misc(configs): prepare trainings for evaluating vicreg hyperparams

configs/simclr_vicreg_0.1.yml

@@ -3,10 +3,10 @@ encoder:
   type: 'thinresnet34'
 model:
   type: 'simclr'
-  loss_factor: 1.0
+  infonce_loss_factor: 1.0
   vic_reg_factor: 0.1
 training:
-  epochs: 150
+  epochs: 300
   batch_size: 256
   learning_rate: 0.001
 dataset:

configs/vicreg_b1024.yml

@@ -4,10 +4,10 @@ encoder:
 model:
   type: 'simclr'
   enable_mlp: true
-  loss_factor: 0.0
-  vic_reg_factor: 25.0
+  infonce_loss_factor: 0.0
+  vic_reg_factor: 1.0
 training:
-  epochs: 500
+  epochs: 300
   batch_size: 1024
   learning_rate: 0.001
 dataset:

configs/vicreg_b2048.yml

@@ -4,10 +4,10 @@ encoder:
 model:
   type: 'simclr'
   enable_mlp: true
-  loss_factor: 0.0
-  vic_reg_factor: 25.0
+  infonce_loss_factor: 0.0
+  vic_reg_factor: 1.0
 training:
-  epochs: 500
+  epochs: 300
   batch_size: 2048
   learning_rate: 0.001
 dataset:

configs/vicreg_b256.yml

@@ -4,10 +4,10 @@ encoder:
 model:
   type: 'simclr'
   enable_mlp: true
-  loss_factor: 0.0
-  vic_reg_factor: 25.0
+  infonce_loss_factor: 0.0
+  vic_reg_factor: 1.0
 training:
-  epochs: 500
+  epochs: 300
   batch_size: 256
   learning_rate: 0.001
 dataset:

configs/vicreg_b256_1_0.5_0.1.yml

@@ -0,0 +1,24 @@
+name: 'vicreg_b256_1_0.5_0.1'
+encoder:
+  type: 'thinresnet34'
+model:
+  type: 'simclr'
+  enable_mlp: true
+  infonce_loss_factor: 0.0
+  vic_reg_factor: 1.0
+  vic_reg_inv_weight: 1.0
+  vic_reg_var_weight: 0.5
+  vic_reg_cov_weight: 0.1
+training:
+  epochs: 300
+  batch_size: 256
+  learning_rate: 0.001
+dataset:
+  frame_length: 32000
+  frame_split: true
+  extract_mfcc: true
+  train: './data/voxceleb1_train_list'
+  val_ratio: 0.0
+  spec_augment: false
+  wav_augment:
+    enable: true

configs/vicreg_b256_1_1_0.04.yml

@@ -0,0 +1,24 @@
+name: 'vicreg_b256_1_1_0.04'
+encoder:
+  type: 'thinresnet34'
+model:
+  type: 'simclr'
+  enable_mlp: true
+  infonce_loss_factor: 0.0
+  vic_reg_factor: 1.0
+  vic_reg_inv_weight: 1.0
+  vic_reg_var_weight: 1.0
+  vic_reg_cov_weight: 0.04
+training:
+  epochs: 300
+  batch_size: 256
+  learning_rate: 0.001
+dataset:
+  frame_length: 32000
+  frame_split: true
+  extract_mfcc: true
+  train: './data/voxceleb1_train_list'
+  val_ratio: 0.0
+  spec_augment: false
+  wav_augment:
+    enable: true

configs/vicreg_b256_4xencoder.yml → configs/vicreg_b256_1_1_0.1.yml

@@ -1,14 +1,16 @@
-name: 'vicreg_b256_4xencoder'
+name: 'vicreg_b256_1_1_0.1'
 encoder:
   type: 'thinresnet34'
-  scale: 2
 model:
   type: 'simclr'
   enable_mlp: true
-  loss_factor: 0.0
-  vic_reg_factor: 25.0
+  infonce_loss_factor: 0.0
+  vic_reg_factor: 1.0
+  vic_reg_inv_weight: 1.0
+  vic_reg_var_weight: 1.0
+  vic_reg_cov_weight: 0.1
 training:
-  epochs: 500
+  epochs: 300
   batch_size: 256
   learning_rate: 0.001
 dataset:

configs/vicreg_b256_2xmse.yml → configs/vicreg_b256_1_1_0.yml

@@ -1,14 +1,16 @@
-name: 'vicreg_b256_2xmse'
+name: 'vicreg_b256_1_1_0'
 encoder:
   type: 'thinresnet34'
 model:
   type: 'simclr'
   enable_mlp: true
-  enable_mse_clean_aug: true
-  loss_factor: 0.0
-  vic_reg_factor: 25.0
+  infonce_loss_factor: 0.0
+  vic_reg_factor: 1.0
+  vic_reg_inv_weight: 1.0
+  vic_reg_var_weight: 1.0
+  vic_reg_cov_weight: 0.0
 training:
-  epochs: 500
+  epochs: 300
   batch_size: 256
   learning_rate: 0.001
 dataset:
@@ -17,7 +19,6 @@ dataset:
   extract_mfcc: true
   train: './data/voxceleb1_train_list'
   val_ratio: 0.0
-  provide_clean_and_aug: true
   spec_augment: false
   wav_augment:
     enable: true

configs/vicreg_b512.yml

@@ -4,10 +4,10 @@ encoder:
 model:
   type: 'simclr'
   enable_mlp: true
-  loss_factor: 0.0
-  vic_reg_factor: 25.0
+  infonce_loss_factor: 0.0
+  vic_reg_factor: 1.0
 training:
-  epochs: 500
+  epochs: 300
   batch_size: 512
   learning_rate: 0.001
 dataset: