def default_shuffled_filepath(paysage_path, filepath):
shuffled_filepath = os.path.join(paysage_path, 'mnist',
'shuffled_mnist.h5')
if not os.path.exists(shuffled_filepath):
print("Shuffled file does not exist, creating a shuffled dataset.")
shuffler = batch.DataShuffler(filepath, shuffled_filepath, complevel=0)
shuffler.shuffle()
return shuffled_filepath
def test_shuffle():
# create temporary files
file_original = tempfile.NamedTemporaryFile()
file_shuffle = tempfile.NamedTemporaryFile()
# create data
num_rows = 10000
num_cols_A = 100
num_cols_B = 1
df_A = pd.DataFrame(
np.arange(num_rows * num_cols_A).reshape(num_rows, num_cols_A),
columns=['col_{}'.format(i) for i in np.arange(num_cols_A)],
index=['ix_{}'.format(i) for i in np.arange(num_rows)])
df_B = pd.DataFrame(
np.arange(num_rows * num_cols_B).reshape(num_rows, num_cols_B),
columns=['col_{}'.format(i) for i in np.arange(num_cols_B)],
index=['ix_{}'.format(i) for i in np.arange(num_rows)])
# save it
store = pd.HDFStore(file_original.name, mode='w', format='table')
store.append("A", df_A)
store.append("B", df_B)
store.close()
# shuffle it, with an artificially low memory limit
shuffler = batch.DataShuffler(file_original.name,
file_shuffle.name,
allowed_mem=0.001)
shuffler.shuffle()
# read the shuffled data
df_As = pd.read_hdf(file_shuffle.name, "A")
df_Bs = pd.read_hdf(file_shuffle.name, "B")
# check the two shuffles are consistent
assert (df_As.index == df_Bs.index).all()
assert (df_As['col_0'] // num_cols_A == df_Bs['col_0'] // num_cols_B).all()
# check that the shuffles preserve the index
ix_A_orig = sorted(list(df_A.index))
ix_A_shuffled = sorted(list(df_As.index))
assert ix_A_orig == ix_A_shuffled
# check a couple of statistics
vals_B = df_B['col_0'].values
vals_Bs = df_Bs['col_0'].values
# the number of fixed points tends to a Poisson distribution with e.v. = 1
assert (vals_B == vals_Bs).sum() < 5
# the difference between values (using the natural numbers as values)
# is a triangular distribution centered at 0. Can check the variance.
diff_dist_std = (vals_B - vals_Bs).std()
assert np.abs(diff_dist_std / (num_rows / np.sqrt(6)) - 1) < 0.05
def default_paths(file = "shuffled"):
files = {"shuffled": {"input": "mnist.h5", "output": "shuffled_mnist.h5"},
}
file_path = os.path.abspath(__file__)
mnist_path = os.path.join(os.path.dirname(file_path), files[file]["input"])
if not os.path.exists(mnist_path):
raise IOError("{} does not exist. run download_mnist.py to fetch from the web"
.format(mnist_path))
shuffled_path = os.path.join(os.path.dirname(file_path), files[file]["output"])
if not os.path.exists(shuffled_path):
print("Shuffled file does not exist, creating a shuffled dataset.")
shuffler = batch.DataShuffler(mnist_path, shuffled_path, complevel=0)
shuffler.shuffle()
return shuffled_path
def test_rbm(paysage_path=None):
num_hidden_units = 50
batch_size = 50
num_epochs = 1
learning_rate = schedules.PowerLawDecay(initial=0.01, coefficient=0.1)
mc_steps = 1
if not paysage_path:
paysage_path = os.path.dirname(
os.path.dirname(os.path.abspath(__file__)))
filepath = os.path.join(paysage_path, 'examples', 'mnist', 'mnist.h5')
if not os.path.exists(filepath):
raise IOError(
"{} does not exist. run mnist/download_mnist.py to fetch from the web"
.format(filepath))
shuffled_filepath = os.path.join(paysage_path, 'examples', 'mnist',
'shuffled_mnist.h5')
# shuffle the data
if not os.path.exists(shuffled_filepath):
shuffler = batch.DataShuffler(filepath, shuffled_filepath, complevel=0)
shuffler.shuffle()
# set a seed for the random number generator
be.set_seed()
import pandas
samples = pre.binarize_color(
be.float_tensor(
pandas.read_hdf(shuffled_filepath,
key='train/images').values[:10000]))
samples_train, samples_validate = batch.split_tensor(samples, 0.95)
data = batch.Batch({
'train':
batch.InMemoryTable(samples_train, batch_size),
'validate':
batch.InMemoryTable(samples_validate, batch_size)
})
# set up the model and initialize the parameters
vis_layer = layers.BernoulliLayer(data.ncols)
hid_layer = layers.BernoulliLayer(num_hidden_units)
rbm = BoltzmannMachine([vis_layer, hid_layer])
rbm.initialize(data)
# obtain initial estimate of the reconstruction error
perf = ProgressMonitor()
untrained_performance = perf.epoch_update(data,
rbm,
store=True,
show=False)
# set up the optimizer and the fit method
opt = optimizers.RMSProp(stepsize=learning_rate)
cd = fit.SGD(rbm, data)
# fit the model
print('training with contrastive divergence')
cd.train(opt, num_epochs, method=fit.pcd, mcsteps=mc_steps)
# obtain an estimate of the reconstruction error after 1 epoch
trained_performance = cd.monitor.memory[-1]
assert (trained_performance['ReconstructionError'] <
untrained_performance['ReconstructionError']), \
"Reconstruction error did not decrease"
# close the HDF5 store
data.close()
def test_rbm(paysage_path=None):
num_hidden_units = 50
batch_size = 50
num_epochs = 1
learning_rate = schedules.PowerLawDecay(initial=0.01, coefficient=0.1)
mc_steps = 1
if not paysage_path:
paysage_path = os.path.dirname(
os.path.dirname(os.path.abspath(__file__)))
filepath = os.path.join(paysage_path, 'mnist', 'mnist.h5')
if not os.path.exists(filepath):
raise IOError(
"{} does not exist. run mnist/download_mnist.py to fetch from the web"
.format(filepath))
shuffled_filepath = os.path.join(paysage_path, 'mnist',
'shuffled_mnist.h5')
# shuffle the data
if not os.path.exists(shuffled_filepath):
shuffler = batch.DataShuffler(filepath, shuffled_filepath, complevel=0)
shuffler.shuffle()
# set a seed for the random number generator
be.set_seed()
# set up the reader to get minibatches
data = batch.HDFBatch(shuffled_filepath,
'train/images',
batch_size,
transform=pre.binarize_color,
train_fraction=0.99)
# set up the model and initialize the parameters
vis_layer = layers.BernoulliLayer(data.ncols)
hid_layer = layers.BernoulliLayer(num_hidden_units)
rbm = model.Model([vis_layer, hid_layer])
rbm.initialize(data)
# obtain initial estimate of the reconstruction error
perf = fit.ProgressMonitor(data, metrics=['ReconstructionError'])
untrained_performance = perf.check_progress(rbm)
# set up the optimizer and the fit method
opt = optimizers.RMSProp(stepsize=learning_rate)
sampler = fit.DrivenSequentialMC.from_batch(rbm, data)
cd = fit.SGD(rbm,
data,
opt,
num_epochs,
sampler,
method=fit.pcd,
mcsteps=mc_steps,
monitor=perf)
# fit the model
print('training with contrastive divergence')
cd.train()
# obtain an estimate of the reconstruction error after 1 epoch
trained_performance = perf.check_progress(rbm)
assert (trained_performance['ReconstructionError'] <
untrained_performance['ReconstructionError']), \
"Reconstruction error did not decrease"
# close the HDF5 store
data.close()
def test_tap_machine(paysage_path=None):
num_hidden_units = 10
batch_size = 50
num_epochs = 1
learning_rate = 0.01
if not paysage_path:
paysage_path = os.path.dirname(
os.path.dirname(os.path.abspath(__file__)))
filepath = os.path.join(paysage_path, 'mnist', 'mnist.h5')
if not os.path.exists(filepath):
raise IOError(
"{} does not exist. run mnist/download_mnist.py to fetch from the web"
.format(filepath))
shuffled_filepath = os.path.join(paysage_path, 'mnist',
'shuffled_mnist.h5')
# shuffle the data
if not os.path.exists(shuffled_filepath):
shuffler = batch.DataShuffler(filepath, shuffled_filepath, complevel=0)
shuffler.shuffle()
# set a seed for the random number generator
be.set_seed()
# set up the reader to get minibatches
data = batch.Batch(shuffled_filepath,
'train/images',
batch_size,
transform=batch.binarize_color,
train_fraction=0.1)
# set up the model and initialize the parameters
vis_layer = layers.BernoulliLayer(data.ncols)
hid_layer = layers.BernoulliLayer(num_hidden_units)
rbm = tap_machine.TAP_rbm([vis_layer, hid_layer],
tolerance_EMF=1e-2,
max_iters_EMF=50)
rbm.initialize(data)
# obtain initial estimate of the reconstruction error
perf = fit.ProgressMonitor(data, metrics=['ReconstructionError'])
untrained_performance = perf.check_progress(rbm)
# set up the optimizer and the fit method
opt = optimizers.Gradient(stepsize=learning_rate,
scheduler=optimizers.PowerLawDecay(0.1),
tolerance=1e-3,
ascent=True)
solver = fit.SGD(rbm, data, opt, num_epochs, method=fit.tap, monitor=perf)
# fit the model
print('training with stochastic gradient ascent')
solver.train()
# obtain an estimate of the reconstruction error after 1 epoch
trained_performance = perf.check_progress(rbm)
assert (trained_performance['ReconstructionError'] <
untrained_performance['ReconstructionError']), \
"Reconstruction error did not decrease"
# close the HDF5 store
data.close()
def test_tap_machine(paysage_path=None):
num_hidden_units = 10
batch_size = 100
num_epochs = 5
learning_rate = schedules.PowerLawDecay(initial=0.1, coefficient=1.0)
if not paysage_path:
paysage_path = os.path.dirname(
os.path.dirname(os.path.abspath(__file__)))
filepath = os.path.join(paysage_path, 'examples', 'mnist', 'mnist.h5')
if not os.path.exists(filepath):
raise IOError(
"{} does not exist. run mnist/download_mnist.py to fetch from the web"
.format(filepath))
shuffled_filepath = os.path.join(paysage_path, 'examples', 'mnist',
'shuffled_mnist.h5')
# shuffle the data
if not os.path.exists(shuffled_filepath):
shuffler = batch.DataShuffler(filepath, shuffled_filepath, complevel=0)
shuffler.shuffle()
# set a seed for the random number generator
be.set_seed()
# set up the reader to get minibatches
samples = pre.binarize_color(
be.float_tensor(
pandas.read_hdf(shuffled_filepath,
key='train/images').as_matrix()[:10000]))
samples_train, samples_validate = batch.split_tensor(samples, 0.95)
data = batch.Batch({
'train':
batch.InMemoryTable(samples_train, batch_size),
'validate':
batch.InMemoryTable(samples_validate, batch_size)
})
# set up the model and initialize the parameters
vis_layer = layers.BernoulliLayer(data.ncols)
hid_layer = layers.BernoulliLayer(num_hidden_units)
rbm = BoltzmannMachine([vis_layer, hid_layer])
rbm.initialize(data)
# obtain initial estimate of the reconstruction error
perf = ProgressMonitor(generator_metrics = \
[ReconstructionError(), TAPLogLikelihood(10), TAPFreeEnergy(10)])
untrained_performance = perf.epoch_update(data,
rbm,
store=True,
show=False)
# set up the optimizer and the fit method
opt = optimizers.Gradient(stepsize=learning_rate, tolerance=1e-5)
tap = fit.TAP(True, 0.1, 0.01, 25, True, 0.5, 0.001, 0.0)
solver = fit.SGD(rbm, data)
solver.monitor.generator_metrics.append(TAPLogLikelihood(10))
solver.monitor.generator_metrics.append(TAPFreeEnergy(10))
# fit the model
print('training with stochastic gradient ascent')
solver.train(opt, num_epochs, method=tap.tap_update)
# obtain an estimate of the reconstruction error after 1 epoch
trained_performance = solver.monitor.memory[-1]
assert (trained_performance['TAPLogLikelihood'] >
untrained_performance['TAPLogLikelihood']), \
"TAP log-likelihood did not increase"
assert (trained_performance['ReconstructionError'] <
untrained_performance['ReconstructionError']), \
"Reconstruction error did not decrease"
# close the HDF5 store
data.close()
def test_rbm(paysage_path=None):
"""TODO : this is just a placeholder, need to clean up & simplifiy setup. Also
need to figure how to deal with consistent random seeding throughout the
codebase to obtain deterministic checkable results.
"""
num_hidden_units = 50
batch_size = 50
num_epochs = 1
learning_rate = 0.01
mc_steps = 1
if not paysage_path:
paysage_path = os.path.dirname(
os.path.dirname(os.path.abspath(__file__)))
filepath = os.path.join(paysage_path, 'mnist', 'mnist.h5')
if not os.path.exists(filepath):
raise IOError(
"{} does not exist. run mnist/download_mnist.py to fetch from the web"
.format(filepath))
shuffled_filepath = os.path.join(paysage_path, 'mnist',
'shuffled_mnist.h5')
# shuffle the data
if not os.path.exists(shuffled_filepath):
shuffler = batch.DataShuffler(filepath, shuffled_filepath, complevel=0)
shuffler.shuffle()
# set a seed for the random number generator
be.set_seed()
# set up the reader to get minibatches
data = batch.Batch(shuffled_filepath,
'train/images',
batch_size,
transform=batch.binarize_color,
train_fraction=0.99)
# set up the model and initialize the parameters
vis_layer = layers.BernoulliLayer(data.ncols)
hid_layer = layers.BernoulliLayer(num_hidden_units)
rbm = hidden.Model([vis_layer, hid_layer])
rbm.initialize(data)
# obtain initial estimate of the reconstruction error
perf = fit.ProgressMonitor(0, data, metrics=[M.ReconstructionError()])
untrained_performance = perf.check_progress(rbm, 0)
# set up the optimizer and the fit method
opt = optimizers.RMSProp(rbm,
stepsize=learning_rate,
scheduler=optimizers.PowerLawDecay(0.1))
sampler = fit.DrivenSequentialMC.from_batch(rbm, data, method='stochastic')
cd = fit.PCD(rbm,
data,
opt,
sampler,
num_epochs,
mcsteps=mc_steps,
skip=200,
metrics=[M.ReconstructionError()])
# fit the model
print('training with contrastive divergence')
cd.train()
# obtain an estimate of the reconstruction error after 1 epoch
trained_performance = perf.check_progress(rbm, 0)
assert (trained_performance['ReconstructionError'] <
untrained_performance['ReconstructionError']), \
"Reconstruction error did not decrease"
# close the HDF5 store
data.close()