[project]
neptune_key = eyJhcGlfYWRkcmVzcyI6Imh0dHBzOi8vdWkubmVwdHVuZS5haSIsImFwaV91cmwiOiJodHRwczovL3VpLm5lcHR1bmUuYWkiLCJhcGlfa2V5IjoiZmI0OGMzNzUtOTg1NS00Yzg2LThjMzYtMWFiYjUwMDUyMjVlIn0=
debug = False
eval = True
seed = 69
owner = si11ium
model_name = CNNBaseline
data_name = CCSLibrosaDatamodule

[data]
num_worker = 10
data_root = data
variable_length = False
target_mel_length_in_seconds = 0.7

n_mels = 128
sr = 16000
hop_length = 128
n_fft = 512

random_apply_chance = 0.7
loudness_ratio = 0.0
shift_ratio = 0.3
noise_ratio = 0.0
mask_ratio = 0.0

[Tester]
weight_init = xavier_normal_
activation = gelu
use_bias = True
use_norm = True
use_residual = True
dropout = 0.21

lat_dim = 32
patch_size = 8
attn_depth = 12
heads = 4
embedding_size = 128
mlp_dim = 32


[CNNBaseline]
weight_init = xavier_normal_
activation = gelu
use_bias = True
use_norm = True
dropout = 0.2
lat_dim = 32
filters = [16, 32, 64, 128]

[BandwiseConvClassifier]
weight_init = xavier_normal_
activation = gelu
use_bias = True
use_norm = True
dropout = 0.2
lat_dim = 32
filters = [16, 32, 64, 128]

[VisualTransformer]
weight_init = xavier_normal_
activation = gelu
use_bias = True
use_norm = True
use_residual = True
dropout = 0.2

lat_dim = 32
mlp_dim = 32
head_dim = 32
patch_size = 8
attn_depth = 12
heads = 4
embedding_size = 33

[VerticalVisualTransformer]
weight_init = xavier_normal_
activation = gelu
use_bias = True
use_norm = True
use_residual = True
dropout = 0.2

mlp_dim = 6
lat_dim = 6
head_dim = 6
patch_size = 8
attn_depth = 6
heads = 4
embedding_size = 30

[HorizontalVisualTransformer]
weight_init = xavier_normal_
activation = gelu
use_bias = True
use_norm = True
use_residual = True
dropout = 0.3

lat_dim = 256
patch_size = 8
attn_depth = 12
heads = 6
embedding_size = 32

[VisualPerformer]
weight_init = xavier_normal_
activation = gelu
use_bias = True
use_norm = True
use_residual = True
dropout = 0.2

lat_dim = 32
patch_size = 8
attn_depth = 12
heads = 4
embedding_size = 30

[train]
outpath = output
version = None
sampler = WeightedRandomSampler
loss = ce_loss
sto_weight_avg = False
weight_decay = 0
opt_reset_interval = 0
max_epochs = 150
batch_size = 30
lr = 0.001
scheduler='LambdaLR'
use_residual = True
lr_scheduler_parameter = 0.97
num_sanity_val_steps = 2
check_val_every_n_epoch = 5
checkpoint_callback = True
gradient_clip_val = 0