cvfinetune

Fine-Tune Framework based on Chainer

Chainer framework is an easy to use DL framework. It is designed in a hierarchical manner and provides usefull implementations for all of the parts required to train a network:

We developed cvmodelz for fast and easy way of initializing commonly used models ("Model" box), and cvdatasets provides methods to load dataset annotations and create a dataset object ("Dataset" box) that can be further passed to the iterator.

An example training script might be looking like this:

#!/usr/bin/env python
if __name__ != '__main__': raise Exception("Do not import me!")

import chainer as ch

from chainer import training
from chainer.training import extensions

from cvargparse import GPUParser
from cvargparse import Arg

from cvdatasets import FileListAnnotations
from cvdatasets import Dataset as BaseDataset

from cvmodelz.models import ModelFactory
from cvmodelz.classifiers import Classifier


class Dataset(BaseDataset):

    def __init__(self, *args, prepare, **kw):
        super().__init__(*args, **kw)
        self._prepare = prepare

    def get_example(self, i):
        im, _, label = super().get_example(i)
        im = self._prepare(im, size=(224, 224))
        return im, label


def main(args):

    model = ModelFactory.new(args.model_type)
    clf = Classifier(model)

    device = ch.get_device(args.gpu[0])
    device.use()

    annot = FileListAnnotations(root_or_infofile=args.data_root)
    train, test = annot.new_train_test_datasets(
        dataset_cls=Dataset, prepare=model.prepare)

    train_iter = ch.iterators.MultiprocessIterator(train, batch_size=32, n_processes=4)
    test_iter = ch.iterators.MultiprocessIterator(test, batch_size=32, n_processes=4,
        repeat=False, shuffle=False)

    # Setup an optimizer
    optimizer = ch.optimizers.AdamW(alpha=1e-3).setup(clf)

    # Create the updater, using the optimizer
    updater = training.StandardUpdater(train_iter, optimizer, device=device)

    # Set up a trainer
    trainer = training.Trainer(updater, (50, 'epoch'), out='result')
    # Evaluate the model with the test dataset for each epoch
    trainer.extend(extensions.Evaluator(test_iter, clf,
        progress_bar=True,
        device=device))

    trainer.extend(extensions.ProgressBar(update_interval=1))
    # Write a log of evaluation statistics for each epoch
    trainer.extend(extensions.LogReport())

    # Print selected entries of the log to stdout
    trainer.extend(extensions.PrintReport(
        ['epoch', 'main/loss', 'validation/main/loss',
         'main/accuracy', 'validation/main/accuracy']))


    #  Run the training
    trainer.run()

parser = GPUParser([
    Arg("data_root"),
    Arg("model_type", choices=ModelFactory.get_models(["cvmodelz"])),
])

main(parser.parse_args())

# start it with 'python train.py path/to/dataset cvmodelz.ResNet50'

Everything after the first two lines in the main function is chainer-related and has to be done every time one needs to write an experiment. Hence, cvfinetune simplifies everything and abstracts the initializations:

from chainer.training.updaters import StandardUpdater

from cvfinetune.finetuner import FinetunerFactory
from cvfinetune.training.trainer import Trainer
from cvfinetune.parser import default_factory
from cvfinetune.parser import FineTuneParser

from cvmodelz.classifiers import Classifier

from cvdatasets.dataset import AnnotationsReadMixin
from cvdatasets.dataset import TransformMixin

parser = FineTuneParser(default_factory())

class Dataset(TransformMixin, AnnotationsReadMixin):
    def __init__(self, *args, prepare, center_crop_on_val: bool = True,  **kwargs):
        super().__init__(*args, **kwargs)
        self.prepare = prepare
        self.center_crop_on_val = center_crop_on_val

    def transform(self, im_obj):
        im, parts, lab = im_obj.as_tuple()
        return self.prepare(im), lab + self.label_shift

def main(args):
    factory = FinetunerFactory(mpi=False)

    tuner = factory(args,
        classifier_cls=Classifier,
        dataset_cls=Dataset,
        updater_cls=StandardUpdater,

        no_sacred=True,
    )
    tuner.run(trainer_cls=Trainer, opts=args)


main(parser.parse_args())

This example can be run with

python example.py data.yml CUB200 GLOBAL -mt cvmodelz.ResNet50 -g0 --input_size infer_big --batch_size 24

using this data config (data.yml):

BASE_DIR: /home/korsch_data/

DATA_DIR: datasets
MODEL_DIR: models

MODELS:
  cvmodelz.InceptionV3:     &inception
    folder: inception
    weights:
      inat: model.inat.ckpt.npz
      imagenet: model.imagenet.ckpt.npz

  cvmodelz.ResNet50:     &resnet50
    folder: resnet
    weights:
      imagenet: model.npz

DATASETS:
  CUB200:         &cub200
    folder: birds/cub200
    annotations: "ORIGINAL"
    annotation_type: FILE_LIST
    n_classes: 200

PART_TYPES:
  GLOBAL:         &parts_global
    feature_suffix: ""

PARTS: {}