Update README.md

Author: rwightman

README.md

@@ -229,7 +229,7 @@ NOTE:
 #### EfficientDet-D0
 
 Latest training run with .336 for D0 (on 4x 1080ti):
-`./distributed_train.sh 4 --model efficientdet_d0 -b 22 --amp --lr .12 --sync-bn --opt fusedmomentum --warmup-epochs 5 --lr-noise 0.4 0.9 --model-ema --model-ema-decay 0.9999`
+`./distributed_train.sh 4 /mscoco --model efficientdet_d0 -b 22 --amp --lr .12 --sync-bn --opt fusedmomentum --warmup-epochs 5 --lr-noise 0.4 0.9 --model-ema --model-ema-decay 0.9999`
 
 These hparams above resulted in a good model, a few points:
 * the mAP peaked very early (epoch 200 of 300) and then appeared to overfit, so likely still room for improvement
@@ -238,6 +238,13 @@ These hparams above resulted in a good model, a few points:
 * drop_path (aka survival_prob / drop_connect) rate of 0.1, which is higher than the suggested 0.0 for D0 in official, but lower than the 0.2 for the other models
 * longer EMA period than default
 
+#### EfficientDet-D1 
+
+Latest run with .394 mAP (on 4x 1080ti):
+`./distributed_train.sh 4 /mscoco --model efficientdet_d1 -b 10 --amp --lr .06 --sync-bn --opt fusedmomentum --warmup-epochs 5 --lr-noise 0.4 0.9 --model-ema --model-ema-decay 0.99995`
+
+For this run I used some improved augmentations, still experimenting so not ready for release, should work well without them but will likely start overfitting a bit sooner and possibly end up a in the .385-.39 range.
+
 VAL2017
 ```
  Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.336251