def get_indexes_da(source_folder, target_folder, nTrain, transferImages):
get_logger().info("Loading indexes")
allowed = ['backpack.hdf5', 'headphones.hdf5', 'monitor.hdf5', 'bike.hdf5', 'keyboard.hdf5', 'mouse.hdf5', 'projector.hdf5', 'calculator.hdf5', 'laptop.hdf5', 'mug.hdf5']
train = []
test = []
MAX_TEST = 100
for categ in allowed:
#support_filename = join(".", basename(filename))
s_filename = join(source_folder, categ)
t_filename = join(target_folder, categ)
get_logger().info("Loading " + s_filename)
shfile = HDF5File(s_filename, 'r')
thfile = HDF5File(t_filename, 'r')
siid = shfile["image_index"][:]
tiid = thfile["image_index"][:]
np.random.shuffle(siid)
np.random.shuffle(tiid)
trainIdx = siid[0:nTrain]
trainTIdx =tiid[0:transferImages]
testIdx = tiid[transferImages:transferImages+MAX_TEST]
train.append((ClassIndexes(s_filename, trainIdx), ClassIndexes(t_filename, trainTIdx))) #data is actually loaded only when needed
test.append([ClassIndexes(t_filename, testIdx)])
shfile.close()
thfile.close()
return Data(train, test)
def load_model(self, ubm_file):
"""Loads the projector (UBM) from a file."""
hdf5file = HDF5File(ubm_file, "r")
logger.debug("Loading model from file '%s'", ubm_file)
# Read the UBM
self.ubm = GMMMachine.from_hdf5(hdf5file)
self.ubm.variance_thresholds = self.variance_threshold
def from_hdf5(cls, hdf5):
"""Creates a new GMMStats object from an `HDF5File` object."""
if isinstance(hdf5, str):
hdf5 = HDF5File(hdf5, "r")
try:
version_major, version_minor = hdf5.attrs["file_version"].split(".")
logger.debug(
f"Reading a GMMStats HDF5 file of version {version_major}.{version_minor}"
)
except (KeyError, RuntimeError):
version_major, version_minor = 0, 0
if int(version_major) >= 1:
if hdf5.attrs["writer_class"] != str(cls):
logger.warning(f"{hdf5.attrs['writer_class']} is not {cls}.")
self = cls(
n_gaussians=hdf5["n_gaussians"][()],
n_features=hdf5["n_features"][()],
)
self.log_likelihood = hdf5["log_likelihood"][()]
self.t = hdf5["T"][()]
self.n = hdf5["n"][...]
self.sum_px = hdf5["sumPx"][...]
self.sum_pxx = hdf5["sumPxx"][...]
else: # Legacy file version
logger.info("Loading a legacy HDF5 stats file.")
self = cls(
n_gaussians=int(hdf5["n_gaussians"][()]),
n_features=int(hdf5["n_inputs"][()]),
)
self.log_likelihood = hdf5["log_liklihood"][()]
self.t = int(hdf5["T"][()])
self.n = np.reshape(hdf5["n"], (self.n_gaussians,))
self.sum_px = np.reshape(hdf5["sumPx"], (self.shape))
self.sum_pxx = np.reshape(hdf5["sumPxx"], (self.shape))
return self
def get_support(filename, size):
hfile = HDF5File(filename, 'r')
ds = hfile['support']
patches = ds[:min(size, ds.shape[0]), :]
hfile.close()
return patches.astype('float')
def test_gmm_ML_1():
"""Trains a GMMMachine with ML_GMMTrainer"""
ar = load_array(
resource_filename("bob.learn.em", "data/faithful.torch3_f64.hdf5"))
gmm_ref = GMMMachine.from_hdf5(
HDF5File(resource_filename("bob.learn.em", "data/gmm_ML.hdf5"), "r"))
for transform in (to_numpy, to_dask_array):
ar = transform(ar)
gmm = loadGMM()
# test rng handling
gmm.convergence_threshold = 0.001
gmm.update_means = True
gmm.update_variances = True
gmm.update_weights = True
gmm.random_state = np.random.RandomState(seed=12345)
gmm = gmm.fit(ar)
gmm = loadGMM()
gmm.convergence_threshold = 0.001
gmm.update_means = True
gmm.update_variances = True
gmm.update_weights = True
# Generate reference
# gmm.save(HDF5File(resource_filename("bob.learn.em", "data/gmm_ML.hdf5"), "w"))
gmm = gmm.fit(ar)
assert_gmm_equal(gmm, gmm_ref)
def test_GMMMachine_stats():
"""Tests a GMMMachine (statistics)"""
arrayset = load_array(
resource_filename("bob.learn.em", "data/faithful.torch3_f64.hdf5"))
gmm = GMMMachine(n_gaussians=2)
gmm.weights = np.array([0.5, 0.5], "float64")
gmm.means = np.array([[3, 70], [4, 72]], "float64")
gmm.variances = np.array([[1, 10], [2, 5]], "float64")
gmm.variance_thresholds = np.array([[0, 0], [0, 0]], "float64")
stats = gmm_module.e_step(
arrayset,
gmm,
)
stats_ref = GMMStats(n_gaussians=2, n_features=2)
stats_ref.load(
HDF5File(resource_filename("bob.learn.em", "data/stats.hdf5"), "r"))
np.testing.assert_equal(stats.t, stats_ref.t)
np.testing.assert_almost_equal(stats.n, stats_ref.n, decimal=10)
# np.testing.assert_equal(stats.sum_px, stats_ref.sum_px)
# Note AA: precision error above
np.testing.assert_almost_equal(stats.sum_px, stats_ref.sum_px, decimal=10)
np.testing.assert_almost_equal(stats.sum_pxx,
stats_ref.sum_pxx,
decimal=10)
def write_out_file(infile, outfile, tree=None):
f = uproot.open(infile)["fancy_tree;1"]
#f = root_open(infile)
#T = f[tree]
names = f.keys()
cells = list(filter(lambda x: x.startswith(b'cell'), names))
assert len(cells) == sum(map(np.prod, LAYER_SPECS)) + OVERFLOW_BINS
for df in uproot.pandas.iterate(infile, "fancy_tree;1", branches = cells):
X = df
for df in uproot.pandas.iterate(infile, "fancy_tree;1", branches = b'TotalEnergy'):
E = df
X = X.values
E = E.values.ravel()
#X = tree.pd.DataFrame(tree2array(T, branches=cells)).values
#E = tree.pd.DataFrame(tree2array(T, branches=['TotalEnergy'])).values.ravel()
print(X.shape)
with HDF5File(outfile, 'w') as h5:
for layer, (sh, (l, u)) in enumerate(zip(LAYER_SPECS, LAYER_DIV)):
h5['layer_{}'.format(layer)] = X[:, l:u].reshape((-1, ) + sh)
print(u)
h5['overflow'] = X[:, -OVERFLOW_BINS:]
h5['energy'] = E.reshape(-1, 1)
def select_random_support(train_dir, support_dir, num_train_images,
support_size, position_influence):
log = get_logger()
train_files = [
f for f in glob(join(train_dir, '*'))
if splitext(f.lower())[1] == '.hdf5'
]
try:
os.makedirs(support_dir)
except:
pass
for target_file in train_files:
log.info('Extracting random support from "%s"...',
basename(target_file))
#(patches, _)= get_standardized_patches(target_file, num_train_images, position_influence)
(patches, _) = get_patches(target_file, num_train_images,
position_influence)
rand_ix = random.sample(range(patches.shape[0]),
min(patches.shape[0], support_size))
patches = patches[np.array(rand_ix), :]
fh = HDF5File(join(support_dir, basename(target_file)), 'w')
ds = fh.create_dataset('support', patches.shape, dtype='float')
ds[:] = patches
ds.attrs['cursor'] = patches.shape[0]
fh.close()
def save_model(self, ubm_file):
"""Saves the projector (UBM) to file."""
# Saves the UBM to file
logger.debug("Saving model to file '%s'", ubm_file)
hdf5 = (ubm_file if isinstance(ubm_file, HDF5File) else HDF5File(
ubm_file, "w"))
self.ubm.save(hdf5)
def from_hdf5(cls, hdf5, ubm=None):
"""Creates a new GMMMachine object from an `HDF5File` object."""
if isinstance(hdf5, str):
hdf5 = HDF5File(hdf5, "r")
try:
version_major, version_minor = hdf5.attrs["file_version"].split(".")
logger.debug(
f"Reading a GMMMachine HDF5 file of version {version_major}.{version_minor}"
)
except (KeyError, RuntimeError):
version_major, version_minor = 0, 0
if int(version_major) >= 1:
if hdf5.attrs["writer_class"] != str(cls):
logger.warning(f"{hdf5.attrs['writer_class']} is not {cls}.")
if hdf5["trainer"] == "map" and ubm is None:
raise ValueError(
"The UBM is needed when loading a MAP machine."
)
self = cls(
n_gaussians=hdf5["n_gaussians"][()],
trainer=hdf5["trainer"][()],
ubm=ubm,
convergence_threshold=1e-5,
max_fitting_steps=hdf5["max_fitting_steps"][()],
weights=hdf5["weights"][...],
k_means_trainer=None,
update_means=hdf5["update_means"][()],
update_variances=hdf5["update_variances"][()],
update_weights=hdf5["update_weights"][()],
)
gaussians_group = hdf5["gaussians"]
self.means = gaussians_group["means"][...]
self.variances = gaussians_group["variances"][...]
self.variance_thresholds = gaussians_group["variance_thresholds"][
...
]
else: # Legacy file version
logger.info("Loading a legacy HDF5 machine file.")
n_gaussians = hdf5["m_n_gaussians"][()]
g_means = []
g_variances = []
g_variance_thresholds = []
for i in range(n_gaussians):
gaussian_group = hdf5[f"m_gaussians{i}"]
g_means.append(gaussian_group["m_mean"][...])
g_variances.append(gaussian_group["m_variance"][...])
g_variance_thresholds.append(
gaussian_group["m_variance_thresholds"][...]
)
weights = np.reshape(hdf5["m_weights"], (n_gaussians,))
self = cls(n_gaussians=n_gaussians, ubm=ubm, weights=weights)
self.means = np.array(g_means).reshape(n_gaussians, -1)
self.variances = np.array(g_variances).reshape(n_gaussians, -1)
self.variance_thresholds = np.array(g_variance_thresholds).reshape(
n_gaussians, -1
)
return self
def save(self, filepath):
with HDF5File(filepath, "w") as hdf5:
hdf5.create_dataset("size", data=self.size)
hdf5.create_dataset("depth", data=self.depth)
hdf5.create_dataset("levels", data=self.levels)
hdf5.create_dataset("first_channels", data=self.first_channels)
hdf5.create_dataset("last_channels", data=self.last_channels)
hdf5.create_dataset("categories", data=self.categories)
HDF5Serializer(hdf5.create_group("weights")).save(self)
def test_gmm_MAP_1():
# Train a GMMMachine with MAP_GMMTrainer
ar = load_array(
resource_filename("bob.learn.em", "data/faithful.torch3_f64.hdf5"))
# test with rng
gmmprior = GMMMachine.from_hdf5(
HDF5File(resource_filename("bob.learn.em", "data/gmm_ML.hdf5"), "r"))
gmm = GMMMachine.from_hdf5(
HDF5File(resource_filename("bob.learn.em", "data/gmm_ML.hdf5"), "r"),
ubm=gmmprior,
)
gmm.update_means = True
gmm.update_variances = False
gmm.update_weights = False
rng = np.random.RandomState(seed=12345)
gmm.random_state = rng
gmm = gmm.fit(ar)
gmmprior = GMMMachine.from_hdf5(
HDF5File(resource_filename("bob.learn.em", "data/gmm_ML.hdf5"), "r"))
gmm = GMMMachine.from_hdf5(
HDF5File(resource_filename("bob.learn.em", "data/gmm_ML.hdf5"), "r"),
ubm=gmmprior,
)
gmm.update_means = True
gmm.update_variances = False
gmm.update_weights = False
# Generate reference
# gmm.save(HDF5File(resource_filename("bob.learn.em", "data/gmm_MAP.hdf5"), "w"))
gmm_ref = GMMMachine.from_hdf5(
HDF5File(resource_filename("bob.learn.em", "data/gmm_MAP.hdf5"), "r"))
for transform in (to_numpy, to_dask_array):
ar = transform(ar)
gmm = gmm.fit(ar)
np.testing.assert_almost_equal(gmm.means, gmm_ref.means, decimal=3)
np.testing.assert_almost_equal(gmm.variances,
gmm_ref.variances,
decimal=3)
np.testing.assert_almost_equal(gmm.weights, gmm_ref.weights, decimal=3)
def load_states(self, filepath):
with HDF5File(filepath, "r") as hdf5:
self.averaged_path_length = float(hdf5["averaged_path_length"][()])
self.augumentation_probability = float(
hdf5["augumentation_probability"][()])
HDF5Deserializer(hdf5["generator"]).load(self.generator)
HDF5Deserializer(hdf5["averaged_generator"]).load(
self.averaged_generator)
HDF5Deserializer(hdf5["discriminator"]).load(self.discriminator)
for key, optimizer in dict(self.optimizers).items():
HDF5Deserializer(hdf5["optimizers"][key]).load(optimizer)
def load(filepath): with HDF5File(filepath, "r") as hdf5: size = int(hdf5["size"][()]) depth = int(hdf5["depth"][()]) levels = int(hdf5["levels"][()]) first_channels = int(hdf5["first_channels"][()]) last_channels = int(hdf5["last_channels"][()]) categories = int(hdf5["categories"][()]) generator = Generator(size, depth, levels, first_channels, last_channels, categories) HDF5Deserializer(hdf5["weights"]).load(generator) return generator
def save(self, hdf5):
"""Saves the current statistsics in an `HDF5File` object."""
if isinstance(hdf5, str):
hdf5 = HDF5File(hdf5, "w")
hdf5.attrs["file_version"] = "1.0"
hdf5.attrs["writer_class"] = str(self.__class__)
hdf5["n_gaussians"] = self.n_gaussians
hdf5["n_features"] = self.n_features
hdf5["log_likelihood"] = float(self.log_likelihood)
hdf5["T"] = int(self.t)
hdf5["n"] = np.array(self.n)
hdf5["sumPx"] = np.array(self.sum_px)
hdf5["sumPxx"] = np.array(self.sum_pxx)
def doSnapshots():
"""
Read the snapshots and plot the corresponding variables.
"""
# get all the filenames
filenames = glob("simple_orbits_*.hdf5")
N = len(filenames)
filenames.sort()
# generate the output arrays
E = np.zeros((N, makeIC.num_part))
t = np.zeros(N)
p = np.zeros((N, 3))
v = np.zeros((N, 3))
for i, f in enumerate(filenames):
# get the data from the file
f = HDF5File(f, "r")
ids = f["PartType1/ParticleIDs"][:]
sort = np.argsort(ids)
ids = ids[sort]
pos = f["PartType1/Coordinates"][sort, :]
pos -= center
vel = f["PartType1/Velocities"][sort, :]
t[i] = f["Header"].attrs["Time"]
r = np.sum(pos**2, axis=1)**0.5
v2 = np.sum(vel**2, axis=1)
E[i, :] = 0.5 * v2 - G * M / r
# Get the pos / vel of the required particle
ind = ids == id_focus
p[i, :] = pos[ind, :]
v[i, :] = vel[ind, :]
# Compute the solution
y0 = np.zeros(4)
y0[:2] = p[0, :2]
y0[2:] = v[0, :2]
# compute the plotting variables
plt.figure(fig_1.number)
plotRelative(t, E, ".", label="Snapshot")
plt.figure(fig_2.number)
plt.plot(p[:, 0], p[:, 1], "-", label="Snapshot", lw=1.)
plt.figure(fig_3.number)
plt.plot(v[:, 0], v[:, 1], "-", label="Snapshot", lw=1.)
def create_hdf5_dataset(output_filename, patches, positions):
log = get_logger()
log.debug('Saving extracted descriptors to %s', output_filename)
hfile = HDF5File(output_filename, 'w', compression='gzip', fillvalue=0.0)
hpatches = hfile.create_dataset('patches',
patches.shape,
dtype="float32",
chunks=True)
hpositions = hfile.create_dataset('positions',
positions.shape,
dtype="uint16",
chunks=True)
hpatches[:] = patches
hpositions[:] = positions
hfile.close()
def save_states(self, filepath):
with HDF5File(filepath, "w") as hdf5:
hdf5.create_dataset("averaged_path_length",
data=self.averaged_path_length)
hdf5.create_dataset("augumentation_probability",
data=self.augumentation_probability)
HDF5Serializer(hdf5.create_group("generator")).save(self.generator)
HDF5Serializer(hdf5.create_group("averaged_generator")).save(
self.averaged_generator)
HDF5Serializer(hdf5.create_group("discriminator")).save(
self.discriminator)
optimizer_group = hdf5.create_group("optimizers")
for key, optimizer in dict(self.optimizers).items():
HDF5Serializer(
optimizer_group.create_group(key)).save(optimizer)
def get_num_patches(filename, num_images):
hfile = HDF5File(filename, 'r')
total_num_patches = hfile[PATCH_TYPE].attrs['cursor']
dim = hfile[PATCH_TYPE].shape[1]
if num_images == 0:
num_images = hfile['image_index'].shape[0]
image_index = hfile['image_index'][:min(num_images, hfile['image_index'].
shape[0])]
# Getting patches only from desired number of images
num_patches = min(image_index[-1, 1], total_num_patches)
hfile.close()
return (num_patches, dim)
def __init__(self, raw_files, comm=None, blocksize=2**16*20*2*4):
"""
Initialize a lofar observation, tracking/joining the two polarizations.
We also parse the corresponding HDF5 files to initialize:
nchan, samplerate, fwidth
"""
# read the HDF5 file and get useful data
h0 = HDF5File(raw_files[0].replace('.raw', '.h5'), 'r')
saps = sorted([i for i in h0.keys() if 'SUB_ARRAY_POINTING' in i])
s0 = h0[saps[0]]
time0 = Time(s0.attrs['EXPTIME_START_UTC'].replace('Z',''),
scale='utc')
beams = sorted([i for i in s0.keys() if 'BEAM' in i])
b0 = s0[beams[0]]
frequencies = (b0['COORDINATES']['COORDINATE_1']
.attrs['AXIS_VALUES_WORLD'] * u.Hz).to(u.MHz)
fbottom = frequencies[0]
stokes = sorted([i for i in b0.keys()
if 'STOKES' in i and 'i2f' not in i])
st0 = b0[stokes[0]]
dtype = _lofar_dtypes[st0.attrs['DATATYPE']]
nchan = len(frequencies) # = st0.attrs['NOF_SUBBANDS']
# can also get from np.diff(frequencies.diff).mean()
fwidth = (b0.attrs['SUBBAND_WIDTH'] *
u.__dict__[b0.attrs['CHANNEL_WIDTH_UNIT']]).to(u.MHz)
samplerate = (b0.attrs['SAMPLING_RATE'] *
u.__dict__[b0.attrs['SAMPLING_RATE_UNIT']]).to(u.MHz)
h0.close()
self.time0 = time0
self.samplerate = samplerate
self.fwidth = fwidth
self.frequencies = frequencies
self.fedge = fbottom
self.fedge_at_top = False
self.dtsample = (1./self.fwidth).to(u.s)
super(LOFARdata, self).__init__(raw_files, blocksize, dtype, nchan,
comm=comm)
# update some of the hdu data
self['PRIMARY'].header['DATE-OBS'] = self.time0.isot
self[0].header.update('TBIN', (1./samplerate).to('s').value)
def write_out_file(infile, outfile, tree=None):
f = root_open(infile)
T = f[tree]
cells = filter(lambda x: x.startswith('cell'), T.branchnames)
assert len(cells) == sum(map(np.prod, LAYER_SPECS)) + OVERFLOW_BINS
X = pd.DataFrame(tree2array(T, branches=cells)).values
E = pd.DataFrame(tree2array(T, branches=['TotalEnergy'])).values.ravel()
with HDF5File(outfile, 'w') as h5:
for layer, (sh, (l, u)) in enumerate(zip(LAYER_SPECS, LAYER_DIV)):
h5['layer_{}'.format(layer)] = X[:, l:u].reshape((-1, ) + sh)
h5['overflow'] = X[:, -OVERFLOW_BINS:]
h5['energy'] = E.reshape(-1, 1)
def save(self, hdf5):
"""Saves the current statistics in an `HDF5File` object."""
if isinstance(hdf5, str):
hdf5 = HDF5File(hdf5, "w")
hdf5.attrs["file_version"] = "1.0"
hdf5.attrs["writer_class"] = str(self.__class__)
hdf5["n_gaussians"] = self.n_gaussians
hdf5["trainer"] = self.trainer
hdf5["convergence_threshold"] = self.convergence_threshold
hdf5["max_fitting_steps"] = self.max_fitting_steps
hdf5["weights"] = self.weights
hdf5["update_means"] = self.update_means
hdf5["update_variances"] = self.update_variances
hdf5["update_weights"] = self.update_weights
gaussians_group = hdf5.create_group("gaussians")
gaussians_group["means"] = self.means
gaussians_group["variances"] = self.variances
gaussians_group["variance_thresholds"] = self.variance_thresholds
def imread_mat(filename, name=None):
"""
Read an 'image' from a MATLAB .MAT file. The file can be any version. Files
that are v7.3 require the h5py module. If no name is given, the first
variable is taken.
"""
try:
# Try general first (doesn't work for v7.3+ files)
# SciPy has this built in
# Supports loading just the given variable name
# Otherwise have to load all variables and skip special keys starting with "__" to find the variable to load
# Loaded matracies are already arrays
from scipy.io import loadmat
if name == None:
try:
# Try to get first variable name without loading entire file (only supported in SciPy 0.12+)
from scipy.io import whosmat
keys in whosmat(file_name)
if len(keys) == 0: raise KeyError()
name = keys[0][0]
except:
pass
x = loadmat(filename, variable_names=name)
if name == None:
name = '__' # we need to find first
for name in x.iterkeys():
if name[:2] != '__': break
if name[:2] == '__': raise KeyError() # no variables
return x[name] # can raise key error
except NotImplementedError:
# Try v7.3 file which is an HDF5 file
# We have to use h5py for this (or PyTables...)
# Always loads entire metadata (not just specific variable) but none of the data
# Data needs to be actually loaded (.value) and transposed (.T)
from h5py import File as HDF5File # TODO: if import error try using PyTables
with HDF5File(
filename, 'r'
) as x: # IOError if it doesn't exist or is the wrong format
if name == None:
try:
name = x.iterkeys().next()
except StopIteration:
raise KeyError() # no variables
return x[name].value.T # can raise key error
def __init__(self,
output_name,
output_dir,
num_files,
patches,
feature_type,
patch_dim=128,
patch_type='uint8',
pos_type='uint16'):
self.log = get_logger()
output_subdir = output_dir
try:
makedirs(output_subdir)
except:
pass
output_filename = join(output_subdir, basename(output_name))
self.log.debug('Saving extracted descriptors to %s', output_filename)
self.mode = 'creating'
dt = special_dtype(vlen=bytes)
patches += 10 #for safety
self.hfile = HDF5File(output_filename,
'w',
compression='gzip',
fillvalue=0.0)
self.patches = self.hfile.create_dataset(
'patches', (num_files * patches, patch_dim),
dtype=patch_type,
chunks=True)
self.positions = self.hfile.create_dataset('positions',
(num_files * patches, 2),
dtype=pos_type,
chunks=True)
self.image_index = self.hfile.create_dataset(
'image_index', (num_files, 2),
dtype='uint64') # Start, End positions of an image
self.keys = self.hfile.create_dataset('keys', (num_files, ), dtype=dt)
self.key_set = set()
self.patches.attrs['cursor'] = 0
self.patches.attrs['feature_type'] = feature_type
self.output_filename = output_filename
def get_patches(filename, num_images, position_influence=0):
hfile = HDF5File(filename, 'r')
total_num_patches = hfile[PATCH_TYPE].attrs['cursor']
if num_images == 0:
num_images = hfile['image_index'].shape[0]
image_index = hfile['image_index'][:min(num_images, hfile['image_index'].
shape[0])]
# Getting patches only from desired number of images
num_patches = min(image_index[-1, 1], total_num_patches)
patches = hfile[PATCH_TYPE][:num_patches, :]
# patches = patches.astype(float)
# norms = (patches**2).sum(axis=1)**0.5
# patches /= norms.max()
#patches = patches.astype(float)
#patches /= patches.max()
feature_type = hfile[PATCH_TYPE].attrs.get('feature_type', None)
# if feature_type == 'DECAF':
# norms = (patches**2).sum(axis=1)**0.5
# patches /= norms.max()
if position_influence > 0:
pos = hfile['positions'][:num_patches, :]
pos = pos.astype(float)
max_x = pos[:, 0].max()
max_y = pos[:, 1].max()
if max_x > 0:
pos[:, 0] /= max_x
if max_y > 0:
pos[:, 1] /= max_y
patches = np.hstack([patches, position_influence * pos])
hfile.close()
return (patches, image_index)
def load(self):
get_logger().info("Loading patches for " + self.file_name)
hfile = HDF5File(self.file_name, 'r')
patches = hfile[self.patch_name]
feature_dim = patches.shape[1]
indexes = self.indexes
num_patches = (indexes[:, 1] - indexes[:, 0]).sum()
self.patches = np.empty([num_patches, feature_dim])
self.new_index = np.empty([indexes.shape[0], 2])
patch_start = n_image = 0
for iid in indexes:
n_patches = iid[1] - iid[0]
self.patches[patch_start:patch_start +
n_patches, :] = patches[iid[0]:iid[1], :]
self.new_index[n_image] = [patch_start, patch_start + n_patches]
patch_start += n_patches
n_image += 1
hfile.close()
get_logger().info("Loaded " + str(num_patches) + " patches")
def getIndexes(patch_folder, nTrain, nTest, position_influence):
files = sorted(glob(join(patch_folder, '*.hdf5')), key=basename)
train = []
test = []
for (classNumber, filename) in enumerate(files):
#support_filename = join(".", basename(filename))
hfile = HDF5File(filename, 'r')
iid = hfile["image_index"][:]
nImages = iid.shape[0]
assert nImages >= (nTrain + nTest), "Not enough images!"
np.random.shuffle(iid)
trainIdx = iid[0:nTrain]
testIdx = iid[nTrain:nTrain + nTest]
trainData = ClassPatches(filename, trainIdx, PATCH_TYPE)
testData = ClassPatches(filename, testIdx, PATCH_TYPE)
test.append(testData)
train.append(
trainData) #train data is actually loaded only when needed
hfile.close()
Data = namedtuple("Data", "Train Test")
return Data(train, test)
def make_split(output_name, output_subdir, _patches, _positions, _image_indexes):
log = get_logger()
#import pdb; pdb.set_trace()
try:
makedirs(output_subdir)
except:
pass
output_filename = join(output_subdir, basename(output_name))
log.debug('Saving extracted descriptors to %s', output_filename)
hfile = HDF5File(output_filename, 'w', compression='gzip', fillvalue=0.0)
patches = hfile.create_dataset('patches', _patches.shape, dtype="float32", chunks=True)
positions = hfile.create_dataset('positions', _positions.shape, dtype="uint16", chunks=True)
image_index = hfile.create_dataset('image_index', _image_indexes.shape, dtype='uint64') # Start, End positions of an image
patches[:]= _patches
positions[:]=_positions
image_index[:]=_image_indexes
patches.attrs['cursor'] = 0
patches.attrs['feature_type'] = "CAFFE"
hfile.close()
def load_split_whole_image_only(input_folder, nTrain, nTest):
logger = get_logger()
files = sorted(glob(join(input_folder, '*.hdf5')), key=basename)
nClasses = len(files)
logger.info("Loading " + str(nClasses) + " classes")
train_patches = np.empty([nClasses * nTrain, patchOptions.size
]) # nClasses*nSamples x nFeatures
test_patches = np.empty([nClasses * nTest, patchOptions.size])
train_labels = np.empty([nClasses * nTrain])
test_labels = np.empty([nClasses * nTest])
start = time.clock()
train_patch_count = test_patch_count = 0
for (classNumber, filename) in enumerate(files):
hfile = HDF5File(filename, 'r')
iid = hfile["image_index"][:]
nImages = iid.shape[0]
assert nImages >= (nTrain + nTest), "Not enough images!"
np.random.shuffle(iid)
trainIdx = iid[0:nTrain]
testIdx = iid[nTrain:nTrain + nTest]
patches = hfile[patchOptions.patch_name]
for iid in trainIdx:
train_patches[train_patch_count] = patches[iid[0]]
train_patch_count += 1
train_labels[classNumber * nTrain:(classNumber + 1) *
nTrain] = classNumber * np.ones(nTrain)
for iid in testIdx:
test_patches[test_patch_count] = patches[iid[0]]
test_patch_count += 1
test_labels[classNumber * nTest:(classNumber + 1) *
nTest] = classNumber * np.ones(nTest)
logger.info("Patch count: " + str(train_patch_count) +
" training and " + str(test_patch_count) +
" test patches for class " + filename)
hfile.close()
end = time.clock()
logger.info("It took " + str((end - start)) + " seconds")
LoadedData = namedtuple(
"LoadedData", "train_patches train_labels test_patches test_labels")
return LoadedData(train_patches, train_labels, test_patches, test_labels)
def load_patches(class_data):
hfile = HDF5File(class_data.filename, 'r')
patches = hfile[patchOptions.patch_name][:]
positions = hfile[patchOptions.position_name][:]
feature_dim = patchOptions.patch_dim
indexes = class_data.index
num_patches=(indexes[:,1]-indexes[:,0]).sum()
loaded_patches = np.empty([num_patches, feature_dim])
loaded_positions = np.empty([num_patches, 2])
loaded_idx = np.empty([class_data.index.shape[0], 2])
tags = np.zeros([num_patches,1])
patch_start = n_image = 0
#import pdb; pdb.set_trace()
for n, iid in enumerate(indexes):
n_patches = iid[1]-iid[0]
loaded_patches[patch_start:patch_start+n_patches,:] = patches[iid[0]:iid[1],:]
loaded_positions[patch_start:patch_start+n_patches,:] = positions[iid[0]:iid[1],:]
loaded_idx[n,:] = np.array([patch_start, patch_start+n_patches])
tags[patch_start] = 1
patch_start += n_patches
n_image += 1
hfile.close()
return ClassData(loaded_patches, loaded_positions, tags, num_patches, loaded_idx)