_channels,
|
96 |
+
kernel_size = 3,
|
97 |
+
stride = 2,
|
98 |
+
padding = 1),
|
99 |
+
nn.BatchNorm2d(out_channels)
|
100 |
+
)
|
101 |
+
|
102 |
+
def _make_layer(self, in_channels: int, out_channels: int, stride: int) -> nn.Module:
|
103 |
+
"""Create sequential basic block."""
|
104 |
+
identity_downsample = None
|
105 |
+
|
106 |
+
# Add downsampling function
|
107 |
+
if stride != 1:
|
108 |
+
identity_downsample = self.identity_downsample(in_channels, out_channels)
|
109 |
+
|
110 |
+
return nn.Sequential(
|
111 |
+
BasicBlock(in_channels, out_channels, identity_downsample=identity_downsample, stride=stride),
|
112 |
+
BasicBlock(out_channels, out_channels)
|
113 |
+
)
|
114 |
+
|
115 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
116 |
+
x = self.conv1(x)
|
117 |
+
x = self.bn1(x)
|
118 |
+
x = self.relu(x)
|
119 |
+
x = self.maxpool(x)
|
120 |
+
|
121 |
+
x = self.layer1(x)
|
122 |
+
x = self.layer2(x)
|
123 |
+
x = self.layer3(x)
|
124 |
+
x = self.layer4(x)
|
125 |
+
|
126 |
+
x = self.avgpool(x)
|
127 |
+
x = x.view(x.shape[0], -1)
|
128 |
+
x = self.fc(x)
|
129 |
+
return x
|