multi-GPU support still WIP

[anttiryt]

According to this post, is needed:

strategy = tf.distribute.MirroredStrategy()
with strategy.scope():

It is still not enough.

File "/home/rac/slomo/multifix/super-slomo/train.py", line 207, in train_step  *
    loss_values = loss_obj.compute_losses(
File "/home/rac/slomo/multifix/super-slomo/models/losses.py", line 122, in compute_losses  *
    p_rec_loss += self.reconstruction_loss(true, pred)
File "/home/rac/slomo/multifix/super-slomo/models/losses.py", line 26, in reconstruction_loss  *
    return self.mae(y_true, y_pred)
File "/home/rac/slomo/tf-cuda.env/lib/python3.10/site-packages/keras/losses.py", line 166, in __call__  **
    reduction = self._get_reduction()
File "/home/rac/slomo/tf-cuda.env/lib/python3.10/site-packages/keras/losses.py", line 217, in _get_reduction
    raise ValueError(

ValueError: Please use `tf.keras.losses.Reduction.SUM` or `tf.keras.losses.Reduction.NONE` for loss reduction when losses are used with `tf.distribute.Strategy` outside of the built-in training loops. You can implement `tf.keras.losses.Reduction.SUM_OVER_BATCH_SIZE` using global batch size like:
```
with strategy.scope():
    loss_obj = tf.keras.losses.CategoricalCrossentropy(reduction=tf.keras.losses.Reduction.NONE)
....
    loss = tf.reduce_sum(loss_obj(labels, predictions)) * (1. / global_batch_size)
```