d169.py

from tensorflow.keras.layers import Conv2D, UpSampling2D, LeakyReLU, Concatenate
from tensorflow.keras import Model
from tensorflow.keras.applications import DenseNet169

class UpscaleBlock(Model):
    def __init__(self, filters, name):      
        super(UpscaleBlock, self).__init__()
        self.up = UpSampling2D(size=(2, 2), interpolation='bilinear', name=name+'_upsampling2d')
        self.concat = Concatenate(name=name+'_concat') # Skip connection        
        self.convA = Conv2D(filters=filters, kernel_size=3, strides=1, padding='same', name=name+'_convA')
        self.reluA = LeakyReLU(alpha=0.2)
        self.convB = Conv2D(filters=filters, kernel_size=3, strides=1, padding='same', name=name+'_convB')
        self.reluB = LeakyReLU(alpha=0.2)
    
    def call(self, x):        
        b = self.reluB( self.convB( self.reluA( self.convA( self.concat( [self.up(x[0]), x[1]] ) ) ) ) )
        return b 

class Encoder(Model):
    def __init__(self):
        super(Encoder, self).__init__()                
        self.base_model = DenseNet169(input_shape=(None, None, 3), include_top=False, weights='imagenet')   
        print('Base model loaded {}'.format(DenseNet169.__name__))
        
        # Create encoder model that produce final features along with multiple intermediate features
        outputs = [self.base_model.outputs[-1]]
        for name in ['pool1', 'pool2_pool', 'pool3_pool', 'conv1/relu'] : outputs.append( self.base_model.get_layer(name).output )        
        self.encoder = Model(inputs=self.base_model.inputs, outputs=outputs)
        
    def call(self, x):
        return self.encoder(x)
    
class Decoder(Model):
    def __init__(self, decode_filters):
        super(Decoder, self).__init__()        
        self.conv2 =  Conv2D(filters=decode_filters, kernel_size=1, padding='same', name='conv2')        
        self.up1 = UpscaleBlock(filters=decode_filters//2,  name='up1')
        self.up2 = UpscaleBlock(filters=decode_filters//4,  name='up2')
        self.up3 = UpscaleBlock(filters=decode_filters//8,  name='up3')
        self.up4 = UpscaleBlock(filters=decode_filters//16, name='up4')
        self.up5 = UpSampling2D(size=(2, 2), name='up5')
        self.conv3 = Conv2D(filters=1, kernel_size=3, strides=1, padding='same', name='conv3')       

    def call(self, features):        
        x, pool1, pool2, pool3, conv1 = features[0], features[1], features[2], features[3], features[4]
        up0 = self.conv2(x)        
        up1 = self.up1([up0, pool3])        
        up2 = self.up2([up1, pool2])        
        up3 = self.up3([up2, pool1])        
        up4 = self.up4([up3, conv1])
        up5 = self.up5(up4)
        return self.conv3( up5 )
    
class DepthEstimate(Model):
    def __init__(self):
        super(DepthEstimate, self).__init__()
        self.encoder = Encoder()
        self.decoder = Decoder( decode_filters = int(self.encoder.layers[-1].output[0].shape[-1] // 2 ) )
        print('\nModel created.')

    def call(self, x):
        return self.decoder( self.encoder(x) )