Upload folder using huggingface_hub

config.gin

@@ -0,0 +1,142 @@
+# Macros:
+# ==============================================================================
+new_freq = 24000
+
+# Parameters for AudioDataModule:
+# ==============================================================================
+AudioDataModule.num_workers = 20
+
+# Parameters for AudioDataset:
+# ==============================================================================
+AudioDataset.half_precision = True
+AudioDataset.mono = True
+AudioDataset.new_freq = %new_freq
+AudioDataset.num_frames = 480000
+AudioDataset.orig_freq = 16000
+
+# Parameters for build_dev_datamodule:
+# ==============================================================================
+build_dev_datamodule.datamodule = @discotube
+
+# Parameters for build_module:
+# ==============================================================================
+build_module.ckpt_path = 'model.ckpt'
+build_module.module = @modules.maskingmodel.MaskingModel
+build_module.net = @nets.conformer.Conformer
+build_module.representation = \
+    [@nets.cqt.CQT,
+     @nets.encodec.EnCodec,
+     @nets.melspectrogram.MelSpectrogram,
+     @nets.waveform.Waveform]
+
+# Parameters for Conformer:
+# ==============================================================================
+Conformer.alpha_deepnorm = 2.6321480259049848
+Conformer.beta_deepnorm = 0.022386873579657126
+Conformer.conv_kernel_size = 5
+Conformer.depth = 24
+Conformer.dropout = 0.2
+Conformer.embed_dim = 1024
+Conformer.input_dropout = 0.0
+Conformer.mlp_ratio = 4.0
+Conformer.mlp_residual_factor = 4.0
+Conformer.num_heads = 8
+Conformer.num_patches = None
+Conformer.use_deepnorm = True
+Conformer.use_rope = True
+
+# Parameters for CosineAnnealingCallback:
+# ==============================================================================
+CosineAnnealingCallback.eta_min = 1e-07
+CosineAnnealingCallback.warmup_steps = 30000
+
+# Parameters for CQT:
+# ==============================================================================
+CQT.bins_per_octave = 24
+CQT.f_min = 32.703
+CQT.hop_len = 320
+CQT.logC = True
+CQT.magnitude = True
+CQT.n_bins = 188
+CQT.norm_mean = 4.754879065310596
+CQT.norm_std = 1.9055732535255916
+CQT.patch_size = (188, 3)
+CQT.power = 2
+CQT.sr = %new_freq
+
+# Parameters for DiscotubeAudioDataModule:
+# ==============================================================================
+DiscotubeAudioDataModule.batch_size = 20
+DiscotubeAudioDataModule.data_dir = ''
+DiscotubeAudioDataModule.filelist_train = ''
+DiscotubeAudioDataModule.filelist_val = ''
+
+# Parameters for EnCodec:
+# ==============================================================================
+EnCodec.norm_type = 'global'
+EnCodec.orig_sr = %new_freq
+EnCodec.patch_size = (128, 3)
+EnCodec.stats_path = None
+EnCodec.weights_path = 'facebook/encodec_24khz'
+
+# Parameters for FiniteScalarQuantizer:
+# ==============================================================================
+FiniteScalarQuantizer.levels = [6, 6, 6, 6, 6]
+FiniteScalarQuantizer.preserve_symmetry = True
+
+# Parameters for MaskingModel:
+# ==============================================================================
+MaskingModel.codebook_dim = 1
+MaskingModel.codebook_size = 7776
+MaskingModel.diff_input = False
+MaskingModel.input_representation = @nets.waveform.Waveform
+MaskingModel.lr = 0.0001
+MaskingModel.mask_prob = 0.6
+MaskingModel.mask_seconds = 0.4
+MaskingModel.num_codebooks = 1
+MaskingModel.plot_tokens = False
+MaskingModel.quantizer_type = 'finite_scalar_quantizer'
+MaskingModel.seed = 0
+MaskingModel.weight_decay = 0.01
+
+# Parameters for MelSpectrogram:
+# ==============================================================================
+MelSpectrogram.freq_mask_param = 0
+MelSpectrogram.hop_len = 320
+MelSpectrogram.mel_scale = 'slaney'
+MelSpectrogram.n_mel = 96
+MelSpectrogram.norm = 'slaney'
+MelSpectrogram.norm_mean = 2.06755686098554
+MelSpectrogram.norm_std = 1.268292820667291
+MelSpectrogram.patch_size = (96, 3)
+MelSpectrogram.power = 2
+MelSpectrogram.sr = %new_freq
+MelSpectrogram.stretch_factor = 1
+MelSpectrogram.time_mask_param = 0
+MelSpectrogram.win_len = 512
+
+# Parameters for train:
+# ==============================================================================
+train.params = \
+    {'accelerator': 'gpu',
+     'devices': 4,
+     'log_every_n_steps': 50,
+     'max_steps': 400000,
+     'num_nodes': 2,
+     'num_sanity_val_steps': 0,
+     'precision': 'bf16-mixed',
+     'strategy': 'ddp_find_unused_parameters_true'}
+train.wandb_params = \
+    {'entity': 'mtg-upf',
+     'group': 'masking_conformer',
+     'name': 'mask_conf_large_au_to_all_25hz_fsq',
+     'offline': True,
+     'project': 'mtg-ssl',
+     'save_dir': '/gpfs/projects/upf97/logs/'}
+
+# Parameters for Waveform:
+# ==============================================================================
+Waveform.norm_mean = None
+Waveform.norm_std = None
+Waveform.patch_size = (1, 960)
+Waveform.sr = %new_freq

model.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2bf2a7e0ff047a60de8d68e8caa77941ae31e71df5034402e007b34a485e726b
+size 2325292802