h2 = self.bnorm3(self.conv2(h1), finetune=test)
pad_x = F.concat((x, U.zero_pad(x, self.inch, self.outch)))
h3 = h2 + pad_x
return h3
class ResidualNN(C.Classifier):
layer_num = [32, 32, 64, 128, 256]
def __init__(self, initializer, layer_num=layer_num):
super().__init__(
chain.Res4Chain3(initializer, Block, layer_num, initialBN=True))
export_dir = os.path.join(os.getcwd(), "save")
filename, ext = os.path.splitext(__file__)
dataset_dir = os.path.join(os.getcwd(), "../dataset")
train = Trainer(filename, export_dir, dataset_dir)
train.set_networks(ResidualNN, I.HeNormal)
train.model_init()
lr_update = [[0, 200, 350, 500], [0.3, 0.1, 0.01, 0.001]]
train.train_loop(lr_update=lr_update)
train.testing()
train.export()
from pathlib import Path
from keras.optimizers import SGD
from Core import Trainer
from NN.conv import LeNet
def get_model(args):
# initialize the model and the model_filename
model = None
print("[INFO] compiling model...")
# build(width, height, depth, classes, nlf='relu')
model = LeNet.build(32, 32, 3, 3)
model.compile(
loss="categorical_crossentropy",
optimizer=SGD(lr=0.005),
metrics=["accuracy"]
)
model_filename = str(Path(args['model'], 'lenet.h5').expanduser())
return model, model_filename
trainer = Trainer()
model, model_filename = get_model(trainer.args)
trainer.execute(model)
model.save(model_filename)