def train(self,
img_folder,
ann_folder,
nb_epoch,
project_folder,
batch_size=8,
do_augment=False,
learning_rate=1e-4,
train_times=1,
valid_times=1,
valid_img_folder="",
valid_ann_folder="",
first_trainable_layer=None,
ignore_zero_class=False):
if ignore_zero_class:
loss_k = masked_categorical_crossentropy
else:
loss_k = 'categorical_crossentropy'
train_generator = create_batch_generator(img_folder, ann_folder, self._input_size, self._output_size, self._n_classes,
batch_size,train_times,do_augment)
if valid_img_folder:
validation_generator = create_batch_generator(valid_img_folder, valid_ann_folder, self._input_size,self._output_size, self._n_classes,batch_size,valid_times)
return train(self._network,loss_k,train_generator,validation_generator,learning_rate, nb_epoch, project_folder, first_trainable_layer)
def train(self,
img_folder,
ann_folder,
nb_epoch,
project_folder,
batch_size=8,
do_augment=False,
learning_rate=1e-4,
train_times=1,
valid_times=1,
valid_img_folder="",
valid_ann_folder="",
first_trainable_layer=None,
ignore_zero_class=False,
metrics='val_loss'):
if metrics != "val_accuracy" and metrics != "val_loss":
print("Unknown metric for SegNet, valid options are: val_loss or val_accuracy. Defaulting ot val_loss")
metrics = "val_loss"
if ignore_zero_class:
loss_k = masked_categorical_crossentropy
else:
loss_k = 'categorical_crossentropy'
train_generator = create_batch_generator(img_folder, ann_folder, self._input_size, self._output_size, self._n_classes,
batch_size,train_times,do_augment)
if valid_img_folder:
validation_generator = create_batch_generator(valid_img_folder, valid_ann_folder, self._input_size,self._output_size, self._n_classes,batch_size,valid_times,False)
return train(self._network,loss_k,train_generator,validation_generator,learning_rate, nb_epoch, project_folder, first_trainable_layer, self, metrics)
def train(self,
img_folder,
nb_epoch,
project_folder,
batch_size=8,
augumentation=False,
learning_rate=1e-4,
train_times=1,
valid_times=1,
valid_img_folder="",
first_trainable_layer=None,
metrics="val_loss"):
if metrics != "val_accuracy" and metrics != "val_loss":
print(
"Unknown metric for Classifier, valid options are: val_loss or val_accuracy. Defaulting ot val_loss"
)
metrics = "val_loss"
train_generator, validation_generator = create_datagen(
img_folder, valid_img_folder, batch_size, self._input_size,
project_folder, augumentation)
model_layers, model_path = train(self._network,
'categorical_crossentropy',
train_generator, validation_generator,
learning_rate, nb_epoch,
project_folder, first_trainable_layer,
self, metrics)
if self._bottleneck_layer:
self.save_bottleneck(model_path, self._bottleneck_layer)
return model_layers, model_path
def train(self,
img_folder,
ann_folder,
nb_epoch,
project_folder,
batch_size=8,
jitter=True,
learning_rate=1e-4,
train_times=1,
valid_times=1,
valid_img_folder="",
valid_ann_folder="",
first_trainable_layer=None,
metrics="mAP",
validation_freq=1):
# 1. get annotations
train_annotations, valid_annotations = get_train_annotations(
self._labels,
img_folder,
ann_folder,
valid_img_folder,
valid_ann_folder,
is_only_detect=False)
# 1. get batch generator
valid_batch_size = len(valid_annotations) * valid_times
if valid_batch_size < batch_size:
raise ValueError(
"Not enough validation images: batch size {} is larger than {} validation images. Add more validation images or decrease batch size!"
.format(batch_size, valid_batch_size))
train_batch_generator = self._get_batch_generator(train_annotations,
batch_size,
train_times,
jitter=jitter)
valid_batch_generator = self._get_batch_generator(valid_annotations,
batch_size,
valid_times,
jitter=False)
# 2. To train model get keras model instance & loss fucntion
model = self._yolo_network.get_model(first_trainable_layer)
loss = self._get_loss_func(batch_size)
# 3. Run training loop
return train(model,
loss,
train_batch_generator,
valid_batch_generator,
learning_rate=learning_rate,
nb_epoch=nb_epoch,
project_folder=project_folder,
first_trainable_layer=first_trainable_layer,
network=self,
metrics="mAP",
validation_freq=validation_freq)
def train(self,
img_folder,
ann_folder,
nb_epoch,
project_folder,
batch_size=8,
jitter=True,
learning_rate=1e-4,
train_times=1,
valid_times=1,
valid_img_folder="",
valid_ann_folder="",
first_trainable_layer=None,
metrics="mAP"):
# 1. get annotations
train_annotations, valid_annotations = get_train_annotations(
self._labels,
img_folder,
ann_folder,
valid_img_folder,
valid_ann_folder,
is_only_detect=False)
# 1. get batch generator
train_batch_generator = self._get_batch_generator(train_annotations,
batch_size,
train_times,
jitter=jitter)
valid_batch_generator = self._get_batch_generator(valid_annotations,
batch_size,
valid_times,
jitter=False)
# 2. To train model get keras model instance & loss fucntion
model = self._yolo_network.get_model(first_trainable_layer)
loss = self._get_loss_func(batch_size)
# 3. Run training loop
return train(model,
loss,
train_batch_generator,
valid_batch_generator,
learning_rate=learning_rate,
nb_epoch=nb_epoch,
project_folder=project_folder,
first_trainable_layer=first_trainable_layer,
network=self,
metrics="mAP")
def train(self,
img_folder,
nb_epoch,
project_folder,
batch_size=8,
augumentation=False,
learning_rate=1e-4,
train_times=1,
valid_times=1,
valid_img_folder="",
first_trainable_layer=None):
train_generator, validation_generator = create_datagen(
img_folder, valid_img_folder, batch_size, self._input_size,
project_folder, augumentation)
return train(self._network, 'categorical_crossentropy',
train_generator, validation_generator, learning_rate,
nb_epoch, project_folder, first_trainable_layer)
