def sample_mlp_vae_family_cfg(seed):
"""Samples a task config for a MLP VAE model on image datasets.
These configs are nested python structures that provide enough information
to create an instance of the problem.
Args:
seed: int Random seed to generate task from.
Returns:
A nested dictionary containing a configuration.
"""
rng = np.random.RandomState(seed)
cfg = {}
enc_n_layers = rng.choice([1, 2, 3, 4])
cfg["enc_hidden_units"] = [
utils.sample_log_int(rng, 32, 128) for _ in range(enc_n_layers)
]
dec_n_layers = rng.choice([1, 2, 3])
cfg["dec_hidden_units"] = [
utils.sample_log_int(rng, 32, 128) for _ in range(dec_n_layers)
]
cfg["activation"] = utils.sample_activation(rng)
cfg["w_init"] = utils.sample_initializer(rng)
cfg["dataset"] = utils.sample_image_dataset(rng)
return cfg
def sample_mlp_ae_family_cfg(seed):
"""Sample a task config for a mlp autoencoder on image data.
These configs are nested python structures that provide enough information
to create an instance of the problem.
Args:
seed: int Random seed to generate task from.
Returns:
A nested dictionary containing a configuration.
"""
rng = np.random.RandomState(seed)
cfg = {}
n_layers = rng.choice([1, 2, 3, 4, 5, 6])
cfg["hidden_units"] = [
utils.sample_log_int(rng, 32, 128) for _ in range(n_layers)
]
cfg["activation"] = utils.sample_activation(rng)
cfg["w_init"] = utils.sample_initializer(rng)
cfg["dataset"] = utils.sample_image_dataset(rng)
# Give relu double weight as this is what is often used in practice.
cfg["output_type"] = rng.choice(
["tanh", "tanh", "sigmoid", "sigmoid", "linear_center", "linear"])
# Give l2 double weight as this is often used
cfg["loss_type"] = rng.choice(["l2", "l2", "l1"])
cfg["reduction_type"] = rng.choice(["reduce_mean", "reduce_sum"])
return cfg
def sample_mlp_family_cfg(seed):
"""Sample a task config for a MLP model on image datasets.
These configs are nested python structures that provide enough information
to create an instance of the problem.
Args:
seed: int Random seed to generate task from.
Returns:
A nested dictionary containing a configuration.
"""
# random offset seed to ensure different task families sample different
# configs.
rng = np.random.RandomState(seed + 99123491)
cfg = collections.OrderedDict()
n_layers = rng.choice([1, 2, 3, 4, 5, 6])
cfg["layer_sizes"] = [
utils.sample_log_int(rng, 16, 128) for _ in range(n_layers)
]
cfg["activation"] = utils.sample_activation(rng)
cfg["w_init"] = utils.sample_initializer(rng)
cfg["dataset"] = utils.sample_image_dataset(rng)
cfg["center_data"] = bool(rng.choice([True, False]))
return cfg
def sample_maf_family_cfg(seed):
"""Sample a task config for a MAF model on image datasets.
These configs are nested python structures that provide enough information
to create an instance of the problem.
Args:
seed: int Random seed to generate task from.
Returns:
A nested dictionary containing a configuration.
"""
# random offset to seed to prevent clashing with other problems.
rng = np.random.RandomState(seed + 6123348)
cfg = {}
cfg["activation"] = utils.sample_activation(rng)
cfg["w_init"] = utils.sample_initializer(rng)
cfg["dataset"] = utils.sample_image_dataset(rng)
n_layers = int(rng.choice([1, 2]))
cfg["hidden_units"] = [
utils.sample_log_int(rng, 16, 128) for _ in range(n_layers)
]
cfg["num_bijectors"] = int(rng.choice([1, 2, 3, 4]))
return cfg
def sample_nvp_family_cfg(seed):
"""Sample a task config for a NVP model on image datasets.
These configs are nested python structures that provide enough information
to create an instance of the problem.
Args:
seed: int Random seed to generate task from.
Returns:
A nested dictionary containing a configuration.
"""
# Random offset on seed to ensure different randomness for task.
rng = np.random.RandomState(seed + 123919936)
cfg = {}
# TODO(lmetz) consider also forwarding per layer activation and per layer
# w_init.
cfg["activation"] = utils.sample_activation(rng)
cfg["w_init"] = utils.sample_initializer(rng)
cfg["dataset"] = utils.sample_image_dataset(rng)
n_layers = int(rng.choice([1, 2]))
cfg["hidden_units"] = [
utils.sample_log_int(rng, 16, 128) for _ in range(n_layers)
]
cfg["num_bijectors"] = int(rng.choice([1, 2, 3, 4]))
return cfg
def sample_conv_fc_family_cfg(seed):
"""Sample a task config for a conv net with a fully connected layer on top.
These configs are nested python structures that provide enough information
to create an instance of the problem.
Args:
seed: int Random seed to generate task from.
Returns:
A nested dictionary containing a configuration.
"""
rng = np.random.RandomState(seed)
cfg = {}
layer_choices = [1, 2, 3, 4, 5]
max_layer = np.max(layer_choices)
n_layers = rng.choice(layer_choices)
# pattern for how strides are chosen. Either all 2s, repeated 1,2
# or repeated 2,1.
stride_pattern = rng.choice(["all_two", "one_two", "two_one"])
if stride_pattern == "all_two":
cfg["strides"] = ([2] * max_layer)[0:n_layers]
elif stride_pattern == "one_two":
cfg["strides"] = ([1, 2] * max_layer)[0:n_layers]
elif stride_pattern == "two_one":
cfg["strides"] = ([2, 1] * max_layer)[0:n_layers]
cfg["strides"] = list(zip(cfg["strides"], cfg["strides"]))
cfg["hidden_units"] = [
utils.sample_log_int(rng, 8, 64) for _ in range(n_layers)
]
cfg["activation"] = utils.sample_activation(rng)
cfg["w_init"] = utils.sample_initializer(rng)
cfg["padding"] = [
str(rng.choice([snt.SAME, snt.VALID])) for _ in range(n_layers)
]
n_fc_layers = rng.choice([0, 1, 2, 3])
cfg["fc_hidden_units"] = [
utils.sample_log_int(rng, 32, 128) for _ in range(n_fc_layers)
]
cfg["use_bias"] = bool(rng.choice([True, False]))
cfg["dataset"] = utils.sample_image_dataset(rng)
cfg["center_data"] = bool(rng.choice([True, False]))
return cfg
def test_sample_get_activation(self):
rng = np.random.RandomState(123)
sampled_acts = []
num = 4000
for _ in range(num):
aname = utils.sample_activation(rng)
sampled_acts.append(aname)
# smoke test to ensure graph builds
out = utils.get_activation(aname)(tf.constant(1.0))
self.assertIsInstance(out, tf.Tensor)
uniques, counts = np.unique(sampled_acts, return_counts=True)
counts_map = {str(u): c for u, c in zip(uniques, counts)}
# 16 is the total sum of unnormalized probs
amount_per_n = num / float(16)
self.assertNear(counts_map["relu"], amount_per_n * 6, 40)
self.assertNear(counts_map["tanh"], amount_per_n * 3, 40)
self.assertNear(counts_map["swish"], amount_per_n, 40)
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
