Deep Learning básico con Keras (Parte 5): DenseNet
Publicado por Jesús Utrera Burgal el
En este artículo vamos a mostrar la arquitectura DenseNet. Ésta fue introducida en el año 2016, consiguiendo en 2017 el premio CVPR 2017 Best Paper Award. El siguiente enlace nos lleva al paper: https://arxiv.org/abs/1608.06993, pero aconsejo el artículo del blog Towards Data Science, pues explica muy bien su funcionamiento.
La idea, muy resumida, se basa en concatenar cada salida de las capas previas hacia las siguientes:
La imagen anterior representa un bloque denso de 5 capas con una tasa de crecimiento k = 4. Cada capa toma como entrada todos los valores de características anteriores.
Como se puede ver en la imagen anterior, donde se representa una red de tres bloques densos, las capas entre dos bloques adyacentes hace referencia a una capa de transición cambiando el tamaño del mapa de características mediante convolución y pooling.
Entrenando la arquitectura DenseNet
Keras tiene a nuestra disposición ésta arquitectura, pero tiene la restricción de que, por defecto, el tamaño de las imágenes debe ser mayor a 187 píxeles, por lo que definiremos una arquitectura más pequeña.
from keras.applications import densenet
from keras.applications import imagenet_utils as imut
def CustomDenseNet(blocks,
include_top=True,
input_tensor=None,
weights=None,
input_shape=None,
pooling=None,
classes=1000):
# Determine proper input shape
input_shape = imut._obtain_input_shape(input_shape,
default_size=224,
min_size=32,
data_format=K.image_data_format(),
require_flatten=include_top,
weights=weights)
if input_tensor is None:
img_input = Input(shape=input_shape)
else:
if not K.is_keras_tensor(input_tensor):
img_input = Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
bn_axis = 3 if K.image_data_format() == 'channels_last' else 1
x = ZeroPadding2D(padding=((3, 3), (3, 3)))(img_input)
x = Conv2D(64, 7, strides=2, use_bias=False, name='conv1/conv')(x)
x = BatchNormalization(axis=bn_axis, epsilon=1.001e-5,
name='conv1/bn')(x)
x = Activation('relu', name='conv1/relu')(x)
x = ZeroPadding2D(padding=((1, 1), (1, 1)))(x)
x = MaxPooling2D(3, strides=2, name='pool1')(x)
x = densenet.dense_block(x, blocks[0], name='conv2')
x = densenet.transition_block(x, 0.5, name='pool2')
x = densenet.dense_block(x, blocks[1], name='conv3')
x = densenet.transition_block(x, 0.5, name='pool3')
x = densenet.dense_block(x, blocks[2], name='conv4')
x = densenet.transition_block(x, 0.5, name='pool4')
x = densenet.dense_block(x, blocks[3], name='conv5')
x = BatchNormalization(axis=bn_axis, epsilon=1.001e-5,
name='bn')(x)
if include_top:
x = GlobalAveragePooling2D(name='avg_pool')(x)
x = Dense(classes, activation='softmax', name='fc1000')(x)
else:
if pooling == 'avg':
x = GlobalAveragePooling2D(name='avg_pool')(x)
elif pooling == 'max':
x = GlobalMaxPooling2D(name='max_pool')(x)
# Ensure that the model takes into account
# any potential predecessors of `input_tensor`.
if input_tensor is not None:
inputs = imut.get_source_inputs(input_tensor)
else:
inputs = img_input
# Create model.
if blocks == [6, 12, 24, 16]:
model = Model(inputs, x, name='densenet121')
elif blocks == [6, 12, 32, 32]:
model = Model(inputs, x, name='densenet169')
elif blocks == [6, 12, 48, 32]:
model = Model(inputs, x, name='densenet201')
else:
model = Model(inputs, x, name='densenet')
# Load weights.
if weights == 'imagenet':
if include_top:
if blocks == [6, 12, 24, 16]:
weights_path = get_file(
'densenet121_weights_tf_dim_ordering_tf_kernels.h5',
DENSENET121_WEIGHT_PATH,
cache_subdir='models',
file_hash='0962ca643bae20f9b6771cb844dca3b0')
elif blocks == [6, 12, 32, 32]:
weights_path = get_file(
'densenet169_weights_tf_dim_ordering_tf_kernels.h5',
DENSENET169_WEIGHT_PATH,
cache_subdir='models',
file_hash='bcf9965cf5064a5f9eb6d7dc69386f43')
elif blocks == [6, 12, 48, 32]:
weights_path = get_file(
'densenet201_weights_tf_dim_ordering_tf_kernels.h5',
DENSENET201_WEIGHT_PATH,
cache_subdir='models',
file_hash='7bb75edd58cb43163be7e0005fbe95ef')
else:
if blocks == [6, 12, 24, 16]:
weights_path = get_file(
'densenet121_weights_tf_dim_ordering_tf_kernels_notop.h5',
DENSENET121_WEIGHT_PATH_NO_TOP,
cache_subdir='models',
file_hash='4912a53fbd2a69346e7f2c0b5ec8c6d3')
elif blocks == [6, 12, 32, 32]:
weights_path = get_file(
'densenet169_weights_tf_dim_ordering_tf_kernels_notop.h5',
DENSENET169_WEIGHT_PATH_NO_TOP,
cache_subdir='models',
file_hash='50662582284e4cf834ce40ab4dfa58c6')
elif blocks == [6, 12, 48, 32]:
weights_path = get_file(
'densenet201_weights_tf_dim_ordering_tf_kernels_notop.h5',
DENSENET201_WEIGHT_PATH_NO_TOP,
cache_subdir='models',
file_hash='1c2de60ee40562448dbac34a0737e798')
model.load_weights(weights_path)
elif weights is not None:
model.load_weights(weights)
return model
def create_densenet():
base_model = CustomDenseNet([6, 12, 16, 8], include_top=False, weights=None, input_tensor=None, input_shape=(32,32,3), pooling=None, classes=100)
x = base_model.output
x = GlobalAveragePooling2D(name='avg_pool')(x)
x = Dense(500)(x)
x = Activation('relu')(x)
x = Dropout(0.5)(x)
predictions = Dense(100, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
return model
Compilamos como hasta ahora...
custom_dense_model = create_densenet()
custom_dense_model.compile(loss='categorical_crossentropy', optimizer='sgd', metrics=['acc', 'mse'])
Una vez hecho esto, vamos a ver un resumen del modelo creado:
custom_dense_model.summary()
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) (None, 32, 32, 3) 0
__________________________________________________________________________________________________
zero_padding2d_1 (ZeroPadding2D (None, 38, 38, 3) 0 input_1[0][0]
__________________________________________________________________________________________________
conv1/conv (Conv2D) (None, 16, 16, 64) 9408 zero_padding2d_1[0][0]
__________________________________________________________________________________________________
conv1/bn (BatchNormalization) (None, 16, 16, 64) 256 conv1/conv[0][0]
__________________________________________________________________________________________________
conv1/relu (Activation) (None, 16, 16, 64) 0 conv1/bn[0][0]
__________________________________________________________________________________________________
zero_padding2d_2 (ZeroPadding2D (None, 18, 18, 64) 0 conv1/relu[0][0]
__________________________________________________________________________________________________
pool1 (MaxPooling2D) (None, 8, 8, 64) 0 zero_padding2d_2[0][0]
__________________________________________________________________________________________________
conv2_block1_0_bn (BatchNormali (None, 8, 8, 64) 256 pool1[0][0]
__________________________________________________________________________________________________
conv2_block1_0_relu (Activation (None, 8, 8, 64) 0 conv2_block1_0_bn[0][0]
__________________________________________________________________________________________________
conv2_block1_1_conv (Conv2D) (None, 8, 8, 128) 8192 conv2_block1_0_relu[0][0]
__________________________________________________________________________________________________
conv2_block1_1_bn (BatchNormali (None, 8, 8, 128) 512 conv2_block1_1_conv[0][0]
__________________________________________________________________________________________________
conv2_block1_1_relu (Activation (None, 8, 8, 128) 0 conv2_block1_1_bn[0][0]
__________________________________________________________________________________________________
conv2_block1_2_conv (Conv2D) (None, 8, 8, 32) 36864 conv2_block1_1_relu[0][0]
__________________________________________________________________________________________________
conv2_block1_concat (Concatenat (None, 8, 8, 96) 0 pool1[0][0]
conv2_block1_2_conv[0][0]
__________________________________________________________________________________________________
conv2_block2_0_bn (BatchNormali (None, 8, 8, 96) 384 conv2_block1_concat[0][0]
__________________________________________________________________________________________________
conv2_block2_0_relu (Activation (None, 8, 8, 96) 0 conv2_block2_0_bn[0][0]
__________________________________________________________________________________________________
conv2_block2_1_conv (Conv2D) (None, 8, 8, 128) 12288 conv2_block2_0_relu[0][0]
__________________________________________________________________________________________________
conv2_block2_1_bn (BatchNormali (None, 8, 8, 128) 512 conv2_block2_1_conv[0][0]
__________________________________________________________________________________________________
conv2_block2_1_relu (Activation (None, 8, 8, 128) 0 conv2_block2_1_bn[0][0]
__________________________________________________________________________________________________
conv2_block2_2_conv (Conv2D) (None, 8, 8, 32) 36864 conv2_block2_1_relu[0][0]
__________________________________________________________________________________________________
conv2_block2_concat (Concatenat (None, 8, 8, 128) 0 conv2_block1_concat[0][0]
conv2_block2_2_conv[0][0]
__________________________________________________________________________________________________
conv2_block3_0_bn (BatchNormali (None, 8, 8, 128) 512 conv2_block2_concat[0][0]
__________________________________________________________________________________________________
conv2_block3_0_relu (Activation (None, 8, 8, 128) 0 conv2_block3_0_bn[0][0]
__________________________________________________________________________________________________
conv2_block3_1_conv (Conv2D) (None, 8, 8, 128) 16384 conv2_block3_0_relu[0][0]
__________________________________________________________________________________________________
conv2_block3_1_bn (BatchNormali (None, 8, 8, 128) 512 conv2_block3_1_conv[0][0]
__________________________________________________________________________________________________
conv2_block3_1_relu (Activation (None, 8, 8, 128) 0 conv2_block3_1_bn[0][0]
__________________________________________________________________________________________________
conv2_block3_2_conv (Conv2D) (None, 8, 8, 32) 36864 conv2_block3_1_relu[0][0]
__________________________________________________________________________________________________
conv2_block3_concat (Concatenat (None, 8, 8, 160) 0 conv2_block2_concat[0][0]
conv2_block3_2_conv[0][0]
__________________________________________________________________________________________________
conv2_block4_0_bn (BatchNormali (None, 8, 8, 160) 640 conv2_block3_concat[0][0]
__________________________________________________________________________________________________
conv2_block4_0_relu (Activation (None, 8, 8, 160) 0 conv2_block4_0_bn[0][0]
__________________________________________________________________________________________________
conv2_block4_1_conv (Conv2D) (None, 8, 8, 128) 20480 conv2_block4_0_relu[0][0]
__________________________________________________________________________________________________
conv2_block4_1_bn (BatchNormali (None, 8, 8, 128) 512 conv2_block4_1_conv[0][0]
__________________________________________________________________________________________________
conv2_block4_1_relu (Activation (None, 8, 8, 128) 0 conv2_block4_1_bn[0][0]
__________________________________________________________________________________________________
conv2_block4_2_conv (Conv2D) (None, 8, 8, 32) 36864 conv2_block4_1_relu[0][0]
__________________________________________________________________________________________________
conv2_block4_concat (Concatenat (None, 8, 8, 192) 0 conv2_block3_concat[0][0]
conv2_block4_2_conv[0][0]
__________________________________________________________________________________________________
conv2_block5_0_bn (BatchNormali (None, 8, 8, 192) 768 conv2_block4_concat[0][0]
__________________________________________________________________________________________________
conv2_block5_0_relu (Activation (None, 8, 8, 192) 0 conv2_block5_0_bn[0][0]
__________________________________________________________________________________________________
conv2_block5_1_conv (Conv2D) (None, 8, 8, 128) 24576 conv2_block5_0_relu[0][0]
__________________________________________________________________________________________________
conv2_block5_1_bn (BatchNormali (None, 8, 8, 128) 512 conv2_block5_1_conv[0][0]
__________________________________________________________________________________________________
conv2_block5_1_relu (Activation (None, 8, 8, 128) 0 conv2_block5_1_bn[0][0]
__________________________________________________________________________________________________
conv2_block5_2_conv (Conv2D) (None, 8, 8, 32) 36864 conv2_block5_1_relu[0][0]
__________________________________________________________________________________________________
conv2_block5_concat (Concatenat (None, 8, 8, 224) 0 conv2_block4_concat[0][0]
conv2_block5_2_conv[0][0]
__________________________________________________________________________________________________
conv2_block6_0_bn (BatchNormali (None, 8, 8, 224) 896 conv2_block5_concat[0][0]
__________________________________________________________________________________________________
conv2_block6_0_relu (Activation (None, 8, 8, 224) 0 conv2_block6_0_bn[0][0]
__________________________________________________________________________________________________
conv2_block6_1_conv (Conv2D) (None, 8, 8, 128) 28672 conv2_block6_0_relu[0][0]
__________________________________________________________________________________________________
conv2_block6_1_bn (BatchNormali (None, 8, 8, 128) 512 conv2_block6_1_conv[0][0]
__________________________________________________________________________________________________
conv2_block6_1_relu (Activation (None, 8, 8, 128) 0 conv2_block6_1_bn[0][0]
__________________________________________________________________________________________________
conv2_block6_2_conv (Conv2D) (None, 8, 8, 32) 36864 conv2_block6_1_relu[0][0]
__________________________________________________________________________________________________
conv2_block6_concat (Concatenat (None, 8, 8, 256) 0 conv2_block5_concat[0][0]
conv2_block6_2_conv[0][0]
__________________________________________________________________________________________________
pool2_bn (BatchNormalization) (None, 8, 8, 256) 1024 conv2_block6_concat[0][0]
__________________________________________________________________________________________________
pool2_relu (Activation) (None, 8, 8, 256) 0 pool2_bn[0][0]
__________________________________________________________________________________________________
pool2_conv (Conv2D) (None, 8, 8, 128) 32768 pool2_relu[0][0]
__________________________________________________________________________________________________
pool2_pool (AveragePooling2D) (None, 4, 4, 128) 0 pool2_conv[0][0]
__________________________________________________________________________________________________
conv3_block1_0_bn (BatchNormali (None, 4, 4, 128) 512 pool2_pool[0][0]
__________________________________________________________________________________________________
conv3_block1_0_relu (Activation (None, 4, 4, 128) 0 conv3_block1_0_bn[0][0]
__________________________________________________________________________________________________
conv3_block1_1_conv (Conv2D) (None, 4, 4, 128) 16384 conv3_block1_0_relu[0][0]
__________________________________________________________________________________________________
conv3_block1_1_bn (BatchNormali (None, 4, 4, 128) 512 conv3_block1_1_conv[0][0]
__________________________________________________________________________________________________
conv3_block1_1_relu (Activation (None, 4, 4, 128) 0 conv3_block1_1_bn[0][0]
__________________________________________________________________________________________________
conv3_block1_2_conv (Conv2D) (None, 4, 4, 32) 36864 conv3_block1_1_relu[0][0]
__________________________________________________________________________________________________
conv3_block1_concat (Concatenat (None, 4, 4, 160) 0 pool2_pool[0][0]
conv3_block1_2_conv[0][0]
__________________________________________________________________________________________________
conv3_block2_0_bn (BatchNormali (None, 4, 4, 160) 640 conv3_block1_concat[0][0]
__________________________________________________________________________________________________
conv3_block2_0_relu (Activation (None, 4, 4, 160) 0 conv3_block2_0_bn[0][0]
__________________________________________________________________________________________________
conv3_block2_1_conv (Conv2D) (None, 4, 4, 128) 20480 conv3_block2_0_relu[0][0]
__________________________________________________________________________________________________
conv3_block2_1_bn (BatchNormali (None, 4, 4, 128) 512 conv3_block2_1_conv[0][0]
__________________________________________________________________________________________________
conv3_block2_1_relu (Activation (None, 4, 4, 128) 0 conv3_block2_1_bn[0][0]
__________________________________________________________________________________________________
conv3_block2_2_conv (Conv2D) (None, 4, 4, 32) 36864 conv3_block2_1_relu[0][0]
__________________________________________________________________________________________________
conv3_block2_concat (Concatenat (None, 4, 4, 192) 0 conv3_block1_concat[0][0]
conv3_block2_2_conv[0][0]
__________________________________________________________________________________________________
conv3_block3_0_bn (BatchNormali (None, 4, 4, 192) 768 conv3_block2_concat[0][0]
__________________________________________________________________________________________________
conv3_block3_0_relu (Activation (None, 4, 4, 192) 0 conv3_block3_0_bn[0][0]
__________________________________________________________________________________________________
conv3_block3_1_conv (Conv2D) (None, 4, 4, 128) 24576 conv3_block3_0_relu[0][0]
__________________________________________________________________________________________________
conv3_block3_1_bn (BatchNormali (None, 4, 4, 128) 512 conv3_block3_1_conv[0][0]
__________________________________________________________________________________________________
conv3_block3_1_relu (Activation (None, 4, 4, 128) 0 conv3_block3_1_bn[0][0]
__________________________________________________________________________________________________
conv3_block3_2_conv (Conv2D) (None, 4, 4, 32) 36864 conv3_block3_1_relu[0][0]
__________________________________________________________________________________________________
conv3_block3_concat (Concatenat (None, 4, 4, 224) 0 conv3_block2_concat[0][0]
conv3_block3_2_conv[0][0]
__________________________________________________________________________________________________
conv3_block4_0_bn (BatchNormali (None, 4, 4, 224) 896 conv3_block3_concat[0][0]
__________________________________________________________________________________________________
conv3_block4_0_relu (Activation (None, 4, 4, 224) 0 conv3_block4_0_bn[0][0]
__________________________________________________________________________________________________
conv3_block4_1_conv (Conv2D) (None, 4, 4, 128) 28672 conv3_block4_0_relu[0][0]
__________________________________________________________________________________________________
conv3_block4_1_bn (BatchNormali (None, 4, 4, 128) 512 conv3_block4_1_conv[0][0]
__________________________________________________________________________________________________
conv3_block4_1_relu (Activation (None, 4, 4, 128) 0 conv3_block4_1_bn[0][0]
__________________________________________________________________________________________________
conv3_block4_2_conv (Conv2D) (None, 4, 4, 32) 36864 conv3_block4_1_relu[0][0]
__________________________________________________________________________________________________
conv3_block4_concat (Concatenat (None, 4, 4, 256) 0 conv3_block3_concat[0][0]
conv3_block4_2_conv[0][0]
__________________________________________________________________________________________________
conv3_block5_0_bn (BatchNormali (None, 4, 4, 256) 1024 conv3_block4_concat[0][0]
__________________________________________________________________________________________________
conv3_block5_0_relu (Activation (None, 4, 4, 256) 0 conv3_block5_0_bn[0][0]
__________________________________________________________________________________________________
conv3_block5_1_conv (Conv2D) (None, 4, 4, 128) 32768 conv3_block5_0_relu[0][0]
__________________________________________________________________________________________________
conv3_block5_1_bn (BatchNormali (None, 4, 4, 128) 512 conv3_block5_1_conv[0][0]
__________________________________________________________________________________________________
conv3_block5_1_relu (Activation (None, 4, 4, 128) 0 conv3_block5_1_bn[0][0]
__________________________________________________________________________________________________
conv3_block5_2_conv (Conv2D) (None, 4, 4, 32) 36864 conv3_block5_1_relu[0][0]
__________________________________________________________________________________________________
conv3_block5_concat (Concatenat (None, 4, 4, 288) 0 conv3_block4_concat[0][0]
conv3_block5_2_conv[0][0]
__________________________________________________________________________________________________
conv3_block6_0_bn (BatchNormali (None, 4, 4, 288) 1152 conv3_block5_concat[0][0]
__________________________________________________________________________________________________
conv3_block6_0_relu (Activation (None, 4, 4, 288) 0 conv3_block6_0_bn[0][0]
__________________________________________________________________________________________________
conv3_block6_1_conv (Conv2D) (None, 4, 4, 128) 36864 conv3_block6_0_relu[0][0]
__________________________________________________________________________________________________
conv3_block6_1_bn (BatchNormali (None, 4, 4, 128) 512 conv3_block6_1_conv[0][0]
__________________________________________________________________________________________________
conv3_block6_1_relu (Activation (None, 4, 4, 128) 0 conv3_block6_1_bn[0][0]
__________________________________________________________________________________________________
conv3_block6_2_conv (Conv2D) (None, 4, 4, 32) 36864 conv3_block6_1_relu[0][0]
__________________________________________________________________________________________________
conv3_block6_concat (Concatenat (None, 4, 4, 320) 0 conv3_block5_concat[0][0]
conv3_block6_2_conv[0][0]
__________________________________________________________________________________________________
conv3_block7_0_bn (BatchNormali (None, 4, 4, 320) 1280 conv3_block6_concat[0][0]
__________________________________________________________________________________________________
conv3_block7_0_relu (Activation (None, 4, 4, 320) 0 conv3_block7_0_bn[0][0]
__________________________________________________________________________________________________
conv3_block7_1_conv (Conv2D) (None, 4, 4, 128) 40960 conv3_block7_0_relu[0][0]
__________________________________________________________________________________________________
conv3_block7_1_bn (BatchNormali (None, 4, 4, 128) 512 conv3_block7_1_conv[0][0]
__________________________________________________________________________________________________
conv3_block7_1_relu (Activation (None, 4, 4, 128) 0 conv3_block7_1_bn[0][0]
__________________________________________________________________________________________________
conv3_block7_2_conv (Conv2D) (None, 4, 4, 32) 36864 conv3_block7_1_relu[0][0]
__________________________________________________________________________________________________
conv3_block7_concat (Concatenat (None, 4, 4, 352) 0 conv3_block6_concat[0][0]
conv3_block7_2_conv[0][0]
__________________________________________________________________________________________________
conv3_block8_0_bn (BatchNormali (None, 4, 4, 352) 1408 conv3_block7_concat[0][0]
__________________________________________________________________________________________________
conv3_block8_0_relu (Activation (None, 4, 4, 352) 0 conv3_block8_0_bn[0][0]
__________________________________________________________________________________________________
conv3_block8_1_conv (Conv2D) (None, 4, 4, 128) 45056 conv3_block8_0_relu[0][0]
__________________________________________________________________________________________________
conv3_block8_1_bn (BatchNormali (None, 4, 4, 128) 512 conv3_block8_1_conv[0][0]
__________________________________________________________________________________________________
conv3_block8_1_relu (Activation (None, 4, 4, 128) 0 conv3_block8_1_bn[0][0]
__________________________________________________________________________________________________
conv3_block8_2_conv (Conv2D) (None, 4, 4, 32) 36864 conv3_block8_1_relu[0][0]
__________________________________________________________________________________________________
conv3_block8_concat (Concatenat (None, 4, 4, 384) 0 conv3_block7_concat[0][0]
conv3_block8_2_conv[0][0]
__________________________________________________________________________________________________
conv3_block9_0_bn (BatchNormali (None, 4, 4, 384) 1536 conv3_block8_concat[0][0]
__________________________________________________________________________________________________
conv3_block9_0_relu (Activation (None, 4, 4, 384) 0 conv3_block9_0_bn[0][0]
__________________________________________________________________________________________________
conv3_block9_1_conv (Conv2D) (None, 4, 4, 128) 49152 conv3_block9_0_relu[0][0]
__________________________________________________________________________________________________
conv3_block9_1_bn (BatchNormali (None, 4, 4, 128) 512 conv3_block9_1_conv[0][0]
__________________________________________________________________________________________________
conv3_block9_1_relu (Activation (None, 4, 4, 128) 0 conv3_block9_1_bn[0][0]
__________________________________________________________________________________________________
conv3_block9_2_conv (Conv2D) (None, 4, 4, 32) 36864 conv3_block9_1_relu[0][0]
__________________________________________________________________________________________________
conv3_block9_concat (Concatenat (None, 4, 4, 416) 0 conv3_block8_concat[0][0]
conv3_block9_2_conv[0][0]
__________________________________________________________________________________________________
conv3_block10_0_bn (BatchNormal (None, 4, 4, 416) 1664 conv3_block9_concat[0][0]
__________________________________________________________________________________________________
conv3_block10_0_relu (Activatio (None, 4, 4, 416) 0 conv3_block10_0_bn[0][0]
__________________________________________________________________________________________________
conv3_block10_1_conv (Conv2D) (None, 4, 4, 128) 53248 conv3_block10_0_relu[0][0]
__________________________________________________________________________________________________
conv3_block10_1_bn (BatchNormal (None, 4, 4, 128) 512 conv3_block10_1_conv[0][0]
__________________________________________________________________________________________________
conv3_block10_1_relu (Activatio (None, 4, 4, 128) 0 conv3_block10_1_bn[0][0]
__________________________________________________________________________________________________
conv3_block10_2_conv (Conv2D) (None, 4, 4, 32) 36864 conv3_block10_1_relu[0][0]
__________________________________________________________________________________________________
conv3_block10_concat (Concatena (None, 4, 4, 448) 0 conv3_block9_concat[0][0]
conv3_block10_2_conv[0][0]
__________________________________________________________________________________________________
conv3_block11_0_bn (BatchNormal (None, 4, 4, 448) 1792 conv3_block10_concat[0][0]
__________________________________________________________________________________________________
conv3_block11_0_relu (Activatio (None, 4, 4, 448) 0 conv3_block11_0_bn[0][0]
__________________________________________________________________________________________________
conv3_block11_1_conv (Conv2D) (None, 4, 4, 128) 57344 conv3_block11_0_relu[0][0]
__________________________________________________________________________________________________
conv3_block11_1_bn (BatchNormal (None, 4, 4, 128) 512 conv3_block11_1_conv[0][0]
__________________________________________________________________________________________________
conv3_block11_1_relu (Activatio (None, 4, 4, 128) 0 conv3_block11_1_bn[0][0]
__________________________________________________________________________________________________
conv3_block11_2_conv (Conv2D) (None, 4, 4, 32) 36864 conv3_block11_1_relu[0][0]
__________________________________________________________________________________________________
conv3_block11_concat (Concatena (None, 4, 4, 480) 0 conv3_block10_concat[0][0]
conv3_block11_2_conv[0][0]
__________________________________________________________________________________________________
conv3_block12_0_bn (BatchNormal (None, 4, 4, 480) 1920 conv3_block11_concat[0][0]
__________________________________________________________________________________________________
conv3_block12_0_relu (Activatio (None, 4, 4, 480) 0 conv3_block12_0_bn[0][0]
__________________________________________________________________________________________________
conv3_block12_1_conv (Conv2D) (None, 4, 4, 128) 61440 conv3_block12_0_relu[0][0]
__________________________________________________________________________________________________
conv3_block12_1_bn (BatchNormal (None, 4, 4, 128) 512 conv3_block12_1_conv[0][0]
__________________________________________________________________________________________________
conv3_block12_1_relu (Activatio (None, 4, 4, 128) 0 conv3_block12_1_bn[0][0]
__________________________________________________________________________________________________
conv3_block12_2_conv (Conv2D) (None, 4, 4, 32) 36864 conv3_block12_1_relu[0][0]
__________________________________________________________________________________________________
conv3_block12_concat (Concatena (None, 4, 4, 512) 0 conv3_block11_concat[0][0]
conv3_block12_2_conv[0][0]
__________________________________________________________________________________________________
pool3_bn (BatchNormalization) (None, 4, 4, 512) 2048 conv3_block12_concat[0][0]
__________________________________________________________________________________________________
pool3_relu (Activation) (None, 4, 4, 512) 0 pool3_bn[0][0]
__________________________________________________________________________________________________
pool3_conv (Conv2D) (None, 4, 4, 256) 131072 pool3_relu[0][0]
__________________________________________________________________________________________________
pool3_pool (AveragePooling2D) (None, 2, 2, 256) 0 pool3_conv[0][0]
__________________________________________________________________________________________________
conv4_block1_0_bn (BatchNormali (None, 2, 2, 256) 1024 pool3_pool[0][0]
__________________________________________________________________________________________________
conv4_block1_0_relu (Activation (None, 2, 2, 256) 0 conv4_block1_0_bn[0][0]
__________________________________________________________________________________________________
conv4_block1_1_conv (Conv2D) (None, 2, 2, 128) 32768 conv4_block1_0_relu[0][0]
__________________________________________________________________________________________________
conv4_block1_1_bn (BatchNormali (None, 2, 2, 128) 512 conv4_block1_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block1_1_relu (Activation (None, 2, 2, 128) 0 conv4_block1_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block1_2_conv (Conv2D) (None, 2, 2, 32) 36864 conv4_block1_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block1_concat (Concatenat (None, 2, 2, 288) 0 pool3_pool[0][0]
conv4_block1_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block2_0_bn (BatchNormali (None, 2, 2, 288) 1152 conv4_block1_concat[0][0]
__________________________________________________________________________________________________
conv4_block2_0_relu (Activation (None, 2, 2, 288) 0 conv4_block2_0_bn[0][0]
__________________________________________________________________________________________________
conv4_block2_1_conv (Conv2D) (None, 2, 2, 128) 36864 conv4_block2_0_relu[0][0]
__________________________________________________________________________________________________
conv4_block2_1_bn (BatchNormali (None, 2, 2, 128) 512 conv4_block2_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block2_1_relu (Activation (None, 2, 2, 128) 0 conv4_block2_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block2_2_conv (Conv2D) (None, 2, 2, 32) 36864 conv4_block2_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block2_concat (Concatenat (None, 2, 2, 320) 0 conv4_block1_concat[0][0]
conv4_block2_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block3_0_bn (BatchNormali (None, 2, 2, 320) 1280 conv4_block2_concat[0][0]
__________________________________________________________________________________________________
conv4_block3_0_relu (Activation (None, 2, 2, 320) 0 conv4_block3_0_bn[0][0]
__________________________________________________________________________________________________
conv4_block3_1_conv (Conv2D) (None, 2, 2, 128) 40960 conv4_block3_0_relu[0][0]
__________________________________________________________________________________________________
conv4_block3_1_bn (BatchNormali (None, 2, 2, 128) 512 conv4_block3_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block3_1_relu (Activation (None, 2, 2, 128) 0 conv4_block3_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block3_2_conv (Conv2D) (None, 2, 2, 32) 36864 conv4_block3_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block3_concat (Concatenat (None, 2, 2, 352) 0 conv4_block2_concat[0][0]
conv4_block3_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block4_0_bn (BatchNormali (None, 2, 2, 352) 1408 conv4_block3_concat[0][0]
__________________________________________________________________________________________________
conv4_block4_0_relu (Activation (None, 2, 2, 352) 0 conv4_block4_0_bn[0][0]
__________________________________________________________________________________________________
conv4_block4_1_conv (Conv2D) (None, 2, 2, 128) 45056 conv4_block4_0_relu[0][0]
__________________________________________________________________________________________________
conv4_block4_1_bn (BatchNormali (None, 2, 2, 128) 512 conv4_block4_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block4_1_relu (Activation (None, 2, 2, 128) 0 conv4_block4_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block4_2_conv (Conv2D) (None, 2, 2, 32) 36864 conv4_block4_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block4_concat (Concatenat (None, 2, 2, 384) 0 conv4_block3_concat[0][0]
conv4_block4_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block5_0_bn (BatchNormali (None, 2, 2, 384) 1536 conv4_block4_concat[0][0]
__________________________________________________________________________________________________
conv4_block5_0_relu (Activation (None, 2, 2, 384) 0 conv4_block5_0_bn[0][0]
__________________________________________________________________________________________________
conv4_block5_1_conv (Conv2D) (None, 2, 2, 128) 49152 conv4_block5_0_relu[0][0]
__________________________________________________________________________________________________
conv4_block5_1_bn (BatchNormali (None, 2, 2, 128) 512 conv4_block5_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block5_1_relu (Activation (None, 2, 2, 128) 0 conv4_block5_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block5_2_conv (Conv2D) (None, 2, 2, 32) 36864 conv4_block5_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block5_concat (Concatenat (None, 2, 2, 416) 0 conv4_block4_concat[0][0]
conv4_block5_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block6_0_bn (BatchNormali (None, 2, 2, 416) 1664 conv4_block5_concat[0][0]
__________________________________________________________________________________________________
conv4_block6_0_relu (Activation (None, 2, 2, 416) 0 conv4_block6_0_bn[0][0]
__________________________________________________________________________________________________
conv4_block6_1_conv (Conv2D) (None, 2, 2, 128) 53248 conv4_block6_0_relu[0][0]
__________________________________________________________________________________________________
conv4_block6_1_bn (BatchNormali (None, 2, 2, 128) 512 conv4_block6_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block6_1_relu (Activation (None, 2, 2, 128) 0 conv4_block6_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block6_2_conv (Conv2D) (None, 2, 2, 32) 36864 conv4_block6_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block6_concat (Concatenat (None, 2, 2, 448) 0 conv4_block5_concat[0][0]
conv4_block6_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block7_0_bn (BatchNormali (None, 2, 2, 448) 1792 conv4_block6_concat[0][0]
__________________________________________________________________________________________________
conv4_block7_0_relu (Activation (None, 2, 2, 448) 0 conv4_block7_0_bn[0][0]
__________________________________________________________________________________________________
conv4_block7_1_conv (Conv2D) (None, 2, 2, 128) 57344 conv4_block7_0_relu[0][0]
__________________________________________________________________________________________________
conv4_block7_1_bn (BatchNormali (None, 2, 2, 128) 512 conv4_block7_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block7_1_relu (Activation (None, 2, 2, 128) 0 conv4_block7_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block7_2_conv (Conv2D) (None, 2, 2, 32) 36864 conv4_block7_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block7_concat (Concatenat (None, 2, 2, 480) 0 conv4_block6_concat[0][0]
conv4_block7_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block8_0_bn (BatchNormali (None, 2, 2, 480) 1920 conv4_block7_concat[0][0]
__________________________________________________________________________________________________
conv4_block8_0_relu (Activation (None, 2, 2, 480) 0 conv4_block8_0_bn[0][0]
__________________________________________________________________________________________________
conv4_block8_1_conv (Conv2D) (None, 2, 2, 128) 61440 conv4_block8_0_relu[0][0]
__________________________________________________________________________________________________
conv4_block8_1_bn (BatchNormali (None, 2, 2, 128) 512 conv4_block8_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block8_1_relu (Activation (None, 2, 2, 128) 0 conv4_block8_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block8_2_conv (Conv2D) (None, 2, 2, 32) 36864 conv4_block8_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block8_concat (Concatenat (None, 2, 2, 512) 0 conv4_block7_concat[0][0]
conv4_block8_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block9_0_bn (BatchNormali (None, 2, 2, 512) 2048 conv4_block8_concat[0][0]
__________________________________________________________________________________________________
conv4_block9_0_relu (Activation (None, 2, 2, 512) 0 conv4_block9_0_bn[0][0]
__________________________________________________________________________________________________
conv4_block9_1_conv (Conv2D) (None, 2, 2, 128) 65536 conv4_block9_0_relu[0][0]
__________________________________________________________________________________________________
conv4_block9_1_bn (BatchNormali (None, 2, 2, 128) 512 conv4_block9_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block9_1_relu (Activation (None, 2, 2, 128) 0 conv4_block9_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block9_2_conv (Conv2D) (None, 2, 2, 32) 36864 conv4_block9_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block9_concat (Concatenat (None, 2, 2, 544) 0 conv4_block8_concat[0][0]
conv4_block9_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block10_0_bn (BatchNormal (None, 2, 2, 544) 2176 conv4_block9_concat[0][0]
__________________________________________________________________________________________________
conv4_block10_0_relu (Activatio (None, 2, 2, 544) 0 conv4_block10_0_bn[0][0]
__________________________________________________________________________________________________
conv4_block10_1_conv (Conv2D) (None, 2, 2, 128) 69632 conv4_block10_0_relu[0][0]
__________________________________________________________________________________________________
conv4_block10_1_bn (BatchNormal (None, 2, 2, 128) 512 conv4_block10_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block10_1_relu (Activatio (None, 2, 2, 128) 0 conv4_block10_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block10_2_conv (Conv2D) (None, 2, 2, 32) 36864 conv4_block10_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block10_concat (Concatena (None, 2, 2, 576) 0 conv4_block9_concat[0][0]
conv4_block10_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block11_0_bn (BatchNormal (None, 2, 2, 576) 2304 conv4_block10_concat[0][0]
__________________________________________________________________________________________________
conv4_block11_0_relu (Activatio (None, 2, 2, 576) 0 conv4_block11_0_bn[0][0]
__________________________________________________________________________________________________
conv4_block11_1_conv (Conv2D) (None, 2, 2, 128) 73728 conv4_block11_0_relu[0][0]
__________________________________________________________________________________________________
conv4_block11_1_bn (BatchNormal (None, 2, 2, 128) 512 conv4_block11_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block11_1_relu (Activatio (None, 2, 2, 128) 0 conv4_block11_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block11_2_conv (Conv2D) (None, 2, 2, 32) 36864 conv4_block11_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block11_concat (Concatena (None, 2, 2, 608) 0 conv4_block10_concat[0][0]
conv4_block11_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block12_0_bn (BatchNormal (None, 2, 2, 608) 2432 conv4_block11_concat[0][0]
__________________________________________________________________________________________________
conv4_block12_0_relu (Activatio (None, 2, 2, 608) 0 conv4_block12_0_bn[0][0]
__________________________________________________________________________________________________
conv4_block12_1_conv (Conv2D) (None, 2, 2, 128) 77824 conv4_block12_0_relu[0][0]
__________________________________________________________________________________________________
conv4_block12_1_bn (BatchNormal (None, 2, 2, 128) 512 conv4_block12_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block12_1_relu (Activatio (None, 2, 2, 128) 0 conv4_block12_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block12_2_conv (Conv2D) (None, 2, 2, 32) 36864 conv4_block12_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block12_concat (Concatena (None, 2, 2, 640) 0 conv4_block11_concat[0][0]
conv4_block12_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block13_0_bn (BatchNormal (None, 2, 2, 640) 2560 conv4_block12_concat[0][0]
__________________________________________________________________________________________________
conv4_block13_0_relu (Activatio (None, 2, 2, 640) 0 conv4_block13_0_bn[0][0]
__________________________________________________________________________________________________
conv4_block13_1_conv (Conv2D) (None, 2, 2, 128) 81920 conv4_block13_0_relu[0][0]
__________________________________________________________________________________________________
conv4_block13_1_bn (BatchNormal (None, 2, 2, 128) 512 conv4_block13_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block13_1_relu (Activatio (None, 2, 2, 128) 0 conv4_block13_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block13_2_conv (Conv2D) (None, 2, 2, 32) 36864 conv4_block13_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block13_concat (Concatena (None, 2, 2, 672) 0 conv4_block12_concat[0][0]
conv4_block13_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block14_0_bn (BatchNormal (None, 2, 2, 672) 2688 conv4_block13_concat[0][0]
__________________________________________________________________________________________________
conv4_block14_0_relu (Activatio (None, 2, 2, 672) 0 conv4_block14_0_bn[0][0]
__________________________________________________________________________________________________
conv4_block14_1_conv (Conv2D) (None, 2, 2, 128) 86016 conv4_block14_0_relu[0][0]
__________________________________________________________________________________________________
conv4_block14_1_bn (BatchNormal (None, 2, 2, 128) 512 conv4_block14_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block14_1_relu (Activatio (None, 2, 2, 128) 0 conv4_block14_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block14_2_conv (Conv2D) (None, 2, 2, 32) 36864 conv4_block14_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block14_concat (Concatena (None, 2, 2, 704) 0 conv4_block13_concat[0][0]
conv4_block14_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block15_0_bn (BatchNormal (None, 2, 2, 704) 2816 conv4_block14_concat[0][0]
__________________________________________________________________________________________________
conv4_block15_0_relu (Activatio (None, 2, 2, 704) 0 conv4_block15_0_bn[0][0]
__________________________________________________________________________________________________
conv4_block15_1_conv (Conv2D) (None, 2, 2, 128) 90112 conv4_block15_0_relu[0][0]
__________________________________________________________________________________________________
conv4_block15_1_bn (BatchNormal (None, 2, 2, 128) 512 conv4_block15_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block15_1_relu (Activatio (None, 2, 2, 128) 0 conv4_block15_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block15_2_conv (Conv2D) (None, 2, 2, 32) 36864 conv4_block15_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block15_concat (Concatena (None, 2, 2, 736) 0 conv4_block14_concat[0][0]
conv4_block15_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block16_0_bn (BatchNormal (None, 2, 2, 736) 2944 conv4_block15_concat[0][0]
__________________________________________________________________________________________________
conv4_block16_0_relu (Activatio (None, 2, 2, 736) 0 conv4_block16_0_bn[0][0]
__________________________________________________________________________________________________
conv4_block16_1_conv (Conv2D) (None, 2, 2, 128) 94208 conv4_block16_0_relu[0][0]
__________________________________________________________________________________________________
conv4_block16_1_bn (BatchNormal (None, 2, 2, 128) 512 conv4_block16_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block16_1_relu (Activatio (None, 2, 2, 128) 0 conv4_block16_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block16_2_conv (Conv2D) (None, 2, 2, 32) 36864 conv4_block16_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block16_concat (Concatena (None, 2, 2, 768) 0 conv4_block15_concat[0][0]
conv4_block16_2_conv[0][0]
__________________________________________________________________________________________________
pool4_bn (BatchNormalization) (None, 2, 2, 768) 3072 conv4_block16_concat[0][0]
__________________________________________________________________________________________________
pool4_relu (Activation) (None, 2, 2, 768) 0 pool4_bn[0][0]
__________________________________________________________________________________________________
pool4_conv (Conv2D) (None, 2, 2, 384) 294912 pool4_relu[0][0]
__________________________________________________________________________________________________
pool4_pool (AveragePooling2D) (None, 1, 1, 384) 0 pool4_conv[0][0]
__________________________________________________________________________________________________
conv5_block1_0_bn (BatchNormali (None, 1, 1, 384) 1536 pool4_pool[0][0]
__________________________________________________________________________________________________
conv5_block1_0_relu (Activation (None, 1, 1, 384) 0 conv5_block1_0_bn[0][0]
__________________________________________________________________________________________________
conv5_block1_1_conv (Conv2D) (None, 1, 1, 128) 49152 conv5_block1_0_relu[0][0]
__________________________________________________________________________________________________
conv5_block1_1_bn (BatchNormali (None, 1, 1, 128) 512 conv5_block1_1_conv[0][0]
__________________________________________________________________________________________________
conv5_block1_1_relu (Activation (None, 1, 1, 128) 0 conv5_block1_1_bn[0][0]
__________________________________________________________________________________________________
conv5_block1_2_conv (Conv2D) (None, 1, 1, 32) 36864 conv5_block1_1_relu[0][0]
__________________________________________________________________________________________________
conv5_block1_concat (Concatenat (None, 1, 1, 416) 0 pool4_pool[0][0]
conv5_block1_2_conv[0][0]
__________________________________________________________________________________________________
conv5_block2_0_bn (BatchNormali (None, 1, 1, 416) 1664 conv5_block1_concat[0][0]
__________________________________________________________________________________________________
conv5_block2_0_relu (Activation (None, 1, 1, 416) 0 conv5_block2_0_bn[0][0]
__________________________________________________________________________________________________
conv5_block2_1_conv (Conv2D) (None, 1, 1, 128) 53248 conv5_block2_0_relu[0][0]
__________________________________________________________________________________________________
conv5_block2_1_bn (BatchNormali (None, 1, 1, 128) 512 conv5_block2_1_conv[0][0]
__________________________________________________________________________________________________
conv5_block2_1_relu (Activation (None, 1, 1, 128) 0 conv5_block2_1_bn[0][0]
__________________________________________________________________________________________________
conv5_block2_2_conv (Conv2D) (None, 1, 1, 32) 36864 conv5_block2_1_relu[0][0]
__________________________________________________________________________________________________
conv5_block2_concat (Concatenat (None, 1, 1, 448) 0 conv5_block1_concat[0][0]
conv5_block2_2_conv[0][0]
__________________________________________________________________________________________________
conv5_block3_0_bn (BatchNormali (None, 1, 1, 448) 1792 conv5_block2_concat[0][0]
__________________________________________________________________________________________________
conv5_block3_0_relu (Activation (None, 1, 1, 448) 0 conv5_block3_0_bn[0][0]
__________________________________________________________________________________________________
conv5_block3_1_conv (Conv2D) (None, 1, 1, 128) 57344 conv5_block3_0_relu[0][0]
__________________________________________________________________________________________________
conv5_block3_1_bn (BatchNormali (None, 1, 1, 128) 512 conv5_block3_1_conv[0][0]
__________________________________________________________________________________________________
conv5_block3_1_relu (Activation (None, 1, 1, 128) 0 conv5_block3_1_bn[0][0]
__________________________________________________________________________________________________
conv5_block3_2_conv (Conv2D) (None, 1, 1, 32) 36864 conv5_block3_1_relu[0][0]
__________________________________________________________________________________________________
conv5_block3_concat (Concatenat (None, 1, 1, 480) 0 conv5_block2_concat[0][0]
conv5_block3_2_conv[0][0]
__________________________________________________________________________________________________
conv5_block4_0_bn (BatchNormali (None, 1, 1, 480) 1920 conv5_block3_concat[0][0]
__________________________________________________________________________________________________
conv5_block4_0_relu (Activation (None, 1, 1, 480) 0 conv5_block4_0_bn[0][0]
__________________________________________________________________________________________________
conv5_block4_1_conv (Conv2D) (None, 1, 1, 128) 61440 conv5_block4_0_relu[0][0]
__________________________________________________________________________________________________
conv5_block4_1_bn (BatchNormali (None, 1, 1, 128) 512 conv5_block4_1_conv[0][0]
__________________________________________________________________________________________________
conv5_block4_1_relu (Activation (None, 1, 1, 128) 0 conv5_block4_1_bn[0][0]
__________________________________________________________________________________________________
conv5_block4_2_conv (Conv2D) (None, 1, 1, 32) 36864 conv5_block4_1_relu[0][0]
__________________________________________________________________________________________________
conv5_block4_concat (Concatenat (None, 1, 1, 512) 0 conv5_block3_concat[0][0]
conv5_block4_2_conv[0][0]
__________________________________________________________________________________________________
conv5_block5_0_bn (BatchNormali (None, 1, 1, 512) 2048 conv5_block4_concat[0][0]
__________________________________________________________________________________________________
conv5_block5_0_relu (Activation (None, 1, 1, 512) 0 conv5_block5_0_bn[0][0]
__________________________________________________________________________________________________
conv5_block5_1_conv (Conv2D) (None, 1, 1, 128) 65536 conv5_block5_0_relu[0][0]
__________________________________________________________________________________________________
conv5_block5_1_bn (BatchNormali (None, 1, 1, 128) 512 conv5_block5_1_conv[0][0]
__________________________________________________________________________________________________
conv5_block5_1_relu (Activation (None, 1, 1, 128) 0 conv5_block5_1_bn[0][0]
__________________________________________________________________________________________________
conv5_block5_2_conv (Conv2D) (None, 1, 1, 32) 36864 conv5_block5_1_relu[0][0]
__________________________________________________________________________________________________
conv5_block5_concat (Concatenat (None, 1, 1, 544) 0 conv5_block4_concat[0][0]
conv5_block5_2_conv[0][0]
__________________________________________________________________________________________________
conv5_block6_0_bn (BatchNormali (None, 1, 1, 544) 2176 conv5_block5_concat[0][0]
__________________________________________________________________________________________________
conv5_block6_0_relu (Activation (None, 1, 1, 544) 0 conv5_block6_0_bn[0][0]
__________________________________________________________________________________________________
conv5_block6_1_conv (Conv2D) (None, 1, 1, 128) 69632 conv5_block6_0_relu[0][0]
__________________________________________________________________________________________________
conv5_block6_1_bn (BatchNormali (None, 1, 1, 128) 512 conv5_block6_1_conv[0][0]
__________________________________________________________________________________________________
conv5_block6_1_relu (Activation (None, 1, 1, 128) 0 conv5_block6_1_bn[0][0]
__________________________________________________________________________________________________
conv5_block6_2_conv (Conv2D) (None, 1, 1, 32) 36864 conv5_block6_1_relu[0][0]
__________________________________________________________________________________________________
conv5_block6_concat (Concatenat (None, 1, 1, 576) 0 conv5_block5_concat[0][0]
conv5_block6_2_conv[0][0]
__________________________________________________________________________________________________
conv5_block7_0_bn (BatchNormali (None, 1, 1, 576) 2304 conv5_block6_concat[0][0]
__________________________________________________________________________________________________
conv5_block7_0_relu (Activation (None, 1, 1, 576) 0 conv5_block7_0_bn[0][0]
__________________________________________________________________________________________________
conv5_block7_1_conv (Conv2D) (None, 1, 1, 128) 73728 conv5_block7_0_relu[0][0]
__________________________________________________________________________________________________
conv5_block7_1_bn (BatchNormali (None, 1, 1, 128) 512 conv5_block7_1_conv[0][0]
__________________________________________________________________________________________________
conv5_block7_1_relu (Activation (None, 1, 1, 128) 0 conv5_block7_1_bn[0][0]
__________________________________________________________________________________________________
conv5_block7_2_conv (Conv2D) (None, 1, 1, 32) 36864 conv5_block7_1_relu[0][0]
__________________________________________________________________________________________________
conv5_block7_concat (Concatenat (None, 1, 1, 608) 0 conv5_block6_concat[0][0]
conv5_block7_2_conv[0][0]
__________________________________________________________________________________________________
conv5_block8_0_bn (BatchNormali (None, 1, 1, 608) 2432 conv5_block7_concat[0][0]
__________________________________________________________________________________________________
conv5_block8_0_relu (Activation (None, 1, 1, 608) 0 conv5_block8_0_bn[0][0]
__________________________________________________________________________________________________
conv5_block8_1_conv (Conv2D) (None, 1, 1, 128) 77824 conv5_block8_0_relu[0][0]
__________________________________________________________________________________________________
conv5_block8_1_bn (BatchNormali (None, 1, 1, 128) 512 conv5_block8_1_conv[0][0]
__________________________________________________________________________________________________
conv5_block8_1_relu (Activation (None, 1, 1, 128) 0 conv5_block8_1_bn[0][0]
__________________________________________________________________________________________________
conv5_block8_2_conv (Conv2D) (None, 1, 1, 32) 36864 conv5_block8_1_relu[0][0]
__________________________________________________________________________________________________
conv5_block8_concat (Concatenat (None, 1, 1, 640) 0 conv5_block7_concat[0][0]
conv5_block8_2_conv[0][0]
__________________________________________________________________________________________________
bn (BatchNormalization) (None, 1, 1, 640) 2560 conv5_block8_concat[0][0]
__________________________________________________________________________________________________
avg_pool (GlobalAveragePooling2 (None, 640) 0 bn[0][0]
__________________________________________________________________________________________________
dense_1 (Dense) (None, 500) 320500 avg_pool[0][0]
__________________________________________________________________________________________________
activation_1 (Activation) (None, 500) 0 dense_1[0][0]
__________________________________________________________________________________________________
dropout_1 (Dropout) (None, 500) 0 activation_1[0][0]
__________________________________________________________________________________________________
dense_2 (Dense) (None, 100) 50100 dropout_1[0][0]
==================================================================================================
Total params: 4,584,424
Trainable params: 4,536,360
Non-trainable params: 48,064
__________________________________________________________________________________________________
Recordemos que la arquitectura ResNet tenía aproximadamente 25 millones de parámetros a entrenar. Esto quiere decir que hemos aumentado la profundidad pero hemos reducido el número de parámetros a entrenar con 4 millones y medio.
Bien, dicho esto, pasamos a entrenar el modelo:
cdense = custom_dense_model.fit(x=x_train, y=y_train, batch_size=32, epochs=10, verbose=1, validation_data=(x_test, y_test), shuffle=True)
Train on 50000 samples, validate on 10000 samples
Epoch 1/10
50000/50000 [==============================] - 328s 7ms/step - loss: 4.2465 - acc: 0.0707 - mean_squared_error: 0.0098 - val_loss: 3.9126 - val_acc: 0.1075 - val_mean_squared_error: 0.0096
Epoch 2/10
50000/50000 [==============================] - 321s 6ms/step - loss: 3.7411 - acc: 0.1331 - mean_squared_error: 0.0095 - val_loss: 3.6431 - val_acc: 0.1434 - val_mean_squared_error: 0.0094
Epoch 3/10
50000/50000 [==============================] - 322s 6ms/step - loss: 3.4654 - acc: 0.1751 - mean_squared_error: 0.0092 - val_loss: 3.4077 - val_acc: 0.1869 - val_mean_squared_error: 0.0091
Epoch 4/10
50000/50000 [==============================] - 325s 7ms/step - loss: 3.2441 - acc: 0.2161 - mean_squared_error: 0.0090 - val_loss: 3.2245 - val_acc: 0.2252 - val_mean_squared_error: 0.0089
Epoch 5/10
50000/50000 [==============================] - 324s 6ms/step - loss: 3.0505 - acc: 0.2472 - mean_squared_error: 0.0087 - val_loss: 3.0746 - val_acc: 0.2480 - val_mean_squared_error: 0.0087
Epoch 6/10
50000/50000 [==============================] - 327s 7ms/step - loss: 2.8631 - acc: 0.2835 - mean_squared_error: 0.0084 - val_loss: 2.9652 - val_acc: 0.2710 - val_mean_squared_error: 0.0085
Epoch 7/10
50000/50000 [==============================] - 326s 7ms/step - loss: 2.7206 - acc: 0.3130 - mean_squared_error: 0.0082 - val_loss: 2.8479 - val_acc: 0.2935 - val_mean_squared_error: 0.0083
Epoch 8/10
50000/50000 [==============================] - 329s 7ms/step - loss: 2.5673 - acc: 0.3409 - mean_squared_error: 0.0079 - val_loss: 2.7761 - val_acc: 0.3077 - val_mean_squared_error: 0.0082
Epoch 9/10
50000/50000 [==============================] - 325s 7ms/step - loss: 2.4398 - acc: 0.3682 - mean_squared_error: 0.0077 - val_loss: 2.7609 - val_acc: 0.3194 - val_mean_squared_error: 0.0082
Epoch 10/10
50000/50000 [==============================] - 327s 7ms/step - loss: 2.3162 - acc: 0.3963 - mean_squared_error: 0.0074 - val_loss: 2.6835 - val_acc: 0.3373 - val_mean_squared_error: 0.0080
Veamos las métricas obtenidas para el entrenamiento y validación gráficamente:
plt.figure(0)
plt.plot(cdense.history['acc'],'r')
plt.plot(cdense.history['val_acc'],'g')
plt.xticks(np.arange(0, 11, 2.0))
plt.rcParams['figure.figsize'] = (8, 6)
plt.xlabel("Num of Epochs")
plt.ylabel("Accuracy")
plt.title("Training Accuracy vs Validation Accuracy")
plt.legend(['train','validation'])
plt.figure(1)
plt.plot(cdense.history['loss'],'r')
plt.plot(cdense.history['val_loss'],'g')
plt.xticks(np.arange(0, 11, 2.0))
plt.rcParams['figure.figsize'] = (8, 6)
plt.xlabel("Num of Epochs")
plt.ylabel("Loss")
plt.title("Training Loss vs Validation Loss")
plt.legend(['train','validation'])
plt.show()
El entrenamiento ha dado muy buenos resultados y aunque no ha generalizado bien, los resultados a priori parecen mejores. Veámoslo a continuación.
Matriz de confusión
Pasemos ahora a ver la matriz de confusión y las métricas de Accuracy, Recall y F1-score.
Vamos a hacer una predicción sobre el dataset de validación y, a partir de ésta, generamos la matriz de confusión y mostramos las métricas mencionadas anteriormente:
cdense_pred = custom_dense_model.predict(x_test, batch_size=32, verbose=1)
cdense_predicted = np.argmax(cdense_pred, axis=1)
cdense_cm = confusion_matrix(np.argmax(y_test, axis=1), cdense_predicted)
# Visualizing of confusion matrix
cdense_df_cm = pd.DataFrame(cdense_cm, range(100), range(100))
plt.figure(figsize = (20,14))
sn.set(font_scale=1.4) #for label size
sn.heatmap(cdense_df_cm, annot=True, annot_kws={"size": 12}) # font size
plt.show()
Y por último, mostramos las métricas:
cdense_report = classification_report(np.argmax(y_test, axis=1), cdense_predicted)
print(cdense_report)
precision recall f1-score support
0 0.55 0.64 0.59 100
1 0.41 0.33 0.36 100
2 0.29 0.19 0.23 100
3 0.15 0.16 0.16 100
4 0.12 0.09 0.10 100
5 0.25 0.39 0.31 100
6 0.39 0.38 0.38 100
7 0.46 0.34 0.39 100
8 0.37 0.48 0.42 100
9 0.44 0.54 0.48 100
10 0.25 0.16 0.19 100
11 0.24 0.16 0.19 100
12 0.39 0.34 0.36 100
13 0.23 0.14 0.17 100
14 0.24 0.30 0.27 100
15 0.20 0.17 0.18 100
16 0.43 0.41 0.42 100
17 0.52 0.42 0.46 100
18 0.30 0.28 0.29 100
19 0.30 0.29 0.30 100
20 0.63 0.56 0.59 100
21 0.40 0.42 0.41 100
22 0.32 0.23 0.27 100
23 0.50 0.50 0.50 100
24 0.50 0.56 0.53 100
25 0.21 0.20 0.20 100
26 0.24 0.25 0.25 100
27 0.24 0.19 0.21 100
28 0.45 0.45 0.45 100
29 0.30 0.26 0.28 100
30 0.31 0.42 0.36 100
31 0.25 0.37 0.30 100
32 0.27 0.23 0.25 100
33 0.32 0.30 0.31 100
34 0.17 0.14 0.15 100
35 0.19 0.13 0.15 100
36 0.35 0.27 0.31 100
37 0.26 0.22 0.24 100
38 0.18 0.16 0.17 100
39 0.54 0.40 0.46 100
40 0.32 0.31 0.31 100
41 0.58 0.57 0.57 100
42 0.19 0.25 0.22 100
43 0.28 0.27 0.27 100
44 0.09 0.04 0.06 100
45 0.16 0.14 0.15 100
46 0.19 0.15 0.17 100
47 0.42 0.46 0.44 100
48 0.45 0.51 0.48 100
49 0.54 0.48 0.51 100
50 0.08 0.06 0.07 100
51 0.30 0.33 0.32 100
52 0.43 0.73 0.54 100
53 0.60 0.56 0.58 100
54 0.49 0.48 0.49 100
55 0.09 0.07 0.08 100
56 0.52 0.49 0.51 100
57 0.32 0.36 0.34 100
58 0.29 0.25 0.27 100
59 0.28 0.15 0.20 100
60 0.64 0.69 0.67 100
61 0.40 0.44 0.42 100
62 0.37 0.51 0.43 100
63 0.31 0.40 0.35 100
64 0.09 0.09 0.09 100
65 0.27 0.14 0.18 100
66 0.25 0.21 0.23 100
67 0.33 0.23 0.27 100
68 0.58 0.67 0.62 100
69 0.51 0.53 0.52 100
70 0.34 0.37 0.35 100
71 0.52 0.64 0.58 100
72 0.05 0.04 0.05 100
73 0.27 0.21 0.24 100
74 0.14 0.14 0.14 100
75 0.50 0.61 0.55 100
76 0.54 0.62 0.58 100
77 0.14 0.16 0.15 100
78 0.18 0.11 0.14 100
79 0.29 0.24 0.26 100
80 0.09 0.10 0.09 100
81 0.24 0.22 0.23 100
82 0.57 0.74 0.64 100
83 0.26 0.27 0.26 100
84 0.17 0.14 0.15 100
85 0.43 0.57 0.49 100
86 0.37 0.34 0.35 100
87 0.37 0.48 0.42 100
88 0.18 0.29 0.23 100
89 0.26 0.41 0.32 100
90 0.23 0.28 0.25 100
91 0.43 0.53 0.48 100
92 0.22 0.23 0.22 100
93 0.17 0.21 0.19 100
94 0.49 0.70 0.57 100
95 0.35 0.44 0.39 100
96 0.25 0.19 0.22 100
97 0.21 0.29 0.25 100
98 0.10 0.09 0.10 100
99 0.25 0.28 0.26 100
avg / total 0.32 0.33 0.32 10000
Curva ROC (tasas de verdaderos positivos y falsos positivos)
Vamos a codificar la curva ROC
from sklearn.datasets import make_classification
from sklearn.preprocessing import label_binarize
from scipy import interp
from itertools import cycle
n_classes = 100
from sklearn.metrics import roc_curve, auc
# Plot linewidth.
lw = 2
# Compute ROC curve and ROC area for each class
fpr = dict()
tpr = dict()
roc_auc = dict()
for i in range(n_classes):
fpr[i], tpr[i], _ = roc_curve(y_test[:, i], cdense_pred[:, i])
roc_auc[i] = auc(fpr[i], tpr[i])
# Compute micro-average ROC curve and ROC area
fpr["micro"], tpr["micro"], _ = roc_curve(y_test.ravel(), cdense_pred.ravel())
roc_auc["micro"] = auc(fpr["micro"], tpr["micro"])
# Compute macro-average ROC curve and ROC area
# First aggregate all false positive rates
all_fpr = np.unique(np.concatenate([fpr[i] for i in range(n_classes)]))
# Then interpolate all ROC curves at this points
mean_tpr = np.zeros_like(all_fpr)
for i in range(n_classes):
mean_tpr += interp(all_fpr, fpr[i], tpr[i])
# Finally average it and compute AUC
mean_tpr /= n_classes
fpr["macro"] = all_fpr
tpr["macro"] = mean_tpr
roc_auc["macro"] = auc(fpr["macro"], tpr["macro"])
# Plot all ROC curves
plt.figure(1)
plt.plot(fpr["micro"], tpr["micro"],
label='micro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["micro"]),
color='deeppink', linestyle=':', linewidth=4)
plt.plot(fpr["macro"], tpr["macro"],
label='macro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["macro"]),
color='navy', linestyle=':', linewidth=4)
colors = cycle(['aqua', 'darkorange', 'cornflowerblue'])
for i, color in zip(range(n_classes-97), colors):
plt.plot(fpr[i], tpr[i], color=color, lw=lw,
label='ROC curve of class {0} (area = {1:0.2f})'
''.format(i, roc_auc[i]))
plt.plot([0, 1], [0, 1], 'k--', lw=lw)
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Some extension of Receiver operating characteristic to multi-class')
plt.legend(loc="lower right")
plt.show()
# Zoom in view of the upper left corner.
plt.figure(2)
plt.xlim(0, 0.2)
plt.ylim(0.8, 1)
plt.plot(fpr["micro"], tpr["micro"],
label='micro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["micro"]),
color='deeppink', linestyle=':', linewidth=4)
plt.plot(fpr["macro"], tpr["macro"],
label='macro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["macro"]),
color='navy', linestyle=':', linewidth=4)
colors = cycle(['aqua', 'darkorange', 'cornflowerblue'])
for i, color in zip(range(3), colors):
plt.plot(fpr[i], tpr[i], color=color, lw=lw,
label='ROC curve of class {0} (area = {1:0.2f})'
''.format(i, roc_auc[i]))
plt.plot([0, 1], [0, 1], 'k--', lw=lw)
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Some extension of Receiver operating characteristic to multi-class')
plt.legend(loc="lower right")
plt.show()
El resultado para tres clases se muestra en los siguientes gráficos:
Salvaremos los datos del histórico de entrenamiento para compararlos con otros modelos. Además, vamos a salvar el modelo con los pesos entrenados para usarlos en el futuro:
#Modelo
custom_dense_model.save(path_base + '/cdense.h5')
#Histórico
with open(path_base + '/cdense_history.txt', 'wb') as file_pi:
pickle.dump(cdense.history, file_pi)
A continuación, vamos a comparar las métricas con los modelos anteriores:
plt.figure(0)
plt.plot(snn_history['val_acc'],'r')
plt.plot(scnn_history['val_acc'],'g')
plt.plot(vgg16_history['val_acc'],'b')
plt.plot(cvgg16_history['val_acc'],'y')
plt.plot(crn50_history['val_acc'],'gold')
plt.plot(cdense_history['val_acc'],'m')
plt.xticks(np.arange(0, 11, 2.0))
plt.rcParams['figure.figsize'] = (8, 6)
plt.xlabel("Num of Epochs")
plt.ylabel("Accuracy")
plt.title("Models")
plt.legend(['simple NN','CNN','VGG 16','Custom VGG','Custom ResNet', 'Custom DenseNet'])
plt.figure(1)
plt.plot(snn_history['val_loss'],'r')
plt.plot(scnn_history['val_loss'],'g')
plt.plot(vgg16_history['val_loss'],'b')
plt.plot(cvgg16_history['val_loss'],'y')
plt.plot(crn50_history['val_loss'],'gold')
plt.plot(cdense_history['val_loss'],'m')
plt.xticks(np.arange(0, 11, 2.0))
plt.rcParams['figure.figsize'] = (8, 6)
plt.xlabel("Num of Epochs")
plt.ylabel("Loss")
plt.title("Models")
plt.legend(['simple NN','CNN','VGG 16','Custom VGG','Custom ResNet', 'Custom DenseNet'])
plt.figure(2)
plt.plot(snn_history['val_mean_squared_error'],'r')
plt.plot(scnn_history['val_mean_squared_error'],'g')
plt.plot(vgg16_history['val_mean_squared_error'],'b')
plt.plot(cvgg16_history['val_mean_squared_error'],'y')
plt.plot(crn50_history['val_mean_squared_error'],'gold')
plt.plot(cdense_history['val_mean_squared_error'],'m')
plt.xticks(np.arange(0, 11, 2.0))
plt.rcParams['figure.figsize'] = (8, 6)
plt.xlabel("Num of Epochs")
plt.ylabel("Mean Squared Error")
plt.title("Models")
plt.legend(['simple NN','CNN','VGG 16','Custom VGG','Custom ResNet', 'Custom DenseNet'])
Conclusión sobre el experimento
La arquitectura DenseNet da unos resultados mucho mejores reduciendo el número de parámetros a aprender, con lo que también puede llegar a reducir los tiempos de entrenamiento.
En el siguiente artículo, presentaremos la arquitectura: NASNet.
¡Síguenos en Twitter para estar al día de nuestros próximos posts!