def load_trainer(base_params, train_params, model, data):
# type: (BaseParameters, TrainParameters, md.AbstractGraphModule, dt.DataCost) -> tr.Train
return tr.Train(
model, data, tr.PredictionType.REGRESSION, ls.mse_loss, 1,
batch_size=train_params.batch_size, clip=None, opt=train_params.optimizer,
lr=train_params.initial_lr, weight_decay=train_params.weight_decay,
predict_log=base_params.predict_log, momentum=train_params.momentum,
nesterov=train_params.nesterov,
)
def graph_model_validate(base_params, model_file, iaca_only):
# type: (BaseParameters, str, bool) -> None
data = load_data(base_params)
if iaca_only:
cnx = ut.create_connection()
legal_code_ids = set(
pd.read_sql('SELECT time_id, code_id FROM times WHERE kind="iaca"', cnx)
.set_index('time_id')
.code_id
)
data.test = [datum for datum in data.test if datum.code_id in legal_code_ids]
model = load_model(base_params, data)
train = tr.Train(
model, data, tr.PredictionType.REGRESSION, ls.mse_loss, 1,
batch_size=1000, clip=None, predict_log=base_params.predict_log,
)
resultfile = os.environ['ITHEMAL_HOME'] + '/learning/pytorch/results/realtime_results.txt'
(actual, predicted) = train.validate(resultfile=resultfile, loadfile=model_file)
def graph_model_benchmark(base_params, benchmark_params):
# type: (BaseParameters, BenchmarkParameters) -> None
data = load_data(base_params)
model = load_model(base_params, data)
train = tr.Train(
model, data, tr.PredictionType.REGRESSION, ls.mse_loss, 1,
batch_size=benchmark_params.batch_size, clip=None, opt=tr.OptimizerType.ADAM_PRIVATE, lr=0.01,
)
model.share_memory()
mp_config = MPConfig(benchmark_params.threads)
partition_size = benchmark_params.examples // benchmark_params.trainers
processes = []
start_time = time.time()
with mp_config:
for rank in range(benchmark_params.trainers):
mp_config.set_env(rank)
partition = (rank * partition_size, (rank + 1) * partition_size)
p = mp.Process(target=train, args=(rank, partition))
p.daemon = True
p.start()
processes.append(p)
for p in processes:
p.join()
end_time = time.time()
print('Time to process {} examples: {} seconds'.format(
benchmark_params.examples,
end_time - start_time,
))
from models import train
if __name__ == '__main__':
train = train.Train()
train.main()