def build_model():
logging.info('building model')
img_prep = ImagePreprocessing()
img_prep.add_featurewise_zero_center()
img_prep.add_featurewise_stdnorm()
encoder = input_data(shape=(None, IMAGE_INPUT_SIZE[0], IMAGE_INPUT_SIZE[1],
3), data_preprocessing=img_prep)
encoder = conv_2d(encoder, 16, 7, activation='relu')
encoder = dropout(encoder, 0.25) # you can have noisy input instead
encoder = max_pool_2d(encoder, 2)
encoder = conv_2d(encoder, 16, 7, activation='relu')
encoder = max_pool_2d(encoder, 2)
encoder = conv_2d(encoder, 8, 7, activation='relu')
encoder = max_pool_2d(encoder, 2)
decoder = conv_2d(encoder, 8, 7, activation='relu')
decoder = upsample_2d(decoder, 2)
decoder = conv_2d(decoder, 16, 7, activation='relu')
decoder = upsample_2d(decoder, 2)
decoder = conv_2d(decoder, 16, 7, activation='relu')
decoder = upsample_2d(decoder, 2)
decoder = conv_2d(decoder, 3, 7)
encoded_str = re.search(r', (.*)\)', str(encoder.get_shape)).group(1)
encoded_size = np.prod([int(o) for o in encoded_str.split(', ')])
original_img_size = np.prod(IMAGE_INPUT_SIZE) * 3
percentage = round(encoded_size / original_img_size, 2) * 100
logging.debug('the encoded representation is {}% of the original \
image'.format(percentage))
return regression(decoder, optimizer='adadelta',
loss='binary_crossentropy', learning_rate=0.005)
def build():
input_img = input_data(shape=(HEIGHT, WIDTH, CHANNELS), name='input');print(input_img.shape)
INPUT = input_img
x = conv_2d(input_img, 16, (3, 3), activation='relu', padding='same');print(x.shape)
x = max_pool_2d(x, (2, 2), padding='same');print(x.shape)
x = conv_2d(x, 8, (3, 3), activation='relu', padding='same');print(x.shape)
x = max_pool_2d(x, (2, 2), padding='same');print(x.shape)
x = conv_2d(x, 8, (3, 3), activation='relu', padding='same');print(x.shape)
encoded = max_pool_2d(x, (2, 2), padding='same');print(encoded.shape) # at this point the representation is (4, 4, 8) i.e. 128-dimensional
HIDDEN_STATE = encoded
print("middle")
x = conv_2d(encoded, 8, (3, 3), activation='relu', padding='same', name='input2');print(x.shape)
x = upsample_2d(x, (2, 2));print(x.shape)
x = conv_2d(x, 8, (3, 3), activation='relu', padding='same');print(x.shape)
x = upsample_2d(x, (2, 2));print(x.shape)
x = conv_2d(x, 16, (3, 3), activation='relu', padding='same');print(x.shape)
x = upsample_2d(x, (2, 2));print(x.shape)
decoded = conv_2d(x, CHANNELS, (3, 3), activation='sigmoid', padding='same');print(decoded.shape)
OUTPUT = decoded
autoencoder = regression(decoded, optimizer='momentum', loss='mean_square',
learning_rate=0.005, name='targets')
model = tflearn.DNN(autoencoder, checkpoint_path=MODEL_NAME,
max_checkpoints=1, tensorboard_verbose=0, tensorboard_dir='log')
return model, INPUT, HIDDEN_STATE, OUTPUT
def build_fcn_all(network):
#Pool1
conv1 = conv_2d(network, 8, 7, activation='relu')
pool1 = max_pool_2d(conv1, 2)
#Pool2
conv2 = conv_2d(pool1, 16, 5, activation='relu')
pool2 = max_pool_2d(conv2, 2)
#Pool3
conv3 = conv_2d(pool2, 32, 5, activation='relu')
pool3 = max_pool_2d(conv3, 2) # output 8x_downsampled
#Pool4
conv4 = conv_2d(pool3, 64, 3, activation='relu')
pool4 = max_pool_2d(conv4, 2) # output 16x_downsampled
#start FCN-32s -----------------------------------
#Pool5
conv5 = conv_2d(pool4, 128, 3, activation='relu')
pool5 = max_pool_2d(conv5, 2)
#Conv6-7
conv6 = conv_2d(pool5, 128, 3, activation='relu')
conv7 = conv_2d(conv6, 128, 3, activation='relu') # output 32x_downsampled
#end FCN-32s -----------------------------------
##start FCN-16s -----------------------------------
#network_32_UP2 = upsample_2d(network_32, 2)
#network_16 = merge([network_32_UP2, network_4], mode='concat', axis=3)
#network_16 = conv_2d(network_16, 3, 128, activation='relu') # output 16x_downsampled
##end FCN-16s -----------------------------------
##start FCN-8s -----------------------------------
#network_32_UP4 = upsample_2d(network_32, 4)
#network_16_UP2 = upsample_2d(network_16, 2)
#network_3_UP8 = upsample_2d(network_3, 8)
pool4_x2 = upsample_2d(pool4, 2)
conv7_x4 = upsample_2d(conv7, 4)
#network_8 = merge([network_32_UP4, network_4_UP2, network_3], mode='concat', axis=3)
fcn_8s = merge([pool3, pool4_x2, conv7_x4], mode='concat', axis=3)
fcn_8s = conv_2d(fcn_8s, 3, 1, activation='relu')
##end FCN-8s -----------------------------------
out = conv_2d(fcn_8s, CHANNELS, 1, activation='relu')
out = upsample_2d(out, 8)
#network_8 = upscore_layer(network_8, num_classes=3, kernel_size=2, strides=8, shape=[384, 1216, 3])
network = tflearn.regression(
out,
loss='mean_square',
#loss='weak_cross_entropy_2d',
)
return network
def add_upsample2d_layer(self,
window_size,
prev_layer_name=None,
exclude_from_path=True,
**kwargs):
prev_name = prev_layer_name or self._curr_layer_name
prev = self._layers[prev_name]
layer = {
"layer": conv.upsample_2d(prev, window_size, **kwargs),
"type": "Upsample2D",
"categories": ["hidden"]
}
return self._add_layer(layer, exclude_from_path)
def CNN(input, reuse=False):
with tf.variable_scope('CNN'):
width = 96
height = 96
n_channels = 1
n_classes = 2
################Create Model######################
with tf.variable_scope('Block1'):
conv1 = conv_2d(input,
32,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv1')
conv1 = conv_2d(conv1,
32,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv2')
pool1 = max_pool_2d(conv1, 2)
with tf.variable_scope('Block2'):
conv2 = conv_2d(pool1,
64,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv3')
conv2 = conv_2d(conv2,
64,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv4')
pool2 = max_pool_2d(conv2, 2)
with tf.variable_scope('Block3'):
conv3 = conv_2d(pool2,
128,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv5')
conv3 = conv_2d(conv3,
128,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv6')
pool3 = max_pool_2d(conv3, 2)
with tf.variable_scope('Block4'):
conv4 = conv_2d(pool3,
256,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv7')
conv4 = conv_2d(conv4,
256,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv8')
pool4 = max_pool_2d(conv4, 2)
with tf.variable_scope('Block5'):
conv5 = conv_2d(pool4,
512,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv9')
conv5 = conv_2d(conv5,
512,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv10')
with tf.variable_scope('Block6'):
up6 = upsample_2d(conv5, 2)
up6 = tflearn.layers.merge_ops.merge([up6, conv4],
'concat',
axis=3)
conv6 = conv_2d(up6,
256,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv11')
conv6 = conv_2d(conv6,
256,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv12')
with tf.variable_scope('Block7'):
up7 = upsample_2d(conv6, 2)
up7 = tflearn.layers.merge_ops.merge([up7, conv3],
'concat',
axis=3)
conv7 = conv_2d(up7,
128,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv13')
conv7 = conv_2d(conv7,
128,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv14')
with tf.variable_scope('Block8'):
up8 = upsample_2d(conv7, 2)
up8 = tflearn.layers.merge_ops.merge([up8, conv2],
'concat',
axis=3)
conv8 = conv_2d(up8,
64,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv15')
conv8 = conv_2d(conv8,
64,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv16')
with tf.variable_scope('Block9'):
up9 = upsample_2d(conv8, 2)
up9 = tflearn.layers.merge_ops.merge([up9, conv1],
'concat',
axis=3)
conv9 = conv_2d(up9,
32,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv17')
conv9 = conv_2d(conv9,
32,
3,
activation='relu',
padding='same',
regularizer="L2",
reuse=reuse,
scope='conv18')
with tf.variable_scope('Output'):
pred = conv_2d(conv9,
2,
1,
activation='linear',
padding='valid',
reuse=reuse,
scope='conv19')
pred_reshape = tf.reshape(pred, [-1, width, height, n_classes])
return pred_reshape
conv2 = conv_2d(pool1, 64, 3, activation='relu', padding='same', regularizer="L2") conv2 = conv_2d(conv2, 64, 3, activation='relu', padding='same', regularizer="L2") pool2 = max_pool_2d(conv2, 2) conv3 = conv_2d(pool2, 128, 3, activation='relu', padding='same', regularizer="L2") conv3 = conv_2d(conv3, 128, 3, activation='relu', padding='same', regularizer="L2") pool3 = max_pool_2d(conv3, 2) conv4 = conv_2d(pool3, 256, 3, activation='relu', padding='same', regularizer="L2") conv4 = conv_2d(conv4, 256, 3, activation='relu', padding='same', regularizer="L2") pool4 = max_pool_2d(conv4, 2) conv5 = conv_2d(pool4, 512, 3, activation='relu', padding='same', regularizer="L2") conv5 = conv_2d(conv5, 512, 3, activation='relu', padding='same', regularizer="L2") up6 = upsample_2d(conv5,2) up6 = tflearn.layers.merge_ops.merge([up6, conv4], 'concat',axis=3) conv6 = conv_2d(up6, 256, 3, activation='relu', padding='same', regularizer="L2") conv6 = conv_2d(conv6, 256, 3, activation='relu', padding='same', regularizer="L2") up7 = upsample_2d(conv6,2) up7 = tflearn.layers.merge_ops.merge([up7, conv3],'concat', axis=3) conv7 = conv_2d(up7, 128, 3, activation='relu', padding='same', regularizer="L2") conv7 = conv_2d(conv7, 128, 3, activation='relu', padding='same', regularizer="L2") up8 = upsample_2d(conv7,2) up8 = tflearn.layers.merge_ops.merge([up8, conv2],'concat', axis=3) conv8 = conv_2d(up8, 64, 3, activation='relu', padding='same', regularizer="L2") conv8 = conv_2d(conv8, 64, 3, activation='relu', padding='same', regularizer="L2") up9 = upsample_2d(conv8,2)
def extractBrain(dataPath, modelCkptLoc, thresholdLoc, modelCkptSeg,
thresholdSeg, bidsDir, out_postfix):
# Step1: Main part brain localization
normalize = "local_max"
width = 128
height = 128
border_x = 15
border_y = 15
n_channels = 1
img_nib = nibabel.load(os.path.join(dataPath))
image_data = img_nib.get_data()
images = np.zeros((image_data.shape[2], width, height, n_channels))
pred3dFinal = np.zeros((image_data.shape[2], image_data.shape[0],
image_data.shape[1], n_channels))
slice_counter = 0
for ii in range(image_data.shape[2]):
img_patch = cv2.resize(image_data[:, :, ii],
dsize=(width, height),
fx=width,
fy=height)
if normalize:
if normalize == "local_max":
images[slice_counter, :, :, 0] = img_patch / np.max(img_patch)
elif normalize == "mean_std":
images[slice_counter, :, :,
0] = (img_patch -
np.mean(img_patch)) / np.std(img_patch)
else:
raise ValueError('Please select a valid normalization')
else:
images[slice_counter, :, :, 0] = img_patch
slice_counter += 1
# Tensorflow graph
g = tf.Graph()
with g.as_default():
with tf.name_scope('inputs'):
x = tf.placeholder(tf.float32, [None, width, height, n_channels])
conv1 = conv_2d(x,
32,
3,
activation='relu',
padding='same',
regularizer="L2")
conv1 = conv_2d(conv1,
32,
3,
activation='relu',
padding='same',
regularizer="L2")
pool1 = max_pool_2d(conv1, 2)
conv2 = conv_2d(pool1,
64,
3,
activation='relu',
padding='same',
regularizer="L2")
conv2 = conv_2d(conv2,
64,
3,
activation='relu',
padding='same',
regularizer="L2")
pool2 = max_pool_2d(conv2, 2)
conv3 = conv_2d(pool2,
128,
3,
activation='relu',
padding='same',
regularizer="L2")
conv3 = conv_2d(conv3,
128,
3,
activation='relu',
padding='same',
regularizer="L2")
pool3 = max_pool_2d(conv3, 2)
conv4 = conv_2d(pool3,
256,
3,
activation='relu',
padding='same',
regularizer="L2")
conv4 = conv_2d(conv4,
256,
3,
activation='relu',
padding='same',
regularizer="L2")
pool4 = max_pool_2d(conv4, 2)
conv5 = conv_2d(pool4,
512,
3,
activation='relu',
padding='same',
regularizer="L2")
conv5 = conv_2d(conv5,
512,
3,
activation='relu',
padding='same',
regularizer="L2")
up6 = upsample_2d(conv5, 2)
up6 = tflearn.layers.merge_ops.merge([up6, conv4], 'concat', axis=3)
conv6 = conv_2d(up6,
256,
3,
activation='relu',
padding='same',
regularizer="L2")
conv6 = conv_2d(conv6,
256,
3,
activation='relu',
padding='same',
regularizer="L2")
up7 = upsample_2d(conv6, 2)
up7 = tflearn.layers.merge_ops.merge([up7, conv3], 'concat', axis=3)
conv7 = conv_2d(up7,
128,
3,
activation='relu',
padding='same',
regularizer="L2")
conv7 = conv_2d(conv7,
128,
3,
activation='relu',
padding='same',
regularizer="L2")
up8 = upsample_2d(conv7, 2)
up8 = tflearn.layers.merge_ops.merge([up8, conv2], 'concat', axis=3)
conv8 = conv_2d(up8,
64,
3,
activation='relu',
padding='same',
regularizer="L2")
conv8 = conv_2d(conv8,
64,
3,
activation='relu',
padding='same',
regularizer="L2")
up9 = upsample_2d(conv8, 2)
up9 = tflearn.layers.merge_ops.merge([up9, conv1], 'concat', axis=3)
conv9 = conv_2d(up9,
32,
3,
activation='relu',
padding='same',
regularizer="L2")
conv9 = conv_2d(conv9,
32,
3,
activation='relu',
padding='same',
regularizer="L2")
pred = conv_2d(conv9, 2, 1, activation='linear', padding='valid')
# Thresholding parameter to binarize predictions
percentileLoc = thresholdLoc * 100
im = np.zeros((1, width, height, n_channels))
pred3d = []
with tf.Session(graph=g) as sess_test_loc:
# Restore the model
tf_saver = tf.train.Saver()
tf_saver.restore(sess_test_loc, modelCkptLoc)
for idx in range(images.shape[0]):
im = np.reshape(images[idx, :, :, :],
[1, width, height, n_channels])
feed_dict = {x: im}
pred_ = sess_test_loc.run(pred, feed_dict=feed_dict)
theta = np.percentile(pred_, percentileLoc)
pred_bin = np.where(pred_ > theta, 1, 0)
pred3d.append(pred_bin[0, :, :, 0].astype('float64'))
#####
pred3d = np.asarray(pred3d)
heights = []
widths = []
coms_x = []
coms_y = []
# Apply PPP
ppp = True
if ppp:
pred3d = post_processing(pred3d)
pred3d = [
cv2.resize(elem,
dsize=(image_data.shape[1], image_data.shape[0]),
interpolation=cv2.INTER_NEAREST) for elem in pred3d
]
pred3d = np.asarray(pred3d)
for i in range(np.asarray(pred3d).shape[0]):
if np.sum(pred3d[i, :, :]) != 0:
pred3d[i, :, :] = extractLargestCC(
pred3d[i, :, :].astype('uint8'))
contours, hierarchy = cv2.findContours(
pred3d[i, :, :].astype('uint8'), cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE)
area = cv2.minAreaRect(np.squeeze(contours))
heights.append(area[1][0])
widths.append(area[1][1])
bbox = cv2.boxPoints(area).astype('int')
coms_x.append(
int((np.max(bbox[:, 1]) + np.min(bbox[:, 1])) / 2))
coms_y.append(
int((np.max(bbox[:, 0]) + np.min(bbox[:, 0])) / 2))
# Saving localization points
med_x = int(np.median(coms_x))
med_y = int(np.median(coms_y))
half_max_x = int(np.max(heights) / 2)
half_max_y = int(np.max(widths) / 2)
x_beg = med_x - half_max_x - border_x
x_end = med_x + half_max_x + border_x
y_beg = med_y - half_max_y - border_y
y_end = med_y + half_max_y + border_y
# Step2: Brain segmentation
width = 96
height = 96
images = np.zeros((image_data.shape[2], width, height, n_channels))
slice_counter = 0
for ii in range(image_data.shape[2]):
img_patch = cv2.resize(image_data[x_beg:x_end, y_beg:y_end, ii],
dsize=(width, height))
if normalize:
if normalize == "local_max":
images[slice_counter, :, :, 0] = img_patch / np.max(img_patch)
elif normalize == "global_max":
images[slice_counter, :, :, 0] = img_patch / max_val
elif normalize == "mean_std":
images[slice_counter, :, :,
0] = (img_patch -
np.mean(img_patch)) / np.std(img_patch)
else:
raise ValueError('Please select a valid normalization')
else:
images[slice_counter, :, :, 0] = img_patch
slice_counter += 1
g = tf.Graph()
with g.as_default():
with tf.name_scope('inputs'):
x = tf.placeholder(tf.float32, [None, width, height, n_channels])
conv1 = conv_2d(x,
32,
3,
activation='relu',
padding='same',
regularizer="L2")
conv1 = conv_2d(conv1,
32,
3,
activation='relu',
padding='same',
regularizer="L2")
pool1 = max_pool_2d(conv1, 2)
conv2 = conv_2d(pool1,
64,
3,
activation='relu',
padding='same',
regularizer="L2")
conv2 = conv_2d(conv2,
64,
3,
activation='relu',
padding='same',
regularizer="L2")
pool2 = max_pool_2d(conv2, 2)
conv3 = conv_2d(pool2,
128,
3,
activation='relu',
padding='same',
regularizer="L2")
conv3 = conv_2d(conv3,
128,
3,
activation='relu',
padding='same',
regularizer="L2")
pool3 = max_pool_2d(conv3, 2)
conv4 = conv_2d(pool3,
256,
3,
activation='relu',
padding='same',
regularizer="L2")
conv4 = conv_2d(conv4,
256,
3,
activation='relu',
padding='same',
regularizer="L2")
pool4 = max_pool_2d(conv4, 2)
conv5 = conv_2d(pool4,
512,
3,
activation='relu',
padding='same',
regularizer="L2")
conv5 = conv_2d(conv5,
512,
3,
activation='relu',
padding='same',
regularizer="L2")
up6 = upsample_2d(conv5, 2)
up6 = tflearn.layers.merge_ops.merge([up6, conv4], 'concat', axis=3)
conv6 = conv_2d(up6,
256,
3,
activation='relu',
padding='same',
regularizer="L2")
conv6 = conv_2d(conv6,
256,
3,
activation='relu',
padding='same',
regularizer="L2")
up7 = upsample_2d(conv6, 2)
up7 = tflearn.layers.merge_ops.merge([up7, conv3], 'concat', axis=3)
conv7 = conv_2d(up7,
128,
3,
activation='relu',
padding='same',
regularizer="L2")
conv7 = conv_2d(conv7,
128,
3,
activation='relu',
padding='same',
regularizer="L2")
up8 = upsample_2d(conv7, 2)
up8 = tflearn.layers.merge_ops.merge([up8, conv2], 'concat', axis=3)
conv8 = conv_2d(up8,
64,
3,
activation='relu',
padding='same',
regularizer="L2")
conv8 = conv_2d(conv8,
64,
3,
activation='relu',
padding='same',
regularizer="L2")
up9 = upsample_2d(conv8, 2)
up9 = tflearn.layers.merge_ops.merge([up9, conv1], 'concat', axis=3)
conv9 = conv_2d(up9,
32,
3,
activation='relu',
padding='same',
regularizer="L2")
conv9 = conv_2d(conv9,
32,
3,
activation='relu',
padding='same',
regularizer="L2")
pred = conv_2d(conv9, 2, 1, activation='linear', padding='valid')
with tf.Session(graph=g) as sess_test_seg:
# Restore the model
tf_saver = tf.train.Saver()
tf_saver.restore(sess_test_seg, modelCkptSeg)
for idx in range(images.shape[0]):
im = np.reshape(images[idx, :, :], [1, width, height, n_channels])
feed_dict = {x: im}
pred_ = sess_test_seg.run(pred, feed_dict=feed_dict)
percentileSeg = thresholdSeg * 100
theta = np.percentile(pred_, percentileSeg)
pred_bin = np.where(pred_ > theta, 1, 0)
#Map predictions to original indices and size
pred_bin = cv2.resize(pred_bin[0, :, :, 0],
dsize=(y_end - y_beg, x_end - x_beg),
interpolation=cv2.INTER_NEAREST)
pred3dFinal[idx, x_beg:x_end, y_beg:y_end,
0] = pred_bin.astype('float64')
pppp = True
if pppp:
pred3dFinal = post_processing(np.asarray(pred3dFinal))
pred3d = [
cv2.resize(elem,
dsize=(image_data.shape[1], image_data.shape[0]),
interpolation=cv2.INTER_NEAREST) for elem in pred3dFinal
]
pred3d = np.asarray(pred3d)
upsampled = np.swapaxes(
np.swapaxes(pred3d, 1, 2), 0,
2) #if Orient module applied, no need for this line(?)
up_mask = nibabel.Nifti1Image(upsampled, img_nib.affine)
# Save
_, name, ext = split_filename(os.path.abspath(dataPath))
save_file = os.path.join(os.getcwd().replace(bidsDir, '/fetaldata'),
''.join((name, out_postfix, ext)))
nibabel.save(up_mask, save_file)
from tflearn.layers.estimator import regression
from tflearn.layers.normalization import local_response_normalization
from tflearn.layers.recurrent import gru
input_img = input_data(shape=(28, 28, 1), name='input') ; print(input_img.shape)
x = conv_2d(input_img, 16, (3, 3), activation='relu', padding='same') ; print(x.shape)
x = max_pool_2d(x, (2, 2), padding='same') ; print(x.shape)
x = conv_2d(x, 8, (3, 3), activation='relu', padding='same') ; print(x.shape)
x = max_pool_2d(x, (2, 2), padding='same') ; print(x.shape)
x = conv_2d(x, 8, (3, 3), activation='relu', padding='same') ; print(x.shape)
print("here")
encoded = max_pool_2d(x, (2, 2), padding='same') ; print(x.shape)# at this point the representation is (4, 4, 8) i.e. 128-dimensional
x = conv_2d(encoded, 8, (3, 3), activation='relu', padding='same') ; print(x.shape)
print("here")
x = upsample_2d (x, (2, 2)) ; print(x.shape)
x = conv_2d(x, 8, (3, 3), activation='relu', padding='same') ; print(x.shape)
x = upsample_2d (x, (2, 2)) ; print(x.shape)
x = conv_2d(x, 16, (3, 3), activation='relu') ; print(x.shape)
x = upsample_2d (x, (2, 2)) ; print(x.shape)
decoded = conv_2d(x, 1, (3, 3), activation='sigmoid', padding='same') ; print(x.shape)
autoencoder = regression(decoded, optimizer='momentum',loss='categorical_crossentropy',
learning_rate=0.001, name='targets')
model = tflearn.DNN(autoencoder, checkpoint_path='autoencoder',
max_checkpoints=1, tensorboard_verbose=0, tensorboard_dir='log')
from keras.datasets import mnist
import numpy as np
conv2_1 = conv_2d(pool1, 64, 3, 3, activation='relu', name="conv2_1") conv2_2 = conv_2d(conv2_1, 64, 3, 3, activation='relu', name="conv2_2") pool2 = max_pool_2d(conv2_2, 2) conv3_1 = conv_2d(pool2, 128, 3, 3, activation='relu', name="conv3_1") conv3_2 = conv_2d(conv3_1, 128, 3, 3, activation='relu', name="conv3_2") pool3 = max_pool_2d(conv3_2, 2) conv4_1 = conv_2d(pool3, 256, 3, 3, activation='relu', name="conv4_1") conv4_2 = conv_2d(conv4_1, 256, 3, 3, activation='relu', name="conv4_2") pool4 = max_pool_2d(conv4_2, 2) conv5_1 = conv_2d(pool4, 512, 3, 3, activation='relu', name="conv5_1") conv5_2 = conv_2d(conv5_1, 512, 3, 3, activation='relu', name="conv5_2") up6 = merge([upsample_2d(conv5_2, 2), conv4_2], mode='concat', axis=1, name='upsamle-5-merge-4') conv6_1 = conv_2d(up6, 256, 3, 3, activation='relu', name="conv6_1") conv6_2 = conv_2d(conv6_1, 256, 3, 3, activation='relu', name="conv6_2") up7 = merge([upsample_2d(conv6_2, 2), conv3_2], mode='concat', axis=1, name='upsamle-6-merge-3') conv7_1 = conv_2d(up7, 128, 3, 3, activation='relu', name="conv7_1") conv7_2 = conv_2d(conv7_1, 128, 3, 3, activation='relu', name="conv7_2") up8 = merge([upsample_2d(conv7_2, 2), conv2_2], mode='concat', axis=1, name='upsamle-7-merge-2') conv8_1 = conv_2d(up8, 64, 3, 3, activation='relu', name="conv8_1") conv8_2 = conv_2d(conv8_1, 64, 3, 3, activation='relu', name="conv8_2") up9 = merge([upsample_2d(conv8_2, 2), conv1_2], mode='concat', axis=1, name='upsamle-8-merge-1') conv9_1 = conv_2d(up9, 32, 3, 3, activation='relu', name="conv9_1") conv9_2 = conv_2d(conv9_1, 32, 3, 3, activation='relu', name="conv9_2")
encoder = conv_2d(encoder, 1, 7, activation='crelu') #1x10
#encoder = max_pool_2d(encoder, [2,1])
print(encoder.get_shape, file=arq)
encoder = tflearn.layers.normalization.batch_normalization(encoder)
# 4x4
# importante: 5 filtros
print(encoder.get_shape, file=arq)
#encoder = max_pool_2d(encoder, [2,2])
# 4x4
# importante: 5 filtros
print(encoder.get_shape, file=arq)
decoder = conv_2d(encoder, 8, 7, activation='crelu')
decoder = upsample_2d(decoder, [2, 2]) # 16x20
# 8x8
print(decoder.get_shape, file=arq)
decoder = conv_2d(decoder, 16, 7, activation='crelu')
decoder = upsample_2d(decoder, [1, 1]) #16x20
# 16x16
print(decoder.get_shape, file=arq)
decoder = conv_2d(decoder, 16, 7, activation='crelu')
decoder = upsample_2d(decoder, [1, 2]) #16x20
# 16x64
decoder = conv_2d(decoder, 1, 7)
print(decoder.get_shape, file=arq)
network = regression(decoder,
optimizer='adam',