When training jukebox on notebooks, why is our training data is stuck at 0 for Sample length and min duration? - pytorch

Question

Hps data:
/home/dotty/anaconda3/envs/JukVersionRoy/jukebox/train.py --hps=small_vqvae --name=small_vqvae --sample_length=9999999999999 --bs=2 --audio_files_dir=/home/dotty/anaconda3/envs/JukVersionRoy/jukebox/my-training-data --labels=False --train --aug_shift --aug_blend
Here
/
HPS ***************
{'rcall_command': '<unknown_rcall_command>', 'git_commit': '<unknown_git_commit>', 'name': '', 'debug_mem': False, 'debug_eval_files': False, 'debug_speed': False, 'debug_iters': 100, 'debug_batch': False, 'debug_grad_accum': False, 'debug_inputs': False, 'local_path': '', 'local_logdir': 'logs', 'max_len': 24, 'max_log': 32, 'save': True, 'save_iters': 20000, 'seed': 0, 'prior': False, 'log_steps': 100, 'func': '', 'audio_files_dir': '', 'finetune': '', 'english_only': False, 'bs': 1, 'bs_sample': 1, 'nworkers': 1, 'aug_shift': False, 'aug_blend': False, 'train_test_split': 0.9, 'train_shrink_factor': 1.0, 'test_shrink_factor': 1.0, 'p_unk': 0.1, 'min_duration': None, 'max_duration': None, 'n_tokens': 0, 'n_vocab': 0, 'use_tokens': False, 'curr_epoch': -1, 'restore_vqvae': '', 'levels': 2, 'downs_t': (1, 1), 'strides_t': (2, 2), 'hvqvae_multipliers': None, 'revival_threshold': 1.0, 'emb_width': 64, 'l_bins': 512, 'l_mu': 0.99, 'commit': 1.0, 'spectral': 0.0, 'multispectral': 1.0, 'loss_fn': 'l2', 'linf_k': 2048, 'lmix_l1': 0.0, 'lmix_l2': 0.0, 'lmix_linf': 0.0, 'use_bottleneck': True, 'depth': 3, 'width': 128, 'm_conv': 1.0, 'dilation_growth_rate': 1, 'dilation_cycle': None, 'vqvae_reverse_decoder_dilation': True, 'restore_prior': '', 'restore_prior_ddp': False, 'max_bow_genre_size': None, 'y_bins': 0, 'level': 0, 'cond_levels': None, 't_bins': 64, 'y_cond_as_bias': False, 'copy_input': False, 'merged_decoder': False, 'single_enc_dec': False, 'alignment_layer': None, 'alignment_head': None, 'n_ctx': 1024, 'prior_depth': 3, 'prior_width': 128, 'heads': 1, 'attn_order': 0, 'blocks': None, 'spread': None, 'attn_dropout': 0.0, 'resid_dropout': 0.0, 'emb_dropout': 0.0, 'zero_out': False, 'res_scale': False, 'pos_init': False, 'init_scale': 1.0, 'm_attn': 0.25, 'm_mlp': 1.0, 'c_res': 0, 'c_attn': 0, 'c_mlp': 0, 'cond_depth': 3, 'cond_width': 128, 'cond_m_conv': 1.0, 'cond_zero_out': False, 'cond_res_scale': False, 'cond_dilation_growth_rate': 1, 'cond_dilation_cycle': None, 'cond_c_res': 0, 'primed_chunk_size': None, 'selected_artists': '', 'temp_top': 1.0, 'temp_rest': 0.99, 'sample_length_in_seconds': 24, 'total_sample_length_in_seconds': 240, 'prime_loss_fraction': 0.1, 'restore_decoder': '', 'prime_depth': 3, 'prime_width': 128, 'prime_heads': 1, 'prime_attn_order': 0, 'prime_blocks': None, 'prime_spread': None, 'prime_attn_dropout': 0.0, 'prime_resid_dropout': 0.0, 'prime_emb_dropout': 0.0, 'prime_zero_out': False, 'prime_res_scale': False, 'prime_pos_init': False, 'prime_init_scale': 1.0, 'prime_m_attn': 0.25, 'prime_m_mlp': 1.0, 'prime_c_res': 0, 'prime_c_attn': 0, 'prime_c_mlp': 0, 'prime_rel_attn': False, 'prime_posemb_timescale': 10000, 'epochs': 10000, 'lr': 0.0003, 'clip': 1.0, 'beta1': 0.9, 'beta2': 0.999, 'ignore_grad_norm': 0, 'weight_decay': 0.0, 'eps': 1e-08, 'lr_warmup': 100.0, 'lr_decay': 10000000000.0, 'lr_gamma': 1.0, 'lr_scale': 1.0, 'lr_use_linear_decay': False, 'lr_start_linear_decay': 0, 'lr_use_cosine_decay': False, 'fp16': False, 'fp16_params': False, 'fp16_loss_scale': None, 'fp16_scale_window': 1000.0, 'fp16_opt': False, 'labels': True, 'labels_v3': False, 'dump': False, 'ema': True, 'ema_fused': True, 'cpu_ema': False, 'cpu_ema_freq': 100, 'reset_best_loss': False, 'reset_step': False, 'reset_opt': False, 'reset_shd': False, 'train': False, 'test': False, 'sample': False, 'sampler': 'ancestral', 'codes_logdir': '', 'date': None, 'labeller': 'top_genres', 'label_line': 0, 'iters_before_update': 1, 'grad_accum_iters': 0, 'mu': None, 'piped': False, 'pipe_depth': 8, 'break_train': 10000000000.0, 'break_test': 10000000000.0, 'exit_train': 10000000000.0, 'n_fft': 1024, 'hop_length': 256, 'window_size': 1024, 'sr': 44100, 'channels': 2, 'wav': '', 'n_inps': 1, 'n_hops': 2, 'n_segment': 1, 'n_total_segment': 1, 'n_segment_each': 1, 'prime_chunks': 4, 'sample_length': 0, 'sample_hop_length': 30000, 'max_silence_pad_length': 0, 'ignore_boundaries': False, 'use_nonrelative_specloss': True, 'multispec_loss_n_fft': (2048, 1024, 512), 'multispec_loss_hop_length': (240, 120, 50), 'multispec_loss_window_size': (1200, 600, 240), 'bucket': 128, 'ngpus': 16, 'argv': '/home/dotty/anaconda3/envs/JukVersionRoy/jukebox/train.py --hps=small_vqvae --name=small_vqvae --sample_length=9999999999999 --bs=2 --audio_files_dir=/home/dotty/anaconda3/envs/JukVersionRoy/jukebox/my-training-data --labels=False --train --aug_shift --aug_blend'}
SAMPLE LENGTH:
0
SAMPLE RATE:
44100
SAMPLE LENGTH / SAMPLE RATE:
0.0
MIN DURATION:
0
AssertionError Traceback (most recent call last)
/tmp/ipykernel_14312/1841762803.py in <module>
325
326
--> 327 run()
/tmp/ipykernel_14312/1841762803.py in run(hps, port, **kwargs)
283 print("HPS ***************")
284 print(hps)
--> 285 data_processor = DataProcessor(hps)
286
287 # Setup models
~/anaconda3/envs/JukVersionRoy/jukebox/data/data_processor.py in init(self, hps)
24 class DataProcessor():
25 def init(self, hps):
---> 26 self.dataset = FilesAudioDataset(hps)
27 duration = 1 if hps.prior else 600
28 hps.bandwidth = calculate_bandwidth(self.dataset, hps, duration=duration)
~/anaconda3/envs/JukVersionRoy/jukebox/data/files_dataset.py in init(self, hps)
24 print("MIN DURATION:")
25 print(self.min_duration)
---> 26 assert hps.sample_length / hps.sr < self.min_duration, f'Sample length {hps.sample_length} per sr {hps.sr} ({hps.sample_length / hps.sr:.2f}) should be shorter than min duration {self.min_duration}'
27 self.aug_shift = hps.aug_shift
28 self.labels = hps.labels
AssertionError: Sample length 0 per sr 44100 (0.00) should be shorter than min duration 0
Our files are not being read in, as the sample stayed at 44100 when we changed the files being read in. we went through all the files dataset.py, dataprocessor.py and audio_utils.py and could not finds where exactly the input files is being registered. Our path is correct. We are trying to sample length and min duration to be read in correctly.