Entropy model in hific codebase

[mharshe]

I'm looking at using the code from model.py to create a model that we can retrain. Our idea is to use tensorflow 2.0 keras layers where possible, but to port the code I see that the archs.py has this block of code:

class FactorizedPriorLayer(tf.keras.layers.Layer):
"""Factorized prior to code a discrete tensor."""

def init(self):
"""Instantiate layer."""
super(FactorizedPriorLayer, self).init(name="FactorizedPrior")
self._entropy_model = tfc.EntropyBottleneck(
name="entropy_model")

and it the class Hyperprior

self._side_entropy_model = FactorizedPriorLayer()

The EntropyBottleneck class was removed in commit 89eefc8 so I'm not sure how this code will run? Should I replace that side_entropy_model with NoisyDeepFactorized?

[jonarchist]

Hi, in general, the functionality of EntropyBottleneck is now found in ContinuousBatchedEntropyModel with an instance of NoisyDeepFactorized as the prior argument, and the equivalent of GaussianConditional is ContinuousIndexedEntropyModel with NoisyGaussian used in the prior_fn argument. But regarding the details in the HiFiC code base, @fab-jul is probably a better contact to ask.

[fab-jul]

The code should work with vanilla TFC classes, so using what you mentioned sounds good.

[mharshe]

I'm looking into changing the GaussianConditional to ContinuousIndexedEntropyModel but the input args are quite different.
The current code is :

 conditional_entropy_model = tfc.GaussianConditional(
        latent_scales, scale_table, mean=latent_means,
        name="conditional_entropy_model")

whereas the ContinuousIndexedEntropyModel expects the arguments : prior_fn, index_ranges, parameter_fns, coding_rank I'm not sure what the arguments need to be changed to.

[jonarchist]

In this case, try the location-scale entropy model.

[mharshe]

Ok, but still the question of the arguments remains -- GaussianConditional is given latent_scales & latent_means which are both the outputs of a Sequential block of 3 SignalConv2D layers. It isn't clear how these get translated to the arguments num_scales, scale_fn (which takes in indexes as a input parameter if I undertand right) and coding_rank of LocationScaleIndexedEntropyModel

[jonarchist]

Please take a look at this code example: https://github.com/tensorflow/compression/blob/master/models/bmshj2018.py#L157
num_scales and scale_fn represent the new way the scale_table is set up. coding_rank should be 3 since you presumably want to encode each HWC block of the NHWC tensor into its own bit sequence.