[fbsync] Clarification for training resnext101_32x8d on ImageNet (#4390)

Summary:
* Fix training resuming in references/segmentation

* Clarification for training resnext101_32x8d

* Update references/classification/README.md

Reviewed By: kazhang

Differential Revision: D30898330

fbshipit-source-id: 195c24c57ad3abe2e23e08b3b9251db68790914c

Co-authored-by: Nicolas Hug <contact@nicolas-hug.com>
Co-authored-by: Nicolas Hug <contact@nicolas-hug.com>

references/classification/README.md

@@ -40,12 +40,17 @@ python -m torch.distributed.launch --nproc_per_node=8 --use_env train.py\
 
 ### ResNext-101 32x8d
 
-On 8 nodes, each with 8 GPUs (for a total of 64 GPUS)
 ```
 python -m torch.distributed.launch --nproc_per_node=8 --use_env train.py\
     --model resnext101_32x8d --epochs 100
 ```
 
+Note that the above command corresponds to a single node with 8 GPUs. If you use
+a different number of GPUs and/or a different batch size, then the learning rate
+should be scaled accordingly. For example, the pretrained model provided by
+`torchvision` was trained on 8 nodes, each with 8 GPUs (for a total of 64 GPUs),
+with `--batch_size 16` and `--lr 0.4`, instead of the current defaults
+which are respectively batch_size=32 and lr=0.1
 
 ### MobileNetV2
 ```