Skip to main content

Browse the archive

Revision 6210494748e601c23cf6edf9ff1ae5adf47a3c4d authored by Madian Khabsa on 11 March 2021, 03:13:27 UTC, committed by Facebook GitHub Bot on 11 March 2021, 03:15:10 UTC 
<fix> support providing configs that contain equal character (#808)
 
Summary:
Pull Request resolved: https://github.com/facebookresearch/mmf/pull/808

The current parser would fail if the opt value has `=` which is needed when providing filters. This diff takes the first occurrence of  `=` as the separator

Reviewed By: apsdehal

Differential Revision: D26955390

fbshipit-source-id: cedfb358de254a0c6d2a88744a0660984b045f47
1 parent c78a408
Raw File
sweep_visual_bert.py 
#!/usr/bin/env python

# Copyright (c) Facebook, Inc. and its affiliates.

import lib as sweep
from lib import hyperparam


def get_grid(args):
    max_update = 22000

    return [
        hyperparam("run_type", "train_val"),
        hyperparam("config", "projects/visual_bert/configs/vqa2/defaults.yaml"),
        # hyperparam("--fp16", save_dir_key=lambda val: "fp16"),
        hyperparam("training.num_workers", 5),
        hyperparam("dataset", "vqa2"),
        hyperparam("model", "visual_bert", save_dir_key=lambda val: val),
        # For nlvr2, we are able to fit batch of size 16 on single GPU with 16GB
        # memory. Same number is 32 for VQA2, so scale accordingly
        hyperparam(
            "training.batch_size", [512, 256], save_dir_key=lambda val: f"bs{val}"
        ),
        hyperparam("training.seed", 1, save_dir_key=lambda val: f"s{val}"),
        hyperparam("scheduler.type", ["warmup_cosine"]),
        hyperparam("scheduler.params.num_warmup_steps", 2000),
        hyperparam("scheduler.params.num_training_steps", max_update),
        hyperparam("optimizer.type", "adam_w", save_dir_key=lambda val: val),
        hyperparam(
            "optimizer.params.lr", [5e-5, 1e-5], save_dir_key=lambda val: f"lr{val}"
        ),
        hyperparam("optimizer.params.eps", 1e-8),
        hyperparam(
            "training.max_updates", max_update, save_dir_key=lambda val: f"mu{val}"
        ),
        hyperparam("training.log_format", "json"),
        hyperparam("training.pin_memory", True),
        hyperparam("training.log_interval", 1000),
        hyperparam("training.checkpoint_interval", 1000),
        hyperparam("training.evaluation_interval", 4000),
        hyperparam("training.find_unused_parameters", True),
        hyperparam(
            "model_config.visual_bert.freeze_base",
            [False],
            save_dir_key=lambda val: f"fb{val}",
        ),
    ]


def postprocess_hyperparams(args, config):
    """Postprocess a given hyperparameter configuration."""
    pass


if __name__ == "__main__":
    sweep.main(get_grid, postprocess_hyperparams)
back to top