mxnet.recordio
MXRecordIO
Reads/writes RecordIO data format, supporting sequential read and write.
record = mx.recordio.MXRecordIO('tmp.rec', 'w')
for i in range(5):
record.write('record_%d'%i)
record.close()
record = mx.recordio.MXRecordIO('tmp.rec', 'r')
for i in range(5):
item = record.read()
print(item)
record_0
record_1
record_2
record_3
record_4
record.close()
MXIndexedRecordIO
Reads/writes RecordIO data format, supporting random access.
record = mx.recordio.MXIndexedRecordIO('tmp.idx', 'tmp.rec', 'w')
for i in range(5):
record.write_idx(i, 'record_%d'%i)
record.close()
record = mx.recordio.MXIndexedRecordIO('tmp.idx', 'tmp.rec', 'r')
record.read_idx(3)
record_3
IRHeader
An alias for HEADER. Used to store metadata (e.g. labels) accompanying a record.
Parameters:
- flag (int) – Available for convenience, can be set arbitrarily.
- label (float or an array of float) – Typically used to store label(s) for a record.
- id (int) – Usually a unique id representing record.
- id2 (int) – Higher order bits of the unique id, should be set to 0 (in most cases).
pack(header, s)
Pack a string into MXImageRecord.
label = 4 # label can also be a 1-D array, for example: label = [1,2,3]
id = 2574
header = mx.recordio.IRHeader(0, label, id, 0)
with open(path, 'r') as file:
s = file.read()
packed_s = mx.recordio.pack(header, s)
unpack(s)
Unpack a MXImageRecord to string.
record = mx.recordio.MXRecordIO('test.rec', 'r')
item = record.read()
header, s = mx.recordio.unpack(item)
header
HEADER(flag=0, label=14.0, id=20129312, id2=0)
unpack_img(s, iscolor=-1)
record = mx.recordio.MXRecordIO('test.rec', 'r')
item = record.read()
header, img = mx.recordio.unpack_img(item)
header
HEADER(flag=0, label=14.0, id=20129312, id2=0)
img
array([[[ 23, 27, 45],
[ 28, 32, 50],
...,
[168, 169, 167],
[166, 167, 165]]], dtype=uint8)
pack_img(header, img, quality=95, img_fmt='.jpg')[source]
Pack an image into MXImageRecord.
label = 4 # label can also be a 1-D array, for example: label = [1,2,3]
id = 2574
header = mx.recordio.IRHeader(0, label, id, 0)
img = cv2.imread('test.jpg')
packed_s = mx.recordio.pack_img(header, img)
we use the Gluon API to define a Dataset and use a DataLoader to iterate through the dataset in mini-batches.
Introduction to Datasets
Dataset objects are used to represent collections of data, and include methods to load and parse the data.
we’ll use the ArrayDataset to introduce the idea of a Dataset.
import mxnet as mx
import os
import tarfile
mx.random.seed(42) # Fix the seed for reproducibility
X = mx.random.uniform(shape=(10, 3))
y = mx.random.uniform(shape=(10, 1))
dataset = mx.gluon.data.dataset.ArrayDataset(X, y)
A key feature of a Dataset is the ability to retrieve a single sample given an index.
Our random data and labels were generated in memory, so this ArrayDataset doesn’t have to load anything from disk, but the interface is the same for all Datasets.
sample_idx = 4
sample = dataset[sample_idx]
assert len(sample) == 2
assert sample[0].shape == (3, )
assert sample[1].shape == (1, )
We don’t usually retrieve individual samples from Dataset objects though (unless we’re quality checking the output samples). Instead we use a DataLoader.
Introduction to DataLoader
A DataLoader is used to create mini-batches of samples from a Dataset, and provides a convenient iterator interface for looping these batches.
A required parameter of DataLoader is the size of the mini-batches you want to create, called batch_size.
Another benefit of using DataLoader is the ability to easily load data in parallel using multiprocessing. You can set the num_workers parameter to the number of CPUs avalaible on your machine for maximum performance.
from multiprocessing import cpu_count
CPU_COUNT = cpu_count()
data_loader = mx.gluon.data.DataLoader(dataset, batch_size=5, num_workers=CPU_COUNT)
for X_batch, y_batch in data_loader:
print("X_batch has shape {}, and y_batch has shape {}".format(X_batch.shape, y_batch.shape))
Our data_loader loop will stop when every sample of dataset has been returned as part of a batch.
Sometimes the dataset length isn’t divisible by the mini-batch size, leaving a final batch with a smaller number of samples. DataLoader‘s default behavior is to return this smaller mini-batch, but this can be changed by setting the last_batch parameter to discard (which ignores the last batch) or rollover (which starts the next epoch with the remaining samples).
Machine learning with Datasets and DataLoaders
Common use cases for loading data are covered already (e.g. mxnet.gluon.data.vision.datasets.ImageFolderDataset), but it’s simple to create your own custom Dataset classes for other types of data.
You can even use included Dataset objects for common datasets if you want to experiment quickly.
Many of the image Datasets accept a function (via the optional transform parameter) which is applied to each sample returned by the Dataset. It’s useful for performing data augmentation, but can also be used for more simple data type conversion and pixel value scaling as seen below.
def transform(data, label):
data = data.astype('float32')/255
return data, label
train_dataset = mx.gluon.data.vision.datasets.FashionMNIST(train=True, transform=transform)
valid_dataset = mx.gluon.data.vision.datasets.FashionMNIST(train=False, transform=transform)
sample_idx = 234
sample = train_dataset[sample_idx]
data = sample[0]
label = sample[1]
When training machine learning models it is important to shuffle the training samples every time you pass through the dataset (i.e. each epoch). Sometimes the order of your samples will have a spurious relationship with the target variable, and shuffling the samples helps remove this. With DataLoader it’s as simple as adding shuffle=True. You don’t need to shuffle the validation and testing data though.
If you have more complex shuffling requirements (e.g. when handling sequential data), take a look at mxnet.gluon.data.BatchSampler and pass this to your DataLoader instead.
batch_size = 32
train_data_loader = mx.gluon.data.DataLoader(train_dataset, batch_size, shuffle=True, num_workers=CPU_COUNT)
valid_data_loader = mx.gluon.data.DataLoader(valid_dataset, batch_size, num_workers=CPU_COUNT)