Skip to content

Latest commit

 

History

History
78 lines (65 loc) · 1.71 KB

CONFIG_DESCRIPTION.md

File metadata and controls

78 lines (65 loc) · 1.71 KB
Raw

Description of .toml config for TabDDPM

First of all, train.T and eval.T denote preprocessing for training and for evaluation, respectively.

Here we list non-obvious parameters.

Main part:

  • seed = 0 -- evaluation seed (and training, but for training it is fixed to 0)
  • parent_dir = "exp/abalone/check" -- exp folder
  • real_data_path = "data/abalone/"
  • model_type = "mlp" -- model type that approximates the reverse process
  • num_numerical_features -- a number of numerical features in dataset
  • device = "cuda:0"

Model params:

  • is_y_cond -- false for regression, true for classification
  • d_in -- input dimension (not necessary, since scripts calculate it automatically)
  • num_calsses -- zero for regression, a number of classes for classification
  • rtdl_params -- MLP parameters
seed = 0
parent_dir = "exp/abalone/check"
real_data_path = "data/abalone/"
model_type = "mlp"
num_numerical_features = 7
device = "cuda:0"

[model_params]
is_y_cond = false
d_in = 11
num_classes = 0

[model_params.rtdl_params]
d_layers = [
    256,
    256,
]
dropout = 0.0

[diffusion_params]
num_timesteps = 1000
gaussian_loss_type = "mse"
scheduler = "cosine"

[train.main]
steps = 1000
lr = 0.001
weight_decay = 1e-05
batch_size = 4096

[train.T]
seed = 0
normalization = "quantile"
num_nan_policy = "__none__"
cat_nan_policy = "__none__"
cat_min_frequency = "__none__"
cat_encoding = "__none__"
y_policy = "default"

[sample]
num_samples = 20800
batch_size = 10000
seed = 0

[eval.type]
eval_model = "catboost"
eval_type = "synthetic"

[eval.T]
seed = 0
normalization = "__none__"
num_nan_policy = "__none__"
cat_nan_policy = "__none__"
cat_min_frequency = "__none__"
cat_encoding = "__none__"
y_policy = "default"