Spaces:
Runtime error
Runtime error
First demo commit, v2
Browse files- README.md +1 -0
- app.py +26 -0
- basic_ops.py +91 -0
- models.py +276 -0
- network.py +117 -0
- network_configure.py +143 -0
- requirements.txt +7 -0
- resnet_module.py +101 -0
- trained_models/README.md +3 -0
- utils/evaluation_utils.py +70 -0
- utils/predict_utils.py +152 -0
- utils/train_utils.py +60 -0
README.md
CHANGED
@@ -7,6 +7,7 @@ sdk: gradio
|
|
7 |
sdk_version: 3.1.3
|
8 |
app_file: app.py
|
9 |
pinned: false
|
|
|
10 |
---
|
11 |
|
12 |
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
|
|
7 |
sdk_version: 3.1.3
|
8 |
app_file: app.py
|
9 |
pinned: false
|
10 |
+
python_version: 3.7
|
11 |
---
|
12 |
|
13 |
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
app.py
ADDED
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
from models import Noise2Same
|
3 |
+
import gradio as gr
|
4 |
+
|
5 |
+
os.system("mkdir trained_models/denoising_ImageNet")
|
6 |
+
os.system("cd trained_models/denoising_ImageNet; gdown https://drive.google.com/uc?id=1asrwULW1lDFasystBc3UfShh5EeTHpkW; gdown https://drive.google.com/uc?id=1Re1ER7KtujBunN0-74QmYrrOx77WpVXK; gdown https://drive.google.com/uc?id=1QdlyUPUKyyGtqD0zBrj5F7qQZtmUELSu; gdown https://drive.google.com/uc?id=1LQsYR26ldHebcdQtP2zt4Mh-ZH9vXQ2S; gdown https://drive.google.com/uc?id=1AxTDD4dS0DtzmBywjGyeJYgDrw-XjYbc; gdown https://drive.google.com/uc?id=1w4UdNAbOjvWSL0Jgbq8_hCniaxqsbLaQ; cd ../..")
|
7 |
+
os.system("wget -O arch.png https://i.imgur.com/NruRABn.png")
|
8 |
+
os.system("wget -O parrot.png https://i.imgur.com/zdji3xv.png")
|
9 |
+
os.system("wget -O lion.png https://i.imgur.com/qNT0lJJ.png")
|
10 |
+
|
11 |
+
model = Noise2Same('trained_models/', 'denoising_ImageNet', dim=2, in_channels=3)
|
12 |
+
|
13 |
+
def norm(x):
|
14 |
+
x = (x-x.min())/(x.max()-x.min())
|
15 |
+
return x
|
16 |
+
|
17 |
+
def predict(img):
|
18 |
+
pred = model.predict(img.astype('float32'))
|
19 |
+
return norm(pred)
|
20 |
+
|
21 |
+
img = gr.inputs.Image()
|
22 |
+
|
23 |
+
title = "Noise2Same: Optimizing A Self-Supervised Bound for Image Denoising"
|
24 |
+
description = "Interactive demo of Noise2Same, an image denoising method developed by Yaochen Xie"
|
25 |
+
|
26 |
+
gr.Interface(predict, "image", "image", examples=[["lion.png"], ["arch.png"], ["parrot.png"]], title=title, description=description).launch()
|
basic_ops.py
ADDED
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import logging, os
|
2 |
+
logging.disable(logging.WARNING)
|
3 |
+
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
|
4 |
+
|
5 |
+
import tensorflow as tf
|
6 |
+
from network_configure import conf_basic_ops
|
7 |
+
|
8 |
+
|
9 |
+
"""This script defines basic operaters.
|
10 |
+
"""
|
11 |
+
|
12 |
+
|
13 |
+
def convolution_2D(inputs, filters, kernel_size, strides, use_bias, name=None):
|
14 |
+
"""Performs 2D convolution without activation function.
|
15 |
+
If followed by batch normalization, set use_bias=False.
|
16 |
+
"""
|
17 |
+
return tf.layers.conv2d(
|
18 |
+
inputs=inputs,
|
19 |
+
filters=filters,
|
20 |
+
kernel_size=kernel_size,
|
21 |
+
strides=strides,
|
22 |
+
padding='same',
|
23 |
+
use_bias=use_bias,
|
24 |
+
kernel_initializer=conf_basic_ops['kernel_initializer'],
|
25 |
+
name=name,
|
26 |
+
)
|
27 |
+
|
28 |
+
def convolution_3D(inputs, filters, kernel_size, strides, use_bias, name=None):
|
29 |
+
"""Performs 3D convolution without activation function.
|
30 |
+
If followed by batch normalization, set use_bias=False.
|
31 |
+
"""
|
32 |
+
return tf.layers.conv3d(
|
33 |
+
inputs=inputs,
|
34 |
+
filters=filters,
|
35 |
+
kernel_size=kernel_size,
|
36 |
+
strides=strides,
|
37 |
+
padding='same',
|
38 |
+
use_bias=use_bias,
|
39 |
+
kernel_initializer=conf_basic_ops['kernel_initializer'],
|
40 |
+
name=name,
|
41 |
+
)
|
42 |
+
|
43 |
+
def transposed_convolution_2D(inputs, filters, kernel_size, strides, use_bias, name=None):
|
44 |
+
"""Performs 2D transposed convolution without activation function.
|
45 |
+
If followed by batch normalization, set use_bias=False.
|
46 |
+
"""
|
47 |
+
return tf.layers.conv2d_transpose(
|
48 |
+
inputs=inputs,
|
49 |
+
filters=filters,
|
50 |
+
kernel_size=kernel_size,
|
51 |
+
strides=strides,
|
52 |
+
padding='same',
|
53 |
+
use_bias=use_bias,
|
54 |
+
kernel_initializer=conf_basic_ops['kernel_initializer'],
|
55 |
+
name=name,
|
56 |
+
)
|
57 |
+
|
58 |
+
def transposed_convolution_3D(inputs, filters, kernel_size, strides, use_bias, name=None):
|
59 |
+
"""Performs 3D transposed convolution without activation function.
|
60 |
+
If followed by batch normalization, set use_bias=False.
|
61 |
+
"""
|
62 |
+
return tf.layers.conv3d_transpose(
|
63 |
+
inputs=inputs,
|
64 |
+
filters=filters,
|
65 |
+
kernel_size=kernel_size,
|
66 |
+
strides=strides,
|
67 |
+
padding='same',
|
68 |
+
use_bias=use_bias,
|
69 |
+
kernel_initializer=conf_basic_ops['kernel_initializer'],
|
70 |
+
name=name,
|
71 |
+
)
|
72 |
+
|
73 |
+
def batch_norm(inputs, training, name=None):
|
74 |
+
"""Performs a batch normalization.
|
75 |
+
We set fused=True for a significant performance boost.
|
76 |
+
See https://www.tensorflow.org/performance/performance_guide#common_fused_ops
|
77 |
+
"""
|
78 |
+
return tf.layers.batch_normalization(
|
79 |
+
inputs=inputs,
|
80 |
+
momentum=conf_basic_ops['momentum'],
|
81 |
+
epsilon=conf_basic_ops['epsilon'],
|
82 |
+
center=True,
|
83 |
+
scale=True,
|
84 |
+
training=training,
|
85 |
+
fused=True,
|
86 |
+
name=name,
|
87 |
+
)
|
88 |
+
|
89 |
+
def relu(inputs, name=None):
|
90 |
+
return tf.nn.relu(inputs, name=name) if conf_basic_ops['relu_type'] == 'relu' \
|
91 |
+
else tf.nn.relu6(inputs, name=name)
|
models.py
ADDED
@@ -0,0 +1,276 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os, cv2
|
2 |
+
import numpy as np
|
3 |
+
from network_configure import conf_unet
|
4 |
+
from network import *
|
5 |
+
from utils.predict_utils import get_coord, PercentileNormalizer, PadAndCropResizer
|
6 |
+
from utils.train_utils import augment_patch
|
7 |
+
from utils import train_utils
|
8 |
+
|
9 |
+
# UNet implementation inherited from GVTNets: https://github.com/zhengyang-wang/GVTNets
|
10 |
+
training_config = {'base_learning_rate': 0.0004,
|
11 |
+
'lr_decay_steps':5e3,
|
12 |
+
'lr_decay_rate':0.5,
|
13 |
+
'lr_staircase':True}
|
14 |
+
|
15 |
+
class Noise2Same(object):
|
16 |
+
|
17 |
+
def __init__(self, base_dir, name,
|
18 |
+
dim=2, in_channels=1, lmbd=None,
|
19 |
+
masking='gaussian', mask_perc=0.5,
|
20 |
+
opt_config=training_config, **kwargs):
|
21 |
+
|
22 |
+
self.base_dir = base_dir # model direction
|
23 |
+
self.name = name # model name
|
24 |
+
self.dim = dim # image dimension
|
25 |
+
self.in_channels = in_channels # image channels
|
26 |
+
self.lmbd = lmbd # lambda in loss fn
|
27 |
+
self.masking = masking
|
28 |
+
self.mask_perc = mask_perc
|
29 |
+
|
30 |
+
self.opt_config = opt_config
|
31 |
+
conf_unet['dimension'] = '%dD'%dim
|
32 |
+
self.net = UNet(conf_unet)
|
33 |
+
|
34 |
+
def _model_fn(self, features, labels, mode):
|
35 |
+
conv_op = convolution_2D if self.dim==2 else convolution_3D
|
36 |
+
axis = {3:[1,2,3,4], 2:[1,2,3]}[self.dim]
|
37 |
+
|
38 |
+
def image_summary(img):
|
39 |
+
return tf.reduce_max(img, axis=1) if self.dim == 3 else img
|
40 |
+
|
41 |
+
# Local average excluding the center pixel (donut)
|
42 |
+
def mask_kernel(features):
|
43 |
+
kernel = (np.array([[0.5, 1.0, 0.5], [1.0, 0.0, 1.0], [0.5, 1.0, 0.5]])
|
44 |
+
if self.dim == 2 else
|
45 |
+
np.array([[[0, 0.5, 0], [0.5, 1.0, 0.5], [0, 0.5, 0]],
|
46 |
+
[[0.5, 1.0, 0.5], [1.0, 0.0, 1.0], [0.5, 1.0, 0.5]],
|
47 |
+
[[0, 0.5, 0], [0.5, 1.0, 0.5], [0, 0.5, 0]]]))
|
48 |
+
kernel = (kernel/kernel.sum())
|
49 |
+
kernels = np.empty([3, 3, self.in_channels, self.in_channels])
|
50 |
+
for i in range(self.in_channels):
|
51 |
+
kernels[:,:,i,i] = kernel
|
52 |
+
nn_conv_op = tf.nn.conv2d if self.dim == 2 else tf.nn.conv3d
|
53 |
+
return nn_conv_op(features, tf.constant(kernels.astype('float32')),
|
54 |
+
[1]*self.dim+[1,1], padding='SAME')
|
55 |
+
|
56 |
+
if not mode == tf.estimator.ModeKeys.PREDICT:
|
57 |
+
noise, mask = tf.split(labels, [self.in_channels, self.in_channels], -1)
|
58 |
+
|
59 |
+
if self.masking == 'gaussian':
|
60 |
+
masked_features = (1 - mask) * features + mask * noise
|
61 |
+
elif self.masking == 'donut':
|
62 |
+
masked_features = (1 - mask) * features + mask * mask_kernel(features)
|
63 |
+
else:
|
64 |
+
raise NotImplementedError
|
65 |
+
|
66 |
+
# Prediction from masked input
|
67 |
+
with tf.variable_scope('main_unet', reuse=tf.compat.v1.AUTO_REUSE):
|
68 |
+
out = self.net(masked_features, mode == tf.estimator.ModeKeys.TRAIN)
|
69 |
+
out = batch_norm(out, mode == tf.estimator.ModeKeys.TRAIN, 'unet_out')
|
70 |
+
out = relu(out)
|
71 |
+
preds = conv_op(out, self.in_channels, 1, 1, False, name = 'out_conv')
|
72 |
+
|
73 |
+
# Prediction from full input
|
74 |
+
with tf.variable_scope('main_unet', reuse=tf.compat.v1.AUTO_REUSE):
|
75 |
+
rawout = self.net(features, mode == tf.estimator.ModeKeys.TRAIN)
|
76 |
+
rawout = batch_norm(rawout, mode == tf.estimator.ModeKeys.TRAIN, 'unet_out')
|
77 |
+
rawout = relu(rawout)
|
78 |
+
rawpreds = conv_op(rawout, self.in_channels, 1, 1, False, name = 'out_conv')
|
79 |
+
|
80 |
+
# Loss components
|
81 |
+
rec_mse = tf.reduce_mean(tf.square(rawpreds - features), axis=None)
|
82 |
+
inv_mse = tf.reduce_sum(tf.square(rawpreds - preds) * mask) / tf.reduce_sum(mask)
|
83 |
+
bsp_mse = tf.reduce_sum(tf.square(features - preds) * mask) / tf.reduce_sum(mask)
|
84 |
+
|
85 |
+
# Tensorboard display
|
86 |
+
tf.summary.image('1_inputs', image_summary(features), max_outputs=3)
|
87 |
+
tf.summary.image('2_raw_predictions', image_summary(rawpreds), max_outputs=3)
|
88 |
+
tf.summary.image('3_mask', image_summary(mask), max_outputs=3)
|
89 |
+
tf.summary.image('4_masked_predictions', image_summary(preds), max_outputs=3)
|
90 |
+
tf.summary.image('5_difference', image_summary(rawpreds-preds), max_outputs=3)
|
91 |
+
tf.summary.image('6_rec_error', image_summary(preds-features), max_outputs=3)
|
92 |
+
tf.summary.scalar('reconstruction', rec_mse, family='loss_metric')
|
93 |
+
tf.summary.scalar('invariance', inv_mse, family='loss_metric')
|
94 |
+
tf.summary.scalar('blind_spot', bsp_mse, family='loss_metric')
|
95 |
+
|
96 |
+
else:
|
97 |
+
with tf.variable_scope('main_unet'):
|
98 |
+
out = self.net(features, mode == tf.estimator.ModeKeys.TRAIN)
|
99 |
+
out = batch_norm(out, mode == tf.estimator.ModeKeys.TRAIN, 'unet_out')
|
100 |
+
out = relu(out)
|
101 |
+
preds = conv_op(out, self.in_channels, 1, 1, False, name = 'out_conv')
|
102 |
+
return tf.estimator.EstimatorSpec(mode=mode, predictions=preds)
|
103 |
+
|
104 |
+
lmbd = 2 if self.lmbd is None else self.lmbd
|
105 |
+
loss = rec_mse + lmbd*tf.sqrt(inv_mse)
|
106 |
+
|
107 |
+
if mode == tf.estimator.ModeKeys.TRAIN:
|
108 |
+
global_step = tf.train.get_or_create_global_step()
|
109 |
+
learning_rate = tf.train.exponential_decay(self.opt_config['base_learning_rate'],
|
110 |
+
global_step,
|
111 |
+
self.opt_config['lr_decay_steps'],
|
112 |
+
self.opt_config['lr_decay_rate'],
|
113 |
+
self.opt_config['lr_staircase'])
|
114 |
+
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
|
115 |
+
|
116 |
+
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope='main_unet')
|
117 |
+
with tf.control_dependencies(update_ops):
|
118 |
+
train_op = optimizer.minimize(loss, global_step)
|
119 |
+
else:
|
120 |
+
train_op = None
|
121 |
+
|
122 |
+
metrics = {'loss_metric/invariance':tf.metrics.mean(inv_mse),
|
123 |
+
'loss_metric/blind_spot':tf.metrics.mean(bsp_mse),
|
124 |
+
'loss_metric/reconstruction':tf.metrics.mean(rec_mse)}
|
125 |
+
|
126 |
+
return tf.estimator.EstimatorSpec(mode=mode, predictions=preds, loss=loss, train_op=train_op,
|
127 |
+
eval_metric_ops=metrics)
|
128 |
+
|
129 |
+
|
130 |
+
def _input_fn(self, sources, patch_size, batch_size, is_train=True):
|
131 |
+
# Stratified sampling inherited from Noise2Void: https://github.com/juglab/n2v
|
132 |
+
get_stratified_coords = getattr(train_utils, 'get_stratified_coords%dD'%self.dim)
|
133 |
+
rand_float_coords = getattr(train_utils, 'rand_float_coords%dD'%self.dim)
|
134 |
+
|
135 |
+
def generator():
|
136 |
+
while(True):
|
137 |
+
source = sources[np.random.randint(len(sources))]
|
138 |
+
valid_shape = source.shape[:-1] - np.array(patch_size)
|
139 |
+
if any([s<=0 for s in valid_shape]):
|
140 |
+
source_patch = augment_patch(source)
|
141 |
+
else:
|
142 |
+
coords = [np.random.randint(0, shape_i+1) for shape_i in valid_shape]
|
143 |
+
s = tuple([slice(coord, coord+size) for coord, size in zip(coords, patch_size)])
|
144 |
+
source_patch = augment_patch(source[s])
|
145 |
+
|
146 |
+
mask = np.zeros_like(source_patch)
|
147 |
+
for c in range(self.in_channels):
|
148 |
+
boxsize = np.round(np.sqrt(100/self.mask_perc)).astype(np.int)
|
149 |
+
maskcoords = get_stratified_coords(rand_float_coords(boxsize),
|
150 |
+
box_size=boxsize, shape=tuple(patch_size))
|
151 |
+
indexing = maskcoords + (c,)
|
152 |
+
mask[indexing] = 1.0
|
153 |
+
|
154 |
+
noise_patch = np.concatenate([np.random.normal(0, 0.2, source_patch.shape), mask], axis=-1)
|
155 |
+
yield source_patch, noise_patch
|
156 |
+
|
157 |
+
def generator_val():
|
158 |
+
for idx in range(len(sources)):
|
159 |
+
source_patch = sources[idx]
|
160 |
+
patch_size = source_patch.shape[:-1]
|
161 |
+
boxsize = np.round(np.sqrt(100/self.mask_perc)).astype(np.int)
|
162 |
+
maskcoords = get_stratified_coords(rand_float_coords(boxsize),
|
163 |
+
box_size=boxsize, shape=tuple(patch_size))
|
164 |
+
indexing = maskcoords + (0,)
|
165 |
+
mask = np.zeros_like(source_patch)
|
166 |
+
mask[indexing] = 1.0
|
167 |
+
noise_patch = np.concatenate([np.random.normal(0, 0.2, source_patch.shape), mask], axis=-1)
|
168 |
+
yield source_patch, noise_patch
|
169 |
+
|
170 |
+
output_types = (tf.float32, tf.float32)
|
171 |
+
output_shapes = (tf.TensorShape(list(patch_size) + [self.in_channels]),
|
172 |
+
tf.TensorShape(list(patch_size) + [self.in_channels*2]))
|
173 |
+
gen = generator if is_train else generator_val
|
174 |
+
dataset = tf.data.Dataset.from_generator(gen, output_types=output_types, output_shapes=output_shapes)
|
175 |
+
dataset = dataset.batch(batch_size).prefetch(tf.data.experimental.AUTOTUNE)
|
176 |
+
|
177 |
+
return dataset
|
178 |
+
|
179 |
+
|
180 |
+
def train(self, source_lst, patch_size, validation=None, batch_size=64, save_steps=1000, log_steps=200, steps=50000):
|
181 |
+
assert len(patch_size)==self.dim
|
182 |
+
assert len(source_lst[0].shape)==self.dim + 1
|
183 |
+
assert source_lst[0].shape[-1]==self.in_channels
|
184 |
+
|
185 |
+
ses_config = tf.ConfigProto()
|
186 |
+
ses_config.gpu_options.allow_growth = True
|
187 |
+
|
188 |
+
run_config = tf.estimator.RunConfig(model_dir=self.base_dir+'/'+self.name,
|
189 |
+
save_checkpoints_steps=save_steps,
|
190 |
+
session_config=ses_config,
|
191 |
+
log_step_count_steps=log_steps,
|
192 |
+
save_summary_steps=log_steps,
|
193 |
+
keep_checkpoint_max=2)
|
194 |
+
|
195 |
+
estimator = tf.estimator.Estimator(model_fn=self._model_fn,
|
196 |
+
model_dir=self.base_dir+'/'+self.name,
|
197 |
+
config=run_config)
|
198 |
+
|
199 |
+
input_fn = lambda: self._input_fn(source_lst, patch_size, batch_size=batch_size)
|
200 |
+
|
201 |
+
if validation is not None:
|
202 |
+
train_spec = tf.estimator.TrainSpec(input_fn=input_fn, max_steps=steps)
|
203 |
+
val_input_fn = lambda: self._input_fn(validation.astype('float32'),
|
204 |
+
validation.shape[1:-1],
|
205 |
+
batch_size=4,
|
206 |
+
is_train=False)
|
207 |
+
eval_spec = tf.estimator.EvalSpec(input_fn=val_input_fn, throttle_secs=120)
|
208 |
+
tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)
|
209 |
+
else:
|
210 |
+
estimator.train(input_fn=input_fn, steps=steps)
|
211 |
+
|
212 |
+
|
213 |
+
# Used for single image prediction
|
214 |
+
def predict(self, image, resizer=PadAndCropResizer(), checkpoint_path=None,
|
215 |
+
im_mean=None, im_std=None):
|
216 |
+
|
217 |
+
tf.logging.set_verbosity(tf.logging.ERROR)
|
218 |
+
estimator = tf.estimator.Estimator(model_fn=self._model_fn,
|
219 |
+
model_dir=self.base_dir+'/'+self.name)
|
220 |
+
|
221 |
+
im_mean, im_std = ((image.mean(), image.std()) if im_mean is None or im_std is None else (im_mean, im_std))
|
222 |
+
image = (image - im_mean)/im_std
|
223 |
+
if self.in_channels == 1:
|
224 |
+
image = resizer.before(image, 2 ** (self.net.depth), exclude=None)
|
225 |
+
input_fn = tf.estimator.inputs.numpy_input_fn(x=image[None, ..., None], batch_size=1, num_epochs=1, shuffle=False)
|
226 |
+
image = list(estimator.predict(input_fn=input_fn, checkpoint_path=checkpoint_path))[0][..., 0]
|
227 |
+
image = resizer.after(image, exclude=None)
|
228 |
+
else:
|
229 |
+
image = resizer.before(image, 2 ** (self.net.depth), exclude=-1)
|
230 |
+
input_fn = tf.estimator.inputs.numpy_input_fn(x=image[None], batch_size=1, num_epochs=1, shuffle=False)
|
231 |
+
image = list(estimator.predict(input_fn=input_fn, checkpoint_path=checkpoint_path))[0]
|
232 |
+
image = resizer.after(image, exclude=-1)
|
233 |
+
image = image*im_std + im_mean
|
234 |
+
|
235 |
+
return image
|
236 |
+
|
237 |
+
# Used for batch images prediction
|
238 |
+
def batch_predict(self, images, resizer=PadAndCropResizer(), checkpoint_path=None,
|
239 |
+
im_mean=None, im_std=None, batch_size=32):
|
240 |
+
|
241 |
+
tf.logging.set_verbosity(tf.logging.ERROR)
|
242 |
+
estimator = tf.estimator.Estimator(model_fn=self._model_fn,
|
243 |
+
model_dir=self.base_dir+'/'+self.name)
|
244 |
+
|
245 |
+
im_mean, im_std = ((images.mean(), images.std()) if im_mean is None or im_std is None else (im_mean, im_std))
|
246 |
+
|
247 |
+
images = (images - im_mean)/im_std
|
248 |
+
images = resizer.before(images, 2 ** (self.net.depth), exclude=0)
|
249 |
+
input_fn = tf.estimator.inputs.numpy_input_fn(x=images[ ..., None], batch_size=batch_size, num_epochs=1, shuffle=False)
|
250 |
+
images = np.stack(list(estimator.predict(input_fn=input_fn, checkpoint_path=checkpoint_path)))[..., 0]
|
251 |
+
images = resizer.after(images, exclude=0)
|
252 |
+
images = images*im_std + im_mean
|
253 |
+
|
254 |
+
return images
|
255 |
+
|
256 |
+
# Used for extremely large input images
|
257 |
+
def crop_predict(self, image, size, margin, resizer=PadAndCropResizer(), checkpoint_path=None,
|
258 |
+
im_mean=None, im_std=None):
|
259 |
+
|
260 |
+
tf.logging.set_verbosity(tf.logging.ERROR)
|
261 |
+
estimator = tf.estimator.Estimator(model_fn=self._model_fn,
|
262 |
+
model_dir=self.base_dir+'/'+self.name)
|
263 |
+
|
264 |
+
im_mean, im_std = ((image.mean(), image.std()) if im_mean is None or im_std is None else (im_mean, im_std))
|
265 |
+
image = (image - im_mean)/im_std
|
266 |
+
out_image = np.empty(image.shape, dtype='float32')
|
267 |
+
for src_s, trg_s, mrg_s in get_coord(image.shape, size, margin):
|
268 |
+
patch = resizer.before(image[src_s], 2 ** (self.net.depth), exclude=None)
|
269 |
+
input_fn = tf.estimator.inputs.numpy_input_fn(x=patch[None, ..., None], batch_size=1, num_epochs=1, shuffle=False)
|
270 |
+
patch = list(estimator.predict(input_fn=input_fn, checkpoint_path=checkpoint_path))[0][..., 0]
|
271 |
+
patch = resizer.after(patch, exclude=None)
|
272 |
+
out_image[trg_s] = patch[mrg_s]
|
273 |
+
|
274 |
+
image = out_image*im_std + im_mean
|
275 |
+
|
276 |
+
return image
|
network.py
ADDED
@@ -0,0 +1,117 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import logging, os
|
2 |
+
logging.disable(logging.WARNING)
|
3 |
+
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
|
4 |
+
|
5 |
+
import tensorflow as tf
|
6 |
+
from basic_ops import *
|
7 |
+
from resnet_module import *
|
8 |
+
|
9 |
+
|
10 |
+
"""This script generates the U-Net architecture according to conf_unet.
|
11 |
+
"""
|
12 |
+
|
13 |
+
class UNet(object):
|
14 |
+
def __init__(self, conf_unet):
|
15 |
+
self.depth = conf_unet['depth']
|
16 |
+
self.dimension = conf_unet['dimension']
|
17 |
+
self.first_output_filters = conf_unet['first_output_filters']
|
18 |
+
self.encoding_block_sizes = conf_unet['encoding_block_sizes']
|
19 |
+
self.downsampling = conf_unet['downsampling']
|
20 |
+
self.decoding_block_sizes = conf_unet['decoding_block_sizes']
|
21 |
+
self.skip_method = conf_unet['skip_method']
|
22 |
+
|
23 |
+
def __call__(self, inputs, training):
|
24 |
+
"""Add operations to classify a batch of input images.
|
25 |
+
|
26 |
+
Args:
|
27 |
+
inputs: A Tensor representing a batch of input images.
|
28 |
+
training: A boolean. Set to True to add operations required only when
|
29 |
+
training the classifier.
|
30 |
+
|
31 |
+
Returns:
|
32 |
+
A logits Tensor with shape [<batch_size>, self.num_classes].
|
33 |
+
"""
|
34 |
+
|
35 |
+
return self._build_network(inputs, training)
|
36 |
+
|
37 |
+
|
38 |
+
################################################################################
|
39 |
+
# Composite blocks building the network
|
40 |
+
################################################################################
|
41 |
+
def _build_network(self, inputs, training):
|
42 |
+
# first_convolution
|
43 |
+
if self.dimension == '2D':
|
44 |
+
convolution = convolution_2D
|
45 |
+
elif self.dimension == '3D':
|
46 |
+
convolution = convolution_3D
|
47 |
+
inputs = convolution(inputs, self.first_output_filters, 3, 1, False, 'first_convolution')
|
48 |
+
|
49 |
+
# encoding_block_1
|
50 |
+
with tf.variable_scope('encoding_block_1'):
|
51 |
+
for block_index in range(0, self.encoding_block_sizes[0]):
|
52 |
+
inputs = res_block(inputs, self.first_output_filters, training, self.dimension,
|
53 |
+
'res_%d' % block_index)
|
54 |
+
|
55 |
+
# encoding_block_i (down) = downsampling + zero or more res_block, i = 2, 3, ..., depth
|
56 |
+
skip_inputs = [] # for identity skip connections
|
57 |
+
for i in range(2, self.depth+1):
|
58 |
+
skip_inputs.append(inputs)
|
59 |
+
with tf.variable_scope('encoding_block_%d' % i):
|
60 |
+
output_filters = self.first_output_filters * (2**(i-1))
|
61 |
+
|
62 |
+
# downsampling
|
63 |
+
downsampling_func = self._get_downsampling_function(self.downsampling[i-2])
|
64 |
+
inputs = downsampling_func(inputs, output_filters, training, self.dimension,
|
65 |
+
'downsampling')
|
66 |
+
|
67 |
+
for block_index in range(0, self.encoding_block_sizes[i-1]):
|
68 |
+
inputs = res_block(inputs, output_filters, training, self.dimension,
|
69 |
+
'res_%d' % block_index)
|
70 |
+
|
71 |
+
# bottom_block = a combination of same_gto and res_block
|
72 |
+
with tf.variable_scope('bottom_block'):
|
73 |
+
output_filters = self.first_output_filters * (2**(self.depth-1))
|
74 |
+
for block_index in range(0, 1):
|
75 |
+
current_func = res_block
|
76 |
+
inputs = current_func(inputs, output_filters, training, self.dimension,
|
77 |
+
'block_%d' % block_index)
|
78 |
+
|
79 |
+
"""
|
80 |
+
Note: Identity skip connections are between the output of encoding_block_i and
|
81 |
+
the output of upsampling in decoding_block_i, i = 1, 2, ..., depth-1.
|
82 |
+
skip_inputs[i] is the output of encoding_block_i now.
|
83 |
+
len(skip_inputs) == depth - 1
|
84 |
+
skip_inputs[depth-2] should be combined during decoding_block_depth-1
|
85 |
+
skip_inputs[0] should be combined during decoding_block_1
|
86 |
+
"""
|
87 |
+
|
88 |
+
# decoding_block_j (up) = upsampling + zero or more res_block, j = depth-1, depth-2, ..., 1
|
89 |
+
for j in range(self.depth-1, 0, -1):
|
90 |
+
with tf.variable_scope('decoding_block_%d' % j):
|
91 |
+
output_filters = self.first_output_filters * (2**(j-1))
|
92 |
+
|
93 |
+
# upsampling
|
94 |
+
upsampling_func = up_transposed_convolution
|
95 |
+
inputs = upsampling_func(inputs, output_filters, training, self.dimension,
|
96 |
+
'upsampling')
|
97 |
+
|
98 |
+
# combine with skip connections
|
99 |
+
if self.skip_method == 'add':
|
100 |
+
inputs = tf.add(inputs, skip_inputs[j-1])
|
101 |
+
elif self.skip_method == 'concat':
|
102 |
+
inputs = tf.concat([inputs, skip_inputs[j-1]], axis=-1)
|
103 |
+
|
104 |
+
for block_index in range(0, self.decoding_block_sizes[self.depth-1-j]):
|
105 |
+
inputs = res_block(inputs, output_filters, training, self.dimension,
|
106 |
+
'res_%d' % block_index)
|
107 |
+
|
108 |
+
return inputs
|
109 |
+
|
110 |
+
|
111 |
+
def _get_downsampling_function(self, name):
|
112 |
+
if name == 'down_res_block':
|
113 |
+
return down_res_block
|
114 |
+
elif name == 'convolution':
|
115 |
+
return down_convolution
|
116 |
+
else:
|
117 |
+
raise ValueError("Unsupported function: %s." % (name))
|
network_configure.py
ADDED
@@ -0,0 +1,143 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import logging, os
|
2 |
+
logging.disable(logging.WARNING)
|
3 |
+
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
|
4 |
+
|
5 |
+
import tensorflow as tf
|
6 |
+
|
7 |
+
|
8 |
+
"""This is the configuration file.
|
9 |
+
"""
|
10 |
+
|
11 |
+
|
12 |
+
################################################################################
|
13 |
+
# Settings for Basic Operaters
|
14 |
+
################################################################################
|
15 |
+
|
16 |
+
conf_basic_ops = dict()
|
17 |
+
|
18 |
+
# kernel_initializer for convolutions and transposed convolutions
|
19 |
+
# If None, the default initializer is the Glorot (Xavier) normal initializer.
|
20 |
+
conf_basic_ops['kernel_initializer'] = tf.glorot_uniform_initializer()
|
21 |
+
|
22 |
+
# momentum for batch normalization
|
23 |
+
conf_basic_ops['momentum'] = 0.997
|
24 |
+
|
25 |
+
# epsilon for batch normalization
|
26 |
+
conf_basic_ops['epsilon'] = 1e-5
|
27 |
+
|
28 |
+
# String options: 'relu', 'relu6'
|
29 |
+
conf_basic_ops['relu_type'] = 'relu'
|
30 |
+
|
31 |
+
################################################################################
|
32 |
+
# Settings for Attention Modules
|
33 |
+
################################################################################
|
34 |
+
|
35 |
+
# Set the attention in same_gto
|
36 |
+
conf_attn_same = dict()
|
37 |
+
|
38 |
+
# Define the relationship between total_key_filters and output_filters.
|
39 |
+
# total_key_filters = output_filters // key_ratio
|
40 |
+
conf_attn_same['key_ratio'] = 1
|
41 |
+
|
42 |
+
# Define the relationship between total_value_filters and output_filters.
|
43 |
+
# total_key_filters = output_filters // value_ratio
|
44 |
+
conf_attn_same['value_ratio'] = 1
|
45 |
+
|
46 |
+
# number of heads
|
47 |
+
conf_attn_same['num_heads'] = 2
|
48 |
+
|
49 |
+
# dropout rate, 0.0 means no dropout
|
50 |
+
conf_attn_same['dropout_rate'] = 0.0
|
51 |
+
|
52 |
+
# whether to use softmax on attention_weights
|
53 |
+
conf_attn_same['use_softmax'] = False
|
54 |
+
|
55 |
+
# whether to use bias terms in input/output transformations
|
56 |
+
conf_attn_same['use_bias'] = True
|
57 |
+
|
58 |
+
# Set the attention in up_gto
|
59 |
+
conf_attn_up = dict()
|
60 |
+
|
61 |
+
conf_attn_up['key_ratio'] = 1
|
62 |
+
conf_attn_up['value_ratio'] = 1
|
63 |
+
conf_attn_up['num_heads'] = 2
|
64 |
+
conf_attn_up['dropout_rate'] = 0
|
65 |
+
conf_attn_up['use_softmax'] = False
|
66 |
+
conf_attn_up['use_bias'] = True
|
67 |
+
|
68 |
+
# Set the attention in down_gto
|
69 |
+
conf_attn_down = dict()
|
70 |
+
|
71 |
+
conf_attn_down['key_ratio'] = 1
|
72 |
+
conf_attn_down['value_ratio'] = 1
|
73 |
+
conf_attn_down['num_heads'] = 2
|
74 |
+
conf_attn_down['dropout_rate'] = 0.0
|
75 |
+
conf_attn_down['use_softmax'] = False
|
76 |
+
conf_attn_down['use_bias'] = True
|
77 |
+
|
78 |
+
################################################################################
|
79 |
+
# Describing the U-net
|
80 |
+
################################################################################
|
81 |
+
|
82 |
+
conf_unet = dict()
|
83 |
+
|
84 |
+
"""
|
85 |
+
Describe your U-Net under the following framework:
|
86 |
+
|
87 |
+
********************************************************************************************
|
88 |
+
layers | output_filters
|
89 |
+
|
|
90 |
+
first_convolution + encoding_block_1 (same) | first_output_filters
|
91 |
+
+ encoding_block_i, i = 2, 3, ..., depth. (down) | first_output_filters*(2**(i-1))
|
92 |
+
+ bottom_block | first_output_filters*(2**(depth-1))
|
93 |
+
+ decoding_block_j, j = depth-1, depth-2, ..., 1 (up) | first_output_filters*(2**(j-1))
|
94 |
+
+ output_layer
|
95 |
+
********************************************************************************************
|
96 |
+
|
97 |
+
Specifically,
|
98 |
+
encoding_block_1 (same) = one or more res_block
|
99 |
+
encoding_block_i (down) = downsampling + zero or more res_block, i = 2, 3, ..., depth-1
|
100 |
+
encoding_block_depth (down) = downsampling
|
101 |
+
bottom_block = a combination of same_gto and res_block
|
102 |
+
decoding_block_j (up) = upsampling + zero or more res_block, j = depth-1, depth-2, ..., 1
|
103 |
+
|
104 |
+
Identity skip connections are between the output of encoding_block_i and
|
105 |
+
the output of upsampling in decoding_block_i, i = 1, 2, ..., depth-1.
|
106 |
+
The combination method could be 'add' or 'concat'.
|
107 |
+
"""
|
108 |
+
|
109 |
+
# Set U-Net depth.
|
110 |
+
conf_unet['depth'] = 3
|
111 |
+
|
112 |
+
# Set the output_filters for first_convolution and encoding_block_1 (same).
|
113 |
+
conf_unet['first_output_filters'] = 96
|
114 |
+
|
115 |
+
# Set the encoding block sizes, i.e., number of res_block in encoding_block_i, i = 1, 2, ..., depth.
|
116 |
+
# It is an integer list whose length equals to depth.
|
117 |
+
# The first entry should be positive since encoding_block_1 = one or more res_block.
|
118 |
+
# The last entry should be zero since encoding_block_depth (down) = downsampling.
|
119 |
+
conf_unet['encoding_block_sizes'] = [1, 1, 0]
|
120 |
+
|
121 |
+
# Set the decoding block sizes, i.e., number of res_block in decoding_block_j, j = depth-1, depth-2, ..., 1.
|
122 |
+
# It is an integer list whose length equals to depth-1.
|
123 |
+
conf_unet['decoding_block_sizes'] = [1, 1]
|
124 |
+
|
125 |
+
# Set the downsampling methods for each encoding_block_i, i = 2, 3, ..., depth.
|
126 |
+
# It is an string list whose length equals to depth-1.
|
127 |
+
# String options: 'down_gto_v1', 'down_gto_v2', 'down_res_block', 'convolution'
|
128 |
+
conf_unet['downsampling'] = ['convolution', 'convolution']
|
129 |
+
|
130 |
+
# Set the combination method for identity skip connections
|
131 |
+
# String options: 'add', 'concat'
|
132 |
+
conf_unet['skip_method'] = 'concat'
|
133 |
+
|
134 |
+
# Set the output layer
|
135 |
+
|
136 |
+
|
137 |
+
# Check
|
138 |
+
assert conf_unet['depth'] == len(conf_unet['encoding_block_sizes'])
|
139 |
+
assert conf_unet['encoding_block_sizes'][0] > 0
|
140 |
+
assert conf_unet['encoding_block_sizes'][-1] == 0
|
141 |
+
assert conf_unet['depth'] == len(conf_unet['decoding_block_sizes']) + 1
|
142 |
+
assert conf_unet['depth'] == len(conf_unet['downsampling']) + 1
|
143 |
+
assert conf_unet['skip_method'] in ['add', 'concat']
|
requirements.txt
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
tensorflow==1.15
|
2 |
+
scipy
|
3 |
+
scikit-image
|
4 |
+
tifffile
|
5 |
+
gdown
|
6 |
+
opencv-python
|
7 |
+
numpy
|
resnet_module.py
ADDED
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import logging, os
|
2 |
+
logging.disable(logging.WARNING)
|
3 |
+
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
|
4 |
+
|
5 |
+
import tensorflow as tf
|
6 |
+
from basic_ops import *
|
7 |
+
|
8 |
+
|
9 |
+
"""This script defines non-attention same-, up-, down- modules.
|
10 |
+
Note that pre-activation is used for residual-like blocks.
|
11 |
+
Note that the residual block could be used for downsampling.
|
12 |
+
"""
|
13 |
+
|
14 |
+
|
15 |
+
def res_block(inputs, output_filters, training, dimension, name):
|
16 |
+
"""Standard residual block with pre-activation.
|
17 |
+
|
18 |
+
Args:
|
19 |
+
inputs: a Tensor with shape [batch, (d,) h, w, channels]
|
20 |
+
output_filters: an integer
|
21 |
+
training: a boolean for batch normalization and dropout
|
22 |
+
dimension: a string, dimension of inputs/outputs -- 2D, 3D
|
23 |
+
name: a string
|
24 |
+
|
25 |
+
Returns:
|
26 |
+
A Tensor of shape [batch, (_d,) _h, _w, output_filters]
|
27 |
+
"""
|
28 |
+
if dimension == '2D':
|
29 |
+
convolution = convolution_2D
|
30 |
+
kernel_size = 3
|
31 |
+
elif dimension == '3D':
|
32 |
+
convolution = convolution_3D
|
33 |
+
kernel_size = 3
|
34 |
+
else:
|
35 |
+
raise ValueError("Dimension (%s) must be 2D or 3D." % (dimension))
|
36 |
+
|
37 |
+
with tf.variable_scope(name):
|
38 |
+
if inputs.shape[-1] == output_filters:
|
39 |
+
shortcut = inputs
|
40 |
+
inputs = batch_norm(inputs, training, 'batch_norm_1')
|
41 |
+
inputs = relu(inputs, 'relu_1')
|
42 |
+
else:
|
43 |
+
inputs = batch_norm(inputs, training, 'batch_norm_1')
|
44 |
+
inputs = relu(inputs, 'relu_1')
|
45 |
+
shortcut = convolution(inputs, output_filters, 1, 1, False, 'projection_shortcut')
|
46 |
+
inputs = convolution(inputs, output_filters, kernel_size, 1, False, 'convolution_1')
|
47 |
+
inputs = batch_norm(inputs, training, 'batch_norm_2')
|
48 |
+
inputs = relu(inputs, 'relu_2')
|
49 |
+
inputs = convolution(inputs, output_filters, kernel_size, 1, False, 'convolution_2')
|
50 |
+
return tf.add(shortcut, inputs)
|
51 |
+
|
52 |
+
|
53 |
+
def down_res_block(inputs, output_filters, training, dimension, name):
|
54 |
+
"""Standard residual block with pre-activation for downsampling."""
|
55 |
+
if dimension == '2D':
|
56 |
+
convolution = convolution_2D
|
57 |
+
projection_shortcut = convolution_2D
|
58 |
+
elif dimension == '3D':
|
59 |
+
convolution = convolution_3D
|
60 |
+
projection_shortcut = convolution_3D
|
61 |
+
else:
|
62 |
+
raise ValueError("Dimension (%s) must be 2D or 3D." % (dimension))
|
63 |
+
|
64 |
+
with tf.variable_scope(name):
|
65 |
+
# The projection_shortcut should come after the first batch norm and ReLU.
|
66 |
+
inputs = batch_norm(inputs, training, 'batch_norm_1')
|
67 |
+
inputs = relu(inputs, 'relu_1')
|
68 |
+
shortcut = projection_shortcut(inputs, output_filters, 1, 2, False, 'projection_shortcut')
|
69 |
+
inputs = convolution(inputs, output_filters, 2, 2, False, 'convolution_1')
|
70 |
+
inputs = batch_norm(inputs, training, 'batch_norm_2')
|
71 |
+
inputs = relu(inputs, 'relu_2')
|
72 |
+
inputs = convolution(inputs, output_filters, 3, 1, False, 'convolution_2')
|
73 |
+
return tf.add(shortcut, inputs)
|
74 |
+
|
75 |
+
def down_convolution(inputs, output_filters, training, dimension, name):
|
76 |
+
"""Use a single stride 2 convolution for downsampling."""
|
77 |
+
if dimension == '2D':
|
78 |
+
convolution = convolution_2D
|
79 |
+
pool = tf.layers.max_pooling2d
|
80 |
+
elif dimension == '3D':
|
81 |
+
convolution = convolution_3D
|
82 |
+
pool = tf.layers.max_pooling3d
|
83 |
+
else:
|
84 |
+
raise ValueError("Dimension (%s) must be 2D or 3D." % (dimension))
|
85 |
+
|
86 |
+
with tf.variable_scope(name):
|
87 |
+
inputs = convolution(inputs, output_filters, 2, 2, True, 'convolution')
|
88 |
+
return inputs
|
89 |
+
|
90 |
+
def up_transposed_convolution(inputs, output_filters, training, dimension, name):
|
91 |
+
"""Use a single stride 2 transposed convolution for upsampling."""
|
92 |
+
if dimension == '2D':
|
93 |
+
transposed_convolution = transposed_convolution_2D
|
94 |
+
elif dimension == '3D':
|
95 |
+
transposed_convolution = transposed_convolution_3D
|
96 |
+
else:
|
97 |
+
raise ValueError("Dimension (%s) must be 2D or 3D." % (dimension))
|
98 |
+
|
99 |
+
with tf.variable_scope(name):
|
100 |
+
inputs = transposed_convolution(inputs, output_filters, 2, 2, True, 'transposed_convolution')
|
101 |
+
return inputs
|
trained_models/README.md
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
Checkpoints for the trained model
|
2 |
+
-----
|
3 |
+
Due to the limitation of repo size, we upload the model files to Google Drive. You can manually download them [here](https://drive.google.com/drive/folders/1VYMo1OoaGxoOLNx6-qIt2Wg03lsZw_kA?usp=sharing).
|
utils/evaluation_utils.py
ADDED
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
from scipy.misc import ascent
|
3 |
+
from skimage.measure import compare_psnr, compare_mse, compare_ssim
|
4 |
+
from .predict_utils import normalize_mi_ma
|
5 |
+
|
6 |
+
def normalize(x, pmin=2, pmax=99.8, axis=None, clip=False, eps=1e-20, dtype=np.float32):
|
7 |
+
"""Percentile-based image normalization."""
|
8 |
+
|
9 |
+
mi = np.percentile(x,pmin,axis=axis,keepdims=True)
|
10 |
+
ma = np.percentile(x,pmax,axis=axis,keepdims=True)
|
11 |
+
return normalize_mi_ma(x, mi, ma, clip=clip, eps=eps, dtype=dtype)
|
12 |
+
|
13 |
+
def norm_minmse(gt, x, normalize_gt=True):
|
14 |
+
"""
|
15 |
+
normalizes and affinely scales an image pair such that the MSE is minimized
|
16 |
+
|
17 |
+
Parameters
|
18 |
+
----------
|
19 |
+
gt: ndarray
|
20 |
+
the ground truth image
|
21 |
+
x: ndarray
|
22 |
+
the image that will be affinely scaled
|
23 |
+
normalize_gt: bool
|
24 |
+
set to True of gt image should be normalized (default)
|
25 |
+
Returns
|
26 |
+
-------
|
27 |
+
gt_scaled, x_scaled
|
28 |
+
"""
|
29 |
+
if normalize_gt:
|
30 |
+
gt = normalize(gt, 0.1, 99.9, clip=False).astype(np.float32, copy = False)
|
31 |
+
x = x.astype(np.float32, copy=False) - np.mean(x)
|
32 |
+
gt = gt.astype(np.float32, copy=False) - np.mean(gt)
|
33 |
+
scale = np.cov(x.flatten(), gt.flatten())[0, 1] / np.var(x.flatten())
|
34 |
+
return gt, scale * x
|
35 |
+
|
36 |
+
|
37 |
+
def get_scores(gt, x, multichan=False):
|
38 |
+
|
39 |
+
gt_, x_ = norm_minmse(gt, x)
|
40 |
+
|
41 |
+
mse = compare_mse(gt_, x_)
|
42 |
+
psnr = compare_psnr(gt_, x_, data_range = 1.)
|
43 |
+
ssim = compare_ssim(gt_, x_, data_range = 1., multichannel=multichan)
|
44 |
+
|
45 |
+
return np.sqrt(mse), psnr, ssim
|
46 |
+
|
47 |
+
if __name__ == '__main__':
|
48 |
+
|
49 |
+
# ground truth image
|
50 |
+
y = ascent().astype(np.float32)
|
51 |
+
# input image to compare to
|
52 |
+
x1 = y + 30*np.random.normal(0,1,y.shape)
|
53 |
+
# a scaled and shifted version of x1
|
54 |
+
x2 = 2*x1+100
|
55 |
+
|
56 |
+
# calulate mse, psnr, and ssim of the normalized/scaled images
|
57 |
+
mse1 = compare_mse(*norm_minmse(y, x1))
|
58 |
+
mse2 = compare_mse(*norm_minmse(y, x2))
|
59 |
+
# should be the same
|
60 |
+
print("MSE1 = %.6f\nMSE2 = %.6f"%(mse1, mse2))
|
61 |
+
|
62 |
+
psnr1 = compare_psnr(*norm_minmse(y, x1), data_range = 1.)
|
63 |
+
psnr2 = compare_psnr(*norm_minmse(y, x2), data_range = 1.)
|
64 |
+
# should be the same
|
65 |
+
print("PSNR1 = %.6f\nPSNR2 = %.6f"%(psnr1,psnr2))
|
66 |
+
|
67 |
+
ssim1 = compare_ssim(*norm_minmse(y, x1), data_range = 1.)
|
68 |
+
ssim2 = compare_ssim(*norm_minmse(y, x2), data_range = 1.)
|
69 |
+
# should be the same
|
70 |
+
print("SSIM1 = %.6f\nSSIM2 = %.6f"%(ssim1,ssim2))
|
utils/predict_utils.py
ADDED
@@ -0,0 +1,152 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from __future__ import print_function, unicode_literals, absolute_import, division
|
2 |
+
from six.moves import range, zip, map, reduce, filter
|
3 |
+
|
4 |
+
import collections
|
5 |
+
import warnings
|
6 |
+
import numpy as np
|
7 |
+
|
8 |
+
|
9 |
+
def get_coord(shape, size, margin):
|
10 |
+
n_tiles_i = int(np.ceil((shape[2]-size)/float(size-2*margin)))
|
11 |
+
n_tiles_j = int(np.ceil((shape[1]-size)/float(size-2*margin)))
|
12 |
+
for i in range(n_tiles_i+1):
|
13 |
+
src_start_i = i*(size-2*margin) if i<n_tiles_i else (shape[2]-size)
|
14 |
+
src_end_i = src_start_i+size
|
15 |
+
left_i = margin if i>0 else 0
|
16 |
+
right_i = margin if i<n_tiles_i else 0
|
17 |
+
for j in range(n_tiles_j+1):
|
18 |
+
src_start_j = j*(size-2*margin) if j<n_tiles_j else (shape[1]-size)
|
19 |
+
src_end_j = src_start_j+size
|
20 |
+
left_j = margin if j>0 else 0
|
21 |
+
right_j = margin if j<n_tiles_j else 0
|
22 |
+
src_s = (slice(None, None),
|
23 |
+
slice(src_start_j, src_end_j),
|
24 |
+
slice(src_start_i, src_end_i))
|
25 |
+
|
26 |
+
trg_s = (slice(None, None),
|
27 |
+
slice(src_start_j+left_j, src_end_j-right_j),
|
28 |
+
slice(src_start_i+left_i, src_end_i-right_i))
|
29 |
+
|
30 |
+
mrg_s = (slice(None, None),
|
31 |
+
slice(left_j, -right_j if right_j else None),
|
32 |
+
slice(left_i, -right_i if right_i else None))
|
33 |
+
|
34 |
+
yield src_s, trg_s, mrg_s
|
35 |
+
|
36 |
+
|
37 |
+
# Below implementation of prediction utils inherited from CARE: https://github.com/CSBDeep/CSBDeep
|
38 |
+
# Content-Aware Image Restoration: Pushing the Limits of Fluorescence Microscopy. Martin Weigert, Uwe Schmidt, Tobias Boothe, Andreas Müller, Alexandr Dibrov, Akanksha Jain, Benjamin Wilhelm, Deborah Schmidt, Coleman Broaddus, Siân Culley, Mauricio Rocha-Martins, Fabián Segovia-Miranda, Caren Norden, Ricardo Henriques, Marino Zerial, Michele Solimena, Jochen Rink, Pavel Tomancak, Loic Royer, Florian Jug, and Eugene W. Myers. Nature Methods 15.12 (2018): 1090–1097.
|
39 |
+
|
40 |
+
def _raise(e):
|
41 |
+
raise e
|
42 |
+
|
43 |
+
def consume(iterator):
|
44 |
+
collections.deque(iterator, maxlen=0)
|
45 |
+
|
46 |
+
def axes_check_and_normalize(axes,length=None,disallowed=None,return_allowed=False):
|
47 |
+
"""
|
48 |
+
S(ample), T(ime), C(hannel), Z, Y, X
|
49 |
+
"""
|
50 |
+
allowed = 'STCZYX'
|
51 |
+
axes is not None or _raise(ValueError('axis cannot be None.'))
|
52 |
+
axes = str(axes).upper()
|
53 |
+
consume(a in allowed or _raise(ValueError("invalid axis '%s', must be one of %s."%(a,list(allowed)))) for a in axes)
|
54 |
+
disallowed is None or consume(a not in disallowed or _raise(ValueError("disallowed axis '%s'."%a)) for a in axes)
|
55 |
+
consume(axes.count(a)==1 or _raise(ValueError("axis '%s' occurs more than once."%a)) for a in axes)
|
56 |
+
length is None or len(axes)==length or _raise(ValueError('axes (%s) must be of length %d.' % (axes,length)))
|
57 |
+
return (axes,allowed) if return_allowed else axes
|
58 |
+
|
59 |
+
|
60 |
+
def axes_dict(axes):
|
61 |
+
"""
|
62 |
+
from axes string to dict
|
63 |
+
"""
|
64 |
+
axes, allowed = axes_check_and_normalize(axes,return_allowed=True)
|
65 |
+
return { a: None if axes.find(a) == -1 else axes.find(a) for a in allowed }
|
66 |
+
|
67 |
+
|
68 |
+
def normalize_mi_ma(x, mi, ma, clip=False, eps=1e-20, dtype=np.float32):
|
69 |
+
if dtype is not None:
|
70 |
+
x = x.astype(dtype,copy=False)
|
71 |
+
mi = dtype(mi) if np.isscalar(mi) else mi.astype(dtype,copy=False)
|
72 |
+
ma = dtype(ma) if np.isscalar(ma) else ma.astype(dtype,copy=False)
|
73 |
+
eps = dtype(eps)
|
74 |
+
try:
|
75 |
+
import numexpr
|
76 |
+
x = numexpr.evaluate("(x - mi) / ( ma - mi + eps )")
|
77 |
+
except ImportError:
|
78 |
+
x = (x - mi) / ( ma - mi + eps )
|
79 |
+
if clip:
|
80 |
+
x = np.clip(x,0,1)
|
81 |
+
return x
|
82 |
+
|
83 |
+
class PercentileNormalizer(object):
|
84 |
+
|
85 |
+
def __init__(self, pmin=2, pmax=99.8, do_after=True, dtype=np.float32, **kwargs):
|
86 |
+
|
87 |
+
(np.isscalar(pmin) and np.isscalar(pmax) and 0 <= pmin < pmax <= 100) or _raise(ValueError())
|
88 |
+
self.pmin = pmin
|
89 |
+
self.pmax = pmax
|
90 |
+
self._do_after = do_after
|
91 |
+
self.dtype = dtype
|
92 |
+
self.kwargs = kwargs
|
93 |
+
|
94 |
+
def before(self, img, axes):
|
95 |
+
|
96 |
+
len(axes) == img.ndim or _raise(ValueError())
|
97 |
+
channel = axes_dict(axes)['C']
|
98 |
+
axes = None if channel is None else tuple((d for d in range(img.ndim) if d != channel))
|
99 |
+
self.mi = np.percentile(img,self.pmin,axis=axes,keepdims=True).astype(self.dtype,copy=False)
|
100 |
+
self.ma = np.percentile(img,self.pmax,axis=axes,keepdims=True).astype(self.dtype,copy=False)
|
101 |
+
return normalize_mi_ma(img, self.mi, self.ma, dtype=self.dtype, **self.kwargs)
|
102 |
+
|
103 |
+
def after(self, img):
|
104 |
+
|
105 |
+
self.do_after or _raise(ValueError())
|
106 |
+
alpha = self.ma - self.mi
|
107 |
+
beta = self.mi
|
108 |
+
return ( alpha*img+beta ).astype(self.dtype,copy=False)
|
109 |
+
|
110 |
+
def do_after(self):
|
111 |
+
|
112 |
+
return self._do_after
|
113 |
+
|
114 |
+
|
115 |
+
class PadAndCropResizer(object):
|
116 |
+
|
117 |
+
def __init__(self, mode='reflect', **kwargs):
|
118 |
+
|
119 |
+
self.mode = mode
|
120 |
+
self.kwargs = kwargs
|
121 |
+
|
122 |
+
def _normalize_exclude(self, exclude, n_dim):
|
123 |
+
"""Return normalized list of excluded axes."""
|
124 |
+
if exclude is None:
|
125 |
+
return []
|
126 |
+
exclude_list = [exclude] if np.isscalar(exclude) else list(exclude)
|
127 |
+
exclude_list = [d%n_dim for d in exclude_list]
|
128 |
+
len(exclude_list) == len(np.unique(exclude_list)) or _raise(ValueError())
|
129 |
+
all(( isinstance(d,int) and 0<=d<n_dim for d in exclude_list )) or _raise(ValueError())
|
130 |
+
return exclude_list
|
131 |
+
|
132 |
+
def before(self, x, div_n, exclude):
|
133 |
+
|
134 |
+
def _split(v):
|
135 |
+
a = v // 2
|
136 |
+
return a, v-a
|
137 |
+
exclude = self._normalize_exclude(exclude, x.ndim)
|
138 |
+
self.pad = [_split((div_n-s%div_n)%div_n) if (i not in exclude) else (0,0) for i,s in enumerate(x.shape)]
|
139 |
+
x_pad = np.pad(x, self.pad, mode=self.mode, **self.kwargs)
|
140 |
+
for i in exclude:
|
141 |
+
del self.pad[i]
|
142 |
+
return x_pad
|
143 |
+
|
144 |
+
def after(self, x, exclude):
|
145 |
+
|
146 |
+
pads = self.pad[:len(x.shape)]
|
147 |
+
crop = [slice(p[0], -p[1] if p[1]>0 else None) for p in self.pad]
|
148 |
+
for i in self._normalize_exclude(exclude, x.ndim):
|
149 |
+
crop.insert(i,slice(None))
|
150 |
+
len(crop) == x.ndim or _raise(ValueError())
|
151 |
+
return x[tuple(crop)]
|
152 |
+
|
utils/train_utils.py
ADDED
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
from tqdm import tqdm
|
3 |
+
|
4 |
+
|
5 |
+
def augment_patch(patch):
|
6 |
+
if len(patch.shape[:-1]) == 2:
|
7 |
+
patch = np.rot90(patch, k=np.random.randint(4), axes=(0, 1))
|
8 |
+
elif len(patch.shape[:-1]) == 3:
|
9 |
+
patch = np.rot90(patch, k=np.random.randint(4), axes=(1, 2))
|
10 |
+
|
11 |
+
patch = np.flip(patch, axis=-2) if np.random.randint(2) else patch
|
12 |
+
return patch
|
13 |
+
|
14 |
+
|
15 |
+
# Below implementation of stratified sampling inherited from Noise2Void: https://github.com/juglab/n2v
|
16 |
+
# Noise2void: learning denoising from single noisy images. Krull, Alexander, Tim-Oliver Buchholz, and Florian Jug. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019.
|
17 |
+
|
18 |
+
def get_stratified_coords2D(coord_gen, box_size, shape):
|
19 |
+
box_count_y = int(np.ceil(shape[0] / box_size))
|
20 |
+
box_count_x = int(np.ceil(shape[1] / box_size))
|
21 |
+
x_coords = []
|
22 |
+
y_coords = []
|
23 |
+
for i in range(box_count_y):
|
24 |
+
for j in range(box_count_x):
|
25 |
+
y, x = next(coord_gen)
|
26 |
+
y = int(i * box_size + y)
|
27 |
+
x = int(j * box_size + x)
|
28 |
+
if (y < shape[0] and x < shape[1]):
|
29 |
+
y_coords.append(y)
|
30 |
+
x_coords.append(x)
|
31 |
+
return (y_coords, x_coords)
|
32 |
+
|
33 |
+
|
34 |
+
def get_stratified_coords3D(coord_gen, box_size, shape):
|
35 |
+
box_count_z = int(np.ceil(shape[0] / box_size))
|
36 |
+
box_count_y = int(np.ceil(shape[1] / box_size))
|
37 |
+
box_count_x = int(np.ceil(shape[2] / box_size))
|
38 |
+
x_coords = []
|
39 |
+
y_coords = []
|
40 |
+
z_coords = []
|
41 |
+
for i in range(box_count_z):
|
42 |
+
for j in range(box_count_y):
|
43 |
+
for k in range(box_count_x):
|
44 |
+
z, y, x = next(coord_gen)
|
45 |
+
z = int(i * box_size + z)
|
46 |
+
y = int(j * box_size + y)
|
47 |
+
x = int(k * box_size + x)
|
48 |
+
if (z < shape[0] and y < shape[1] and x < shape[2]):
|
49 |
+
z_coords.append(z)
|
50 |
+
y_coords.append(y)
|
51 |
+
x_coords.append(x)
|
52 |
+
return (z_coords, y_coords, x_coords)
|
53 |
+
|
54 |
+
def rand_float_coords2D(boxsize):
|
55 |
+
while True:
|
56 |
+
yield (np.random.rand() * boxsize, np.random.rand() * boxsize)
|
57 |
+
|
58 |
+
def rand_float_coords3D(boxsize):
|
59 |
+
while True:
|
60 |
+
yield (np.random.rand() * boxsize, np.random.rand() * boxsize, np.random.rand() * boxsize)
|