1f720ee453871b2ab764f608926281716ef7bf81,tools/train_pl.py,,,#,340

Before Change

lit = LitModel(opt)
// warning grad_clip_mode is ignored.
trainer = pl.Trainer(
    callbacks=[OnEpochStartCallback()],
    default_root_dir=opt.checkpoint_path,
    resume_from_checkpoint=resume_from,
    distributed_backend="ddp",

After Change

    // fast_dev_run=True,
)

if os.getenv("EVALUATE", "0") == "1":
    lit.load_state_dict(torch.load(resume_from)["state_dict"])
    // Because ddp cant work with test
    trainer = pl.Trainer(
        default_root_dir=opt.checkpoint_path,
        resume_from_checkpoint=resume_from,
        distributed_backend="dp",
        gpus=torch.cuda.device_count(),
        checkpoint_callback=checkpoint_callback,
        profiler=True,
        // limit_train_batches=500,
        // progress_bar_refresh_rate=0,
        // fast_dev_run=True,
    )
    trainer.test(lit)
else:
    trainer.fit(lit)

Instances