CODEWITHSUNDEEP
pythontensorflowtensorflow-probability
M-P
M-P

Reputation: 53

Tensorflow probability - Bijector training

I have been trying to follow the example from this tutorial, but I am having trouble training any of the variables.

I wrote a small example, but I haven't been able to make that work either:

# Train a shift bijector
shift = tf.Variable(initial_value=tf.convert_to_tensor([1.0], dtype=tf.float32), trainable=True, name='shift_var')
bijector = tfp.bijectors.Shift(shift=shift)

# Input
x = tf.convert_to_tensor(np.array([0]), dtype=tf.float32)
target = tf.convert_to_tensor(np.array([2]), dtype=tf.float32)

optimizer = tf.optimizers.Adam(learning_rate=0.5)
nsteps = 1

print(bijector(x).numpy(), bijector.shift)
for _ in range(nsteps):

    with tf.GradientTape() as tape:
        out = bijector(x)
        loss = tf.math.square(tf.math.abs(out - target))
        #print(out, loss)
    
        gradients = tape.gradient(loss, bijector.trainable_variables)
    
    optimizer.apply_gradients(zip(gradients, bijector.trainable_variables))
    
print(bijector(x).numpy(), bijector.shift)

For nsteps = 1, the two print statements result in the following output:

[1.] <tf.Variable 'shift_var:0' shape=(1,) dtype=float32, numpy=array([1.], dtype=float32)>
[1.] <tf.Variable 'shift_var:0' shape=(1,) dtype=float32, numpy=array([1.4999993], dtype=float32)>

It seems like the bijector still uses the original shift even though the printed value of bijector.shift has been updated.

I cannot increase nsteps as the gradient is None after the first iteration, and I got this error:

ValueError: No gradients provided for any variable: ['shift_var:0'].

I'm using

tensorflow version 2.3.0
tensorflow-probability version 0.11.0

I also tried it on a colab notebook, so I doubt it's a version problem.

Upvotes: 1

Views: 417

Answers (2)

Brian Patton
Brian Patton

Reputation: 1076

You found a bug. The bijector forward function weakly caches the result->input mapping to make downstream inverses and log-determinants fast. But somehow this is also interfering with the gradient. A workaround is adding a del out, as in https://colab.research.google.com/gist/brianwa84/04249c2e9eb089c2d748d05ee2c32762/bijector-cache-bug.ipynb

Upvotes: 1

M-P
M-P

Reputation: 53

Still not sure I understand exactly what is going on here, but at least I can get my example to work now.

For some reason the behaviour is different if I wrap it in a class that inherits from tf.keras.Model:

class BijectorModel(tf.keras.Model):

    def __init__(self):
        super().__init__()

        self.shift = tf.Variable(initial_value=tf.convert_to_tensor([1.5], dtype=tf.float32), trainable=True, name='shift_var')
        self.bijector = tfp.bijectors.Shift(shift=self.shift)

    def call(self, input):
        return self.bijector(input)

I made a function for a training iteration, although this does not seem to be necessary:

def training_iteration(model, input, target):

    optimizer = tf.optimizers.SGD(learning_rate=0.1)

    with tf.GradientTape() as tape:

        loss = tf.math.square(tf.math.abs(model(input) - target))

        gradients = tape.gradient(loss, model.trainable_variables)

    optimizer.apply_gradients(zip(gradients, model.trainable_variables))

Executing like this

x = tf.convert_to_tensor(np.array([0]), dtype=tf.float32)
target = tf.convert_to_tensor(np.array([2]), dtype=tf.float32)
model = BijectorModel()

nsteps = 10
for _ in range(nsteps):
    training_iteration(model, x, target)
    print('Iteration {}: Output {}'.format(_, model(x)))

produces the expect/desired output:

Iteration 0: Output [1.6]
Iteration 1: Output [1.6800001]
Iteration 2: Output [1.7440001]
Iteration 3: Output [1.7952001]
Iteration 4: Output [1.8361601]
Iteration 5: Output [1.8689281]
Iteration 6: Output [1.8951424]
Iteration 7: Output [1.916114]
Iteration 8: Output [1.9328911]
Iteration 9: Output [1.9463129]

My conclusion is that trainable variables are handled differently when part of a model, compared to being accessed through a bijector-object.

Upvotes: 1

Related Questions