diff --git a/models.py b/models.py index 7c806b4..aba310c 100644 --- a/models.py +++ b/models.py @@ -1,22 +1,12 @@ -from keras.models import * -from keras.layers import * -from keras import layers -from keras.regularizers import l2 +from tensorflow.keras.models import Model +from tensorflow.keras.layers import Conv2D, Concatenate, ZeroPadding2D, BatchNormalization, Activation, MaxPooling2D, UpSampling2D, Input, Layer +from tensorflow.keras import layers +from tensorflow.keras.regularizers import l2 +import tensorflow as tf resnet50_Weights_path='./pretrained_model/resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5' -IMAGE_ORDERING ='channels_last' -MERGE_AXIS=-1 - -def one_side_pad( x ): - x = ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING)(x) - if IMAGE_ORDERING == 'channels_first': - x = Lambda(lambda x : x[: , : , :-1 , :-1 ] )(x) - elif IMAGE_ORDERING == 'channels_last': - x = Lambda(lambda x : x[: , :-1 , :-1 , : ] )(x) - return x - -def identity_block(input_tensor, kernel_size, filters, stage, block): +def identity_block(input_tensor, kernel_size, filters, stage, block, data_format='channels_last'): """The identity block is the block that has no conv layer at shortcut. # Arguments input_tensor: input tensor @@ -29,24 +19,21 @@ def identity_block(input_tensor, kernel_size, filters, stage, block): """ filters1, filters2, filters3 = filters - if IMAGE_ORDERING == 'channels_last': - bn_axis = 3 - else: - bn_axis = 1 + bn_axis = 3 if data_format == 'channels_last' else 1 conv_name_base = 'res' + str(stage) + block + '_branch' bn_name_base = 'bn' + str(stage) + block + '_branch' - x = Conv2D(filters1, (1, 1) , data_format=IMAGE_ORDERING , name=conv_name_base + '2a')(input_tensor) + x = Conv2D(filters1, (1, 1) , data_format=data_format , name=conv_name_base + '2a')(input_tensor) x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x) x = Activation('relu')(x) - x = Conv2D(filters2, kernel_size , data_format=IMAGE_ORDERING , + x = Conv2D(filters2, kernel_size , data_format=data_format , padding='same', name=conv_name_base + '2b')(x) x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x) x = Activation('relu')(x) - x = Conv2D(filters3 , (1, 1), data_format=IMAGE_ORDERING , name=conv_name_base + '2c')(x) + x = Conv2D(filters3 , (1, 1), data_format=data_format , name=conv_name_base + '2c')(x) x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2c')(x) x = layers.add([x, input_tensor]) @@ -54,7 +41,7 @@ def identity_block(input_tensor, kernel_size, filters, stage, block): return x -def conv_block(input_tensor, kernel_size, filters, stage, block, strides=(2, 2)): +def conv_block(input_tensor, kernel_size, filters, stage, block, strides=(2, 2), data_format='channels_last'): """conv_block is the block that has a conv layer at shortcut # Arguments input_tensor: input tensor @@ -69,28 +56,25 @@ def conv_block(input_tensor, kernel_size, filters, stage, block, strides=(2, 2)) """ filters1, filters2, filters3 = filters - if IMAGE_ORDERING == 'channels_last': - bn_axis = 3 - else: - bn_axis = 1 + bn_axis = 3 if data_format == 'channels_last' else 1 conv_name_base = 'res' + str(stage) + block + '_branch' bn_name_base = 'bn' + str(stage) + block + '_branch' - x = Conv2D(filters1, (1, 1) , data_format=IMAGE_ORDERING , strides=strides, + x = Conv2D(filters1, (1, 1) , data_format=data_format, strides=strides, name=conv_name_base + '2a')(input_tensor) x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x) x = Activation('relu')(x) - x = Conv2D(filters2, kernel_size , data_format=IMAGE_ORDERING , padding='same', + x = Conv2D(filters2, kernel_size , data_format=data_format, padding='same', name=conv_name_base + '2b')(x) x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x) x = Activation('relu')(x) - x = Conv2D(filters3, (1, 1) , data_format=IMAGE_ORDERING , name=conv_name_base + '2c')(x) + x = Conv2D(filters3, (1, 1), data_format=data_format, name=conv_name_base + '2c')(x) x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2c')(x) - shortcut = Conv2D(filters3, (1, 1) , data_format=IMAGE_ORDERING , strides=strides, + shortcut = Conv2D(filters3, (1, 1), data_format=data_format, strides=strides, name=conv_name_base + '1')(input_tensor) shortcut = BatchNormalization(axis=bn_axis, name=bn_name_base + '1')(shortcut) @@ -99,219 +83,131 @@ def conv_block(input_tensor, kernel_size, filters, stage, block, strides=(2, 2)) return x -def resnet50_unet_light(n_classes,input_height=224,input_width=224,weight_decay=1e-6,pretraining=False): - assert input_height%32 == 0 - assert input_width%32 == 0 - - - img_input = Input(shape=(input_height,input_width , 3 )) - - if IMAGE_ORDERING == 'channels_last': - bn_axis = 3 - else: - bn_axis = 1 - - x = ZeroPadding2D((3, 3), data_format=IMAGE_ORDERING)(img_input) - x = Conv2D(64, (7, 7), data_format=IMAGE_ORDERING, strides=(2, 2),kernel_regularizer=l2(weight_decay), name='conv1')(x) +class PadMultiple(Layer): + def __init__(self, mods, data_format='channels_last'): + super().__init__() + self.mods = mods + self.data_format = data_format + + def call(self, x): + h, w = self.mods + padding = ( + [(0,0), (0, -tf.shape(x)[1] % h), (0, -tf.shape(x)[2] % w), (0,0)] if self.data_format == 'channels_last' + else [(0,0), (0,0), (0, -tf.shape(x)[1] % h), (0, -tf.shape(x)[2] % w)]) + return tf.pad(x, padding) + + +class CutTo(Layer): + def __init__(self, data_format='channels_last'): + super().__init__() + self.data_format = data_format + + def call(self, inputs): + h, w = (1, 2) if self.data_format == 'channels_last' else (2,4) + h, w = tf.shape(inputs[1])[h], tf.shape(inputs[1])[w] + return inputs[0][:, :h, :w] if self.data_format == 'channels_last' else inputs[0][:, :, :h, :w] + + +def resnet50_unet(n_classes, input_height=None, input_width=None, weight_decay=1e-6, pretraining=False, last_activation='softmax', skip_last_batchnorm=False, light_version=False, data_format='channels_last'): + """ Returns a U-NET model using the keras functional API. """ + img_input = Input(shape=(input_height, input_width, 3 )) + padded_to_multiple = PadMultiple((32,32))(img_input) + + bn_axis = 3 if data_format == 'channels_last' else 1 + merge_axis = 3 if data_format == 'channels_last' else 1 + + x = ZeroPadding2D((3, 3), data_format=data_format)(padded_to_multiple) + x = Conv2D(64, (7, 7), data_format=data_format, strides=(2, 2), kernel_regularizer=l2(weight_decay), name='conv1')(x) f1 = x x = BatchNormalization(axis=bn_axis, name='bn_conv1')(x) x = Activation('relu')(x) - x = MaxPooling2D((3, 3) , data_format=IMAGE_ORDERING , strides=(2, 2))(x) + x = MaxPooling2D((3, 3) , data_format=data_format , strides=(2, 2))(x) - x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1)) - x = identity_block(x, 3, [64, 64, 256], stage=2, block='b') - x = identity_block(x, 3, [64, 64, 256], stage=2, block='c') - f2 = one_side_pad(x ) + x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1), data_format=data_format) + x = identity_block(x, 3, [64, 64, 256], stage=2, block='b', data_format=data_format) + x = identity_block(x, 3, [64, 64, 256], stage=2, block='c', data_format=data_format) + f2 = ZeroPadding2D(((1,0), (1,0)), data_format=data_format)(x) - x = conv_block(x, 3, [128, 128, 512], stage=3, block='a') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='b') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='c') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='d') + x = conv_block(x, 3, [128, 128, 512], stage=3, block='a', data_format=data_format) + x = identity_block(x, 3, [128, 128, 512], stage=3, block='b', data_format=data_format) + x = identity_block(x, 3, [128, 128, 512], stage=3, block='c', data_format=data_format) + x = identity_block(x, 3, [128, 128, 512], stage=3, block='d', data_format=data_format) f3 = x - x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='b') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='c') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='d') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='e') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='f') + x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a', data_format=data_format) + x = identity_block(x, 3, [256, 256, 1024], stage=4, block='b', data_format=data_format) + x = identity_block(x, 3, [256, 256, 1024], stage=4, block='c', data_format=data_format) + x = identity_block(x, 3, [256, 256, 1024], stage=4, block='d', data_format=data_format) + x = identity_block(x, 3, [256, 256, 1024], stage=4, block='e', data_format=data_format) + x = identity_block(x, 3, [256, 256, 1024], stage=4, block='f', data_format=data_format) f4 = x - x = conv_block(x, 3, [512, 512, 2048], stage=5, block='a') - x = identity_block(x, 3, [512, 512, 2048], stage=5, block='b') - x = identity_block(x, 3, [512, 512, 2048], stage=5, block='c') + x = conv_block(x, 3, [512, 512, 2048], stage=5, block='a', data_format=data_format) + x = identity_block(x, 3, [512, 512, 2048], stage=5, block='b', data_format=data_format) + x = identity_block(x, 3, [512, 512, 2048], stage=5, block='c', data_format=data_format) f5 = x - if pretraining: - model=Model( img_input , x ).load_weights(resnet50_Weights_path) + Model(img_input, x).load_weights(resnet50_Weights_path) + if light_version: + v512_2048 = Conv2D(512, (1, 1), padding='same', data_format=data_format, kernel_regularizer=l2(weight_decay))(f5) + v512_2048 = BatchNormalization(axis=bn_axis)(v512_2048) + v512_2048 = Activation('relu')(v512_2048) - v512_2048 = Conv2D( 512 , (1, 1) , padding='same', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay) )( f5 ) - v512_2048 = ( BatchNormalization(axis=bn_axis))(v512_2048) - v512_2048 = Activation('relu')(v512_2048) - - - - v512_1024=Conv2D( 512 , (1, 1) , padding='same', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay) )( f4 ) - v512_1024 = ( BatchNormalization(axis=bn_axis))(v512_1024) - v512_1024 = Activation('relu')(v512_1024) - - - o = ( UpSampling2D( (2,2), data_format=IMAGE_ORDERING))(v512_2048) - o = ( concatenate([ o ,v512_1024],axis=MERGE_AXIS ) ) - o = ( ZeroPadding2D( (1,1), data_format=IMAGE_ORDERING))(o) - o = ( Conv2D(512, (3, 3), padding='valid', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay)))(o) - o = ( BatchNormalization(axis=bn_axis))(o) - o = Activation('relu')(o) - - - o = ( UpSampling2D( (2,2), data_format=IMAGE_ORDERING))(o) - o = ( concatenate([ o ,f3],axis=MERGE_AXIS ) ) - o = ( ZeroPadding2D( (1,1), data_format=IMAGE_ORDERING))(o) - o = ( Conv2D( 256, (3, 3), padding='valid', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay)))(o) - o = ( BatchNormalization(axis=bn_axis))(o) - o = Activation('relu')(o) - - - o = ( UpSampling2D( (2,2), data_format=IMAGE_ORDERING))(o) - o = ( concatenate([o,f2],axis=MERGE_AXIS ) ) - o = ( ZeroPadding2D((1,1) , data_format=IMAGE_ORDERING))(o) - o = ( Conv2D( 128 , (3, 3), padding='valid' , data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay) ) )(o) - o = ( BatchNormalization(axis=bn_axis))(o) - o = Activation('relu')(o) - - - - o = ( UpSampling2D( (2,2), data_format=IMAGE_ORDERING))(o) - o = ( concatenate([o,f1],axis=MERGE_AXIS ) ) - o = ( ZeroPadding2D((1,1) , data_format=IMAGE_ORDERING ))(o) - o = ( Conv2D( 64 , (3, 3), padding='valid' , data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay) ))(o) - o = ( BatchNormalization(axis=bn_axis))(o) - o = Activation('relu')(o) - - - o = ( UpSampling2D( (2,2), data_format=IMAGE_ORDERING))(o) - o = ( concatenate([o,img_input],axis=MERGE_AXIS ) ) - o = ( ZeroPadding2D((1,1) , data_format=IMAGE_ORDERING ))(o) - o = ( Conv2D( 32 , (3, 3), padding='valid' , data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay) ))(o) - o = ( BatchNormalization(axis=bn_axis))(o) - o = Activation('relu')(o) - - - - o = Conv2D( n_classes , (1, 1) , padding='same', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay) )( o ) - o = ( BatchNormalization(axis=bn_axis))(o) - o = (Activation('softmax'))(o) - - - model = Model( img_input , o ) - return model - -def resnet50_unet(n_classes,input_height=224,input_width=224,weight_decay=1e-6,pretraining=False): - assert input_height%32 == 0 - assert input_width%32 == 0 - - - img_input = Input(shape=(input_height,input_width , 3 )) - - if IMAGE_ORDERING == 'channels_last': - bn_axis = 3 + v512_1024 = Conv2D(512, (1, 1), padding='same', data_format=data_format, kernel_regularizer=l2(weight_decay))(f4) + v512_1024 = BatchNormalization(axis=bn_axis)(v512_1024) + v512_1024 = Activation('relu')(v512_1024) + x, c = v512_2048, v512_1024 # continuation and concatenation layers else: - bn_axis = 1 - - x = ZeroPadding2D((3, 3), data_format=IMAGE_ORDERING)(img_input) - x = Conv2D(64, (7, 7), data_format=IMAGE_ORDERING, strides=(2, 2),kernel_regularizer=l2(weight_decay), name='conv1')(x) - f1 = x - - x = BatchNormalization(axis=bn_axis, name='bn_conv1')(x) - x = Activation('relu')(x) - x = MaxPooling2D((3, 3) , data_format=IMAGE_ORDERING , strides=(2, 2))(x) - - - x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1)) - x = identity_block(x, 3, [64, 64, 256], stage=2, block='b') - x = identity_block(x, 3, [64, 64, 256], stage=2, block='c') - f2 = one_side_pad(x ) - - - x = conv_block(x, 3, [128, 128, 512], stage=3, block='a') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='b') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='c') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='d') - f3 = x - - x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='b') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='c') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='d') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='e') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='f') - f4 = x - - x = conv_block(x, 3, [512, 512, 2048], stage=5, block='a') - x = identity_block(x, 3, [512, 512, 2048], stage=5, block='b') - x = identity_block(x, 3, [512, 512, 2048], stage=5, block='c') - f5 = x - - if pretraining: - Model( img_input , x ).load_weights(resnet50_Weights_path) - - v1024_2048 = Conv2D( 1024 , (1, 1) , padding='same', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay) )( f5 ) - v1024_2048 = ( BatchNormalization(axis=bn_axis))(v1024_2048) - v1024_2048 = Activation('relu')(v1024_2048) - - - o = ( UpSampling2D( (2,2), data_format=IMAGE_ORDERING))(v1024_2048) - o = ( concatenate([ o ,f4],axis=MERGE_AXIS ) ) - o = ( ZeroPadding2D( (1,1), data_format=IMAGE_ORDERING))(o) - o = ( Conv2D(512, (3, 3), padding='valid', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay)))(o) - o = ( BatchNormalization(axis=bn_axis))(o) + v1024_2048 = Conv2D(1024, (1, 1), padding='same', data_format=data_format, kernel_regularizer=l2(weight_decay))(f5) + v1024_2048 = BatchNormalization(axis=bn_axis)(v1024_2048) + v1024_2048 = Activation('relu')(v1024_2048) + x, c = v1024_2048, f4 # continuation and concatenation layers + + o = UpSampling2D((2,2), data_format=data_format)(x) + o = Concatenate(axis=merge_axis)([o ,c]) + o = ZeroPadding2D( (1,1), data_format=data_format)(o) + o = Conv2D(512, (3, 3), padding='valid', data_format=data_format, kernel_regularizer=l2(weight_decay))(o) + o = BatchNormalization(axis=bn_axis)(o) o = Activation('relu')(o) - - o = ( UpSampling2D( (2,2), data_format=IMAGE_ORDERING))(o) - o = ( concatenate([ o ,f3],axis=MERGE_AXIS ) ) - o = ( ZeroPadding2D( (1,1), data_format=IMAGE_ORDERING))(o) - o = ( Conv2D( 256, (3, 3), padding='valid', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay)))(o) - o = ( BatchNormalization(axis=bn_axis))(o) + o = UpSampling2D( (2,2), data_format=data_format)(o) + o = Concatenate(axis=merge_axis)([ o ,f3]) + o = ZeroPadding2D( (1,1), data_format=data_format)(o) + o = Conv2D( 256, (3, 3), padding='valid', data_format=data_format, kernel_regularizer=l2(weight_decay))(o) + o = BatchNormalization(axis=bn_axis)(o) o = Activation('relu')(o) - - o = ( UpSampling2D( (2,2), data_format=IMAGE_ORDERING))(o) - o = ( concatenate([o,f2],axis=MERGE_AXIS ) ) - o = ( ZeroPadding2D((1,1) , data_format=IMAGE_ORDERING))(o) - o = ( Conv2D( 128 , (3, 3), padding='valid' , data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay) ) )(o) - o = ( BatchNormalization(axis=bn_axis))(o) + o = UpSampling2D( (2,2), data_format=data_format)(o) + o = Concatenate(axis=merge_axis)([o,f2]) + o = ZeroPadding2D((1,1) , data_format=data_format)(o) + o = Conv2D( 128 , (3, 3), padding='valid', data_format=data_format, kernel_regularizer=l2(weight_decay))(o) + o = BatchNormalization(axis=bn_axis)(o) o = Activation('relu')(o) - - o = ( UpSampling2D( (2,2), data_format=IMAGE_ORDERING))(o) - o = ( concatenate([o,f1],axis=MERGE_AXIS ) ) - o = ( ZeroPadding2D((1,1) , data_format=IMAGE_ORDERING ))(o) - o = ( Conv2D( 64 , (3, 3), padding='valid' , data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay) ))(o) - o = ( BatchNormalization(axis=bn_axis))(o) + o = UpSampling2D( (2,2), data_format=data_format)(o) + o = Concatenate(axis=merge_axis)([o,f1]) + o = ZeroPadding2D((1,1) , data_format=data_format)(o) + o = Conv2D( 64 , (3, 3), padding='valid', data_format=data_format, kernel_regularizer=l2(weight_decay))(o) + o = BatchNormalization(axis=bn_axis)(o) o = Activation('relu')(o) - - o = ( UpSampling2D( (2,2), data_format=IMAGE_ORDERING))(o) - o = ( concatenate([o,img_input],axis=MERGE_AXIS ) ) - o = ( ZeroPadding2D((1,1) , data_format=IMAGE_ORDERING ))(o) - o = ( Conv2D( 32 , (3, 3), padding='valid' , data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay) ))(o) - o = ( BatchNormalization(axis=bn_axis))(o) + o = UpSampling2D( (2,2), data_format=data_format)(o) + o = Concatenate(axis=merge_axis)([o, padded_to_multiple]) + o = ZeroPadding2D((1,1) , data_format=data_format)(o) + o = Conv2D(32, (3, 3), padding='valid', data_format=data_format, kernel_regularizer=l2(weight_decay))(o) + o = BatchNormalization(axis=bn_axis)(o) o = Activation('relu')(o) + o = Conv2D(n_classes, (1, 1), padding='same', data_format=data_format, kernel_regularizer=l2(weight_decay))(o) + if not skip_last_batchnorm: + o = BatchNormalization(axis=bn_axis)(o) - o = Conv2D( n_classes , (1, 1) , padding='same', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay) )( o ) - o = ( BatchNormalization(axis=bn_axis))(o) - o = (Activation('softmax'))(o) - - model = Model( img_input , o ) + o = Activation(last_activation)(o) + o = CutTo()([o, img_input]) - - - - return model + return Model(img_input , o)