def test_lossfun(self):
self.assertEqual(GET.lossfun('mse').__name__, 'mean_squared_error')
self.assertEqual(GET.lossfun('mae').__name__, 'mean_absolute_error')
self.assertEqual(GET.lossfun('ber').__name__, 'bernoulli_nll')
self.assertEqual(
GET.lossfun('gauss_kl').__name__, 'gaussian_kl_divergence')
self.assertEqual(GET.lossfun('test').__name__, 'mean_squared_error')
self.assertEqual(GET.lossfun('').__name__, 'mean_squared_error')
def main(args):
# 各種データをユニークな名前で保存するために時刻情報を取得する
exec_time = GET.datetimeSHA()
# Set up a neural network to train
# Classifier reports softmax cross entropy loss and accuracy at every
# iteration, which will be used by the PrintReport extension below.
# 活性化関数を取得する
actfun1 = GET.actfun(args.actfun1)
actfun2 = GET.actfun(args.actfun2)
# モデルを決定する
if args.network == 0:
from Lib.network import JC_DDUU as JC
else:
from Lib.network2 import JC_UDUD as JC
model = L.Classifier(
JC(n_unit=args.unit, layer=args.layer_num, rate=args.shuffle_rate,
actfun1=actfun1, actfun2=actfun2, dropout=args.dropout,
view=args.only_check),
lossfun=GET.lossfun(args.lossfun)
)
# Accuracyは今回使用しないのでFalseにする
# もしも使用したいのであれば、自分でAccuracyを評価する関数を作成する必要あり？
model.compute_accuracy = False
# Setup an optimizer
optimizer = GET.optimizer(args.optimizer).setup(model)
# Load dataset
train, test, _ = GET.imgData(args.in_path)
train = ResizeImgDataset(train, args.shuffle_rate)
test = ResizeImgDataset(test, args.shuffle_rate)
# predict.pyでモデルを決定する際に必要なので記憶しておく
model_param = F.args2dict(args)
model_param['shape'] = train[0][0].shape
train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
test_iter = chainer.iterators.SerialIterator(test, args.batchsize,
repeat=False, shuffle=False)
# Set up a trainer
updater = training.StandardUpdater(
train_iter, optimizer, device=args.gpu_id
)
trainer = training.Trainer(
updater, (args.epoch, 'epoch'), out=args.out_path
)
# Evaluate the model with the test dataset for each epoch
trainer.extend(extensions.Evaluator(test_iter, model, device=args.gpu_id))
# Dump a computational graph from 'loss' variable at the first iteration
# The "main" refers to the target link of the "main" optimizer.
trainer.extend(
extensions.dump_graph('main/loss', out_name=exec_time + '_graph.dot')
)
# Take a snapshot for each specified epoch
frequency = args.epoch if args.frequency == -1 else max(1, args.frequency)
trainer.extend(
extensions.snapshot(filename=exec_time + '_{.updater.epoch}.snapshot'),
trigger=(frequency, 'epoch')
)
# Write a log of evaluation statistics for each epoch
trainer.extend(extensions.LogReport(log_name=exec_time + '.log'))
# trainer.extend(extensions.observe_lr())
# Save two plot images to the result dir
if args.plot and extensions.PlotReport.available():
trainer.extend(
PlotReportLog(['main/loss', 'validation/main/loss'],
'epoch', file_name='loss.png')
)
# trainer.extend(
# PlotReportLog(['lr'],
# 'epoch', file_name='lr.png', val_pos=(-80, -60))
# )
# Print selected entries of the log to stdout
# Here "main" refers to the target link of the "main" optimizer again, and
# "validation" refers to the default name of the Evaluator extension.
# Entries other than 'epoch' are reported by the Classifier link, called by
# either the updater or the evaluator.
trainer.extend(extensions.PrintReport([
'epoch',
'main/loss',
'validation/main/loss',
# 'lr',
'elapsed_time'
]))
# Print a progress bar to stdout
trainer.extend(extensions.ProgressBar())
# Resume from a snapshot
if args.resume:
chainer.serializers.load_npz(args.resume, trainer)
# Set pruning
# http://tosaka2.hatenablog.com/entry/2017/11/17/194051
masks = pruning.create_model_mask(model, args.pruning, args.gpu_id)
trainer.extend(pruning.pruned(model, masks))
# Make a specified GPU current
if args.gpu_id >= 0:
chainer.backends.cuda.get_device_from_id(args.gpu_id).use()
# Copy the model to the GPU
model.to_gpu()
chainer.global_config.autotune = True
else:
model.to_intel64()
# predict.pyでモデルのパラメータを読み込むjson形式で保存する
if args.only_check is False:
F.dict2json(args.out_path, exec_time + '_train', model_param)
# Run the training
trainer.run()
# 最後にモデルを保存する
# スナップショットを使ってもいいが、
# スナップショットはファイルサイズが大きい
chainer.serializers.save_npz(
F.getFilePath(args.out_path, exec_time, '.model'),
model
)