def build_model(input_shape, num_classes, weights='imagenet'):
# create the base pre-trained model
base_model = DenseNet121(weights=weights, include_top=False, input_shape=input_shape,
backend=keras.backend, layers=keras.layers, models=keras.models,
utils=keras.utils)
# add a global spatial average pooling layer
x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu', name='fc2014_1')(x)
x = Dropout(0.5)(x)
x = Dense(1024, activation='relu', name='fc2014_2')(x)
x = Dropout(0.5)(x)
x = Dense(num_classes, activation='sigmoid', name='fc28')(x)
# this is the model we will train
model = Model(inputs=base_model.input, outputs=x, name='pre_trained_resnet50')
sgd = SGD(lr=0.01, momentum=0.9, nesterov=True, decay=1e-06)
# compile the model (should be done *after* setting layers to non-trainable)
model.compile(optimizer=sgd, loss=binary_crossentropy, metrics=['accuracy'])
return model
def get_densenet121_unet_softmax(input_shape, weights='imagenet'):
blocks = [6, 12, 24, 16]
img_input = Input(input_shape + (4, ))
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)
conv1 = x
x = ZeroPadding2D(padding=((1, 1), (1, 1)))(x)
x = MaxPooling2D(3, strides=2, name='pool1')(x)
x = dense_block(x, blocks[0], name='conv2')
conv2 = x
x = transition_block(x, 0.5, name='pool2')
x = dense_block(x, blocks[1], name='conv3')
conv3 = x
x = transition_block(x, 0.5, name='pool3')
x = dense_block(x, blocks[2], name='conv4')
conv4 = x
x = transition_block(x, 0.5, name='pool4')
x = dense_block(x, blocks[3], name='conv5')
x = BatchNormalization(axis=bn_axis, epsilon=1.001e-5, name='bn')(x)
conv5 = x
conv6 = conv_block(UpSampling2D()(conv5), 320)
conv6 = concatenate([conv6, conv4], axis=-1)
conv6 = conv_block(conv6, 320)
conv7 = conv_block(UpSampling2D()(conv6), 256)
conv7 = concatenate([conv7, conv3], axis=-1)
conv7 = conv_block(conv7, 256)
conv8 = conv_block(UpSampling2D()(conv7), 128)
conv8 = concatenate([conv8, conv2], axis=-1)
conv8 = conv_block(conv8, 128)
conv9 = conv_block(UpSampling2D()(conv8), 96)
conv9 = concatenate([conv9, conv1], axis=-1)
conv9 = conv_block(conv9, 96)
conv10 = conv_block(UpSampling2D()(conv9), 64)
conv10 = conv_block(conv10, 64)
res = Conv2D(3, (1, 1), activation='softmax')(conv10)
model = Model(img_input, res)
if weights == 'imagenet':
densenet = DenseNet121(input_shape=input_shape + (3, ),
weights=weights,
include_top=False)
w0 = densenet.layers[2].get_weights()
w = model.layers[2].get_weights()
w[0][:, :, [0, 1, 2], :] = 0.9 * w0[0][:, :, :3, :]
w[0][:, :, 3, :] = 0.1 * w0[0][:, :, 1, :]
model.layers[2].set_weights(w)
for i in range(3, len(densenet.layers)):
model.layers[i].set_weights(densenet.layers[i].get_weights())
model.layers[i].trainable = False
return model
def build_model(config):
"""
Returns: a model with specified weights
"""
input_size = config['input_size']
input_shape = (input_size, input_size, 3)
if ('pretrain' in config) and config['pretrain']:
assert config['input_size'] == 224
weights = 'imagenet'
else:
weights = None
assert config['image_model'] in ['densenet', 'resnet', 'linear']
if config['image_model'] == 'densenet':
print("Using Densenet.")
base_model = DenseNet121(input_shape=input_shape,
weights=weights,
include_top=False,
pooling='avg')
image_input = base_model.input
layer = base_model.output
elif config['image_model'] == 'resnet':
print("Using Resnet.")
base_model = ResNet50(input_shape=input_shape,
weights=weights,
include_top=False,
pooling='avg')
image_input = base_model.input
layer = base_model.output
elif config['image_model'] == 'linear':
print("Using linear model.")
image_input = Input(shape=input_shape)
layer = Flatten()(image_input)
if ('freeze' in config) and config['freeze']:
for layer in base_model.layers:
try:
layer.trainable = False
print("Freezing {}".format(layer))
except Exception:
print("Not trainable {}".format(layer))
predictions = Dense(14, activation='sigmoid')(layer)
model = Model(inputs=image_input, outputs=predictions)
return model
def get_model(img_size,
num_channels,
num_classes,
model_name='densenet121',
pretrain=False):
weights = 'imagenet' if pretrain else None
if pretrain:
assert img_size == 224
input_tensor = Input(shape=(img_size, img_size, num_channels))
if model_name == 'densenet121':
model = DenseNet121(include_top=False,
weights=weights,
input_tensor=input_tensor,
pooling='avg')
preprocessor = pi_densenet121
elif model_name == 'resnet50':
model = ResNet50(include_top=False,
weights=weights,
input_tensor=input_tensor,
pooling='avg')
preprocessor = pi_resnet50
elif model_name == 'vgg19':
model = VGG19(include_top=False,
weights=weights,
input_tensor=input_tensor,
pooling='avg')
preprocessor = pi_vgg19
else:
raise Exception("Unrecognized model name: %s" % model_name)
output_tensor = model(input_tensor)
output_tensor = Dense(num_classes, activation='sigmoid')(output_tensor)
complete_model = Model(input_tensor, output_tensor)
return complete_model, preprocessor
def get_densenet121_unet_sigmoid_gn(input_shape=(CONFIG.img_h, CONFIG.img_w, CONFIG.img_c),
output_channels=1,
weights='imagenet'):
blocks = [6, 12, 24, 16]
img_input = Input(input_shape)
x = ZeroPadding2D(padding=((3, 3), (3, 3)))(img_input)
x = Conv2D(64, 7, strides=2, use_bias=False, name='conv1/conv')(x)
x = GroupNormalization(axis=GN_AXIS, groups=16,
scale=False,
name='conv1/gn')(x)
x = Activation('relu', name='conv1/relu')(x)
conv1 = x
x = ZeroPadding2D(padding=((1, 1), (1, 1)))(x)
x = MaxPooling2D(3, strides=2, name='pool1')(x)
x = dense_block(x, blocks[0], name='conv2')
conv2 = x
x = transition_block(x, 0.5, name='pool2')
x = dense_block(x, blocks[1], name='conv3')
conv3 = x
x = transition_block(x, 0.5, name='pool3')
x = dense_block(x, blocks[2], name='conv4')
conv4 = x
x = transition_block(x, 0.5, name='pool4')
x = dense_block(x, blocks[3], name='conv5')
x = GroupNormalization(axis=GN_AXIS, groups=32,
scale=False,
name='conv5/gn')(x)
conv5 = x
# squeeze and excite block
conv5 = squeeze_excite_block(conv5)
conv6 = conv_block(UpSampling2D()(conv5), 320)
conv6 = concatenate([conv6, conv4], axis=-1)
conv6 = conv_block(conv6, 320)
conv7 = conv_block(UpSampling2D()(conv6), 256)
conv7 = concatenate([conv7, conv3], axis=-1)
conv7 = conv_block(conv7, 256)
conv8 = conv_block(UpSampling2D()(conv7), 128)
conv8 = concatenate([conv8, conv2], axis=-1)
conv8 = conv_block(conv8, 128)
conv9 = conv_block(UpSampling2D()(conv8), 96)
conv9 = concatenate([conv9, conv1], axis=-1)
conv9 = conv_block(conv9, 96)
conv10 = conv_block(UpSampling2D()(conv9), 64)
conv10 = conv_block(conv10, 64)
res = Conv2D(output_channels, (1, 1), activation='sigmoid')(conv10)
model = Model(img_input, res)
if weights == 'imagenet':
densenet = DenseNet121(input_shape=(input_shape[0], input_shape[1], 3), weights=weights, include_top=False)
print("Loading imagenet weights.")
for i in tqdm(range(2, len(densenet.layers) - 1)):
model.layers[i].set_weights(densenet.layers[i].get_weights())
model.layers[i].trainable = False
return model