def RepresentationMask(mask, serializer, **unused_kwargs):
"""Upsamples a mask to cover the serialized representation."""
# Trax enforces the mask to be of the same size as the target. Get rid of the
# extra dimensions.
mask = np.amax(mask, axis=tuple(range(2, mask.ndim)))
return np.broadcast_to(mask[:, :, None],
mask.shape + (serializer.representation_length, ))
def _update_sketched(self, grads, weights, m, v, opt_params):
"""Update for higher-rank parameters."""
learning_rate = opt_params['learning_rate']
momentum = opt_params['momentum']
shape = weights.shape
rank = len(shape)
reshaped_accumulators = [np.reshape(v[i], self._expanded_shape(shape, i))
for i in range(rank)]
current_accumulator = self._minimum(reshaped_accumulators)
current_accumulator += grads * grads
accumulator_inv_sqrt = np.where(current_accumulator > 0.0,
1.0 / np.sqrt(current_accumulator),
np.zeros_like(current_accumulator))
preconditioned_gradient = grads * accumulator_inv_sqrt
m = (1.0 - momentum) * preconditioned_gradient + momentum * m
weights = weights - (learning_rate * m).astype(weights.dtype)
for i in range(len(v)):
axes = list(range(int(i))) + list(range(int(i) + 1, rank))
dim_accumulator = np.amax(current_accumulator, axis=axes)
v[i] = dim_accumulator
return weights, (m, v)
def representation_mask(mask):
mask = jnp.amax(mask, axis=tuple(range(2, mask.ndim)))
return jnp.broadcast_to(
mask[:, :, None],
mask.shape + (serializer.representation_length, ))
