I reproduced the results from your paper by running train_hclt.py on the MNIST dataset (achieving 1.22 bpd); however, after using backward with the DLTM model, I received the following output:
Hi, many thanks for playing with the code! The backward pass will not be implemented in PICs, as sampling from those energy-based-like neural nets is not straightforward and requires approximate techniques.
I reproduced the results from your paper by running![samples](https://github.com/gengala/pic/assets/55181191/4a1782d6-3a21-4e31-a852-93cbef1b3475)
train_hclt.py
on the MNIST dataset (achieving 1.22 bpd); however, after usingbackward
with theDLTM
model, I received the following output:Are you going to add
backward
also toPIC
?Many thanks.