lose 0.4.3

A helper package for hdf5 data handling

lose

lose, but in particular lose.LOSE(), is a helper class for handling data using hdf5 file format and PyTables

>>> from lose import LOSE
>>> l = LOSE()
>>> l
<hdf5 data handler, fname=None, fmode='r', atom=Float32Atom(shape=(), dflt=0.0)>

installation

pip3 install -U lose

or

pip install -U lose

structure

vars

LOSE.fname is the path to to the .h5 file including the name and extension, default is None.

LOSE.fmode is the mode .h5 file from LOSE.fname will be opened with, 'r' for read(default), 'w' for write, 'a' for append.

LOSE.atom recommended to be left at default, is the dtype for the data to be stored in, default is tables.Float32Atom() which results to arrays with dtype==np.float32.

LOSE.batch_obj default is '[:]', recommended to be left default, specifies the amount of data to be loaded by LOSE.load(), works like python list slicing, must be a string, default loads everything.

LOSE.generator() related vars:

LOSE.batch_size batch size of data getting pulled from the .h5 file, default is 1.

LOSE.limit limits the amount of data loaded by the generator, default is None, if None all available data will be loaded.

LOSE.loopforever bool that allows infinite looping over the data, default is False.

LOSE.iterItems list of X group names and list of Y group names, default is None, required to be user defined for LOSE.generator() to work.

LOSE.iterOutput list of X output names and list of Y output names, default is None, required to be user defined for LOSE.generator() to work.

LOSE.shuffle bool that enables shuffling of the data, default is False, shuffling is affected by LOSE.limit and LOSE.batch_size.

methods

Help on class LOSE in module lose.dataHandler:

class LOSE(builtins.object)
 |  Methods defined here:
 |  
 |  __init__(self)
 |      Initialize self.  See help(type(self)) for accurate signature.
 |  
 |  __repr__(self)
 |      Return repr(self).
 |  
 |  generator(self)
 |  
 |  get_shape(self, arrName)
 |  
 |  load(self, *args)
 |  
 |  make_generator(self, layerNames, limit=None, batch_size=1, shuffle=False, **kwards)
 |  
 |  newGroup(self, **kwards)
 |  
 |  save(self, **kwards)
 |  
 |  ----------------------------------------------------------------------
 |  Data descriptors defined here:
 |  
 |  __dict__
 |      dictionary for instance variables (if defined)
 |  
 |  __weakref__
 |      list of weak references to the object (if defined)

LOSE.newGroup(**groupNames) is used to add/set(depends on the file mode) group(expandable array) names and shapes in the .h5 file.

LOSE.save(**groupNamesAndSahpes) is used to save data in write/append mode(depends on the file mode) into a group into a .h5 file, the data needs to have the same shape as group.shape[1:] the data was passed to, LOSE.get_shape(groupName) can be used to get the group.shape of a group.

LOSE.load(*groupNames) is used to load data(hole group or a slice, to load a slice change LOSE.batch_obj to a string with the desired slice, default is "[:]") from a group, group has to be present in the .h5 file.

LOSE.get_shape(groupName) is used to get the shape of a single group, group has to be present in the .h5 file.

LOSE.generator() check LOSE.generator() details section, LOSE.iterItems and LOSE.iterOutput have to be defined.

LOSE.make_generator(layerNames, limit=None, batch_size=1, shuffle=False, **kwards) again check LOSE.generator() details more details.

example usage

creating/adding new groups to a file in append/write mode

import numpy as np
from lose import LOSE

l = LOSE()
l.fname = 'path/to/you/save/file.h5' # path to the .h5 file, has to be user defined before any methods can be used, default is None
l.fmode = 'w' # 'w' for write mode, 'a' for append mode, default is 'r'

exampleDataX = np.arange(20, dtype=np.float32)
exampleDataY = np.arange(3, dtype=np.float32)

l.newGroup(x=(0, *exampleDataX.shape), y=(0, *exampleDataY.shape)) # creating new groups(ready for data saved to) in a file, if fmode is 'w' all groups in the file will be overwritten

saving data into a group in append/write mode

import numpy as np
from lose import LOSE

l = LOSE()
l.fname = 'path/to/you/save/file.h5' # path to the .h5 file, has to be user defined before any methods can be used, default is None
l.fmode = 'a' # 'w' for write mode, 'a' for append mode, default is 'r', 'a' mode append data to the file, 'w' mode overwrites data for the group in the file

exampleDataX = np.arange(20, dtype=np.float32)
exampleDataY = np.arange(3, dtype=np.float32)

l.save(x=[exampleDataX, exampleDataX], y=[exampleDataY, exampleDataY]) # saving data into groups defined in the previous example, in append mode
l.save(y=[exampleDataY], x=[exampleDataX]) # the same thing

loading data from a group within a file

import numpy as np
from lose import LOSE

l = LOSE()
l.fname = 'path/to/you/save/file.h5' # path to the .h5 file, has to be user defined before any methods can be used, default is None

x, y = l.load('x', 'y') # loading data from the .h5 file(has to be a real file) populated by previous examples
y2compare, x2compare = l.load('y', 'x') # the same thing

print (np.all(x == x2compare), np.all(y == y2compare)) # True True

getting the shape of a group

import numpy as np
from lose import LOSE

l = LOSE()
l.fname = 'path/to/you/save/file.h5' # path to the .h5 file(populated by previous examples), has to be user defined before any methods can be used, default is None

print (l.get_shape('x')) # (3, 20)
print (l.get_shape('y')) # (3, 3)

LOSE.generator() details

LOSE.generator() is a python generator used to access data from a hdf5 file in LOSE.batch_size pieces without loading the hole file/group into memory, also works with tf.keras.model.fit_generator(), have to be used with a with context statement(see examples below).

LOSE.iterItems and LOSE.iterOutput have to be defined by user first.

LOSE.make_generator(layerNames, limit=None, batch_size=1, shuffle=False, **kwards) has the same rules as LOSE.generator(). however the data needs to be passed to it each time it's initialized, data is only stored temporarily, the parameters are passed to it on initialization, layerNames acts like LOSE.iterOutput but every name in it has to match to names of the data passed(see examples below), if file temp.h5 exists it will be overwritten and then deleted.

example LOSE.generator() usage

for this example lets say that file has requested data in it and the model input/output layer names are present.

import numpy as np
from lose import LOSE

l = LOSE()
l.fname = 'path/to/you/save/file.h5' # path to data

l.iterItems = [['x1', 'x2'], ['y']] # names of X and Y groups, all group names need to have batch dim the same and be present in the .h5 file
l.iterOutput = [['input_1', 'input_2'], ['dense_5']] # names of model's layers the data will be cast on, group.shape[1:] needs to match the layer's input shape
l.loopforever = True
l.batch_size = 20 # some batch size, can be bigger then the dataset, but won't output more data, it will just loop over or stop the iteration if LOSE.loopforever is False

l.limit = 10000 # lets say that the file has more data, but you only want to train on first 10000 samples

l.shuffle = True # enable data shuffling for the generator, costs memory and time

with l.generator() as gen:
	some_model.fit_generator(gen(), steps_per_epoch=50, epochs=1000, shuffle=False) # model.fit_generator() still can't shuffle the data, but LOSE.generator() can

example LOSE.make_generator(layerNames, limit=None, batch_size=1, shuffle=False, **kwards) usage

for this example lets say the model's input/output layer names are present and shapes match with the data.

import numpy as np
from lose import LOSE

l = LOSE()

num_samples = 1000

x1 = np.zeros((num_samples, 200)) #data for the model
x2 = np.zeros((num_samples, 150)) #data for the model
y = np.zeros((num_samples, 800)) #data for the model

with l.make_generator([['input_1', 'input_2'], ['dense_5']], shuffle=True, input_2=x2, input_1=x1, dense_5=y) as gen:
	del x1 #remove from memory
	del x2 #remove from memory
	del y #remove from memory

	some_model.fit_generator(gen(), steps_per_epoch=100, epochs=10000, shuffle=False) # again data can't be shuffled by model.fit_generator(), shuffling should be done by the generator

bugs/problems/issues

report them.