ContinuousBatchedEntropyModel.bits/8 differ significantly from compress strings length about compression

[trougnouf]

I can reproduce this early in training:
the side_string bytes are evaluated as above after 400 iterations:

1.400:  mse loss: 0.08312108367681503;  bpp: 0.25097793340682983;       combined loss: 0.25180914998054504
eval: entropy_bottleneck.bits(z, training=False)[0]/8: 1778.6331787109375
train: entropy_bottleneck.bits(z, training=True)[0]/8: 1778.98876953125
compress: len(entropy_bottleneck.compress(z).numpy()[0]): 2104

The bits() function decreases a bit while the length of compress is pretty much constant

Some process with optimizer2.apply_gradients(...) commented out to make sure that the entropy_bottleneck was being optimized:

1.400:  mse loss: 0.072761669754982;    bpp: 0.3175785541534424;        combined loss: 0.31830617785453796                                                                
eval: 2095.109130859375                                                                                                                                                   
train: 2095.53662109375                                                                                                                                                   
compress: 2096   

Here the overall bpp decreases but the size of the side_strings is never optimized.

[trougnouf]

There doesn't seem to be an issue if I instantiate a new ContinuousBatchedEntropyModel with compression=True using a reference to the same prior, and save that model for inference, as is suggested in the code comments.