A quick and easy way to start running pytorch experiments within few minutes.
Installation
- Install latest pytorch and torchvision from Pytorch official website
- pip install easytorch
'How to use?' you ask!
1. Define your trainer
from easytorch import ETTrainer
class MyTrainer(ETTrainer):
def __init__(self, args):
super().__init__(args)
def _init_nn_model(self):
self.nn['model'] = NeuralNetModel(self.args['num_channel'], self.args['num_class'])
def iteration(self, batch):
inputs = batch['input'].to(self.device['gpu']).float()
labels = batch['label'].to(self.device['gpu']).long()
out = self.nn['model'](inputs)
loss = F.cross_entropy(out, labels)
out = F.softmax(out, 1)
_, pred = torch.max(out, 1)
sc = self.new_metrics()
sc.add(pred, labels)
avg = self.new_averages()
avg.add(loss.item(), len(inputs))
return {'loss': loss, 'averages': avg, 'output': out, 'metrics': sc, 'predictions': pred}
def init_experiment_cache(self):
r"""What scores you want to plot."""
r"""By default it plots Loss,Accuracy,F1,Precision,Recall."""
self.cache['log_header'] = 'Loss,Accuracy,F1,Precision,Recall'
r"""This is for best model selection: """
r"""It tells which metrics to monitor and either to maximize(F1 score), minimize(MSE)"""
self.cache.update(monitor_metric='f1', metric_direction='maximize')
2. Use custom or pytorch based Datasets class.
Define specification for your datasets:
import os
sep = os.sep
MYDATA = {
'name': 'mydata',
'data_dir': 'MYDATA' + sep + 'images',
'label_dir': 'MYDATA' + sep + 'labels',
'label_getter': lambda file_name: file_name.split('_')[0] + 'label.csv'
}
MyOTHERDATA = {
'name': 'otherdata',
'data_dir': 'OTHERDATA' + sep + 'images',
'label_dir': 'OTHERDATA' + sep + 'labels',
'label_getter': lambda file_name: file_name.split('_')[0] + 'label.csv'
}
Define how to load each data item by using EasyTorch's base ETDataset class to get extra benefits like limiting, pooling data...
from easytorch import ETDataset
import torchvision
class MyDataset(ETDataset):
def __init__(self, **kw):
super().__init__(**kw)
def __getitem__(self, index):
dataset_name, file = self.indices[index]
dataspec = self.dataspecs[dataset_name]
"""
All the info. (data_dir, label_dir, label_getter...) defined above will be in dataspec.
"""
image = #Todo # Load file/Image.
label = #Todo # Load corresponding label.
# Extra preprocessing, if needed.
# Apply transforms.
return {'indices': self.indices[index],
'input': image,
'label': label}
@property
def transforms(self):
return torchvision.transforms.Compose(["""List of transforms"""])
(If one proceeds with the above (by overriding the ETDataset), can skip directly to 3.)
Or, one can use any other custom datasets as follows:
from easytorch import EasyTorch
class MyExperiment(EasyTorch):
def _load_dataset(self, split_key, split_file, dspec:dict, dataset_cls=None):
return ...
Here, the framework will:
- call _load_dataset(...) with every data split key (train, test, validation).
- So, we just need to write logic to load data for a given key, and return the dataset object.
- Now, just use class MyExperiment class instead of EasyTorch in the entrypoint.
For more advanced cases, one can override the following and directly specify each datasets(train/validation/test):
- The framework will internally call _load_dataset(...) from each of the following methods with corresponding split_key
- So only implement the following if you absolutely have to. Otherwise, implementing _load_dataset() will be enough in most of the cases.
from easytorch import EasyTorch
class MyExperiment(EasyTorch):
def _get_train_dataset(self, split_file, dspec: dict, dataset_cls=None):
return ...
def _get_validation_dataset(self, split_file, dspec: dict, dataset_cls=None):
return ...
def _get_test_dataset(self, split_file, dspec: dict, dataset_cls=None):
return ...
3. Entry point
from easytorch import EasyTorch
runner = EasyTorch([MYDATA, MyOTHERDATA],
phase="train", batch_size=4, epochs=21,
num_channel=1, num_class=2)
if __name__ == "__main__":
runner.run(MyTrainer, MyDataset)
runner.run_pooled(MyTrainer, MyDataset)
Complete Examples
- Vessel segmentation with U-NET example.
- Healthy/Pneumonia/Covid-19 chest x-ray (Multi-label/Binary)classification example.
- DCGAN Example.
Feature Higlights
- For advanced training with multiple networks, and complex training steps, click here:
- Implement custom metrics as here.
- Minimal configuration to setup a new experiment.
- Use your choice of Neural Network architecture.
- Automatic k-fold cross validation/Auto dataset split.
- Automatic logging/plotting, and model checkpointing. ..more features
Default arguments[default-value]. Easily add custom arguments.
- -ph/--phase [Required]
- Which phase to run? 'train' (runs all train, validation, test steps) OR 'test' (runs only test step).
- -b/--batch_size [4]
- -ep/--epochs [51]
- -lr/--learning_rate [0.001]
- -gpus/--gpus [0]
- List of gpus to be used. Eg. [0], [1], [0, 1]
- -nw/--num_workers [4]
- Number of workers for data loading so that cpu can keep-up with GPU speed when loading mini-batches.
- -lim/--load-limit[max]
- Specifies a limit on images/files to load for debug purpose for pipeline debugging.
- -nf/--num_folds [None]
- Number of folds in k-fold cross validation(Integer value like 5, 10).
- -spl/--split_ratio [0.6 0.2 0.2]
- Split ratio for train, validation, test set if two items given| train, test if three items given| train only if one item given.
- ...see more
All the best! for whatever you are working on. Cheers!
Please star or cite if you find it useful.
@misc{easytorch,
author = {Khanal, Aashis},
title = {Easy Torch}
year = {2020},
publisher = {GitHub},
journal = {GitHub repository},
url = {https://github.com/sraashis/easytorch}
}
1 maintainer
Metadata
1 maintainer
Release files for easytorch 2.4.82
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | |
|---|---|---|---|---|
| easytorch-2.4.82.tar.gz | 28.6 kB | Without provenance | Details |
Release files / easytorch-2.4.82.tar.gz
| Download URL | easytorch-2.4.82.tar.gz |
|---|---|
| Size | 28.6 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
8647f783b98943d70a0a53d4ec2f5fb6ad5dd72cc5f89b4580b7c10f150afe56
|
|
BLAKE2b-256 checksum How to use checksums |
de89d2ea4e1e5afc3dbf2c5172c9982be7adab78e19f0632ec7459ee6d1e9406
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/3.3.0 pkginfo/1.7.0 requests/2.25.1 setuptools/49.2.1 requests-toolbelt/0.9.1 tqdm/4.57.0 CPython/3.9.1
|
Security for easytorch 2.4.82
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
