NNmeta is created in order to use NNPackage.
pip3 install nnmeta
git clone https://github.com/AlexanderDKazakov/nnmeta
cd nnmeta
pip3 install -r requirements.txt
pip3 install .
NNmeta support certain convention in data structure and it is required to structure the data in such way:
# /path/to/my/base
base
└── xyz
├── samples.xyz
└── samples_for_cc.xyz
After first run additional folders will be created: models, dbs, splits, tests
from nnmeta import NNClass # this is a main class for NN training
info = dict(
runner = { # network name
# data source [extended xyz file]; used for converting to DB [internal usage]
# "filename" : {"range" ex. [from:to:step], epochs should be done}
"samples.xyz" : {":" : 20}, # train `runner` nn on all samples of `samples.xyz` 20 epochs
},
runner_features = dict(
n_features = 64, # details in NN class [default is 128]
n_filters = 32, #
n_gaussians = 12, # default 25
batch_size = 512, # [parameter for tuning]
lr = 1e-4, # learning rate [parameter for tuning]
db_properties = ("energy", "forces", "dipole_moment"), # what can be found in the `samples.xyz` file
training_properties = ("energy", "forces", "dipole_moment"), # what one wants to train
loss_tradeoff = (0.2, 0.8, 0.6),
n_layers_energy_force = 2, # default 2 [parameter for tuning]
n_neurons_energy_force = None, # default None [parameter for tuning]
n_layers_dipole_moment = 2, # default 2 [parameter for tuning]
n_neurons_dipole_moment = None, # default None [parameter for tuning]
loss_function_choice = "mse", # "mae", "mse", "sae"
train_samples_percent = 70,
valid_samples_percent = 20,
predict_each_epoch = 200,
validate_each_epoch = 30,
# cross-check with next files
check_list_files = {
# this file should lie in the same `xyz` dir
"samples_for_cc.xyz" : dict(num_points = 1000),
}
)
)
nn = NNClass(info=info, network_name="runner",
system_path="/path/to/my/base")
nn.prepare_network()Feel free to contribute to the project, but please create initially an issue with detailed problem and way to resolve it.
MIT