def _sample_min_max_well(rng):
"""Sample a min max well problem."""
is_noise = utils.sample_bool(rng, 0.5)
return {
"dim":
utils.sample_log_int(rng, 10, 1000),
"noise_stdev":
utils.sample_log_float(rng, 0.01, 10.0) if is_noise else 0.0,
}
def _sample_optimization_test_problems(rng):
"""Sample an optimization test function problem."""
is_noise = utils.sample_bool(rng, 0.5)
return {
"problem":
rng.choice(sorted(_opt_test_problems.keys())),
"noise_stdev":
utils.sample_log_float(rng, 0.01, 10.0) if is_noise else 0.0,
}
def _sample_quadratic_problem(rng):
"""Sample a quadratic problem."""
is_noise = utils.sample_bool(rng, 0.5)
return {
"dim":
utils.sample_log_int(rng, 10, 1000),
"noise_stdev":
utils.sample_log_float(rng, 0.01, 10.0) if is_noise else 0.0,
}
def _sample_sparse_softmax_regression(rng):
"""Sample a sparse softmax regression problem."""
is_noise = utils.sample_bool(rng, 0.5)
return {
"n_features": utils.sample_log_int(rng, 2, 100),
"n_classes": 2,
"noise_stdev":
utils.sample_log_float(rng, 0.01, 10.0) if is_noise else 0.0,
"bs": utils.sample_log_int(rng, 1, 50),
"n_samples": utils.sample_log_int(rng, 1, 30),
}
def _sample_bowl_problems(rng):
"""Sample a bowl problem."""
is_noise = utils.sample_bool(rng, 0.5)
return {
"cond":
utils.sample_log_float(rng, 0.01, 100),
"angle":
rng.choice([0, 0, np.pi / 4., np.pi / 3]),
"noise_stdev":
utils.sample_log_float(rng, 0.01, 10.0) if is_noise else 0.0,
}
def sample_char_rnn_language_model_family_cfg(seed):
"""Samples a character NN language modeling task."""
rng = np.random.RandomState(seed)
cfg = {}
cfg["embed_dim"] = utils.sample_log_int(rng, 8, 128)
cfg["w_init"] = utils.sample_initializer(rng)
full_vocab = utils.sample_bool(rng, 0.8)
if full_vocab:
cfg["vocab_size"] = 256
else:
# only operate on some subset of full words.
cfg["vocab_size"] = utils.sample_log_int(rng, 100, 256)
cfg["core"] = utils.sample_rnn_core(rng)
cfg["trainable_init"] = bool(rng.choice([True, False]))
cfg["dataset"] = utils.sample_char_lm_dataset(rng)
return cfg
def _sample_fully_connected(rng):
"""Sample a fully connected problem."""
n_layer = rng.choice([2, 3, 4, 5])
fixed = utils.sample_bool(rng, 0.5)
cfg = {
"n_features": utils.sample_log_int(rng, 1, 16),
"n_classes": 2,
"activation": utils.sample_activation(rng),
"bs": utils.sample_log_int(rng, 1, 200),
"n_samples": utils.sample_log_int(rng, 1, 30),
}
if fixed:
cfg["hidden_sizes"] = [utils.sample_log_int(rng, 4, 32)] * n_layer
else:
cfg["hidden_sizes"] = [
utils.sample_log_int(rng, 4, 32) for _ in range(n_layer)
]
return cfg
def _sample_sparse_problem(rng):
"""Sample a sparse problem.
This problem modifies a sampled base problem by setting some gradients to
zero.
Args:
rng: Random state
Returns:
The sampled config.
"""
is_noise = utils.sample_bool(rng, 0.5)
base_config = rng.choice(_to_modify)
return {
"base": (base_config, _problem_sample_get[base_config][0](rng)),
"zero_probability":
rng.uniform(0.9, 0.99),
"noise_stdev":
utils.sample_log_float(rng, 0.01, 10.0) if is_noise else 0.0,
}
