def get_model(class_names,
base_weights=None,
weights_path=None,
image_dimension=224,
color_mode='grayscale',
weight_decay=1e-4,
class_mode='multiclass',
final_activation="softmax"):
"""
Create model for transfer learning
Arguments:
class_names - list of str
weights_path - str
Returns:
model - Keras model
"""
if weights_path is not None:
base_weights = None
base_model = DenseNet121(include_top=False,
weights=base_weights,
pooling="None")
x = base_model.output
x = GlobalAveragePooling2D()(x)
# dense layers for different class
predictions = []
if class_mode == 'multiclass':
print(f"** Final layer activation function is {final_activation}")
prediction = Dense(4096,
kernel_regularizer=regularizers.l2(weight_decay),
name="fc_hidden_layer1")(x)
predictions = Dense(
len(class_names),
activation=final_activation,
name="fc_output_layer",
kernel_regularizer=regularizers.l2(weight_decay))(prediction)
elif class_mode == 'multibinary':
for i, class_name in enumerate(class_names):
prediction = Dense(
1024, kernel_regularizer=regularizers.l2(weight_decay))(x)
prediction = Dense(
1,
kernel_regularizer=regularizers.l2(weight_decay),
activation="sigmoid",
name=class_name)(prediction)
predictions.append(prediction)
model = Model(inputs=base_model.input, outputs=predictions)
if weights_path is not None:
model.load_weights(weights_path)
model.base_model = base_model
return model
max_score = 0
epoch = 200
for i in xrange(epoch):
model.fit([codes, partial_captions],
next_words,
batch_size=128,
nb_epoch=1,
validation_data=([codesT, partial_captionsT], next_wordsT))
reference_file = "ref_UNK.txt"
predictions = []
for j, id in enumerate(ids):
predictions.append(
str(id) + "\t" + gen.printSentence(sens[j][0][1:-1]) + '\n')
del gen
(goldMap, predictionMap) = computeMaps(predictions, reference_file)
score = bleuFromMaps(goldMap, predictionMap)[0]
print 'eter' + str(i)
print score
if max_score < score and i > 5:
max_score = score
json_string = model.to_json()
open('../model/codenn_512.json', 'w').write(json_string)
model.save_weights('../model/codenn_512.h5')