facebookresearch/ParlAI

Retire custom AdamWithDecay

Open

#2,232 opened on Dec 2, 2019

View on GitHub
 (1 comment) (0 reactions) (0 assignees)Python (10,371 stars) (2,089 forks)batch import
AgentsCode QualityHelp WantedMinordonotreap

Description

We have a custom Adam optimizer that says it could be removed via pytorch 1.2:

https://github.com/facebookresearch/ParlAI/blob/3a5de86667816abbdb0f08049c86d1e2c3aff03e/parlai/agents/bert_ranker/helpers.py#L242-L350

The task is to remove it and upgrade to the official pytorch version.

The helper function minimizes the pain required, though I'm wondering if we'll need to upgrade some state dicts, so make sure you run long tests. https://github.com/facebookresearch/ParlAI/blob/3a5de86667816abbdb0f08049c86d1e2c3aff03e/parlai/agents/bert_ranker/helpers.py#L205

Contributor guide

Tech stack
pythonpytorch
Domain
machine learningbackend
Issue type
refactor
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-2 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
fresh
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
mostly clear
Prerequisites
Familiarity with PyTorch optimizersUnderstanding of state dict migration
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
40
Research direction
Examine the custom AdamWithDecay implementation at https://github.com/facebookresearch/ParlAI/blob/3a5de86667816abbdb0f08049c86d1e2c3aff03e/parlai/agents/bert ranker/helpers.py#L242 L350 and the helper function that minimizes pain. Replace the custom optimizer with torch.optim.Adam, ensuring any state dict differences are handled. Run long tests to verify correctness. Check if any other parts of the codebase depend on the custom implementation.