read training strategy from config (#66)

Summary:
* Read training strategy from config file
* Add `strategy: ddp` to all existing config files
* Can also pass in strategy from the terminal by `python train.py config=configs/pretraining/debug.yaml training.lightning.strategy=ddp`
* Priority: terminal argument > config

Pull Request resolved: #66

Test Plan:
logged `self.trainer.strategy` in `FLAVAPreTrainingLightningModule`

test 1
* command: `python train.py config=configs/pretraining/debug.yaml`
* config: `strategy: ddp`
* result: `DDPStrategy`

test 2
* command: `python train.py config=configs/pretraining/debug.yaml training.lightning.strategy=deepspeed`
* config: `strategy: ddp`
* result: `DeepSpeedStrategy`

Reviewed By: ebsmothers

Differential Revision: D37124420

Pulled By: katrina433

fbshipit-source-id: a02cafba7ad784963a0c27ad543256fd5daaa417

Author: katrina433

examples/flava/configs/finetuning/qnli.yaml

@@ -8,6 +8,7 @@ training:
     progress_bar_refresh_rate: 50
     val_check_interval: 1000
     num_sanity_val_steps: 0
+    strategy: ddp
   lightning_checkpoint:
     dirpath: "."
     filename: flava-{epoch:02d}-{step}

examples/flava/configs/finetuning/rendered_sst2.yaml

@@ -8,6 +8,7 @@ training:
     progress_bar_refresh_rate: 50
     val_check_interval: 100
     num_sanity_val_steps: 0
+    strategy: ddp
   lightning_checkpoint:
     dirpath: "."
     filename: flava-{epoch:02d}-{step}

examples/flava/configs/pretraining/debug.yaml

@@ -7,6 +7,7 @@ training:
     progress_bar_refresh_rate: 50
     val_check_interval: 10000
     num_sanity_val_steps: 0
+    strategy: ddp
   lightning_checkpoint:
     dirpath: "."
     filename: flava-{epoch:02d}-{step}

examples/flava/train.py

@@ -68,7 +68,6 @@ def main():
     trainer = Trainer(
         **OmegaConf.to_container(config.training.lightning),
         callbacks=callbacks,
-        strategy="ddp",
     )
     ckpt_path = config.training.lightning_load_from_checkpoint