forked from divelab/Noise2Same
/
models.py
276 lines (226 loc) · 14.5 KB
/
models.py
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
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
import os, cv2
import numpy as np
from network_configure import conf_unet
from network import *
from utils.predict_utils import get_coord, PercentileNormalizer, PadAndCropResizer
from utils.train_utils import augment_patch
from utils import train_utils
# UNet implementation inherited from GVTNets: https://github.com/zhengyang-wang/GVTNets
training_config = {'base_learning_rate': 0.0004,
'lr_decay_steps':5e3,
'lr_decay_rate':0.5,
'lr_staircase':True}
class Noise2Same(object):
def __init__(self, base_dir, name,
dim=2, in_channels=1, lmbd=None,
masking='gaussian', mask_perc=0.5,
opt_config=training_config, **kwargs):
self.base_dir = base_dir # model direction
self.name = name # model name
self.dim = dim # image dimension
self.in_channels = in_channels # image channels
self.lmbd = lmbd # lambda in loss fn
self.masking = masking
self.mask_perc = mask_perc
self.opt_config = opt_config
conf_unet['dimension'] = '%dD'%dim
self.net = UNet(conf_unet)
def _model_fn(self, features, labels, mode):
conv_op = convolution_2D if self.dim==2 else convolution_3D
axis = {3:[1,2,3,4], 2:[1,2,3]}[self.dim]
def image_summary(img):
return tf.reduce_max(img, axis=1) if self.dim == 3 else img
# Local average excluding the center pixel (donut)
def mask_kernel(features):
kernel = (np.array([[0.5, 1.0, 0.5], [1.0, 0.0, 1.0], [0.5, 1.0, 0.5]])
if self.dim == 2 else
np.array([[[0, 0.5, 0], [0.5, 1.0, 0.5], [0, 0.5, 0]],
[[0.5, 1.0, 0.5], [1.0, 0.0, 1.0], [0.5, 1.0, 0.5]],
[[0, 0.5, 0], [0.5, 1.0, 0.5], [0, 0.5, 0]]]))
kernel = (kernel/kernel.sum())
kernels = np.empty([3, 3, self.in_channels, self.in_channels])
for i in range(self.in_channels):
kernels[:,:,i,i] = kernel
nn_conv_op = tf.nn.conv2d if self.dim == 2 else tf.nn.conv3d
return nn_conv_op(features, tf.constant(kernels.astype('float32')),
[1]*self.dim+[1,1], padding='SAME')
if not mode == tf.estimator.ModeKeys.PREDICT:
noise, mask = tf.split(labels, [self.in_channels, self.in_channels], -1)
if self.masking == 'gaussian':
masked_features = (1 - mask) * features + mask * noise
elif self.masking == 'donut':
masked_features = (1 - mask) * features + mask * mask_kernel(features)
else:
raise NotImplementedError
# Prediction from masked input
with tf.variable_scope('main_unet', reuse=tf.compat.v1.AUTO_REUSE):
out = self.net(masked_features, mode == tf.estimator.ModeKeys.TRAIN)
out = batch_norm(out, mode == tf.estimator.ModeKeys.TRAIN, 'unet_out')
out = relu(out)
preds = conv_op(out, self.in_channels, 1, 1, False, name = 'out_conv')
# Prediction from full input
with tf.variable_scope('main_unet', reuse=tf.compat.v1.AUTO_REUSE):
rawout = self.net(features, mode == tf.estimator.ModeKeys.TRAIN)
rawout = batch_norm(rawout, mode == tf.estimator.ModeKeys.TRAIN, 'unet_out')
rawout = relu(rawout)
rawpreds = conv_op(rawout, self.in_channels, 1, 1, False, name = 'out_conv')
# Loss components
rec_mse = tf.reduce_mean(tf.square(rawpreds - features), axis=None)
inv_mse = tf.reduce_sum(tf.square(rawpreds - preds) * mask) / tf.reduce_sum(mask)
bsp_mse = tf.reduce_sum(tf.square(features - preds) * mask) / tf.reduce_sum(mask)
# Tensorboard display
tf.summary.image('1_inputs', image_summary(features), max_outputs=3)
tf.summary.image('2_raw_predictions', image_summary(rawpreds), max_outputs=3)
tf.summary.image('3_mask', image_summary(mask), max_outputs=3)
tf.summary.image('4_masked_predictions', image_summary(preds), max_outputs=3)
tf.summary.image('5_difference', image_summary(rawpreds-preds), max_outputs=3)
tf.summary.image('6_rec_error', image_summary(preds-features), max_outputs=3)
tf.summary.scalar('reconstruction', rec_mse, family='loss_metric')
tf.summary.scalar('invariance', inv_mse, family='loss_metric')
tf.summary.scalar('blind_spot', bsp_mse, family='loss_metric')
else:
with tf.variable_scope('main_unet'):
out = self.net(features, mode == tf.estimator.ModeKeys.TRAIN)
out = batch_norm(out, mode == tf.estimator.ModeKeys.TRAIN, 'unet_out')
out = relu(out)
preds = conv_op(out, self.in_channels, 1, 1, False, name = 'out_conv')
return tf.estimator.EstimatorSpec(mode=mode, predictions=preds)
lmbd = 2 if self.lmbd is None else self.lmbd
loss = rec_mse + lmbd*tf.sqrt(inv_mse)
if mode == tf.estimator.ModeKeys.TRAIN:
global_step = tf.train.get_or_create_global_step()
learning_rate = tf.train.exponential_decay(self.opt_config['base_learning_rate'],
global_step,
self.opt_config['lr_decay_steps'],
self.opt_config['lr_decay_rate'],
self.opt_config['lr_staircase'])
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope='main_unet')
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(loss, global_step)
else:
train_op = None
metrics = {'loss_metric/invariance':tf.metrics.mean(inv_mse),
'loss_metric/blind_spot':tf.metrics.mean(bsp_mse),
'loss_metric/reconstruction':tf.metrics.mean(rec_mse)}
return tf.estimator.EstimatorSpec(mode=mode, predictions=preds, loss=loss, train_op=train_op,
eval_metric_ops=metrics)
def _input_fn(self, sources, patch_size, batch_size, is_train=True):
# Stratified sampling inherited from Noise2Void: https://github.com/juglab/n2v
get_stratified_coords = getattr(train_utils, 'get_stratified_coords%dD'%self.dim)
rand_float_coords = getattr(train_utils, 'rand_float_coords%dD'%self.dim)
def generator():
while(True):
source = sources[np.random.randint(len(sources))]
valid_shape = source.shape[:-1] - np.array(patch_size)
if any([s<=0 for s in valid_shape]):
source_patch = augment_patch(source)
else:
coords = [np.random.randint(0, shape_i+1) for shape_i in valid_shape]
s = tuple([slice(coord, coord+size) for coord, size in zip(coords, patch_size)])
source_patch = augment_patch(source[s])
mask = np.zeros_like(source_patch)
for c in range(self.in_channels):
boxsize = np.round(np.sqrt(100/self.mask_perc)).astype(np.int)
maskcoords = get_stratified_coords(rand_float_coords(boxsize),
box_size=boxsize, shape=tuple(patch_size))
indexing = maskcoords + (c,)
mask[indexing] = 1.0
noise_patch = np.concatenate([np.random.normal(0, 0.2, source_patch.shape), mask], axis=-1)
yield source_patch, noise_patch
def generator_val():
for idx in range(len(sources)):
source_patch = sources[idx]
patch_size = source_patch.shape[:-1]
boxsize = np.round(np.sqrt(100/self.mask_perc)).astype(np.int)
maskcoords = get_stratified_coords(rand_float_coords(boxsize),
box_size=boxsize, shape=tuple(patch_size))
indexing = maskcoords + (0,)
mask = np.zeros_like(source_patch)
mask[indexing] = 1.0
noise_patch = np.concatenate([np.random.normal(0, 0.2, source_patch.shape), mask], axis=-1)
yield source_patch, noise_patch
output_types = (tf.float32, tf.float32)
output_shapes = (tf.TensorShape(list(patch_size) + [self.in_channels]),
tf.TensorShape(list(patch_size) + [self.in_channels*2]))
gen = generator if is_train else generator_val
dataset = tf.data.Dataset.from_generator(gen, output_types=output_types, output_shapes=output_shapes)
dataset = dataset.batch(batch_size).prefetch(tf.data.experimental.AUTOTUNE)
return dataset
def train(self, source_lst, patch_size, validation=None, batch_size=64, save_steps=1000, log_steps=200, steps=50000):
assert len(patch_size)==self.dim
assert len(source_lst[0].shape)==self.dim + 1
assert source_lst[0].shape[-1]==self.in_channels
ses_config = tf.ConfigProto()
ses_config.gpu_options.allow_growth = True
run_config = tf.estimator.RunConfig(model_dir=self.base_dir+'/'+self.name,
save_checkpoints_steps=save_steps,
session_config=ses_config,
log_step_count_steps=log_steps,
save_summary_steps=log_steps,
keep_checkpoint_max=2)
estimator = tf.estimator.Estimator(model_fn=self._model_fn,
model_dir=self.base_dir+'/'+self.name,
config=run_config)
input_fn = lambda: self._input_fn(source_lst, patch_size, batch_size=batch_size)
if validation is not None:
train_spec = tf.estimator.TrainSpec(input_fn=input_fn, max_steps=steps)
val_input_fn = lambda: self._input_fn(validation.astype('float32'),
validation.shape[1:-1],
batch_size=4,
is_train=False)
eval_spec = tf.estimator.EvalSpec(input_fn=val_input_fn, throttle_secs=120)
tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)
else:
estimator.train(input_fn=input_fn, steps=steps)
# Used for single image prediction
def predict(self, image, resizer=PadAndCropResizer(), checkpoint_path=None,
im_mean=None, im_std=None):
tf.logging.set_verbosity(tf.logging.ERROR)
estimator = tf.estimator.Estimator(model_fn=self._model_fn,
model_dir=self.base_dir+'/'+self.name)
im_mean, im_std = ((image.mean(), image.std()) if im_mean is None or im_std is None else (im_mean, im_std))
image = (image - im_mean)/im_std
if self.in_channels == 1:
image = resizer.before(image, 2 ** (self.net.depth), exclude=None)
input_fn = tf.estimator.inputs.numpy_input_fn(x=image[None, ..., None], batch_size=1, num_epochs=1, shuffle=False)
image = list(estimator.predict(input_fn=input_fn, checkpoint_path=checkpoint_path))[0][..., 0]
image = resizer.after(image, exclude=None)
else:
image = resizer.before(image, 2 ** (self.net.depth), exclude=-1)
input_fn = tf.estimator.inputs.numpy_input_fn(x=image[None], batch_size=1, num_epochs=1, shuffle=False)
image = list(estimator.predict(input_fn=input_fn, checkpoint_path=checkpoint_path))[0]
image = resizer.after(image, exclude=-1)
image = image*im_std + im_mean
return image
# Used for batch images prediction
def batch_predict(self, images, resizer=PadAndCropResizer(), checkpoint_path=None,
im_mean=None, im_std=None, batch_size=32):
tf.logging.set_verbosity(tf.logging.ERROR)
estimator = tf.estimator.Estimator(model_fn=self._model_fn,
model_dir=self.base_dir+'/'+self.name)
im_mean, im_std = ((images.mean(), images.std()) if im_mean is None or im_std is None else (im_mean, im_std))
images = (images - im_mean)/im_std
images = resizer.before(images, 2 ** (self.net.depth), exclude=0)
input_fn = tf.estimator.inputs.numpy_input_fn(x=images[ ..., None], batch_size=batch_size, num_epochs=1, shuffle=False)
images = np.stack(list(estimator.predict(input_fn=input_fn, checkpoint_path=checkpoint_path)))[..., 0]
images = resizer.after(images, exclude=0)
images = images*im_std + im_mean
return images
# Used for extremely large input images
def crop_predict(self, image, size, margin, resizer=PadAndCropResizer(), checkpoint_path=None,
im_mean=None, im_std=None):
tf.logging.set_verbosity(tf.logging.ERROR)
estimator = tf.estimator.Estimator(model_fn=self._model_fn,
model_dir=self.base_dir+'/'+self.name)
im_mean, im_std = ((image.mean(), image.std()) if im_mean is None or im_std is None else (im_mean, im_std))
image = (image - im_mean)/im_std
out_image = np.empty(image.shape, dtype='float32')
for src_s, trg_s, mrg_s in get_coord(image.shape, size, margin):
patch = resizer.before(image[src_s], 2 ** (self.net.depth), exclude=None)
input_fn = tf.estimator.inputs.numpy_input_fn(x=patch[None, ..., None], batch_size=1, num_epochs=1, shuffle=False)
patch = list(estimator.predict(input_fn=input_fn, checkpoint_path=checkpoint_path))[0][..., 0]
patch = resizer.after(patch, exclude=None)
out_image[trg_s] = patch[mrg_s]
image = out_image*im_std + im_mean
return image