Upload ./RepCodec/train_configs/ex_dim768_mse.yaml with huggingface_hub
Browse files
RepCodec/train_configs/ex_dim768_mse.yaml
ADDED
@@ -0,0 +1,74 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
###########################################################
|
2 |
+
# DATA SETTING #
|
3 |
+
###########################################################
|
4 |
+
data:
|
5 |
+
path: "/dir/to/representations/"
|
6 |
+
subset:
|
7 |
+
train: "train_set_name"
|
8 |
+
valid: "valid_set_name"
|
9 |
+
test: "test_set_name"
|
10 |
+
|
11 |
+
###########################################################
|
12 |
+
# MODEL SETTING #
|
13 |
+
###########################################################
|
14 |
+
model_params:
|
15 |
+
input_channels: 768
|
16 |
+
output_channels: 768
|
17 |
+
encode_channels: 768
|
18 |
+
decode_channels: 768
|
19 |
+
code_dim: 768
|
20 |
+
codebook_num: 1
|
21 |
+
codebook_size: 1024
|
22 |
+
bias: true
|
23 |
+
enc_ratios: [1, 1]
|
24 |
+
dec_ratios: [1, 1]
|
25 |
+
enc_strides: [1, 1] # no downsampling
|
26 |
+
dec_strides: [1, 1]
|
27 |
+
enc_kernel_size: 3
|
28 |
+
dec_kernel_size: 3
|
29 |
+
enc_block_dilations: [1, 1]
|
30 |
+
enc_block_kernel_size: 3
|
31 |
+
dec_block_dilations: [1, 1]
|
32 |
+
dec_block_kernel_size: 3
|
33 |
+
|
34 |
+
###########################################################
|
35 |
+
# METRIC LOSS SETTING #
|
36 |
+
###########################################################
|
37 |
+
repr_reconstruct_loss_params:
|
38 |
+
loss_type: l2
|
39 |
+
|
40 |
+
###########################################################
|
41 |
+
# LOSS WEIGHT SETTING #
|
42 |
+
###########################################################
|
43 |
+
lambda_vq_loss: 1.0 # Loss weight of vector quantize loss.
|
44 |
+
lambda_repr_reconstruct_loss: 45.0
|
45 |
+
|
46 |
+
###########################################################
|
47 |
+
# DATA LOADER SETTING #
|
48 |
+
###########################################################
|
49 |
+
batch_size: 32 # Batch size.
|
50 |
+
batch_length: 96 # Length of each audio in batch (training w/o adv).
|
51 |
+
pin_memory: true # Whether to pin memory in Pytorch DataLoader.
|
52 |
+
num_workers: 4 # Number of workers in Pytorch DataLoader.
|
53 |
+
|
54 |
+
###########################################################
|
55 |
+
# OPTIMIZER & SCHEDULER SETTING #
|
56 |
+
###########################################################
|
57 |
+
model_optimizer_type: Adam
|
58 |
+
model_optimizer_params:
|
59 |
+
lr: 1.0e-4
|
60 |
+
betas: [0.5, 0.9]
|
61 |
+
weight_decay: 0.0
|
62 |
+
model_scheduler_type: StepLR
|
63 |
+
model_scheduler_params:
|
64 |
+
step_size: 200000 # Model's scheduler step size.
|
65 |
+
gamma: 1.0
|
66 |
+
grad_norm: -1
|
67 |
+
|
68 |
+
###########################################################
|
69 |
+
# INTERVAL SETTING #
|
70 |
+
###########################################################
|
71 |
+
train_max_steps: 200000 # Number of training steps. (w/o adv)
|
72 |
+
save_interval_steps: 20000 # Interval steps to save checkpoint.
|
73 |
+
eval_interval_steps: 2000 # Interval steps to evaluate the network.
|
74 |
+
log_interval_steps: 100 # Interval steps to record the training log.
|