The core library contains the following files:
- agents.py: this file contains a few basic agents which can be extended by your own model
- Agent: base class for all other agents, implements the act() method which receives an observation table and returns a table in response
- Teacher: child of Agent, also implements the report method for returning metrics. Tasks implement the Teacher class
- MultiTaskTeacher: creates a set of teachers based on a "task string" passed to the Teacher, creating multiple teachers within it and alternating between them
- create_task_teacher: instantiate a teacher from a given task string (e.g. 'babi:task:1' or 'squad')
- build_data.py: basic utilities for setting up data for tasks. you can override if your filesystem needs different functionality.
- dict.py: contains code for building general NLP-style dictionaries from observations
- DictionaryAgent: agent which tracks the index and frequency of words in a dictionary, and can parse a sentence into indices into its dictionary or back
- gpt2_helper.py: byte pair encoding utilities from GPT-2
- image_featurizers.py: provides functionality for loading images
- loader.py: functions for loading world, agents, tasks, and teacher modules
- logs: utilities for logging metrics to tensorboard
- message: contains message object
- Message: class for objects and observations in ParlAI
- metrics.py: computes evaluation metrics, e.g. ranking metrics, etc.
- params.py: uses argparse to interpret command line arguments for ParlAI
- teachers.py: contains teachers that deal with dialogue-based tasks, as well as data classes for storing data
- FixedDialogTeacher: base class for a teacher that utilizes fixed data
- DialogTeacher: base class for a teacher doing dialogue with fixed chat logs
- ParlAIDialogTeacher: a teacher that implements a simple standard text format for many tasks (non-visual tasks only)
- torch_agent: utility code for building PyTorch-based agents in ParlAI
- TorchAgent: class which serves as a useful parent class for other model agents
- Batch: namedtuple which is the input type of the main abstract methods of the TorchAgent class
- Output: namedtuple which is the expected output type of the main abstract methods of the TorchAgent class
- History: class which handles tracking the dialogue state over the course of an episode.
- torch_classifier_agent: abstract agent which extends TorchAgent and contains useful utilities for classification models
- torch_generator_agent: abstract agent which extends TorchAgent and contains useful utilities for generation models
- torch_ranker_agent: abstract agent which extends TorchAgent and contains useful utilities for ranking models
- worlds.py: contains a set of basic worlds for tasks to take place inside
- World: base class for all other worlds, implements
parley,shutdown,__enter__, and__exit__ - DialogPartnerWorld: default world for turn-based two-agent communication
- MultiAgentDialogWorld: round-robin turn-based agent communication for two or more agents
- HogwildWorld: default world for setting up a separate world for every thread when using multiple threads (processes)
- World: base class for all other worlds, implements