Finetuning on custom dataset !

Hi,

Thanks for sharing this amazing code :)

We want to finetune pre-trained X3D model on our custom dataset with 34 classes.

We made the following changes and wanted to check if these changes are sufficient for fine-tuning the complete model on a new dataset.

Should we decrease the learning rate also as we are fine-tuning ?

Any other empirical suggestions for hyperparams are most welcome !!

```
TRAIN.CHECKPOINT_EPOCH_RESET True # Make this True else scaler throws dimension mismatch error
TRAIN.CHECKPOINT_FILE_PATH x3d_l.pyth # Pretrained checkpoint from model zoo
MODEL.NUM_CLASSES 34 # Specify number of classes in our dataset
```

Thanks Again !