Add options for more pretrained models

[vochicong]

Hi,

I'm trying PPLM with XLNet instead of GPT-2 and encountered errors regarding the output format. It seems that the GPT-2 model has 3 output values, but the XLNet model has only 2.

/content/PPLM/run_pplm.py in generate_text_pplm(model, tokenizer, context, past, device, perturb, bow_indices, classifier, class_label, loss_type, length, stepsize, temperature, top_k, sample, num_iterations, grad_length, horizon_length, window_length, decay, gamma, gm_scale, kl_scale, verbosity_level)
    594                 past = model(output_so_far[:, :-1])[1]
    595 
--> 596         unpert_logits, unpert_past, unpert_all_hidden = model(output_so_far)
    597         unpert_last_hidden = unpert_all_hidden[-1]
    598 

ValueError: not enough values to unpack (expected 3, got 2)

What should I try? Could you give me some hints?

My experimental notebook: https://colab.research.google.com/drive/1ywRuiFMC1NTXiIpg0xTihF8ljP0W1ECS

Thank you!

[w4nderlust]

Do you know what is XLNet outputting instead of those 3 tensors?

[CLAassistant]

All committers have signed the CLA.