def train_multiple_networks():
'''
Trains list of CNNs.
'''
metadata = data_preparation.load_metadata()
metadata, labels = data_preparation.preprocess_metadata(metadata)
train, valid = data_preparation.stratify_train_test_split(metadata)
# for these image sizes, we don't need gradient_accumulation to achieve BATCH_SIZE = 256
optimizer = 'adam'
if params.BATCH_SIZE < 256:
optimizer = gradient_accumulation.AdamAccumulate(
lr=params.LEARNING_RATE, accum_iters=params.ACCUMULATION_STEPS)
base_models = [
[MobileNet, params.MOBILENET_IMG_SIZE, MobileNet_preprocess_input],
[InceptionResNetV2, params.INCEPTIONRESNETV2_IMG_SIZE,
InceptionResNetV2_preprocess_input],
[VGG19, params.VGG19_IMG_SIZE, VGG19_preprocess_input],
[InceptionV3, params.INCEPTIONV3_IMG_SIZE, InceptionV3_preprocess_input],
[MobileNetV2, params.MOBILENETV2_IMG_SIZE, MobileNetV2_preprocess_input],
[NASNetLarge, params.NASNETLARGE_IMG_SIZE, NASNetLarge_preprocess_input],
]
# for [_Model, input_shape, preprocess_input] in base_models:
# train_model(_Model, input_shape, preprocess_input,
# train, valid, labels,
# create_simple_model, optimizer, 'simple')
for [_Model, input_shape, preprocess_input] in base_models:
train_model(_Model, input_shape, preprocess_input,
train, valid, labels,
create_attention_model, optimizer, 'attention')
def loop_in_combinations(callback, image_size=None, transfer_learing=True, use_preprocess_input=False):
'''
Trains list of CNNs.
'''
metadata = data_preparation.load_metadata()
metadata, labels = data_preparation.preprocess_metadata(metadata)
train, valid = data_preparation.stratify_train_test_split(metadata)
# for these image sizes, we don't need gradient_accumulation to achieve BATCH_SIZE = 256
optimizer = 'adam'
if params.DEFAULT_OPTIMIZER != optimizer:
optimizer = gradient_accumulation.AdamAccumulate(
lr=params.LEARNING_RATE, accum_iters=params.ACCUMULATION_STEPS)
unfrozen = 'unfrozen_'
if transfer_learing:
unfrozen = ''
custom_layers = [
[create_attention_model, unfrozen+'latest_attention'],
[create_simple_model, unfrozen+'latest_simple'],
]
for [custome_layer, name_prefix] in custom_layers:
for [_Model, input_shape, preprocess_input] in base_models:
_image_size = image_size
if _image_size is None:
_image_size = input_shape
_preprocess_input = preprocess_input
if not use_preprocess_input:
_preprocess_input = None
callback(_Model, _image_size, transfer_learing, _preprocess_input,
train, valid, labels,
custome_layer, optimizer, name_prefix)
def predict():
'''
Predicts the inout score.
'''
metadata = data_preparation.load_metadata(os.path.join(params.DATA_FOLDER, 'demo', 'pictures360new'))
#metadata, labels = data_preparation.preprocess_metadata(metadata)
labels = ['scenic', 'non-scenic']
# for these image sizes, we don't need gradient_accumulation to achieve BATCH_SIZE = 256
optimizer = 'adam'
if False: # params.BATCH_SIZE < 256:
optimizer = gradient_accumulation.AdamAccumulate(
lr=params.LEARNING_RATE, accum_iters=params.ACCUMULATION_STEPS)
base_models = [
[VGG19, params.VGG19_IMG_SIZE, VGG19_preprocess_input],
]
for [_Model, input_shape, preprocess_input] in base_models:
predict_model(_Model, input_shape, preprocess_input,
metadata, labels,
create_attention_model, optimizer, 'attention')