Training is not steady (exploding gradient ) when training with UCF101 datasets and base_channel_size=256.

[shim0114]

--timesteps 256 --loss_type l2 --train_lr 0.0003 --beta2 0.99 --train_num_steps 600000 --train_batch_size 16 --gradient_accumulate_every 4 --ema_decay 0.9999 --base_channel_size 256 --image_size 64