Develop Reference

h5py file subset taking more space than parent file?

I have an existing h5py file that I downloaded which is ~18G in size. It has a number of nested datasets within it:
h5f = h5py.File('input.h5', 'r')
data = h5f['data']
latlong_data = data['lat_long'].value
I want to be able to some basic min/max scaling of the numerical data within latlong, so i want to put it in its own h5py file for easier use and lower memory usage.
However, when i try to write it out to its own file:
out = h5py.File('latlong_only.h5', 'w')
out.create_dataset('latlong', data=latlong)
out.close()
The output file is incredibly large. It's still not done writing to disk and is ~85GB in space. Why is the data being written to the new file not compressed?

Could be h5f['data/lat_long'] is using compression filters (and you aren't). To check the original dataset's compression settings, use this line:
print (h5f['data/latlong'].compression, h5f['data/latlong'].compression_opts)
After writing my answer, it occurred to me that you don't need to copy the data to another file to reduce the memory footprint. Your code reads the dataset into an array, which is not necessary in most use cases. A h5py dataset object behaves similar to a NumPy array. Instead, use this line: ds = h5f1['data/latlong'] to create a dataset object (instead of an array) and use it "like" it's a NumPy array. FYI, .value is a deprecated method to return the dataset as an array. Use this syntax instead arr = h5f1['data/latlong'][()]. Loading the dataset into an array also requires more memory than using an h5py object (which could be an issue with large datasets).
There are other ways to access the data. My suggestion to use dataset objects is 1 way. Your method (extracting data to a new file) is another way. I am not found of that approach because you now have 2 copies of the data; a bookkeeping nightmare. Another alternative is to create external links from the new file to the existing 18GB file. That way you have a small file that links to the big file (and no duplicate data). I describe that method in this post: [How can I combine multiple .h5 file?][1] Method 1: Create External Links.
If you still want to copy the data, here is what I would do. Your code reads the dataset into an array then writes the array to the new file (uncompressed). Instead, copy the dataset using h5py's group .copy() method, it will retain compression settings and attributes.
See below:
with h5py.File('input.h5', 'r') as h5f1, \
h5py.File('latlong_only.h5', 'w') as h5f2:
h5f1.copy(h5f1['data/latlong'], h5f2,'latlong')

Set up data pipeline for video processing

I am newbie to Tensorflow so I would appreciate any constructive help.
I am trying to build a feature extraction and data pipeline with Tensorflow for video processing where multiple folders holding video files with multiple classes (JHMDB database), but kind of stuck.
I have the feature extracted to one folder, at the moment to separate *.npz compressed arrays, in the filename I have stored the class name as well.
First Attempt
First I thought I would use this code from the TF tutorials site, simply reading files from folder method:
jhmdb_path = Path('...')
# Process files in folder
list_ds = tf.data.Dataset.list_files(str(jhmdb_path/'*.npz'))
for f in list_ds.take(5):
print(f.numpy())
def process_path(file_path):
labels = tf.strings.split(file_path, '_')[-1]
features = np.load(file_path)
return features, labels
labeled_ds = list_ds.map(process_path)
for a, b in labeled_ds.take(5):
print(a, b)
TypeError: expected str, bytes or os.PathLike object, not Tensor
..but this not working.
Second Attempt
Then I thought ok I will use generators:
# using generator
jhmdb_path = Path('...')
def generator():
for item in jhmdb_path.glob("*.npz"):
features = np.load(item)
print(item.files)
print(f['PAFs'].shape)
features = features['PAFs']
yield features
dataset = tf.data.Dataset.from_generator(generator, (tf.uint8))
iter(next(dataset))
TypeError: 'FlatMapDataset' object is not an iterator...not working.
In the first case, somehow the path is a byte type, and I could not change it to str to be able to load it with np.load(). (If I point a file directly on np.load(direct_path), then strange, but it works.)
At second case... I am not sure what is wrong.
I looked for hours to find a solution how to build an iterable dataset from list of relatively big and large numbers of 'npz' or 'npy' files, but seems to be this is not mentioned anywhere (or just way too trivial maybe).
Also, as I could not test the model so far, I am not sure if this is the good way to go. I.e. to feed the model with hundreds of files in this way, or just build one huge 3.5 GB npz (that would sill fit in memory) and use that instead. Or use TFrecords, but that looks more complicated, than the usual examples.
What is really annoying here, that TF tutorials and in general all are about how to load a ready dataset directly, or how to load np array(s) or how to load, image, text, dataframe objects, but no way to find any real examples how to process big chunks of data files, e.g. the case of extracting features from audio or video files.
So any suggestions or solutions would be highly appreciated and I would be really, really grateful to have something finally working! :)

Most efficient way to use a large data set for PyTorch?

Perhaps this question has been asked before, but I'm having trouble finding relevant info for my situation.
I'm using PyTorch to create a CNN for regression with image data. I don't have a formal, academic programming background, so many of my approaches are ad-hoc and just terribly inefficient. May times I can go back through my code and clean things up later because the inefficiency is not so drastic that performance is significantly affected. However, in this case, my method for using the image data takes a long time, uses a lot of memory, and it is done every time I want to test a change in the model.
What I've done is essentially loaded the image data into numpy arrays, saved those arrays in an .npy file, and then when I want to use said data for the model I import all of the data in that file. I don't think the data set is really THAT large, as it is comprised of 5000, 3 color channel images of size 64x64. Yet my memory usage shoots up to 70%-80% (out of 16gb) when it is being loaded, and it takes 20-30 seconds to load in every time.
My guess is that I'm being dumb about the way I'm loading it in, but frankly I'm not sure what the standard is. Should I, in some way, put the image data somewhere before I need it, or should the data be loaded directly from the image files? And in either case, what is the best, most efficient way to do that, independent of file structure?
I would really appreciate any help on this.

For speed I would advise to used HDF5 or LMDB:
Reasons to use LMDB:
LMDB uses memory-mapped files, giving much better I/O performance.
Works well with really large datasets. The HDF5 files are always read
entirely into memory, so you can’t have any HDF5 file exceed your
memory capacity. You can easily split your data into several HDF5
files though (just put several paths to h5 files in your text file).
Then again, compared to LMDB’s page caching the I/O performance won’t
be nearly as good.
[http://deepdish.io/2015/04/28/creating-lmdb-in-python/]
If you decide to used LMDB:
ml-pyxis is a tool for creating and reading deep learning datasets using LMDBs.*(I am co author of this tool)
It allows to create binary blobs (LMDB) and they can be read quite fast. The link above comes with some simple examples on how to create and read the data. Including python generators/ iteratos .
This notebook has an example on how to create a dataset and read it paralley while using pytorch.
If you decide to use HDF5:
PyTables is a package for managing hierarchical datasets and designed to efficiently and easily cope with extremely large amounts of data.
https://www.pytables.org/

Here is a concrete example to demonstrate what I meant. This assumes that you've already dumped the images into an hdf5 file (train_images.hdf5) using h5py.
import h5py
hf = h5py.File('train_images.hdf5', 'r')
group_key = list(hf.keys())[0]
ds = hf[group_key]
# load only one example
x = ds[0]
# load a subset, slice (n examples)
arr = ds[:n]
# should load the whole dataset into memory.
# this should be avoided
arr = ds[:]
In simple terms, ds can now be used as an iterator which gives images on the fly (i.e. it doesn't load anything in memory). This should make the whole run time blazing fast.
for idx, img in enumerate(ds):
# do something with `img`

In addition to the above answers, the following may be useful due to some recent advances (2020) in the Pytorch world.
Your question: Should I, in some way, put the image data somewhere before I need it, or should the data be loaded directly from the image files? And in either case, what is the best, most efficient way to do that, independent of file structure?
You can leave the image files in their original format (.jpg, .png, etc.) on your local disk or on the cloud storage, but with one added step - compress the directory as a tar file. Please read this for more details:
Pytorch Blog (Aug 2020): Efficient PyTorch I/O library for Large Datasets, Many Files, Many GPUs (https://pytorch.org/blog/efficient-pytorch-io-library-for-large-datasets-many-files-many-gpus/)
This package is designed for situations where the data files are too large to fit in memory for training. Therefore, you give the URL of the dataset location (local, cloud, ..) and it will bring in the data in batches and in parallel.
The only (current) requirement is that the dataset must be in a tar file format.
The tar file can be on the local disk or on the cloud. With this, you don't have to load the entire dataset into the memory every time. You can use the torch.utils.data.DataLoader to load in batches for stochastic gradient descent.

No need saving image into npy and loading all into memory. Just load a batch of image path and transform then into tensor.
The following code define the MassiveDataset, and pass it into DataLoader, everything goes well.
from torch.utils.data.dataset import Dataset
from typing import Optional, Callable
import os
import multiprocessing
def apply_transform(transform: Callable, data):
try:
if isinstance(data, (list, tuple)):
return [transform(item) for item in data]
return transform(data)
except Exception as e:
raise RuntimeError(f'applying transform {transform}: {e}')
class MassiveDataset(Dataset):
def __init__(self, filename, transform: Optional[Callable] = None):
self.offset = []
self.n_data = 0
if not os.path.exists(filename):
raise ValueError(f'filename does not exist: {filename}')
with open(filename, 'rb') as fp:
self.offset = [0]
while fp.readline():
self.offset.append(fp.tell())
self.offset = self.offset[:-1]
self.n_data = len(self.offset)
self.filename = filename
self.fd = open(filename, 'rb', buffering=0)
self.lock = multiprocessing.Lock()
self.transform = transform
def __len__(self):
return self.n_data
def __getitem__(self, index: int):
if index < 0:
index = self.n_data + index
with self.lock:
self.fd.seek(self.offset[index])
line = self.fd.readline()
data = line.decode('utf-8').strip('\n')
return apply_transform(self.transform, data) if self.transform is not None else data
NB: open file with buffering=0 and multiprocessing.Lock() are used to avoid loading bad data (usually a bit from one part of the file and a bit from the another part of the file).
additionally, if using multiprocessing in DataLoader, one could get such exception TypeError: cannot serialize '_io.BufferedReader' object. This is caused by pickle module used in multiprocessing, it cannot serialize _io.BufferedReader, but dill can. Replacing multiprocessing with multiprocess, things goes okay (major changes compare with multiprocessing, enhanced serialization is done with dill)
same thing was discussed in this issue

how to save torchtext Dataset?

I'm working with text and use torchtext.data.Dataset.
Creating the dataset takes a considerable amount of time.
For just running the program this is still acceptable. But I would like to debug the torch code for the neural network. And if python is started in debug mode, the dataset creation takes roughly 20 minutes (!!). That's just to get a working environment where I can debug-step through the neural network code.
I would like to save the Dataset, for example with pickle. This sample code is taken from here, but I removed everything that is not necessary for this example:
from torchtext import data
from fastai.nlp import *
PATH = 'data/aclImdb/'
TRN_PATH = 'train/all/'
VAL_PATH = 'test/all/'
TRN = f'{PATH}{TRN_PATH}'
VAL = f'{PATH}{VAL_PATH}'
TEXT = data.Field(lower=True, tokenize="spacy")
bs = 64;
bptt = 70
FILES = dict(train=TRN_PATH, validation=VAL_PATH, test=VAL_PATH)
md = LanguageModelData.from_text_files(PATH, TEXT, **FILES, bs=bs, bptt=bptt, min_freq=10)
with open("md.pkl", "wb") as file:
pickle.dump(md, file)
To run the code, you need the aclImdb dataset, it can be downloaded from here. Extract it into a data/ folder next to this code snippet. The code produces an error in the last line, where pickle is used:
Traceback (most recent call last):
File "/home/lhk/programming/fastai_sandbox/lesson4-imdb2.py", line 27, in <module>
pickle.dump(md, file)
TypeError: 'generator' object is not callable
The samples from fastai often use dill instead of pickle. But that doesn't work for me either.

I came up with the following functions for myself:
import dill
from pathlib import Path
import torch
from torchtext.data import Dataset
def save_dataset(dataset, path):
if not isinstance(path, Path):
path = Path(path)
path.mkdir(parents=True, exist_ok=True)
torch.save(dataset.examples, path/"examples.pkl", pickle_module=dill)
torch.save(dataset.fields, path/"fields.pkl", pickle_module=dill)
def load_dataset(path):
if not isinstance(path, Path):
path = Path(path)
examples = torch.load(path/"examples.pkl", pickle_module=dill)
fields = torch.load(path/"fields.pkl", pickle_module=dill)
return Dataset(examples, fields)
Not that actual objects could be a bit different, for example, if you save TabularDataset, then load_dataset returns an instance of class Dataset. This unlikely affect the data pipeline but may require extra diligence for tests.
In the case of a custom tokenizer, it should be serializable as well (e.g. no lambda functions, etc).

You can use dill instead of pickle. It works for me.
You can save a torchtext Field like
TEXT = data.Field(sequential=True, tokenize=tokenizer, lower=True,fix_length=200,batch_first=True)
with open("model/TEXT.Field","wb")as f:
dill.dump(TEXT,f)
And load a Field like
with open("model/TEXT.Field","rb")as f:
TEXT=dill.load(f)
The offical code suppport is under development，you can follow https://github.com/pytorch/text/issues/451 and https://github.com/pytorch/text/issues/73 .

You can always use the pickle to dump the objects, but keep in mind one thing that dumping a list of dictionary or fields objects are not taken care of by the module, so to the best try to decompose the list first
To Store the DataSet Object to a pickle file for later easy loading
def save_to_pickle(dataSetObject,PATH):
with open(PATH,'wb') as output:
for i in dataSetObject:
pickle.dump(vars(i), output, pickle.HIGHEST_PROTOCOL)
The toughest thing is yet to come, Yeah loading the pickle file.... ;)
First, try to look for all field names and field attributes and then go for the kill
To load the pickle file into the DataSetObject
def load_pickle(PATH, FIELDNAMES, FIELD):
dataList = []
with open(PATH, "rb") as input_file:
while True:
try:
# Taking the dictionary instance as the input Instance
inputInstance = pickle.load(input_file)
# plugging it into the list
dataInstance = [inputInstance[FIELDNAMES[0]],inputInstance[FIELDNAMES[1]]]
# Finally creating an example objects list
dataList.append(Example().fromlist(dataInstance,fields=FIELD))
except EOFError:
break
# At last creating a data Set Object
exampleListObject = Dataset(dataList, fields=data_fields)
return exampleListObject
This hackish solution has worked in my case, hope you will find it useful in your case too.
Btw any suggestion is welcome :).

The pickle/dill approach is fine if your dataset is small. But if you are working with large datasets I won't recommend it as it will be too slow.
I simply save the examples (iteratively) as JSON-strings. The reason behind this is because saving the whole Dataset object takes a lot of time, plus you need serialization tricks such a dill, which makes the serialization even slower.
Moreover, these serializers take a lot of memory (some of them even create copies of the dataset) and if they start making use of the swap memory, you're done. That process is gonna take so long that you will probably terminate it before it finishes.
Therefore, I end up with the following approach:
Iterate over the examples
Convert each example (on the fly) to a JSON-string
Write that JSON-string into a text file (one sample per
line)
When loading, add the examples to the Dataset object along with the fields
def save_examples(dataset, savepath):
with open(savepath, 'w') as f:
# Save num. elements (not really need it)
f.write(json.dumps(total)) # Write examples length
f.write("\n")
# Save elements
for pair in dataset.examples:
data = [pair.src, pair.trg]
f.write(json.dumps(data)) # Write samples
f.write("\n")
def load_examples(filename):
examples = []
with open(filename, 'r') as f:
# Read num. elements (not really need it)
total = json.loads(f.readline())
# Save elements
for i in range(total):
line = f.readline()
example = json.loads(line)
# example = data.Example().fromlist(example, fields) # Create Example obj. (you can do it here or later)
examples.append(example)
end = time.time()
print(end - start)
return examples
Then, you can simply rebuild the dataset by:
# Define fields
SRC = data.Field(...)
TRG = data.Field(...)
fields = [('src', SRC), ('trg', TRG)]
# Load examples from JSON and convert them to "Example objects"
examples = load_examples(filename)
examples = [data.Example().fromlist(d, fields) for d in examples]
# Build dataset
mydataset = Dataset(examples, fields)
The reason why I use JSON instead of pickle, dill, msgpack, etc is not arbitrary.
I did some tests and these are the results:
Dataset size: 2x (1,960,641)
Saving times:
- Pickle/Dill*: >30-45 min (...or froze my computer)
- MessagePack (iterative): 123.44 sec
100%|██████████| 1960641/1960641 [02:03<00:00, 15906.52it/s]
- JSON (iterative): 16.33 sec
100%|██████████| 1960641/1960641 [00:15<00:00, 125955.90it/s]
- JSON (bulk): 46.54 sec (memory problems)
Loading times:
- Pickle/Dill*: -
- MessagePack (iterative): 143.79 sec
100%|██████████| 1960641/1960641 [02:23<00:00, 13635.20it/s]
- JSON (iterative): 33.83 sec
100%|██████████| 1960641/1960641 [00:33<00:00, 57956.28it/s]
- JSON (bulk): 27.43 sec
*Similar approach as the other answers

How to bulk write TFRecords?

I have a CSV with approximately 40 million rows. Each row is a training instance. As per the documentation on consuming TFRecords I am trying to encode and save the data in a TFRecord file.
All the examples I have found (even the ones in the TensorFlow repo) show the process of creating a TFRecord is dependant on the class TFRecordWriter. This class has a method write that takes as input a serialised string representation of the data and writes it to disk. However, this appears to be done one training instance at a time.
How do I write a batch of the serialised data?
Let's say I have a funtion:
def write_row(sentiment, text, encoded):
feature = {"one_hot": _float_feature(encoded),
"label": _int64_feature([sentiment]),
"text": _bytes_feature([text.encode()])}
example = tf.train.Example(features=tf.train.Features(feature=feature))
writer.write(example.SerializeToString())
Writing to disk 40 million times (once for each example) is going to be incredibly slow. It would be far more efficient to batch this data and write 50k or 100k examples at a time (as far as the machine's resources will allow). However there does not appear to be any method to do this inside TFRecordWriter.
Something along the lines of:
class MyRecordWriter:
def __init__(self, writer):
self.records = []
self.counter = 0
self.writer = writer
def write_row_batched(self, sentiment, text, encoded):
feature = {"one_hot": _float_feature(encoded),
"label": _int64_feature([sentiment]),
"text": _bytes_feature([text.encode()])}
example = tf.train.Example(features=tf.train.Features(feature=feature))
self.records.append(example.SerializeToString())
self.counter += 1
if self.counter >= 10000:
self.writer.write(os.linesep.join(self.records))
self.counter = 0
self.records = []
But when reading the file created by this method I get the following error:
tensorflow/core/framework/op_kernel.cc:1192] Invalid argument: Could not parse example input, value: '
��
label
��
one_hot����
��
Note: I could change the encoding process so that each example proto contains several thousand examples instead of just one but
I don't want to pre-batch the data when writing to the TFrecord file in this way as it will introduce extra overhead in my training pipeline when I want to use the file for training with different batch sizes.

TFRecords is a binary format. With the following line you are treating it like a text file: self.writer.write(os.linesep.join(self.records))
That is because you are using the operation system depending linesep (either \n or \r\n).
Solution:
Just write the records. You are asking to batch write them. You can use a buffered writer. For 40 million rows you might also want to consider splitting the data up into separate files to allow better parallelisation.
When using TFRecordWriter: The file is already buffered.
Evidence for that is found in the source:
tf_record.py calls pywrap_tensorflow.PyRecordWriter_New
PyRecordWriter calls Env::Default()->NewWritableFile
Env->NewWritableFile calls NewWritableFile on the matching FileSystem
e.g. PosixFileSystem calls fopen
fopen returns a stream which "is fully buffered by default if it is known to not refer to an interactive device"
That will be file system dependent but WritableFile notes "The implementation must provide buffering since callers may append small fragments at a time to the file."