Style-controlled generation

Agents for training and evaluating generative models conditioned on a style token, for instance, the "personality" string attached to each example of the Image-Chat dataset.

• --model projects.style_gen.style_gen:StyleGenAgent: Subclass of TransformerGeneratorAgent for which style tokens can be appended to the history for every training/evaluation example. Found in style_gen.py.
• --model projects.style_gen.classifier:ClassifierAgent: Subclass of TransformerGeneratorAgent that adds a classifier head on top of the base generator. Can be used for fine-tuning a classifier using the weights of a pretrained generator model. Classifier labels can be specified either in the 'labels' field of the observation or in a separate 'personality' field, and all encoder and decoder weights can optionally be frozen when training the classifier head. Found in classifier.py.

Basic examples

Evaluating a style-controlled generation model on the BlendedSkillTalk dataset, labeled with personalities from Image-Chat:

parlai eval_model \
--model-file zoo:style_gen/c75_labeled_dialogue_generator/model \
--model projects.style_gen.style_gen:StyleGenAgent \
--skip-generation True \
--task style_gen:LabeledBlendedSkillTalk \
--datatype test \
--use-style-frac 1.00

Evaluating the Image-Chat-trained classifier on the LabeledBlendedSkillTalk dataset, which it itself labeled with a style for each utterance:

parlai eval_model \
--model-file zoo:style_gen/prev_curr_classifier/model \
--model projects.style_gen.classifier:ClassifierAgent \
--classes-from-file image_chat_personalities_file \
--task style_gen:PrevCurrUttStyle \
--wrapper-task style_gen:LabeledBlendedSkillTalk