Implementing U-net for multi-class road segmentation

i found that Conv2DTranspose works better than UpSampling2D and here is a quick implementation using the same

def conv_block(tensor, nfilters, size=3, padding='same', initializer="he_normal"):
    x = Conv2D(filters=nfilters, kernel_size=(size, size), padding=padding, kernel_initializer=initializer)(tensor)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)
    x = Conv2D(filters=nfilters, kernel_size=(size, size), padding=padding, kernel_initializer=initializer)(x)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)
    return x


def deconv_block(tensor, residual, nfilters, size=3, padding='same', strides=(2, 2)):
    y = Conv2DTranspose(nfilters, kernel_size=(size, size), strides=strides, padding=padding)(tensor)
    y = concatenate([y, residual], axis=3)
    y = conv_block(y, nfilters)
    return y


def Unet(img_height, img_width, nclasses=3, filters=64):
# down
    input_layer = Input(shape=(img_height, img_width, 3), name='image_input')
    conv1 = conv_block(input_layer, nfilters=filters)
    conv1_out = MaxPooling2D(pool_size=(2, 2))(conv1)
    conv2 = conv_block(conv1_out, nfilters=filters*2)
    conv2_out = MaxPooling2D(pool_size=(2, 2))(conv2)
    conv3 = conv_block(conv2_out, nfilters=filters*4)
    conv3_out = MaxPooling2D(pool_size=(2, 2))(conv3)
    conv4 = conv_block(conv3_out, nfilters=filters*8)
    conv4_out = MaxPooling2D(pool_size=(2, 2))(conv4)
    conv4_out = Dropout(0.5)(conv4_out)
    conv5 = conv_block(conv4_out, nfilters=filters*16)
    conv5 = Dropout(0.5)(conv5)
# up
    deconv6 = deconv_block(conv5, residual=conv4, nfilters=filters*8)
    deconv6 = Dropout(0.5)(deconv6)
    deconv7 = deconv_block(deconv6, residual=conv3, nfilters=filters*4)
    deconv7 = Dropout(0.5)(deconv7) 
    deconv8 = deconv_block(deconv7, residual=conv2, nfilters=filters*2)
    deconv9 = deconv_block(deconv8, residual=conv1, nfilters=filters)
# output
    output_layer = Conv2D(filters=nclasses, kernel_size=(1, 1))(deconv9)
    output_layer = BatchNormalization()(output_layer)
    output_layer = Activation('softmax')(output_layer)

    model = Model(inputs=input_layer, outputs=output_layer, name='Unet')
    return model

Now for the data generators, you can use the builtin ImageDataGenerator class
here is the code from Keras docs

# we create two instances with the same arguments
data_gen_args = dict(featurewise_center=True,
                     featurewise_std_normalization=True,
                     rotation_range=90,
                     width_shift_range=0.1,
                     height_shift_range=0.1,
                     zoom_range=0.2)
image_datagen = ImageDataGenerator(**data_gen_args)
mask_datagen = ImageDataGenerator(**data_gen_args)

# Provide the same seed and keyword arguments to the fit and flow methods
seed = 1
image_datagen.fit(images, augment=True, seed=seed)
mask_datagen.fit(masks, augment=True, seed=seed)

image_generator = image_datagen.flow_from_directory(
    'data/images',
    class_mode=None,
    seed=seed)

mask_generator = mask_datagen.flow_from_directory(
    'data/masks',
    class_mode=None,
    seed=seed)

# combine generators into one which yields image and masks
train_generator = zip(image_generator, mask_generator)

model.fit_generator(
    train_generator,
    steps_per_epoch=2000,
    epochs=50)

Another way to go is implement your own generator by extending the Sequence class from Keras

class seg_gen(Sequence):
    def __init__(self, x_set, y_set, batch_size, image_dir, mask_dir):
        self.x, self.y = x_set, y_set
        self.batch_size = batch_size
        self.samples = len(self.x)
        self.image_dir = image_dir
        self.mask_dir = mask_dir

    def __len__(self):
        return int(np.ceil(len(self.x) / float(self.batch_size)))

    def __getitem__(self, idx):
        idx = np.random.randint(0, self.samples, batch_size)
        batch_x, batch_y = [], []
        drawn = 0
        for i in idx:
            _image = image.img_to_array(image.load_img(f'{self.image_dir}/{self.x[i]}', target_size=(img_height, img_width)))/255.   
            mask = image.img_to_array(image.load_img(f'{self.mask_dir}/{self.y[i]}', grayscale=True, target_size=(img_height, img_width)))
#             mask = np.resize(mask,(img_height*img_width, classes))
            batch_y.append(mask)
            batch_x.append(_image)
        return np.array(batch_x), np.array(batch_y)

Here is a sample code to train the model

unet = Unet(256, 256, nclasses=66, filters=64)
print(unet.output_shape)
p_unet = multi_gpu_model(unet, 4)
p_unet.load_weights('models-dr/top_weights.h5')
p_unet.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
tb = TensorBoard(log_dir='logs', write_graph=True)
mc = ModelCheckpoint(mode='max', filepath='models-dr/top_weights.h5', monitor='acc', save_best_only='True', save_weights_only='True', verbose=1)
es = EarlyStopping(mode='max', monitor='acc', patience=6, verbose=1)
callbacks = [tb, mc, es]
train_gen = seg_gen(image_list, mask_list, batch_size)


p_unet.fit_generator(train_gen, steps_per_epoch=steps, epochs=13, callbacks=callbacks, workers=8)

I have tried using the dice loss when i had only two classes, here is the code for it

def dice_coeff(y_true, y_pred):
    smooth = 1.
    y_true_f = K.flatten(y_true)
    y_pred_f = K.flatten(y_pred)
    intersection = K.sum(y_true_f * y_pred_f)
    score = (2. * intersection + smooth) / (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)
    return score

def dice_loss(y_true, y_pred):
    loss = 1 - dice_coeff(y_true, y_pred)
    return loss