from funkyyak import grad, kylist from hypergrad.data import load_data_dicts from hypergrad.nn_utils import make_nn_funs, VectorParser, logit, inv_logit, fill_parser from hypergrad.optimizers import simple_sgd, sgd_parsed from hypergrad.util import RandomState import whetlab parameters = { 'init_log_alphas': { 'min': -3, 'max': 2, 'type': 'float' }, 'init_invlogit_betas': { 'min': inv_logit(0.1), 'max': inv_logit(0.999), 'type': 'float' }, 'init_log_param_scale': { 'min': -5, 'max': 1, 'type': 'float' } } outcome = {'name': 'Training Gain'} scientist = whetlab.Experiment( name="ICML Hypergrad paper - optimize initial values - v2 no randomness", description="Vanilla tuning of hyperparameters. v2", parameters=parameters, outcome=outcome)
from hypergrad.optimizers import sgd4, rms_prop, adam # ----- Fixed params ----- layer_sizes = [784, 10] batch_size = 200 N_iters = 60 N_classes = 10 N_train = 1000 N_valid = 10**3 N_tests = 10**3 N_batches = N_train / batch_size #N_iters = N_epochs * N_batches # ----- Initial values of learned hyper-parameters ----- init_log_L2_reg = 0.0 init_log_alphas = -2.0 init_invlogit_betas = inv_logit(0.9) init_log_param_scale = 0.0 # ----- Superparameters ----- meta_alpha = 0.04 N_meta_iter = 100 global_seed = npr.RandomState(3).randint(1000) def fill_parser(parser, items): partial_vects = [np.full(parser[name].size, items[i]) for i, name in enumerate(parser.names)] return np.concatenate(partial_vects, axis=0) def run(): train_data, valid_data, tests_data = load_data_dicts(N_train, N_valid, N_tests) parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
# ----- Fixed params ----- layer_sizes = [784, 100, 10] batch_size = 200 N_iters = 100 N_classes = 10 N_train = 1000 N_valid = 10000 N_tests = 10000 N_learning_checkpoint = 2 thin = np.ceil(N_iters / N_learning_checkpoint) # ----- Initial values of learned hyper-parameters ----- init_log_L2_reg = 0.0 init_log_L2_reg_noise = 0.0 init_log_alphas = -1.0 init_invlogit_betas = inv_logit(0.5) init_log_param_scale = -3.0 # ----- Superparameters ----- meta_alpha = 0.4 N_meta_iter = 50 seed = 0 def fill_parser(parser, items): partial_vects = [ np.full(parser[name].size, items[i]) for i, name in enumerate(parser.names) ] return np.concatenate(partial_vects, axis=0)
def test_inv_logit(): assert np.allclose(inv_logit(logit(0.5)), 0.5, rtol=1e-3, atol=1e-4) assert np.allclose(inv_logit(logit(0.6)), 0.6, rtol=1e-3, atol=1e-4) assert np.allclose(inv_logit(logit(0.1)), 0.1, rtol=1e-3, atol=1e-4) assert np.allclose(inv_logit(logit(0.2)), 0.2, rtol=1e-3, atol=1e-4)
"""Find good initial values using Whetlab.""" import numpy as np import pickle from collections import defaultdict from funkyyak import grad, kylist from hypergrad.data import load_data_dicts from hypergrad.nn_utils import make_nn_funs, VectorParser, logit, inv_logit, fill_parser from hypergrad.optimizers import simple_sgd, sgd_parsed from hypergrad.util import RandomState import whetlab parameters = { 'init_log_alphas':{'min':-3, 'max':2, 'type':'float'}, 'init_invlogit_betas':{'min':inv_logit(0.1), 'max':inv_logit(0.999),'type':'float'}, 'init_log_param_scale':{'min':-5, 'max':1,'type':'float'}} outcome = {'name':'Training Gain'} scientist = whetlab.Experiment(name="ICML Hypergrad paper - optimize initial values", description="Vanilla tuning of hyperparameters.", parameters=parameters, outcome=outcome) # ----- Fixed params ----- layer_sizes = [784, 50, 50, 50, 10] batch_size = 200 N_iters = 100 N_classes = 10 N_train = 10000 N_valid = 10000 N_tests = 10000 N_learning_checkpoint = 10
"""Find good initial values using Whetlab.""" import pickle from collections import defaultdict import numpy as np import whetlab from funkyyak import grad, kylist from hypergrad.data import load_data_dicts from hypergrad.nn_utils import make_nn_funs, VectorParser, logit, inv_logit, fill_parser from hypergrad.optimizers import sgd_parsed from hypergrad.util import RandomState parameters = { "init_log_alphas": {"min": -3, "max": 2, "type": "float"}, "init_invlogit_betas": {"min": inv_logit(0.1), "max": inv_logit(0.999), "type": "float"}, "init_log_param_scale": {"min": -5, "max": 1, "type": "float"}, } outcome = {"name": "Training Gain"} scientist = whetlab.Experiment( name="ICML Hypergrad paper - optimize initial values", description="Vanilla tuning of hyperparameters.", parameters=parameters, outcome=outcome, ) # ----- Fixed params ----- layer_sizes = [784, 50, 50, 50, 10] batch_size = 200 N_iters = 100 N_classes = 10
def test_inv_logit(): assert np.allclose(inv_logit(logit(0.5)), 0.5, rtol=1e-3, atol=1e-4) assert np.allclose(inv_logit(logit(0.6)), 0.6, rtol=1e-3, atol=1e-4) assert np.allclose(inv_logit(logit(0.1)), 0.1, rtol=1e-3, atol=1e-4) assert np.allclose(inv_logit(logit(0.2)), 0.2, rtol=1e-3, atol=1e-4)