Fix transformer training

[marksgraham]

I've noticed a bug in the way we're training transformers - currently, we are training them to predict token N, given tokens 0-N inclusive!

To fix this we need to shift the targets by one, so

loss = ce_loss(logits, latent)

becomes

loss = ce_loss(logits[:,:,:-1], quantizations_target[:,1:])

Alternatively we could change the return value of the inferer so it provides the shifted versions - any opinions @Warvito ?

[marksgraham]

This also affects the get_likelihood function

[Warvito]

Yes, you are right. It is one thing that is right in the 2d_vqvae_transformer_tutorial tutorial:

GenerativeModels/tutorials/generative/2d_vqvae_transformer/2d_vqvae_transformer_tutorial.py

Line 407 in 7d02598

quantizations_input = convert_tensor(quantizations[:, :-1], device, non_blocking=True)

but need to be fixed in the anomaly detection one:

GenerativeModels/tutorials/generative/anomaly_detection/anomaly_detection_with_transformers.py

Line 363 in 7d02598

loss = ce_loss(logits, quantizations_target)

[marksgraham]

Will also update 2d_vqvae_transformer_tutorial to use the inferer class