#data
dataset_name: cadillac 
test_name: "testcase_train"
gpu_id: 1
epochs: 1
# data_dir: /raid/ansysai/udbhav/alpha_Xdata/data_prep_transformer/cadillac_v2/1_VTK_surface/
data_dir: datasets/examples/testcase/cadillac_v2/1_VTK_surface


json_file: ${data_dir}/params.json

splits_file: ${data_dir}/
data_folder: ${dataset_name}

input_normalization: "shift_axis" # options: "min_max", "std_norm", "none"
normalization: "std_norm"
norm_vars: "pressure"
physical_scale_for_test: True

diff_input_velocity: True # If true, inlet_x_velocity is added as an input feature

# num_points: 40000
num_points: 30000
num_workers: 1

#model
indim: 4
outdim: 1
model: ansysLPFMs
hidden_dim: 256
n_heads: 8
n_decoder: 8
mlp_ratio: 2


#training
val_iter: 1
lr: 0.001
batch_size: 1

optimizer: 
    type: AdamW
scheduler: OneCycleLR #OneCycleLR
loss_type: huber  # options: mse, mae, huber

# scheduler: LinearWarmupCosineAnnealingLR
num_processes: 1
max_grad_norm: 1.0
mixed_precision: True #currently default fp16 is selected by torch.autocast(). Fp16 gave the best results for Transformer based models.
eval: False
save_latent: False


chunked_eval: True # Default with True is evaluation of max chunks of size num_points that can fit in a data sample, to avoid small last chunks 
train_ckpt_load: False ## Will load best model if ckpt_load is false

pos_embed_sincos: False


project_name: ${dataset_name}