def test_format_named_configs():
ingred = Ingredient('ingred')
ex = Experiment(name='experiment', ingredients=[ingred])
@ingred.named_config
def named_config1():
pass
@ex.named_config
def named_config2():
"""named config with doc"""
pass
dict_config = dict(v=42)
ingred.add_named_config('dict_config', dict_config)
named_configs_text = _format_named_configs(
OrderedDict(ex.gather_named_configs()))
assert named_configs_text.startswith('Named Configurations (' + COLOR_DOC +
'doc' + ENDC + '):')
assert 'named_config2' in named_configs_text
assert '# named config with doc' in named_configs_text
assert 'ingred.named_config1' in named_configs_text
assert 'ingred.dict_config' in named_configs_text
def test_format_named_configs():
ingred = Ingredient("ingred")
ex = Experiment(name="experiment", ingredients=[ingred])
@ingred.named_config
def named_config1():
pass
@ex.named_config
def named_config2():
"""named config with doc"""
pass
dict_config = dict(v=42)
ingred.add_named_config("dict_config", dict_config)
named_configs_text = _format_named_configs(
OrderedDict(ex.gather_named_configs()))
assert named_configs_text.startswith("Named Configurations (" + COLOR_DOC +
"doc" + ENDC + "):")
assert "named_config2" in named_configs_text
assert "# named config with doc" in named_configs_text
assert "ingred.named_config1" in named_configs_text
assert "ingred.dict_config" in named_configs_text
pixel_prior = {
'p': 0.0, # probability of success for pixel prior Bernoulli
'mu': 0.0, # mean of pixel prior Gaussian
'sigma': 0.25 # std of pixel prior Gaussian
}
# em
k = 3 # number of components
nr_steps = 10 # number of (RN)N-EM steps
e_sigma = 0.25 # sigma used in the e-step when pixel distributions are Gaussian (acts as a temperature)
pred_init = 0.0 # initial prediction used to compute the input
# named config to be used when processing sequential data
nem.add_named_config('sequential', {
'gradient_gamma': False,
'loss_step_weights': 'all'
})
class NEMCell(RNNCell):
"""A RNNCell like implementation of (RN)N-EM."""
@nem.capture
def __init__(self, cell, input_shape, distribution, pred_init, e_sigma):
self.cell = cell
if not isinstance(input_shape, tf.TensorShape):
input_shape = tf.TensorShape(input_shape)
self.input_shape = input_shape
self.gamma_shape = tf.TensorShape(input_shape.as_list()[:-1] + [1])
self.distribution = distribution
self.pred_init = pred_init
self.e_sigma = e_sigma
ds = Ingredient('dataset')
@ds.config
def cfg():
name = 'shapes'
path = './data'
binary = True
train_size = None # subset of training set (None, int)
valid_size = 1000 # subset of valid set (None, int)
test_size = None # subset of test set (None, int)
queue_capacity = 100 # nr of batches in the queue
ds.add_named_config('shapes', {'name': 'shapes', 'binary': True})
ds.add_named_config('flying_shapes', {'name': 'flying_shapes', 'binary': True})
ds.add_named_config('flying_shapes_4', {
'name': 'flying_shapes_4',
'binary': True
})
ds.add_named_config('flying_shapes_5', {
'name': 'flying_shapes_5',
'binary': True
})
ds.add_named_config('flying_mnist_medium_20_2', {
'name': 'flying_mnist_medium20_2digits',
'binary': False
})
ds.add_named_config('flying_mnist_medium_500_2', {
'name': 'flying_mnist_medium500_2digits',
reconstruction_loss = {'name': 'recons_nll', 'params': {'loss': 'bce'}}
bvae_loss = {'name': 'beta-vae', 'params': {'reconstruction_loss': 'bce',
'beta': 4.0}}
cap_const = {'name': 'constrained-beta-vae', 'params': {'reconstruction_loss':
'bce', 'gamma': 100,
'capacity': 7}}
bxent_loss = {'name': 'bxent', 'params': {}}
xent_loss = {'name': 'xent', 'params': {}}
accuracy = {'name': 'acc',
'params': {'output_transform': thresholded_output_transform}}
mse_loss = {'name': 'mse', 'params': {}}
kl_div = {'name': 'kl-div', 'params': {}}
training = Ingredient('training')
training.add_named_config('vae', loss=vae_loss,
metrics=[reconstruction_loss, kl_div])
training.add_named_config('bvae', loss=bvae_loss,
metrics=[reconstruction_loss, kl_div])
training.add_named_config('capconst', loss=cap_const,
metrics=[reconstruction_loss, kl_div])
training.add_named_config('2afc', loss=bxent_loss,
metrics=[bxent_loss, accuracy])
training.add_named_config('mafc', loss=xent_loss,
metrics=[xent_loss, accuracy])
training.add_named_config('recons_nll', loss=reconstruction_loss,
metrics=[reconstruction_loss])
init_optimizer = training.capture(init_optimizer)
ds = Ingredient('dataset')
@ds.config
def cfg():
name = 'shapes'
path = './data'
binary = True
train_size = None # subset of training set (None, int)
valid_size = 1000 # subset of valid set (None, int)
test_size = None # subset of test set (None, int)
queue_capacity = 100 # nr of batches in the queue
ds.add_named_config('shapes', {'name': 'shapes', 'binary': True})
ds.add_named_config('flying_shapes', {'name': 'flying_shapes', 'binary': True})
ds.add_named_config('flying_shapes_4', {
'name': 'flying_shapes_4',
'binary': True
})
ds.add_named_config('flying_shapes_5', {
'name': 'flying_shapes_5',
'binary': True
})
ds.add_named_config('flying_mnist_medium_20_2', {
'name': 'flying_mnist_medium20_2digits',
'binary': False
})
ds.add_named_config('flying_mnist_medium_500_2', {
'name': 'flying_mnist_medium500_2digits',
net.add_named_config(
'enc_dec_84_atari', {
'input': [
{
'name': 'reshape',
'shape': (84, 84, 1)
},
{
'name': 'conv',
'size': 16,
'act': 'elu',
'stride': [4, 4],
'kernel': (8, 8),
'ln': True
},
{
'name': 'conv',
'size': 32,
'act': 'elu',
'stride': [2, 2],
'kernel': (4, 4),
'ln': True
},
{
'name': 'reshape',
'shape': -1
},
{
'name': 'fc',
'size': 250,
'act': 'elu',
'ln': True
},
],
'output': [
{
'name': 'fc',
'size': 250,
'act': 'relu',
'ln': True
},
{
'name': 'fc',
'size': 10 * 10 * 32,
'act': 'relu',
'ln': True
},
{
'name': 'reshape',
'shape': (10, 10, 32)
},
{
'name': 'r_conv',
'size': 16,
'act': 'relu',
'stride': [2, 2],
'kernel': (4, 4),
'ln': True,
'offset': 1
},
{
'name': 'r_conv',
'size': 1,
'act': 'sigmoid',
'stride': [4, 4],
'kernel': (8, 8)
},
{
'name': 'reshape',
'shape': -1
},
]
})
net.add_named_config(
'flying_mnist', {
'input': [{
'name': 'input_norm'
}, {
'name': 'reshape',
'shape': (24, 24, 1)
}, {
'name': 'conv',
'size': 32,
'act': 'elu',
'stride': [2, 2],
'kernel': (4, 4),
'ln': True
}, {
'name': 'conv',
'size': 64,
'act': 'elu',
'stride': [2, 2],
'kernel': (4, 4),
'ln': True
}, {
'name': 'conv',
'size': 128,
'act': 'elu',
'stride': [2, 2],
'kernel': (4, 4),
'ln': True
}, {
'name': 'reshape',
'shape': -1
}, {
'name': 'fc',
'size': 512,
'act': 'elu',
'ln': True
}],
'recurrent': [{
'name': 'rnn',
'size': 250,
'act': 'sigmoid',
'ln': True
}],
'output': [{
'name': 'fc',
'size': 512,
'act': 'relu',
'ln': True
}, {
'name': 'fc',
'size': 3 * 3 * 128,
'act': 'relu',
'ln': True
}, {
'name': 'reshape',
'shape': (3, 3, 128)
}, {
'name': 'r_conv',
'size': 64,
'act': 'relu',
'stride': [2, 2],
'kernel': (4, 4),
'ln': True
}, {
'name': 'r_conv',
'size': 32,
'act': 'relu',
'stride': [2, 2],
'kernel': (4, 4),
'ln': True
}, {
'name': 'r_conv',
'size': 1,
'act': '*',
'stride': [2, 2],
'kernel': (4, 4),
'ln': False
}, {
'name': 'reshape',
'shape': -1
}]
})
from sacred import Ingredient
ds = Ingredient('dataset')
@ds.config
def cfg():
name = 'balls4mass64'
path = './data'
train_size = None # subset of training set (None, int)
valid_size = 1000 # subset of valid set (None, int)
test_size = None # subset of test set (None, int)
queue_capacity = 100 # nr of batches in the queue
ds.add_named_config('balls4mass64', {'name': 'balls4mass64'})
ds.add_named_config('balls678mass64', {'name': 'balls678mass64'})
ds.add_named_config('balls3curtain64', {'name': 'balls3curtain64'})
ds.add_named_config('atari', {'name': 'atari'})
class InputPipeLine(object):
@ds.capture
def _open_dataset(self, out_list, path, name, train_size, valid_size,
test_size):
# open dataset file
self._hdf5_file = h5py.File(os.path.join(path, name + '.h5'), 'r')
self._data_in_file = {
data_name: self._hdf5_file[self.usage][data_name]
for data_name in out_list
}
# from dataset.tdisc import load_tdata
from dataset.sprites import load_sprites
from dataset.shapes3d import load_shapes3d
from dataset.mpi import load_mpi3d
from dataset.transforms import Triplets
import configs.datasplits as splits
dataset = Ingredient('dataset')
load_sprites = dataset.capture(load_sprites)
load_shapes3d = dataset.capture(load_shapes3d)
load_mpi3d = dataset.capture(load_mpi3d)
load_composition = dataset.capture(Triplets)
dataset.add_config(setting='unsupervised')
dataset.add_named_config('unsupervised', setting='unsupervised')
dataset.add_named_config('supervised', setting='supervised')
@dataset.capture
def get_dataset(dataset):
if dataset == 'dsprites':
dataset_loader = load_sprites
elif dataset == 'shapes3d':
dataset_loader = load_shapes3d
elif dataset == 'mpi3d':
dataset_loader = load_mpi3d
elif dataset == 'composition':
dataset_loader = load_composition
else:
raise ValueError('Unrecognized dataset {}'.format(dataset))