Chainer Training Template

Usage

Copy files

Copy files under "template" directory

Implement some modules

dataset (src/dataset.py)

get_dataset
- returns dictionary of Chainer datasets.
  - Dictionary keys must include "train" for training dataset.
  - Dictionary keys must include "test" for test dataset if you need.
  - Dictionary keys may include "validation" for validation while training.
- example:
```
def get_dataset():
    train, test = chainer.datasets.get_mnist()
    validation, train = chainer.datasets.split_dataset_random(train, 5000)
    return {
        'train': train,
        'validation': validation,
        'test': test,
    }
```

model (src/model.py)

calculate_metrics

returns loss and other metrics for each optimizer.
example:

def calculate_metrics(net, batch):
    x, t = batch
    y = net(x)
    loss = F.softmax_cross_entropy
    return loss

with other metorics:

def calculate_metrics(net, batch):
    x, t = batch
    y = net(x)
    loss = F.softmax_cross_entropy(y, t)
    accuracy = F.accuracy(y, t)
    return {
        'loss': loss,
        'accuracy': accuracy
    }

with two optimizers:

def calculate_metrics(nets, batch):
    gen = nets['gen']
    dis = nets['dis']

    # ... calculate loss
    loss_gen = ...
    loss_dis = ...

    return {
        'gen': loss_gen,
        'dis': loss_dis,
    }
    # or
    return {
        'gen': {},
        'dis': loss_dis,
    }

make_eval_func:
- makes evaluation function for Chainer Evaluator extension.
predict:
- returns prediction of your neural network.

net (src/net/network_name.py)

Implement your neural network class that extends chainer.Chain.

Implement scripts

Training script (src/train.py)

Usually you don't have to modify training script except adding chainer training extensions if you need.

Prediction script (src/predict.py)

Usually you don't have to modify prediction script.

Evaluation script (src/evaluate.py)

Implement your evaluation script.

Configuration files

Application configuration (config/app.json)

Configuration independent of each training/prediction.

train(object): training configuration.
- dump_graph(str): Name of the root for dump_graph extension.
- plot_report(object):
  - key: File name for figure file without extension.
  - value(array of str): y_keys for PlotReport extension.
- print_report(array of str): entries for PrintReport extension.
- max_value_trigger(string or null): The best model is saved when the value associated with this key string becomes maximum. (e.g. "validation/main/accuracy")
- min_value_trigger(string or null): The best model is saved when the value associated with this key string becomes minimum. (e.g. "validation/main/loss") You can use only either max_value_trigger or min_value_trigger.

Configuration (config/config.json)

Configuration depends on each training/prediction.

batch_size(int): Mini batch size for training/prediction.
epoch(int): The number of epochs for training.
gpu(int): Default GPU device index (negative value indicates CPU). This can be changed with command line option.
dataset(object):
- parameter(object, array, single value or null): Parameter for get_dataset. null indicates no parameter.
network(object): Network information.
- key: network identifier.
- value(object):
  - class(str): Network class.
  - parameter(object, array, single value or null): Parameter for network constructor.
  - optimizer(object): Network optimizer information.
    - class(str): Optimizer class.
    - parameter(object, array, single value or null): Parameter for optimizer constructor.
    - hook(array of objects):
      - class(str): Hook class.
      - parameter(object, array, single value or null): Parameter for hook constructor.
output_dir(str): Output directory path for training.

dsanno / chainer-training-template