def test_FullMTL(self):
metrics = ['rmse']
ddataset = SD.SyntheticDataCreator(num_tasks=3,
cellsPerTask=500,
drugsPerTask=10,
function="gauss",
normalize=False,
noise=1,
graph=False,
test_split=0.3)
dataset.prepare_data()
exp_folder = "fullMtlGP_test"
exp = ModelTraining(exp_folder)
methods = [
MtlGP.GPyFullMTL(num_iters=50,
length_scale=20,
noise_covar=.9,
n_inducing_points=500,
num_tasks=3)
]
exp.execute(dataset, methods, metrics, nruns=1)
df = exp.getResultsWrapper()
rmses = df['Value'].values
for rmse in rmses:
assert rmse < 1.5 # arbitrary threshold for reasonable prediction
def synthetic(n_examples, **kwargs):
'''
Run a synthetic experiment using n_examples examples
'''
np.random.seed(42)
config = synthetic_config.copy()
config.update(kwargs)
assert config.get(
'vocab_size') is not None, "Must specify synthetic data vocab_size"
sampler = SyntheticData(**config)
c, a, y = sampler.sample_truth(n_examples)
t = sampler.sample_text((c, a, y))
mask = sampler.mar(a, (c, y, t))
debug = kwargs.get('debug', False)
return experiment(c, a, y, t, mask, debug=debug)
def synthetic(n_examples, n_train, **kwargs):
'''
Run a synthetic experiment using n_examples examples in target dataset p(A*, C, Y)
and n_train examples of external data to estimate p(A*, A)
'''
np.random.seed(42)
proxy_i = 1
confound_i = (
0,
2,
)
config = synthetic_config.copy()
config.update(kwargs)
assert config.get(
'vocab_size') is not None, "Must specify synthetic data vocab_size"
sampler = SyntheticData(**config)
truth = sampler.sample_truth(n_examples)
truth_t = sampler.sample_text(truth[proxy_i])
external = sampler.sample_truth(n_train)
external_t = sampler.sample_text(external[proxy_i])
debug = kwargs.get('debug', False)
return train_adjust(np.concatenate(
(np.array(external), np.transpose(external_t)), axis=0),
np.concatenate(
(np.array(truth), np.transpose(truth_t)), axis=0),
proxy_i,
confound_i,
debug=debug)
def test_NN(self):
ddataset = SD.SyntheticDataCreator(num_tasks=3,
cellsPerTask=400,
drugsPerTask=10,
function="cosine",
normalize=True,
noise=1,
graph=False,
test_split=0.3)
dataset.prepare_data()
exp_folder = "NN_test"
exp = ModelTraining(exp_folder)
methods = [FeedForwardNN([25, 25], 'relu', epochs=60, lr=1e-3)]
exp.execute(dataset, methods, metrics, nruns=1)
df = exp.getResultsWrapper()
rmses = df['Value'].values
for rmse in rmses:
assert rmse < 1.5 # arbitrary threshold for reasonable prediction
import sys
sys.path.append('../')
from design import ModelTraining
from methods.regressor.FFNN import FeedForwardNN
import methods.mtl.MTL_GP as MtlGP
from methods.knn.KNN import KNN_Normalized
from datasets import SyntheticData as SD
import methods.mtl.NCF_MTL as NCF_MTL
from methods.matrix_factorization.MF import SVD_MF, NonNegative_MF
if __name__ == '__main__':
dataset = SD.SyntheticDataCreator(num_tasks=3,cellsPerTask=400, drugsPerTask=10, function="cosine",
normalize=False, noise=.5, graph=False, test_split=0.3)
dataset.prepare_data()
hyperparams_mtlmlp = {'batch_size': 64, 'epochs': 150, 'layers': '[64,32,16,8]', \
'learner': 'adam', 'lr': .001,'mlp_lr': .001, 'num_factors': 10, \
'reg_layers': '[0,0,0,.01]', 'reg_mf': 0.01, 'verbose': 1}
hyperparams_mtlmf = {'batch_size': 64, 'epochs': 150, 'layers': '[64,32,16,8]', \
'learner': 'adam', 'lr': .001,'mf_lr': .001, 'num_factors': 10, \
'reg_layers': '[0,0,0,.01]', 'reg_mf': 0.01, 'verbose': 1}
methods = [MtlGP.HadamardMTL(num_iters=150, length_scale=57, noise_covar=.24, n_inducing_points=1000, \
composite=False, learning_rate=.07, validate=False,bias=False,stabilize=True),
MtlGP.GPyFullMTL(num_iters=72, length_scale=58.828, noise_covar=0.31587, n_inducing_points=500, num_tasks=3, learning_rate=0.02729),
NCF_MTL.Neural_Collaborative_Filtering_FeaturesMTLMLP(hyperparams_mtlmlp,'MTL NCF MLP', 'feature_based'),
NCF_MTL.Neural_Collaborative_Filtering_FeaturesMTLMF(hyperparams_mtlmf,'NCF_MTL_MF', 'feature_based'),
SVD_MF(n_factors=10),