def build_network(self):
'''
This method builds the network architecture for the CNN model.
Modify as needed.
'''
# Image Preprocessing
img_preprocess = DataPreprocessing()
img_preprocess.add_samplewise_zero_center()
img_preprocess.add_featurewise_stdnorm()
# Image Augmentation
img_aug = ImageAugmentation()
img_aug.add_random_flip_leftright()
img_aug.add_random_rotation(max_angle=5.0)
# Input Layer
self.network = input_data(shape=[None, IMG_SIZE, IMG_SIZE, 1],
data_preprocessing=img_preprocess,
data_augmentation=img_aug)
# Convolution Layer 1
self.network = conv_2d(self.network, 64, 5, activation='relu')
self.network = batch_normalization(self.network)
self.network = max_pool_2d(self.network, 3, strides=2)
# Convolution Layer 2
self.network = conv_2d(self.network, 64, 5, activation='relu')
self.network = batch_normalization(self.network)
self.network = max_pool_2d(self.network, 3, strides=2)
# Convolution Layer 3
self.network = conv_2d(self.network, 128, 4, activation='relu')
self.network = batch_normalization(self.network)
self.network = dropout(self.network, 0.2)
# Penultimate FC Layer
self.network = fully_connected(self.network, 3072, activation='relu')
self.network = batch_normalization(self.network)
# Final FC Layer
self.network = fully_connected(self.network,
len(EMOTIONS),
activation='softmax')
# Create network
optimizer = Momentum(learning_rate=0.01, lr_decay=0.99,
decay_step=250) # Learning function
self.network = regression(self.network,
optimizer=optimizer,
loss='categorical_crossentropy')
# Create model
self.model = tflearn.DNN(self.network,
tensorboard_dir=TENSORBOARD_PATH,
checkpoint_path=CHECKPOINT_PATH,
max_checkpoints=1,
tensorboard_verbose=1)
def cnn(self):
images = input_data(shape=[None, 48, 48, 1], name='input')
images = conv_2d(images, 64, 3, padding='same', activation=self.activation)
images = batch_normalization(images)
#max pooling
images = max_pool_2d(images, 3, strides=2)
images = conv_2d(images, 128, 3, padding='same', activation=self.activation)
images = max_pool_2d(images, 3, strides=2)
images = conv_2d(images, 256, 3, padding='same', activation=self.activation)
#maxpooling
images = max_pool_2d(images, 3, strides=2)
images = dropout(images, keep_prob=self.dropout)
images = fully_connected(images, 4096, activation=self.activation)
images = dropout(images, keep_prob=self.dropout)
#fully connected layers
images = fully_connected(images, 1024, activation=self.activation)
images = fully_connected(images, 7, activation='softmax')
if self.optimizer == 'momentum':
optimizers = Momentum(
learning_rate=self.learning_rate,
momentum=self.optimizer_param,
lr_decay=self.learning_rate_decay,
decay_step=self.decay_step
)
elif self.optimizer == 'adam':
optimizers = Adam(
learning_rate=self.learning_rate,
beta1=self.optimizer_param,
beta2=self.learning_rate_decay,
)
else:
print("Error Optimizer")
network = regression(
images,
optimizer=optimizers,
loss='categorical_crossentropy',
learning_rate=self.learning_rate,
name='output'
)
return network
def get_network():
biases = zeros(shape=[9, 19, 1, 192])
biases2 = zeros(shape=[19, 19, 1])
network = input_data(shape=[None, 19, 19, 2], name='input')
network = conv_2d(network, 192, 5, activation='elu', weights_init=truncated_normal(stddev=stddev5), bias=False) + biases[0]
network = conv_2d(network, 192, 3, activation='elu', weights_init=truncated_normal(stddev=stddev3), bias=False) + biases[1]
network = conv_2d(network, 192, 3, activation='elu', weights_init=truncated_normal(stddev=stddev3), bias=False) + biases[2]
network = conv_2d(network, 192, 3, activation='elu', weights_init=truncated_normal(stddev=stddev3), bias=False) + biases[3]
network = conv_2d(network, 192, 3, activation='elu', weights_init=truncated_normal(stddev=stddev3), bias=False) + biases[4]
network = conv_2d(network, 192, 3, activation='elu', weights_init=truncated_normal(stddev=stddev3), bias=False) + biases[5]
network = conv_2d(network, 192, 3, activation='elu', weights_init=truncated_normal(stddev=stddev3), bias=False) + biases[6]
network = conv_2d(network, 192, 3, activation='elu', weights_init=truncated_normal(stddev=stddev3), bias=False) + biases[7]
network = conv_2d(network, 192, 3, activation='elu', weights_init=truncated_normal(stddev=stddev3), bias=False) + biases[8]
network = conv_2d(network, 1, 3, activation='elu', weights_init=truncated_normal(stddev=stddev3), bias=False) + biases2
network = fully_connected(network, 19*19, activation='softmax')
momentum = Momentum(learning_rate=0.002)
network = regression(network, optimizer=momentum, loss='categorical_crossentropy', name='target')
return network
def build_resnet(self):
# Adapted resnet
# Adaptation did not work
print('[+] Building Resnet')
# Residual blocks
n = 5
self.network = input_data(shape=[None, SIZE_FACE, SIZE_FACE, 1])
self.network = conv_2d(self.network,
SIZE_FACE / 2,
3,
regularizer='L2',
weight_decay=0.0001)
self.network = residual_block(self.network, n, SIZE_FACE / 2)
self.network = residual_block(self.network,
1,
SIZE_FACE,
downsample=True)
self.network = residual_block(self.network, n - 1, SIZE_FACE)
self.network = residual_block(self.network,
1,
SIZE_FACE * 2,
downsample=True)
self.network = residual_block(self.network, n - 1, SIZE_FACE * 2)
self.network = batch_normalization(self.network)
self.network = activation(self.network, 'relu')
self.network = global_avg_pool(self.network)
self.network = fully_connected(self.network,
len(EMOTIONS),
activation='softmax')
self.momentum = Momentum(0.1,
lr_decay=0.1,
decay_step=32000,
staircase=True)
self.network = regression(self.network,
optimizer=self.momentum,
loss='categorical_crossentropy')
self.model = tflearn.DNN(self.network,
checkpoint_path=SAVE_DIRECTORY + RUN_NAME,
max_checkpoints=1,
tensorboard_verbose=0)
return self.model
def gamornet_tl_tflearn(training_imgs,
training_labels,
validation_imgs,
validation_labels,
input_shape,
load_layers_bools=[True] * 8,
trainable_bools=[True] * 8,
model_load_path="./",
files_save_path="./",
epochs=100,
max_checkpoints=1,
batch_size=64,
lr=0.00001,
momentum=0.9,
decay=0.0,
nesterov=False,
loss='categorical_crossentropy',
save_model=True,
show_metric=True,
clear_session=False):
"""
Performs Transfer Learning (TL) using a previously trained GaMorNet model.
Parameters
-----------
training_imgs: Numpy ndarray [nsamples,x,y,ndim]
The array of images which are to be used for the TL process. We insist on numpy arrays
as many of the underlying deep learning frameworks work better with numpy arrays compared to
other array-like elements.
training_labels: Numpy ndarray [nsamples,label_arrays]
The truth labels for each of the TL images. The supplied labels must be in the one-hot encoding
format. We reproduce below what each individual label array should look like:-
* Disk-dominated - ``[1,0,0]``
* Indeterminate - ``[0,1,0]``
* Bulge-dominated - ``[0,0,1]``
validation_imgs: Numpy ndarray [nsamples,x,y,ndim]
The array of images which are to be used for the validation process. We insist on numpy arrays
as many of the underlying deep learning frameworks work better with numpy arrays compared to
other array-like elements.
validation_labels: Numpy ndarray [nsamples,label_arrays]
The truth labels for each of the validation images. The supplied labels must be in the one-hot encoding
format. We reproduce below what each individual label array should look like:-
* Disk-dominated - ``[1,0,0]``
* Indeterminate - ``[0,1,0]``
* Bulge-dominated - ``[0,0,1]``
input_shape: tuple of ints (x, y, ndim) or allowed str
The shape of the images being used in the form of a tuple.
This parameter can also take the following special values:-
* ``SDSS`` - Sets the input shape to be (167,167,1) as was used for the SDSS g-band images in Ghosh et. al. (2020)
* ``CANDELS`` - Sets the input shape to be (83,83,1) as was used for the CANDELS H-band images in Ghosh et. al. (2020)
load_layers_bools: array of bools
This variable is used to identify which of the 5 convolutional and 3 fully-connected layers of GaMorNet will be
loaded during the transfer learning process from the supplied starting model. The rest of the layers will be
initialized from scratch.
The order of the bools correspond to the following layer numbers [2, 5, 8, 9, 10, 13, 15, 17] in GaMorNet. Please see
Figure 4 and Table 2 of Ghosh et. al. (2020) to get more details. The first five layers are the convolutional
layers and the last three are the fully connected layers.
This parameter can also take the following special values which are handy when you are using our models to
perform predictions:-
* ``load_bools_SDSS`` - Sets the bools according to what was done for the SDSS data in Ghosh et. al. (2020)
* ``load_bools_CANDELS``- Sets the bools according to what was done for the CANDELS data in Ghosh et. al. (2020)
trainable_bools: array of bools
This variable is used to identify which of the 5 convolutional and 3 fully-connected layers of GaMorNet will be
trainable during the transfer learning process. The rest are frozen at the values loaded from the previous
model.
The order of the bools correspond to the following layer numbers [2, 5, 8, 9, 10, 13, 15, 17] in GaMorNet. Please see
Figure 4 and Table 2 of Ghosh et. al. (2020) to get more details. The first five layers are the convolutional
layers and the last three are the fully connected layers.
This parameter can also take the following special values which are handy when you are using our models to
perform predictions:-
* ``train_bools_SDSS`` - Sets the bools according to what was done for the SDSS data in Ghosh et. al. (2020)
* ``train_bools_CANDELS``- Sets the bools according to what was done for the CANDELS data in Ghosh et. al. (2020)
model_load_path: str
Path to the saved model, which will serve as the starting point for transfer learning. Note that
tflearn models usually consist of three files in the format ``file_name.data``,
``file_name.index``, ``file_name.meta``. For this parameter, simply specify file_path/file_name.
This parameter can also take the following special values
* ``SDSS_sim`` -- Downloads and uses GaMorNet models trained on SDSS g-band simulations at z~0 from Ghosh et. al. (2020)
* ``SDSS_tl`` -- Downloads and uses GaMorNet models trained on SDSS g-band simulations and real data at z~0 from Ghosh et. al. (2020)
* ``CANDELS_sim`` -- Downloads and uses GaMorNet models trained on CANDELS H-band simulations at z~1 from Ghosh et. al. (2020)
* ``CANDELS_tl`` -- Downloads and uses GaMorNet models trained on CANDELS H-band simulations and real data at z~1 from Ghosh et. al. (2020)
files_save_path: str
The full path to the location where the models generated during the training process are to be
saved. The path should end with the name of the file. For eg. ``/path/checkpoint``. This
will result in model files of the form ``checkpoint.meta``, ``checkpoint.data`` and
``checkpoint.index`` being saved.
Set this to ``/dev/null`` on a unix system if you don't want to save the output.
epochs: int
The number of epochs for which you want to train the model.
max_checkpoints: int
TFLearn saves the model at the end of each epoch. This parameter controls how many of the
most recent models are saved. For eg. setting this to 2, will save the model state during the
most recent two epochs.
batch_size: int
This variable specifies how many images will be processed in a single batch. This is a
hyperparameter. The default value is a good starting point
lr: float
This is the learning rate to be used during the training process. This is a
hyperparameter that should be tuned during the training process. The default value is a good
starting point.
momentum: float
The value of the momentum to be used in the gradient descent optimizer that is used to train GaMorNet.
This must always be :math:`\geq 0`. This accelerates the gradient descent process. This is a
hyperparameter. The default value is a good starting point.
decay: float
The amount of learning rate decay to be applied over each update.
nesterov: bool
Whether to apply Nesterov momentum or not.
loss: allowed str or function
The loss function to be used. If using the string option, you need to specify the name of
the loss function. This can be set to be any loss available in ``tflearn``
save_model: bool
Whether you want to save the model files at each epoch during training. This
parameter should be used in conjunction with ``max_checkpoints`` to configure
how many of the saved model files are preserved till the end.
show_metric: bool
Whether to display the training/testing metrics during training.
clear_session: bool
If set to True, this will clear the TensorFlow session currently running. This is handy while running GaMorNet in a
notebook to avoid variable name confusions. (Sometimes, under the hood, TFLearn & Tensorflow reuse the same layer names
leading to conflicts)
Note that if set to True, you will lose access to any other graphs you may have run before.
Returns
--------
Trained TFLearn Model: TFLearn ``models.dnn.DNN`` class
"""
# TFLearn Loads graphs from memory by name, hence it's always advisable to
# set this to True if using in a Notebook.
if clear_session is True:
K.clear_session()
check_imgs_validity(training_imgs)
check_imgs_validity(validation_imgs)
check_labels_validity(training_labels)
check_labels_validity(validation_labels)
model = gamornet_build_model_tflearn(input_shape=input_shape,
trainable_bools=trainable_bools,
load_layers_bools=load_layers_bools)
if nesterov is False:
optimizer = Momentum(momentum=momentum, lr_decay=decay)
else:
optimizer = Nesterov(momentum=momentum, lr_decay=decay)
model = regression(model, optimizer=optimizer, loss=loss, learning_rate=lr)
model = tflearn.DNN(model,
checkpoint_path=files_save_path + "check",
max_checkpoints=max_checkpoints)
model = gamornet_load_model_tflearn(model, model_load_path)
model.fit(training_imgs,
training_labels,
n_epoch=epochs,
validation_set=(validation_imgs, validation_labels),
shuffle=True,
show_metric=show_metric,
batch_size=batch_size,
snapshot_step=None,
snapshot_epoch=save_model)
return model
def build_modelB(optimizer=HYPERPARAMS.optimizer,
optimizer_param=HYPERPARAMS.optimizer_param,
learning_rate=HYPERPARAMS.learning_rate,
keep_prob=HYPERPARAMS.keep_prob,
learning_rate_decay=HYPERPARAMS.learning_rate_decay,
decay_step=HYPERPARAMS.decay_step):
images_network = input_data(
shape=[None, NETWORK.input_size, NETWORK.input_size, 1], name='input1')
images_network = conv_2d(images_network,
64,
3,
activation=NETWORK.activation)
#images_network = local_response_normalization(images_network)
if NETWORK.use_batchnorm_after_conv_layers:
images_network = batch_normalization(images_network)
images_network = max_pool_2d(images_network, 3, strides=2)
images_network = conv_2d(images_network,
128,
3,
activation=NETWORK.activation)
if NETWORK.use_batchnorm_after_conv_layers:
images_network = batch_normalization(images_network)
images_network = max_pool_2d(images_network, 3, strides=2)
images_network = conv_2d(images_network,
256,
3,
activation=NETWORK.activation)
if NETWORK.use_batchnorm_after_conv_layers:
images_network = batch_normalization(images_network)
images_network = max_pool_2d(images_network, 3, strides=2)
images_network = dropout(images_network, keep_prob=keep_prob)
images_network = fully_connected(images_network,
4096,
activation=NETWORK.activation)
images_network = dropout(images_network, keep_prob=keep_prob)
images_network = fully_connected(images_network,
1024,
activation=NETWORK.activation)
if NETWORK.use_batchnorm_after_fully_connected_layers:
images_network = batch_normalization(images_network)
if NETWORK.use_landmarks or NETWORK.use_hog_and_landmarks:
if NETWORK.use_hog_sliding_window_and_landmarks:
landmarks_network = input_data(shape=[None, 2728], name='input2')
elif NETWORK.use_hog_and_landmarks:
landmarks_network = input_data(shape=[None, 208], name='input2')
else:
landmarks_network = input_data(shape=[None, 68, 2], name='input2')
landmarks_network = fully_connected(landmarks_network,
1024,
activation=NETWORK.activation)
if NETWORK.use_batchnorm_after_fully_connected_layers:
landmarks_network = batch_normalization(landmarks_network)
landmarks_network = fully_connected(landmarks_network,
128,
activation=NETWORK.activation)
if NETWORK.use_batchnorm_after_fully_connected_layers:
landmarks_network = batch_normalization(landmarks_network)
images_network = fully_connected(images_network,
128,
activation=NETWORK.activation)
network = merge([images_network, landmarks_network], 'concat', axis=1)
else:
network = images_network
network = fully_connected(network,
NETWORK.output_size,
activation='softmax')
if optimizer == 'momentum':
optimizer = Momentum(learning_rate=learning_rate,
momentum=optimizer_param,
lr_decay=learning_rate_decay,
decay_step=decay_step)
elif optimizer == 'adam':
optimizer = Adam(learning_rate=learning_rate,
beta1=optimizer_param,
beta2=learning_rate_decay)
else:
print("Unknown optimizer: {}".format(optimizer))
network = regression(network,
optimizer=optimizer,
loss=NETWORK.loss,
learning_rate=learning_rate,
name='output')
return network
conv3_1 = conv_2d(pool2, 256, 3, activation='relu', name="conv3_1")
conv3_2 = conv_2d(conv3_1, 256, 3, activation='relu', name="conv3_2")
conv3_3 = conv_2d(conv3_2, 256, 3, activation='relu', name="conv3_3")
pool3 = max_pool_2d(conv3_3, 2, strides=2)
conv4_1 = conv_2d(pool3, 512, 3, activation='relu', name="conv4_1")
conv4_2 = conv_2d(conv4_1, 512, 3, activation='relu', name="conv4_2")
conv4_3 = conv_2d(conv4_2, 512, 3, activation='relu', name="conv4_3")
pool4 = max_pool_2d(conv4_3, 2, strides=2)
conv5_1 = conv_2d(pool4, 512, 3, activation='relu', name="conv5_1")
conv5_2 = conv_2d(conv5_1, 512, 3, activation='relu', name="conv5_2")
conv5_3 = conv_2d(conv5_2, 512, 3, activation='relu', name="conv5_3")
pool5 = max_pool_2d(conv5_3, 2, strides=2)
fc6 = fully_connected(pool5, 4096, activation='relu', name="fc6")
fc6_dropout = dropout(fc6, 0.5)
fc7 = fully_connected(fc6_dropout, 4096, activation='relu', name="fc7")
fc7_droptout = dropout(fc7, 0.5)
fc8 = fully_connected(fc7_droptout, 1000, activation='softmax', name="fc8")
mm = Momentum(learning_rate=0.01, momentum=0.9, lr_decay=0.1, decay_step=1000)
network = regression(fc8, optimizer=mm, loss='categorical_crossentropy', restore=False)
model = tflearn.DNN(network)
model.fit(X, y, n_epoch=74, shuffle=True, show_metric=True, batch_size=256, snapshot_step=500, run_id='vgg16_imagenet')
model.save("vgg16_weights.tflearn")
def build_model(optimizer=HYPERPARAMS.optimizer, optimizer_param=HYPERPARAMS.optimizer_param,
learning_rate=HYPERPARAMS.learning_rate, keep_prob=HYPERPARAMS.keep_prob,
learning_rate_decay=HYPERPARAMS.learning_rate_decay, decay_step=HYPERPARAMS.decay_step):
images_input = input_data(shape=[None, NETWORK.input_size, NETWORK.input_size, 1], name='input1')
images_network = conv_2d(images_input, 16, 3, activation='relu')
if NETWORK.use_batchnorm_after_conv_layers:
images_network = batch_normalization(images_network)
images_network = conv_2d(images_network, 16, 3, activation='relu')
if NETWORK.use_batchnorm_after_conv_layers:
images_network = batch_normalization(images_network)
images_network = max_pool_2d(images_network, 2, strides=2) #24*24*16
images_network = conv_2d(images_network, 32, 3, activation='relu')
if NETWORK.use_batchnorm_after_conv_layers:
images_network = batch_normalization(images_network)
images_network = conv_2d(images_network, 32, 3, activation='relu')
if NETWORK.use_batchnorm_after_conv_layers:
images_network = batch_normalization(images_network)
images_network = max_pool_2d(images_network, 2, strides=2) #12*12*32
images_network=tf.pad(images_network,[[0,0],[18,18],[18,18],[0,0]],'CONSTANT')
images_network = merge([images_network, images_input], 'concat', axis=3) #48*48*33
images_network = conv_2d(images_network, 64, 3, activation='relu')
if NETWORK.use_batchnorm_after_conv_layers:
images_network = batch_normalization(images_network)
images_network = conv_2d(images_network, 64, 3, activation='relu')
if NETWORK.use_batchnorm_after_conv_layers:
images_network = batch_normalization(images_network)
images_network = conv_2d(images_network, 64, 3, activation='relu')
if NETWORK.use_batchnorm_after_conv_layers:
images_network = batch_normalization(images_network)
images_network = max_pool_2d(images_network, 2, strides=2) #24*24*64
images_network = conv_2d(images_network, 128, 3, activation='relu')
if NETWORK.use_batchnorm_after_conv_layers:
images_network = batch_normalization(images_network)
images_network = conv_2d(images_network, 128, 3, activation='relu')
if NETWORK.use_batchnorm_after_conv_layers:
images_network = batch_normalization(images_network)
images_network = conv_2d(images_network, 128, 3, activation='relu')
if NETWORK.use_batchnorm_after_conv_layers:
images_network = batch_normalization(images_network)
images_network = max_pool_2d(images_network, 2, strides=2) #12*12*128
#
images_network = conv_2d(images_network, 128, 3, activation='relu')
if NETWORK.use_batchnorm_after_conv_layers:
images_network = batch_normalization(images_network)
images_network = conv_2d(images_network, 128, 3, activation='relu')
if NETWORK.use_batchnorm_after_conv_layers:
images_network = batch_normalization(images_network)
images_network = conv_2d(images_network, 128, 3, activation='relu')
if NETWORK.use_batchnorm_after_conv_layers:
images_network = batch_normalization(images_network)
images_network = max_pool_2d(images_network, 2, strides=2) #6*6*128
images_network = fully_connected(images_network, 1024, activation='relu')
images_network = dropout(images_network,keep_prob=keep_prob)
if NETWORK.use_batchnorm_after_fully_connected_layers:
images_network = batch_normalization(images_network)
images_network = fully_connected(images_network, 1024, activation='relu')
images_network = dropout(images_network, keep_prob=keep_prob)
if NETWORK.use_batchnorm_after_fully_connected_layers:
images_network = batch_normalization(images_network)
if NETWORK.use_landmarks or NETWORK.use_hog_and_landmarks:
if NETWORK.use_hog_sliding_window_and_landmarks:
landmarks_network = input_data(shape=[None, 2728], name='input2')
elif NETWORK.use_hog_and_landmarks:
landmarks_network = input_data(shape=[None, 208], name='input2')
else:
landmarks_network = input_data(shape=[None, 68, 2], name='input2')
landmarks_network = fully_connected(landmarks_network, 1024, activation=NETWORK.activation)
if NETWORK.use_batchnorm_after_fully_connected_layers:
landmarks_network = batch_normalization(landmarks_network)
landmarks_network = fully_connected(landmarks_network, 40, activation=NETWORK.activation)
if NETWORK.use_batchnorm_after_fully_connected_layers:
landmarks_network = batch_normalization(landmarks_network)
images_network = fully_connected(images_network, 40, activation=NETWORK.activation)
network = merge([images_network, landmarks_network], 'concat', axis=1)
else:
network = images_network
network = fully_connected(network, NETWORK.output_size, activation='softmax')
if optimizer == 'momentum':
# FIXME base_lr * (1 - iter/max_iter)^0.5, base_lr = 0.01
optimizer = Momentum(learning_rate=learning_rate, momentum=optimizer_param,
lr_decay=learning_rate_decay, decay_step=decay_step)
elif optimizer == 'adam':
optimizer = Adam(learning_rate=learning_rate, beta1=optimizer_param, beta2=learning_rate_decay)
else:
print("Unknown optimizer: {}".format(optimizer))
network = regression(network, optimizer=optimizer, loss=NETWORK.loss, learning_rate=learning_rate, name='output')
return network
model = None
if in_args.architecture == 'VGG16':
if not _is_valid_size(224, in_args.patch_size, in_args.extract_level):
raise RuntimeError("Real size of extracted patch less than network input size!")
img_prep.add_random_crop((224, 224))
cnn_network = input_data(shape=[None, 224, 224, 3],
data_preprocessing=img_prep,
data_augmentation=img_aug)
cnn_network = create_vgg16_network( cnn_network, num_classes=n_classes,
normalize_batch=apply_batchnorm,
add_color_transfer=apply_colortransform)
MomOpt = Momentum(learning_rate=0.01, lr_decay=0.8, decay_step=200)
cnn_network = regression(cnn_network, optimizer=MomOpt,
loss='categorical_crossentropy')
elif in_args.architecture == 'simple':
if not _is_valid_size(28, in_args.patch_size, in_args.extract_level):
raise RuntimeError("Real size of extracted patch less than network input size!")
img_prep.add_random_crop((28, 28))
cnn_network = input_data(shape=[None, 28, 28, 3],
data_preprocessing=img_prep,
data_augmentation=img_aug)
cnn_network = create_simple_network( cnn_network, num_classes=n_classes)
