An efficient way to store/load and manage dataset, model and optimizer for deep learning with matorage!
Matorage is tensor(multidimensional matrix) object storage manager for deep learning framework(Pytorch, Tensorflow V2, Keras).
- Boilerplated data pipeline for dataset, model and optimizer.
- High performance on tensor storage
For researchers who need to focus on model training:
- Support storing data in pre-processed Tensor(multidimensional matrix), eliminate training time.
- Reduce storage space through multiple compression methods.
- Manage data and models while training
For AI Developer who need to focus on creating data pipeline:
- Concurrency data save & load
- Compatible with object storage such as MinIO, S3
- Generate pipeline from user endpoints data.
For an example of tensorflow, refer to the detailed document. If you want to see the full code, see below
-
Content
$ pip install matoragequick start with NAS(network access storage) using docker
It can be managed through the web through the address http://127.0.0.1:9000/, and security is managed through MINIO_ACCESS_KEY and MINIO_SECRET_KEY.
$ mkdir ~/shared # create nas storage folder
$ docker run -it -p 9000:9000 \
--restart always -e \
"MINIO_ACCESS_KEY=minio" -e \
"MINIO_SECRET_KEY=miniosecretkey" \
-v ~/shared:/container/vol \
minio/minio gateway nas /container/volFirst, create a DataConfig by importing matorage.
This is an example of pre-processing mnist and storing it in distributed storage.
additional is freely in the form of a dict, and records the shape and type of tensor to be stored in attributes.
from matorage import DataConfig
traindata_config = DataConfig(
endpoint='127.0.0.1:9000',
access_key='minio',
secret_key='miniosecretkey',
dataset_name='mnist',
additional={
"mode": "train",
"framework" : "pytorch",
...
"blah" : "blah"
},
attributes=[
('image', 'float32', (1, 28, 28)),
('target', 'int64', (1))
]
)Now do a simple pre-processing and save the data.
from matorage import DataSaver
traindata_saver = DataSaver(config=traindata_config)
train_loader = DataLoader(dataset, batch_size=60, num_workers=8)
for (image, target) in tqdm(train_loader):
# image shape : torch.Size([64, 1, 28, 28])
# target shape : torch.Size([64])
traindata_saver({
'image': image,
'target': target
})
traindata_saver.disconnect()Now fetch data iteratively from storage with the same config as the saved dataset when training.
from matorage.torch import Dataset
train_dataset = Dataset(config=traindata_config, clear=True)
train_loader = DataLoader(
train_dataset, batch_size=64, num_workers=8, shuffle=True
)
for batch_idx, (image, target) in enumerate(tqdm(train_loader)):
image, target = image.to(device), target.to(device)Only an index can be fetched through lazy load.
train_dataset = Dataset(config=traindata_config, clear=True)
print(train_dataset[0], len(train_dataset))During training, you can save and load models of specific steps or epochs in distributed storage through inmemory. First, make the model config the same as the dataset.
from matorage import ModelConfig
from matorage.torch import ModelManager
model_config = ModelConfig(
endpoint='127.0.0.1:9000',
access_key='minio',
secret_key='miniosecretkey',
model_name='mnist_simple_training',
additional={
"version" : "1.0.1",
...
"blah" : "blah"
}
)
model_manager = ModelManager(config=model_config)
print(model_manager.get_metadata)
model_manager.save(model, epoch=1)
print(model_manager.get_metadata)When an empty model is loaded with specific steps or epochs, the appropriate weight is filled into the model.
print(model.state_dict())
model_manager.load(model, epoch=1)
print(model.state_dict())
# load a layer weight.
print(model_manager.load('net1.0.weight', step=0))Save and load of optimizer is similar to managing model.
from matorage import OptimizerConfig
from matorage.torch import OptimizerManager
optimizer_config = OptimizerConfig(
endpoint='127.0.0.1:9000',
access_key='minio',
secret_key='miniosecretkey',
optimizer_name='adam',
additional={
"model" : "1.0.1",
...
"blah" : "blah"
}
)
optimizer_manager = OptimizerManager(config=optimizer_config)
print(optimizer_manager.get_metadata)
# The optimizer contains information about the step.
optimizer_manager.save(optimizer)
print(optimizer_manager.get_metadata)When an empty optimizer is loaded with specific steps, the appropriate weight is filled into the optimizer.
optimizer = optim.Adam(model.parameters(), lr=0.01)
optimizer_manager.load(optimizer, step=938)$ git clone https://github.com/graykode/matorage && cd matorage
$ python -m tests.test_suite- torch(>=1.0.0), torchvision(>=0.2.2)
- tensorflow(>=2.2), tensorflow_io(>=0.13)
Tae Hwan Jung(@graykode) We are looking for a contributor.
![dependabot[bot] avatar](https://avatars.githubusercontent.com/in/29110?v=4 "dependabot[bot]")
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent