Datasets:

ArXiv:
License:
timewarp / 2aa_config.yaml
andrewfoong's picture
Upload 2aa_config.yaml
f0d29c6 verified
dataset: 2AA-1-big
model_config:
model_type: custom_attention_transformer_nvp
transformer_cvae_config: null
transformer_nvp_config: null
custom_transformer_nvp_config:
atom_embedding_dim: 64
latent_mlp_hidden_dims:
- 256
num_coupling_layers: 12
num_transformer_layers: 6
encoder_layer_config:
d_model: 128
dim_feedforward: 2048
dropout: 0.0
num_heads: 6
attention_type: kernel
lengthscales:
- 0.1
- 0.2
- 0.5
- 0.7
- 1.0
- 1.2
max_radius: null
normalise_kernel_values: true
cheb_order: null
force_asymptotic_zero: null
position_layer_index_mod_2: 0
conditional_flow_density:
scale_requires_grad: true
ignore_conditional_velocity: true
use_displacement_as_target: true
equivariant_nvp_config: null
gaussian_density_transformer_config: null
graphormer_nvp_config: null
step_width: 1000000
batch_size: 256
num_epochs: 100000000
patience: 100
data_augmentation: true
measure_equivariance_discrepancy: false
use_aml_logging: false
loss:
nll: null
nll_and_energy: null
nll_and_acceptance:
openmm_provider:
pdb_dirs: /mnt/simulated-data/trajectory-data/2AA-1-large
device: cpu
cache_size: 8
random_velocs: true
beta: 0.1
clamp: false
num_samples: 1
weights:
- 0.9
- 0.1
pre_softmax_weights: null
loss_schedule: null
run_prefix: ''
optimizer: Lamb
learning_rate: 1.0e-05
warmup_steps: 1000
weight_decay: 0.0
clip_grad_norm: 100.0
seed: 0
randomise_seed: true
data_dir: /mnt/lmdb-simulated-data/trajectory-data/2AA-1-big/step_width_1000000
dataset_cache_dir: .data
dataset_use_lmdb: true
pdb_dir: /mnt/simulated-data/trajectory-data/2AA-1-large
output_folder: /mnt/output/projects/timewarp/amlt-results/7328294309.72749-b6024015-4a4c-4fb3-92d3-f4b6ff86975a
enable_profiler: false
saved_model_path: /mnt/default/projects/timewarp/amlt-results/7329812703.49702-998fb2c0-84c9-4ee6-b8be-cbe29511a372
valid_batch_size: 256
min_check_point_iters: 500
random_velocities: true
warm_start: true
num_pdbs_per_local_batch: 4
equal_data_spacing: false
run_valid_first: false
lr_scheduler:
type: plateau
factor: 0.5
iteration_patience: 10000
logging_period: 100
alpha: 0.2