VAE_ConvNeXt_2D
Select with config.model.set_model('VAE_ConvNeXt_2D'). The table shows the actual defaults for this model variant. A dash in Default search means the parameter stays fixed unless you assign a compatible distribution through config.model.search.
| Parameter | Type | Default | Meaning / values | Default search |
|---|---|---|---|---|
config.model.parameters.n_res_blocks |
int |
4 |
Number of residual blocks per level. | IntRange(low=4, high=5, step=1, log=False) |
config.model.parameters.n_levels |
int |
4 |
Number of encoder and decoder levels. | — |
config.model.parameters.z_channels |
int |
32 |
Channel count in the spatial bottleneck. | Choice(values=(32, 64)) |
config.model.parameters.bottleneck_dim |
int |
64 |
Dimension of the latent vector. | Choice(values=(64, 128)) |
config.model.parameters.recon_weight |
float |
10.0 |
Weight of the reconstruction loss. | — |
config.model.parameters.beta_kl_start |
float |
0.0 |
Initial weight of the KL loss. | — |
config.model.parameters.beta_kl_max |
float |
0.08 |
Maximum weight of the KL loss. | — |
config.model.parameters.beta_kl_warmup_start |
int |
0 |
Epoch at which the KL weight begins to increase. | — |
config.model.parameters.beta_kl_warmup_epochs |
int |
200 |
Number of epochs needed to reach beta_kl_max. | — |
config.model.parameters.free_bits |
float |
0.001 |
KL free-bits allowance for latent dimensions. | — |
config.model.parameters.latent_recon_weight |
float |
0.0 |
Weight of the latent reconstruction Smooth L1 loss; 0 disables the loss. | — |
config.model.parameters.latent_recon_noise_scale |
float |
1.0 |
Scale of Gaussian noise added to the detached latent mean for the reconstruction cycle; positive. | — |
config.model.parameters.latent_recon_image_noise_std |
float |
0.03 |
Standard deviation of Gaussian noise added to the cycle image before re-encoding during training; non-negative. | — |
config.model.parameters.recon_loss |
str |
'smoothl1' |
Reconstruction loss, such as 'mse' or 'smoothl1'. | — |
config.model.parameters.recon_smoothl1_beta |
float |
1.0 |
Transition point of the Smooth L1 loss. | — |
config.model.parameters.use_transpose_conv |
bool |
False |
Use transposed convolutions for upsampling. | — |
config.model.parameters.fg_weight |
float |
1.0 |
Additional weight for foreground pixels in the reconstruction loss. | — |
config.model.parameters.fg_threshold |
float |
0.0 |
Threshold for selecting foreground pixels. | — |
config.model.parameters.drop_path_rate |
float |
0.001 |
Stochastic depth rate in ConvNeXt blocks. | — |
config.model.parameters.dropout |
float |
0.001 |
Dropout probability within the model. | — |
config.model.parameters.skip_dropout_p |
float |
1.0 |
Shared dropout probability for skip connections. | — |
config.model.parameters.skip_dropout_ps |
Optional[List[float]] |
None |
Dropout per skip level; overrides skip_dropout_p and requires n_levels values. | — |
config.model.parameters.skip_alpha |
float |
0.0 |
Shared skip connection scale; range [0, 1]. | — |
config.model.parameters.skip_alphas |
Optional[List[float]] |
None |
Scale per skip level; overrides skip_alpha and requires n_levels values. | — |