I am trying to implement sample- and pixel-dependent dependent loss weighting in tf.Keras (TensorFlow 2.0.0rc0) for a 3-D U-Net with sparse annotation data (Cicek 2016, arxiv:1606.06650).
This is my code:
import numpy as np
import tensorflow as tf
from tensorflow.keras import layers, losses, models
# disabling eager execution makes this example work:
# tf.python.framework_ops.disable_eager_execution()
def get_loss_fcn(w):
def loss_fcn(y_true, y_pred):
loss = w * losses.mse(y_true, y_pred)
return loss
return loss_fcn
data_x = np.random.rand(5, 4, 1)
data_w = np.random.rand(5, 4)
data_y = np.random.rand(5, 4, 1)
x = layers.Input([4, 1])
w = layers.Input([4])
y = layers.Activation('tanh')(x)
model = models.Model(inputs=[x, w], outputs=y)
loss = get_loss_fcn(model.input[1])
# using another loss makes it work, too:
# loss = 'mse'
model.compile(loss=loss)
model.fit((data_x, data_w), data_y)
print('Done.')
This runs fine when disabling eager execution, but one of the points of TensorFlow 2 is to have eager execution by default. What stands between me and that goal is the custom loss function, as you can see (using 'mse' as a loss removes that error, too):
File "MWE.py", line 30, in <module>
model.fit((data_x, data_w), data_y)
[...]
tensorflow.python.eager.core._SymbolicException: Inputs to eager execution function cannot be Keras symbolic tensors, but found [<tf.Tensor 'input_2:0' shape=(None, 4) dtype=float32>]
What can I do to make this kind of structure work with eager execution?
One idea that I had was to concatenate w to the output y and separate y_pred into the original y_pred and w in the loss function, but this is a hack I'd like to avoid. It works, though, with changes marked by # HERE:
import numpy as np
import tensorflow as tf
from tensorflow.keras import layers, losses, models
# HERE
def loss_fcn(y_true, y_pred):
w = y_pred[:, :, -1] # HERE
y_pred = y_pred[:, :, :-1] # HERE
loss = w * losses.mse(y_true, y_pred)
return loss
data_x = np.random.rand(5, 4, 1)
data_w = np.random.rand(5, 4, 1) # HERE
data_y = np.random.rand(5, 4, 1)
x = layers.Input([4, 1])
w = layers.Input([4, 1]) # HERE
y = layers.Activation('tanh')(x)
output = layers.Concatenate()([y, w]) # HERE
model = models.Model(inputs=[x, w], outputs=output) # HERE
loss = loss_fcn # HERE
model.compile(loss=loss)
model.fit((data_x, data_w), data_y)
print('Done.')
Any other ideas?
One alternative solution is to pass weights as additional output features rather than input features.
This keeps the model completely free of anything weights related, and the weights appear only in the loss function and the .fit() call:
import numpy as np
import tensorflow as tf
from tensorflow.keras import layers, losses, models
data_x = 2 * np.ones((7, 11, 15, 3), dtype=float)
data_y = 5 * np.ones((7, 9, 13, 5), dtype=float)
x = layers.Input(data_x.shape[1:])
y = layers.Conv2D(5, kernel_size=3)(x)
model = models.Model(inputs=x, outputs=y)
def loss(y_true, y_pred):
(y_true, w) = tf.split(y_true, num_or_size_splits=[-1, 1], axis=-1)
loss = tf.squeeze(w, axis=-1) * losses.mse(y_true, y_pred)
tf.print(tf.math.reduce_mean(y_true), "== 5")
tf.print(tf.math.reduce_mean(w), "== 3")
return loss
model.compile(loss=loss)
data_w = 3 * np.ones((7, 9, 13, 1), dtype=float)
data_yw = np.concatenate((data_y, data_w), axis=-1)
model.fit(data_x, data_yw)
One drawback still is that you need to manipulate (potentially) large arrays when merging y and w in numpy.stack(), so anymore more TensorFlow-like will be appreciated.