def save_bottleneck_features(X_train, X_val):
# build original VGG16 model
model = get_vgg_model.build_vgg_model(img_cols, img_rows, weights_path)
print('Doing prediction on default VGG weights...')
bottleneck_features_train = model.predict(X_train, batch_size=128, verbose=0)
print('Prediction for train samples completed. Saving to disk...')
np.save(open(w_path + train_feat_name, 'wb'), bottleneck_features_train)
bottleneck_features_validation = model.predict(X_val, batch_size=128, verbose=0)
print('Prediction for validation samples completed. Saving to disk...')
np.save(open(w_path + val_feat_name, 'wb'), bottleneck_features_validation)
print('Bottleneck features are saved at %s' % dt.datetime.now().strftime("%Y-%m-%d %H:%M:%S"))
img_rows, img_cols = he_load_data.IM_ROWS, he_load_data.IM_COLS
# The HumanEva images are RGB.
img_channels = 3
print('Loading training data...')
if data_augmentation:
X_train, Y_train, X_val, Y_val = he_load_data.load_aug_data()
else:
X_train, Y_train, X_val, Y_val = he_load_data.load_data()
print(X_train.shape[0], 'train samples')
print(X_val.shape[0], 'test samples')
# BUILD THE VGG MODEL ==================
print('building vgg model...')
model = get_vgg_model.build_vgg_model(img_cols, img_rows, weights_path)
# BUILD THE TOPMODEL (upon the VGG16 model) ===========================
print('building the top model...')
top_model = get_he_topmodel.build_he_topmodel(model.output_shape[1:],
fcLayers=topModConf,
reg_w=REG_W_FC,
reg_b=REG_B_FC,
bn_eps=BN_EPS,
dropout=DROPOUT_FC)
# note that it is necessary to start with a fully-trained
# classifier, including the top classifier,
# in order to successfully do fine-tuning
top_model.load_weights(
top_model_weights_path
chk_path = '/home/edogan/checkpoints/'
chk_file = chk_path + 'finetune_umpm_16_20170221-1133/weights-improvement-022-129.35.hdf5'
prd_path = '/home/edogan/predictions/'
prd_file = prd_path + 'finetune_umpm_16.npy'
prd_file_val = prd_path + 'finetune_umpm_16_val.npy'
topModConf = [1024]
batch_size = 128
nb_parts = umpm_load_data.NUM_PARTS
# input image dimensions
img_rows, img_cols = umpm_load_data.IM_ROWS, umpm_load_data.IM_COLS
img_channels = 3
model = get_vgg_model.build_vgg_model(img_cols, img_rows)
print('Building the top model...')
top_model = get_umpm_topmodel.build_umpm_topmodel(model.output_shape[1:],
fcLayers=topModConf)
model.add(top_model)
model.summary()
# model.load_weights(chk_file)
# print('Weights loaded: %s' % chk_file)
# model.compile(loss='mse',
# optimizer=SGD,
# metrics=['mean_squared_error'])