# In[118]:


def inception(i, num_layers, w, stride):
    c = tf.layers.conv2d(
        inputs=i,
        filters=num_layers,
        kernel_size=(w, 1),
        padding='same',
        use_bias=False
    )
    c = tf.layers.conv2d(
        inputs=c,
        filters=num_layers,
        kernel_size=(1, w),
        padding='same',
        use_bias=False
    )
    bias = tf.Variable(tf.zeros([num_layers]))
    c = tf.nn.bias_add(c, bias)
    c = tf.nn.relu(c)
    return c


# In[140]:


tf.reset_default_graph()
sess = tf.InteractiveSession()

x = tf.placeholder(tf.float32, shape=(None, None, None, 3))
y = tf.placeholder(tf.float32, shape=(None, None, None, 1))

c = inception(x, 64, 5)
for _ in range(16):
    c = tf.concat([
        inception(c, 16, 7),
        inception(c, 16, 5),
        inception(c, 32, 3),
    ], 3)
o = tf.layers.conv2d(c, 1, (1, 1), activation=tf.nn.sigmoid)

y_flat = tf.layers.flatten(y)
o_flat = tf.layers.flatten(o)
loss_op = tf.reduce_mean(tf.keras.backend.binary_crossentropy(y_flat, o_flat), 1)
train_op = tf.train.AdamOptimizer(1e-4).minimize(loss_op)
iou_op = mean_iou(y_flat, o_flat)