FFRecord
The FFRecord format is a simple format for storing a sequence of binary records developed by HFAiLab, which supports random access and Linux Asynchronous Input/Output (AIO) read.
File Format
Storage Layout:
+-----------------------------------+---------------------------------------+
| checksum | N |
+-----------------------------------+---------------------------------------+
| checksums | offsets |
+---------------------+---------------------+--------+----------------------+
| sample 1 | sample 2 | .... | sample N |
+---------------------+---------------------+--------+----------------------+
Fields:
| field | size (bytes) | description |
|---|---|---|
| checksum | 4 | CRC32 checksum of metadata |
| N | 8 | number of samples |
| checksums | 4 * N | CRC32 checksum of each sample |
| offsets | 8 * N | byte offset of each sample |
| sample i | offsets[i + 1] – offsets[i] | data of the i-th sample |
Get Started
Requirements
- OS: Linux
- Python >= 3.6
- Pytorch >= 1.6
- libaio 0.8.3
- NumPy
Install
pip3 install ffrecord
Usage
We provide ffrecord.FileWriter and ffrecord.FileReader for reading and writing, respectively.
Write
To create a FileWriter object, you need to specify a file name and the total number of samples. And then you could call FileWriter.write_one() to write a sample to the FFRecord file. It accepts bytes or bytearray as input and appends the data to the end of the opened file.
from ffrecord import FileWriter
def serialize(sample):
""" Serialize a sample to bytes or bytearray
You could use anything you like to serialize the sample.
Here we simply use pickle.dumps().
"""
return pickle.dumps(sample)
samples = [i for i in range(100)] # anything you would like to store
fname = 'test.ffr'
n = len(samples) # number of samples to be written
writer = FileWriter(fname, n)
for i in range(n):
data = serialize(samples[i]) # data should be bytes or bytearray
writer.write_one(data)
writer.close()
Read
To create a FileReader object, you only need to specify the file name. And then you could call FileWriter.read() to read multiple samples from the FFReocrd file. It accepts a list of indices as input and outputs the corresponding samples data.
The reader would validate the checksum before returning the data if check_data = True.
from ffrecord import FileReader
def deserialize(data):
""" deserialize bytes data
The deserialize method should be paired with the serialize method above.
"""
return pickle.loads(data)
fname = 'test.ffr'
reader = FileReader(fname, check_data=True)
print(f'Number of samples: {reader.n}')
indices = [3, 6, 0, 10] # indices of each sample
data = reader.read(indices) # return a list of bytes data
for i in range(n):
sample = deserialize(data[i])
# do what you want
reader.close()
Dataset and DataLoader for PyTorch
We also provide ffrecord.torch.Dataset and ffrecord.torch.DataLoader for PyTorch users to train models using FFRecord.
Different from torch.utils.data.Dataset which accepts an index as input and returns one sample, ffrecord.torch.Dataset accepts a batch of indices as input and returns a batch of samples. One advantage of ffrecord.torch.Dataset is that it could read a batch of data at a time using Linux AIO.
We first read a batch of bytes data from the FFReocrd file and then pass the bytes data to process() function. Users need to inherit from ffrecord.torch.Dataset and define their custom process() function.
reader.read(indices) process()
“>
Pipline: indices ----------------------------> bytes -------------> samples
reader.read(indices) process()
For example:
class CustomDataset(ffrecord.torch.Dataset):
def __init__(self, fname, check_data=True, transform=None):
super().__init__(fname, check_data)
self.transform = transform
def process(self, indices, data):
# deserialize data
samples = [pickle.loads(b) for b in data]
# transform data
if self.transform:
samples = [self.transform(s) for s in samples]
return samples
dataset = CustomDataset('train.ffr')
indices = [3, 4, 1, 0]
samples = dataset[indices]
ffrecord.torch.Dataset could be combined with ffrecord.torch.DataLoader just like PyTorch.
dataset = CustomDataset('train.ffr')
loader = ffrecord.torch.DataLoader(dataset,
batch_size=16,
shuffle=True,
num_workers=8)
for i, batch in enumerate(loader):
# training model
GitHub
HFAiLab
