Metadata-Version: 2.0
Name: torchpack
Version: 0.0.9
Summary: A set of interfaces to simplify the usage of PyTorch
Home-page: https://github.com/hellock/torchpack
Author: Kai Chen
Author-email: chenkaidev@gmail.com
License: MIT
Description-Content-Type: UNKNOWN
Keywords: computer vision
Platform: UNKNOWN
Classifier: Development Status :: 3 - Alpha
Classifier: License :: OSI Approved :: MIT License
Classifier: Operating System :: OS Independent
Classifier: Programming Language :: Python :: 2
Classifier: Programming Language :: Python :: 2.7
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.4
Classifier: Programming Language :: Python :: 3.5
Classifier: Programming Language :: Python :: 3.6
Classifier: Topic :: Utilities
Requires-Dist: cvbase
Requires-Dist: six

# torchpack

[![PyPI Version](https://img.shields.io/pypi/v/torchpack.svg)](https://pypi.python.org/pypi/torchpack)

Torchpack is a set of interfaces to simplify the usage of PyTorch.

Documentation is ongoing.


## Installation

- Install with pip. 
```
pip install torchpack
```
- Install from source.
```
git clone https://github.com/hellock/torchpack.git
cd torchpack
python setup.py install
```

**Note**: If you want to use tensorboard to visualize the training process, you need to
install tensorflow([`installation guide`](https://www.tensorflow.org/install/install_linux)) and tensorboardX(`pip install tensorboardX`).

## What can torchpack do

Torchpack aims to help users to start training with less code, while stays
flexible and configurable. It provides a `Runner` with lots of `Hooks`.

## Example

```python
######################## file1: config.py #######################
work_dir = './demo'  # dir to save log file and checkpoints
optimizer = dict(
    algorithm='SGD', args=dict(lr=0.001, momentum=0.9, weight_decay=5e-4))
workflow = [('train', 2), ('val', 1)]  # train 2 epochs and then validate 1 epochs, iteratively
max_epoch = 16
lr_policy = dict(policy='step', step=12)  # decrese learning rate by 10 every 12 epochs
checkpoint_cfg = dict(interval=1)  # save checkpoint at every epoch
log_cfg = dict(
    # log at every 50 iterations
    interval=50,
    # two logging hooks, one for printing in terminal and one for tensorboard visualization
    hooks=[
        ('TextLoggerHook', {}),
        ('TensorboardLoggerHook', dict(log_dir=work_dir + '/log'))
    ])

######################### file2: main.py ########################
import torch
from torchpack import Config, Runner
from collections import OrderedDict

# define how to process a batch and return a dict
def batch_processor(model, data, train_mode):
    img, label = data
    volatile = False if train_mode else True
    img_var = torch.autograd.Variable(img, volatile=volatile)
    label_var = torch.autograd.Variable(label, requires_grad=False)
    pred = model(img)
    loss = F.cross_entropy(pred, label_var)
    accuracy = get_accuracy(pred, label_var)
    log_vars = OrderedDict()
    log_vars['loss'] = loss.data[0]
    log_vars['accuracy'] = accuracy.data[0]
    outputs = dict(
        loss=loss, log_vars=log_vars, num_samples=img.size(0))
    return outputs

cfg = Config.from_file('config.py')  # or config.yaml/config.json
model = resnet18()
runner = Runner(model, cfg.optimizer, batch_processor, cfg.work_dir)
runner.register_default_hooks(cfg.lr_policy, cfg.checkpoint_cfg, cfg.log_cfg)

runner.run([train_loader, val_loader], cfg.workflow, cfg.max_epoch)
```

