-
Notifications
You must be signed in to change notification settings - Fork 18
/
createDeepBackProjectionNetworkModel.R
432 lines (381 loc) · 19.1 KB
/
createDeepBackProjectionNetworkModel.R
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
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
#' 2-D implementation of the deep back-projection network.
#'
#' Creates a keras model of the deep back-project network for image super
#' resolution. More information is provided at the authors' website:
#'
#' \url{https://www.toyota-ti.ac.jp/Lab/Denshi/iim/members/muhammad.haris/projects/DBPN.html}
#'
#' with the paper available here:
#'
#' \url{https://arxiv.org/abs/1803.02735}
#'
#' This particular implementation was influenced by the following keras (python)
#' implementation:
#'
#' \url{https://github.com/rajatkb/DBPN-Keras}
#'
#' with help from the original author's Caffe and Pytorch implementations:
#'
#' \url{https://github.com/alterzero/DBPN-caffe}
#' \url{https://github.com/alterzero/DBPN-Pytorch}
#'
#' @param inputImageSize Used for specifying the input tensor shape. The
#' shape (or dimension) of that tensor is the image dimensions followed by
#' the number of channels (e.g., red, green, and blue).
#' @param numberOfOutputs number of outputs (e.g., 3 for RGB images).
#' @param numberOfFeatureFilters number of feature filters.
#' @param numberOfBaseFilters number of base filters.
#' @param numberOfBackProjectionStages number of up-/down-projection stages. This
#' number includes the final up block.
#' @param convolutionKernelSize kernel size for certain convolutional layers. This
#' and \code{strides} are dependent on the scale factor discussed in
#' original paper. Factors used in the original implementation are as follows:
#' 2x --> \code{convolutionKernelSize = c( 6, 6 )},
#' 4x --> \code{convolutionKernelSize = c( 8, 8 )},
#' 8x --> \code{convolutionKernelSize = c( 12, 12 )}. We default to 8x parameters.
#' @param strides strides for certain convolutional layers. This and the
#' \code{convolutionKernelSize} are dependent on the scale factor discussed in
#' original paper. Factors used in the original implementation are as follows:
#' 2x --> \code{strides = c( 2, 2 )}, 4x --> \code{strides = c( 4, 4 )}, 8x -->
#' \code{strides = c( 8, 8 )}. We default to 8x parameters.
#' @param lastConvolution the kernel size for the last convolutional layer
#' @param numberOfLossFunctions the number of data targets, e.g. 2 for 2 targets
#'
#' @return a keras model defining the deep back-projection network.
#' @author Tustison NJ
#' @examples
#' model = createDeepBackProjectionNetworkModel2D(c(25, 25, 1))
#' rm(model); gc()
#' @import keras
#' @export
createDeepBackProjectionNetworkModel2D <-
function(
inputImageSize,
numberOfOutputs = 1,
numberOfBaseFilters = 64,
numberOfFeatureFilters = 256,
numberOfBackProjectionStages = 7,
convolutionKernelSize = c( 12, 12 ),
strides = c( 8, 8 ),
lastConvolution = c( 3, 3 ),
numberOfLossFunctions = 1
)
{
upBlock2D <- function( L, numberOfFilters = 64, kernelSize = c( 12, 12 ),
strides = c( 8, 8 ), includeDenseConvolutionLayer = TRUE )
{
if( includeDenseConvolutionLayer )
{
L <- L %>% layer_conv_2d( filters = numberOfFilters, use_bias = TRUE,
kernel_size = c( 1, 1 ), strides = c( 1, 1 ), padding = 'same' )
L <- L %>% layer_activation_parametric_relu(
alpha_initializer = 'zero', shared_axes = c( 1, 2 ) )
}
# Scale up
H0 <- L %>% layer_conv_2d_transpose( filters = numberOfFilters,
kernel_size = kernelSize, strides = strides,
kernel_initializer = 'glorot_uniform', padding = 'same' )
H0 <- H0 %>% layer_activation_parametric_relu(
alpha_initializer = 'zero', shared_axes = c( 1, 2 ) )
# Scale down
L0 <- H0 %>% layer_conv_2d( filters = numberOfFilters,
kernel_size = kernelSize, strides = strides,
kernel_initializer = 'glorot_uniform', padding = 'same' )
L0 <- L0 %>% layer_activation_parametric_relu(
alpha_initializer = 'zero', shared_axes = c( 1, 2 ) )
# Residual
E <- layer_subtract( list( L0, L ) )
# Scale residual up
H1 <- E %>% layer_conv_2d_transpose( filters = numberOfFilters,
kernel_size = kernelSize, strides = strides,
kernel_initializer = 'glorot_uniform', padding = 'same' )
H1 <- H1 %>% layer_activation_parametric_relu(
alpha_initializer = 'zero', shared_axes = c( 1, 2 ) )
# Output feature map
upBlock <- layer_add( list( H0, H1 ) )
return( upBlock )
}
downBlock2D <- function( H, numberOfFilters = 64, kernelSize = c( 12, 12 ),
strides = c( 8, 8 ), includeDenseConvolutionLayer = TRUE )
{
if( includeDenseConvolutionLayer )
{
H <- H %>% layer_conv_2d( filters = numberOfFilters, use_bias = TRUE,
kernel_size = c( 1, 1 ), strides = c( 1, 1 ), padding = 'same' )
H <- H %>% layer_activation_parametric_relu(
alpha_initializer = 'zero', shared_axes = c( 1, 2 ) )
}
# Scale down
L0 <- H %>% layer_conv_2d( filters = numberOfFilters,
kernel_size = kernelSize, strides = strides,
kernel_initializer = 'glorot_uniform', padding = 'same' )
L0 <- L0 %>% layer_activation_parametric_relu(
alpha_initializer = 'zero', shared_axes = c( 1, 2 ) )
# Scale up
H0 <- L0 %>% layer_conv_2d_transpose( filters = numberOfFilters,
kernel_size = kernelSize, strides = strides,
kernel_initializer = 'glorot_uniform', padding = 'same' )
H0 <- H0 %>% layer_activation_parametric_relu(
alpha_initializer = 'zero', shared_axes = c( 1, 2 ) )
# Residual
E <- layer_subtract( list( H0, H ) )
# Scale residual down
L1 <- E %>% layer_conv_2d( filters = numberOfFilters,
kernel_size = kernelSize, strides = strides,
kernel_initializer = 'glorot_uniform', padding = 'same' )
L1 <- L1 %>% layer_activation_parametric_relu(
alpha_initializer = 'zero', shared_axes = c( 1, 2 ) )
# Output feature map
downBlock <- layer_add( list( L0, L1 ) )
return( downBlock )
}
inputs <- layer_input( shape = inputImageSize )
# Initial feature extraction
model <- inputs %>% layer_conv_2d( filters = numberOfFeatureFilters,
kernel_size = c( 3, 3 ), strides = c( 1, 1 ), padding = 'same',
kernel_initializer = "glorot_uniform" )
model <- model %>% layer_activation_parametric_relu( alpha_initializer = 'zero',
shared_axes = c( 1, 2 ) )
# Feature smashing
model <- model %>% layer_conv_2d( filters = numberOfBaseFilters,
kernel_size = c( 1, 1 ), strides = c( 1, 1 ), padding = 'same',
kernel_initializer = "glorot_uniform" )
model <- model %>% layer_activation_parametric_relu( alpha_initializer = 'zero',
shared_axes = c( 1, 2 ) )
# Back projection
upProjectionBlocks <- list()
downProjectionBlocks <- list()
model <- upBlock2D( model, numberOfFilters = numberOfBaseFilters,
kernelSize = convolutionKernelSize, strides = strides )
upProjectionBlocks[[1]] <- model
for( i in seq_len( numberOfBackProjectionStages ) )
{
if( i == 1 )
{
model <- downBlock2D( model, numberOfFilters = numberOfBaseFilters,
kernelSize = convolutionKernelSize, strides = strides )
downProjectionBlocks[[i]] <- model
model <- upBlock2D( model, numberOfFilters = numberOfBaseFilters,
kernelSize = convolutionKernelSize, strides = strides )
upProjectionBlocks[[i+1]] <- model
model <- layer_concatenate( upProjectionBlocks, trainable = TRUE )
} else {
model <- downBlock2D( model, numberOfFilters = numberOfBaseFilters,
kernelSize = convolutionKernelSize, strides = strides,
includeDenseConvolutionLayer = TRUE )
downProjectionBlocks[[i]] <- model
model <- layer_concatenate( downProjectionBlocks, trainable = TRUE )
model <- upBlock2D( model, numberOfFilters = numberOfBaseFilters,
kernelSize = convolutionKernelSize, strides = strides,
includeDenseConvolutionLayer = TRUE )
upProjectionBlocks[[i+1]] <- model
model <- layer_concatenate( upProjectionBlocks, trainable = TRUE )
}
}
# Final convolution layer
outputs <- model %>% layer_conv_2d( filters = numberOfOutputs,
kernel_size = lastConvolution, strides = c( 1, 1 ), padding = 'same',
kernel_initializer = "glorot_uniform" )
if( numberOfLossFunctions == 1 )
{
deepBackProjectionNetworkModel <- keras_model( inputs = inputs, outputs = outputs )
} else {
outputList = list()
for( k in seq_len( numberOfLossFunctions ) )
{
outputList[[k]] = outputs
}
deepBackProjectionNetworkModel <- keras_model( inputs = inputs, outputs = outputList )
}
return( deepBackProjectionNetworkModel )
}
#' 3-D implementation of the deep back-projection network.
#'
#' Creates a keras model of the deep back-project network for image super
#' resolution. More information is provided at the authors' website:
#'
#' \url{https://www.toyota-ti.ac.jp/Lab/Denshi/iim/members/muhammad.haris/projects/DBPN.html}
#'
#' with the paper available here:
#'
#' \url{https://arxiv.org/abs/1803.02735}
#'
#' This particular implementation was influenced by the following keras (python)
#' implementation:
#'
#' \url{https://github.com/rajatkb/DBPN-Keras}
#'
#' with help from the original author's Caffe and Pytorch implementations:
#'
#' \url{https://github.com/alterzero/DBPN-caffe}
#' \url{https://github.com/alterzero/DBPN-Pytorch}
#'
#' @param inputImageSize Used for specifying the input tensor shape. The
#' shape (or dimension) of that tensor is the image dimensions followed by
#' the number of channels (e.g., red, green, and blue).
#' @param numberOfOutputs number of outputs (e.g., 3 for RGB images).
#' @param numberOfFeatureFilters number of feature filters.
#' @param numberOfBaseFilters number of base filters.
#' @param numberOfBackProjectionStages number of up-/down-projection stages. This
#' number includes the final up block.
#' @param convolutionKernelSize kernel size for certain convolutional layers. This
#' and \code{strides} are dependent on the scale factor discussed in
#' original paper. Factors used in the original implementation are as follows:
#' 2x --> \code{convolutionKernelSize = c( 6, 6, 6 )},
#' 4x --> \code{convolutionKernelSize = c( 8, 8, 8 )},
#' 8x --> \code{convolutionKernelSize = c( 12, 12, 12 )}. We default to 8x parameters.
#' @param strides strides for certain convolutional layers. This and the
#' \code{convolutionKernelSize} are dependent on the scale factor discussed in
#' original paper. Factors used in the original implementation are as follows:
#' 2x --> \code{strides = c( 2, 2, 2 )}, 4x --> \code{strides = c( 4, 4, 4 )},
#' 8x --> \code{strides = c( 8, 8, 8 )}. We default to 8x parameters.
#' @param lastConvolution the kernel size for the last convolutional layer
#' @param numberOfLossFunctions the number of data targets, e.g. 2 for 2 targets
#'
#' @return a keras model defining the deep back-projection network.
#' @author Tustison NJ
#' @examples
#' model = createDeepBackProjectionNetworkModel3D(c(25, 25, 25, 1))
#' rm(model); gc()
#' @import keras
#' @export
createDeepBackProjectionNetworkModel3D <-
function( inputImageSize,
numberOfOutputs = 1,
numberOfBaseFilters = 64,
numberOfFeatureFilters = 256,
numberOfBackProjectionStages = 7,
convolutionKernelSize = c( 12, 12, 12 ),
strides = c( 8, 8, 8 ),
lastConvolution = c( 3, 3, 3 ),
numberOfLossFunctions = 1
)
{
upBlock3D <- function( L, numberOfFilters = 64, kernelSize = c( 12, 12, 12 ),
strides = c( 8, 8, 8 ), includeDenseConvolutionLayer = TRUE )
{
if( includeDenseConvolutionLayer )
{
L <- L %>% layer_conv_3d( filters = numberOfFilters, use_bias = TRUE,
kernel_size = c( 1, 1, 1 ), strides = c( 1, 1, 1 ), padding = 'same' )
L <- L %>% layer_activation_parametric_relu(
alpha_initializer = 'zero', shared_axes = c( 1, 2, 3 ) )
}
# Scale up
H0 <- L %>% layer_conv_3d_transpose( filters = numberOfFilters,
kernel_size = kernelSize, strides = strides,
kernel_initializer = 'glorot_uniform', padding = 'same' )
H0 <- H0 %>% layer_activation_parametric_relu(
alpha_initializer = 'zero', shared_axes = c( 1, 2, 3 ) )
# Scale down
L0 <- H0 %>% layer_conv_3d( filters = numberOfFilters,
kernel_size = kernelSize, strides = strides,
kernel_initializer = 'glorot_uniform', padding = 'same' )
L0 <- L0 %>% layer_activation_parametric_relu(
alpha_initializer = 'zero', shared_axes = c( 1, 2, 3 ) )
# Residual
E <- layer_subtract( list( L0, L ) )
# Scale residual up
H1 <- E %>% layer_conv_3d_transpose( filters = numberOfFilters,
kernel_size = kernelSize, strides = strides,
kernel_initializer = 'glorot_uniform', padding = 'same' )
H1 <- H1 %>% layer_activation_parametric_relu(
alpha_initializer = 'zero', shared_axes = c( 1, 2, 3 ) )
# Output feature map
upBlock <- layer_add( list( H0, H1 ) )
return( upBlock )
}
downBlock3D <- function( H, numberOfFilters = 64, kernelSize = c( 12, 12, 12 ),
strides = c( 8, 8, 8 ), includeDenseConvolutionLayer = TRUE )
{
if( includeDenseConvolutionLayer )
{
H <- H %>% layer_conv_3d( filters = numberOfFilters, use_bias = TRUE,
kernel_size = c( 1, 1, 1 ), strides = c( 1, 1, 1 ), padding = 'same' )
H <- H %>% layer_activation_parametric_relu(
alpha_initializer = 'zero', shared_axes = c( 1, 2, 3 ) )
}
# Scale down
L0 <- H %>% layer_conv_3d( filters = numberOfFilters,
kernel_size = kernelSize, strides = strides,
kernel_initializer = 'glorot_uniform', padding = 'same' )
L0 <- L0 %>% layer_activation_parametric_relu(
alpha_initializer = 'zero', shared_axes = c( 1, 2, 3 ) )
# Scale up
H0 <- L0 %>% layer_conv_3d_transpose( filters = numberOfFilters,
kernel_size = kernelSize, strides = strides,
kernel_initializer = 'glorot_uniform', padding = 'same' )
H0 <- H0 %>% layer_activation_parametric_relu(
alpha_initializer = 'zero', shared_axes = c( 1, 2, 3 ) )
# Residual
E <- layer_subtract( list( H0, H ) )
# Scale residual down
L1 <- E %>% layer_conv_3d( filters = numberOfFilters,
kernel_size = kernelSize, strides = strides,
kernel_initializer = 'glorot_uniform', padding = 'same' )
L1 <- L1 %>% layer_activation_parametric_relu(
alpha_initializer = 'zero', shared_axes = c( 1, 2, 3 ) )
# Output feature map
downBlock <- layer_add( list( L0, L1 ) )
return( downBlock )
}
inputs <- layer_input( shape = inputImageSize )
# Initial feature extraction
model <- inputs %>% layer_conv_3d( filters = numberOfFeatureFilters,
kernel_size = c( 3, 3, 3 ), strides = c( 1, 1, 1 ), padding = 'same',
kernel_initializer = "glorot_uniform" )
model <- model %>% layer_activation_parametric_relu( alpha_initializer = 'zero',
shared_axes = c( 1, 2, 3 ) )
# Feature smashing
model <- model %>% layer_conv_3d( filters = numberOfBaseFilters,
kernel_size = c( 1, 1, 1 ), strides = c( 1, 1, 1 ), padding = 'same',
kernel_initializer = "glorot_uniform" )
model <- model %>% layer_activation_parametric_relu( alpha_initializer = 'zero',
shared_axes = c( 1, 2, 3 ) )
# Back projection
upProjectionBlocks <- list()
downProjectionBlocks <- list()
model <- upBlock3D( model, numberOfFilters = numberOfBaseFilters,
kernelSize = convolutionKernelSize, strides = strides )
upProjectionBlocks[[1]] <- model
for( i in seq_len( numberOfBackProjectionStages ) )
{
if( i == 1 )
{
model <- downBlock3D( model, numberOfFilters = numberOfBaseFilters,
kernelSize = convolutionKernelSize, strides = strides )
downProjectionBlocks[[i]] <- model
model <- upBlock3D( model, numberOfFilters = numberOfBaseFilters,
kernelSize = convolutionKernelSize, strides = strides )
upProjectionBlocks[[i+1]] <- model
model <- layer_concatenate( upProjectionBlocks, trainable = TRUE )
} else {
model <- downBlock3D( model, numberOfFilters = numberOfBaseFilters,
kernelSize = convolutionKernelSize, strides = strides,
includeDenseConvolutionLayer = TRUE )
downProjectionBlocks[[i]] <- model
model <- layer_concatenate( downProjectionBlocks, trainable = TRUE )
model <- upBlock3D( model, numberOfFilters = numberOfBaseFilters,
kernelSize = convolutionKernelSize, strides = strides,
includeDenseConvolutionLayer = TRUE )
upProjectionBlocks[[i+1]] <- model
model <- layer_concatenate( upProjectionBlocks, trainable = TRUE )
}
}
# Final convolution layer
outputs <- model %>% layer_conv_3d( filters = numberOfOutputs,
kernel_size = lastConvolution, strides = c( 1, 1, 1 ), padding = 'same',
kernel_initializer = "glorot_uniform" )
if( numberOfLossFunctions == 1 )
{
deepBackProjectionNetworkModel <- keras_model( inputs = inputs, outputs = outputs )
} else {
outputList = list()
for( k in seq_len( numberOfLossFunctions ) )
{
outputList[[k]] = outputs
}
deepBackProjectionNetworkModel <- keras_model( inputs = inputs, outputs = outputList )
}
return( deepBackProjectionNetworkModel )
}