p2_train_advanced.py

config = default_config() # model dimensions
config['transpose_range'] = (0, 12) # data augmentation: transpose key
config['encode_position'] = True # enable beat positional encoding
config['mask_steps'] = 4 # attention mask: teacher forcing

model = get_language_model(arch=MusicTransformerXL, vocab_sz=len(data.vocab), config=config.copy())
learn = MusicLearner(data, model)

learn.to_fp16(dynamic=True, clip=0.5); # mixed precision

learn.fit_one_cycle(4) # one cycle learning rates