ddp bug fix

Summary:
1. migrated to d2go lr_scheduler for 100% match
2. fixed the ddp module bug in compute loss, the loss should be computed using a ddp module, not it's child, which cause gradient to not sycn
3. added 2 configs to train from scratch

Reviewed By: srihari2761

Differential Revision: D67433973

fbshipit-source-id: 53a14c02a3915746b262641808ccd3d694afb548

Author: dspgbgjd

detectron2/data/build.py

@@ -490,7 +490,7 @@ def _train_loader_from_config(cfg, mapper=None, *, dataset=None, sampler=None):
         else:
             logger.info("Using training sampler {}".format(sampler_name))
             if sampler_name == "TrainingSampler":
-                sampler = TrainingSampler(len(dataset))
+                sampler = TrainingSampler(len(dataset), seed=cfg.SEED)
             elif sampler_name == "RepeatFactorTrainingSampler":
                 repeat_factors = RepeatFactorTrainingSampler.repeat_factors_from_category_frequency(
                     dataset, cfg.DATALOADER.REPEAT_THRESHOLD, sqrt=cfg.DATALOADER.REPEAT_SQRT