def resnet_decoder_timedist(input_shape=(shape_r, shape_c, 3),
verbose=True,
print_shapes=True,
n_outs=1,
ups=8,
dil_rate=(1, 1)):
inp = Input(shape=input_shape)
### ENCODER ###
dcn = dcn_resnet(input_tensor=inp)
if print_shapes: print('resnet output shape:', dcn.output.shape)
x = Lambda(
lambda x: K.repeat_elements(
K.expand_dims(x, axis=1), nb_timestep, axis=1), lambda s:
(s[0], nb_timestep) + s[1:])(dcn.output)
## DECODER ##
outs_dec = decoder_block_timedist(x,
dil_rate=dil_rate,
print_shapes=print_shapes,
dec_filt=256)
outs_final = [outs_dec] * n_outs
# Building model
m = Model(inp, outs_final)
if verbose:
m.summary()
return m
def sam_resnet_new(input_shape = (shape_r, shape_c, 3),
conv_filters=512,
lstm_filters=512,
att_filters=512,
verbose=True,
print_shapes=True,
n_outs=1,
ups=8,
nb_gaussian=nb_gaussian):
'''SAM-ResNet ported from the original code.'''
inp = Input(shape=input_shape)
dcn = dcn_resnet(input_tensor=inp)
conv_feat = Conv2D(conv_filters, 3, padding='same', activation='relu')(dcn.output)
if print_shapes:
print('Shape after first conv after dcn_resnet:',conv_feat.shape)
# Attentive ConvLSTM
att_convlstm = Lambda(repeat, repeat_shape)(conv_feat)
att_convlstm = AttentiveConvLSTM2D(filters=lstm_filters,
attentive_filters=att_filters,
kernel_size=(3,3),
attentive_kernel_size=(3,3),
padding='same',
return_sequences=False)(att_convlstm)
# Learned Prior (1)
priors1 = LearningPrior(nb_gaussian=nb_gaussian)(att_convlstm)
concat1 = Concatenate(axis=-1)([att_convlstm, priors1])
dil_conv1 = Conv2D(conv_filters, 5, padding='same', activation='relu', dilation_rate=(4, 4))(concat1)
# Learned Prior (2)
priors2 = LearningPrior(nb_gaussian=nb_gaussian)(att_convlstm)
concat2 = Concatenate(axis=-1)([dil_conv1, priors2])
dil_conv2 = Conv2D(conv_filters, 5, padding='same', activation='relu', dilation_rate=(4, 4))(concat2)
# Final conv to get to a heatmap
outs = Conv2D(1, kernel_size=1, padding='same', activation='relu')(dil_conv2)
if print_shapes:
print('Shape after 1x1 conv:',outs.shape)
# Upsampling back to input shape
outs_up = UpSampling2D(size=(ups,ups), interpolation='bilinear')(outs)
if print_shapes:
print('shape after upsampling',outs_up.shape)
outs_final = [outs_up]*n_outs
# Building model
m = Model(inp, outs_final)
if verbose:
m.summary()
return m
def sam_resnet_nopriors(input_shape = (224, 224, 3), conv_filters=128, lstm_filters=512,
att_filters=512, verbose=True, print_shapes=True, n_outs=1, ups=8):
'''Sam ResNet with no priors.'''
inp = Input(shape=input_shape)
dcn = dcn_resnet(input_tensor=inp)
conv_feat = Conv2D(conv_filters, 3, padding='same', activation='relu')(dcn.output)
if print_shapes:
print('Shape after first conv after dcn_resnet:',conv_feat.shape)
# Attentive ConvLSTM
att_convlstm = Lambda(repeat, repeat_shape)(conv_feat)
att_convlstm = AttentiveConvLSTM2D(filters=lstm_filters, attentive_filters=att_filters, kernel_size=(3,3),
attentive_kernel_size=(3,3), padding='same', return_sequences=False)(att_convlstm)
# Dilated convolutions (priors would go here)
dil_conv1 = Conv2D(conv_filters, 5, padding='same', activation='relu', dilation_rate=(4, 4))(att_convlstm)
dil_conv2 = Conv2D(conv_filters, 5, padding='same', activation='relu', dilation_rate=(4, 4))(dil_conv1)
# Final conv to get to a heatmap
outs = Conv2D(1, kernel_size=1, padding='same', activation='relu')(dil_conv2)
if print_shapes:
print('Shape after 1x1 conv:',outs.shape)
# Upsampling back to input shape
outs_up = UpSampling2D(size=(ups,ups), interpolation='bilinear')(outs)
if print_shapes:
print('shape after upsampling',outs_up.shape)
outs_final = [outs_up]*n_outs
# Building model
m = Model(inp, outs_final)
if verbose:
m.summary()
return m
def dcnn_3stream(input_shape=(224, 224, 3),
conv_filters=512,
n_streams=3,
verbose=True,
print_shapes=True):
def out_shape(s):
print("out shape", out_shape)
return (s[0], 2) + s[1:]
def _output_stream(x):
x = Conv2D(filters=conv_filters,
kernel_size=3,
padding='same',
activation='relu')(x)
# convolve with a kernel size of 1 to flatten the tensor (w x h x nfeatures) from
# 40 x 30 x 512 to 40 x 30 x 1
# use kernel size of 1 so you ONLY flatten instead of doing local computations
# use one filter so that you only get one image channel
x = Conv2D(filters=1, kernel_size=1, padding='same',
activation='relu')(x)
# upsample from 40x30 to 320x240 (8x upsample)
x = UpSampling2D(size=(16, 16), interpolation='bilinear')(x)
x = Lambda(
lambda y: K.repeat_elements(K.expand_dims(y, axis=1), 2, axis=1),
output_shape=out_shape)(x)
return x
inp = Input(shape=input_shape)
dcn = dcn_resnet(input_tensor=inp)
outs = [_output_stream(dcn.output) for _ in range(n_streams)]
model = Model(inputs=inp, outputs=outs)
if verbose:
model.summary()
return model
def resnet_decoder(input_shape = (shape_r, shape_c, 3),
verbose=True,
print_shapes=True,
n_outs=1,
ups=8,
dil_rate = (2,2)):
inp = Input(shape=input_shape)
### ENCODER ###
dcn = dcn_resnet(input_tensor=inp)
if print_shapes: print('resnet output shape:',dcn.output.shape)
## DECODER ##
outs_dec = decoder_block(dcn.output, dil_rate=dil_rate, print_shapes=print_shapes, dec_filt=512)
outs_final = [outs_dec]*n_outs
# Building model
m = Model(inp, outs_final)
if verbose:
m.summary()
return m
def sam_resnet_3d(input_shape=(224, 224, 3),
filt_3d=128,
conv_filters=128,
lstm_filters=128,
att_filters=128,
verbose=True,
print_shapes=True,
n_outs=1,
nb_timestep=3,
ups=upsampling_factor):
inp = Input(shape=input_shape)
# Input CNN
dcn = dcn_resnet(input_tensor=inp)
conv_feat = Conv2D(conv_filters, 3, padding='same',
activation='relu')(dcn.output)
if print_shapes:
print('Shape after first conv after dcn_resnet:', conv_feat.shape)
# Attentive ConvLSTM
att_convlstm = Lambda(
lambda x: K.repeat_elements(
K.expand_dims(x, axis=1), nb_timestep, axis=1), lambda s:
(s[0], nb_timestep) + s[1:])(x)
att_convlstm = AttentiveConvLSTM2D(filters=lstm_filters,
attentive_filters=att_filters,
kernel_size=(3, 3),
attentive_kernel_size=(3, 3),
padding='same',
return_sequences=True)(att_convlstm)
if print_shapes: print('att_convlstm output shape', att_convlstm.shape)
# Output flow
x = Conv3D(filt_3d, (3, 3, 3),
strides=(1, 1, 1),
padding='same',
dilation_rate=(4, 4, 1),
activation=None,
kernel_initializer='he_normal')(att_convlstm)
x = BatchNormalization(axis=4)(x)
x = Activation('relu')(x)
x = Conv3D(filt_3d, (3, 3, 1),
strides=(1, 1, 1),
padding='same',
dilation_rate=(4, 4, 1),
activation=None,
kernel_initializer='he_normal')(x)
x = BatchNormalization(axis=4)(x)
x = Activation('relu')(x)
x = Conv3D(1, (1, 1, 1),
strides=(1, 1, 1),
padding='same',
activation='relu',
kernel_initializer='he_normal')(x)
out_final = TimeDistributed(
UpSampling2D(size=(ups, ups), interpolation='bilinear'))(x)
if print_shapes: print('outs_final shape:', outs_final.shape)
outs_final = [outs_final] * n_outs
m = Model(inp, outs_final)
if verbose:
m.summary()
return m