def main(args):
base_model = gen_base_model()
models, image_lists, label_lists = {}, {}, {}
with open('answer.csv', 'w') as answer_file:
answer_writer = csv.writer(answer_file, lineterminator='\n')
with open(args.csv_file, 'r') as csvfile:
csv_reader = csv.reader(csvfile)
for [jpg_file, label_category, label] in csv_reader:
weights_file = os.path.join(args.retrain_base_dir,
label_category, args.weights)
if os.path.exists(weights_file) and label == '?':
print('\ncategory label: ', label_category)
if label_category not in models:
image_lists_path = os.path.join(
args.retrain_base_dir, label_category,
'image_lists.json')
assert os.path.exists(image_lists_path)
with open(image_lists_path, 'r') as f:
image_lists[label_category] = json.load(f)
n_classes = len(image_lists[label_category].keys())
label_lists[label_category] = \
sorted(
[i for i in image_lists[label_category].keys()])
if label_lists[label_category][0] == 'label 0':
# There are some data that never behave as 'label 0'
# and our model doesn't predict
# but we still need to write its prob as 0 later
need_label_0 = False
else:
need_label_0 = True
models[label_category] = make_model(
base_model, add_final_layer, weights_file,
n_classes)
# write answers, eg. '0.4365;0.5180;0.8741;0.2189;0.3938;0.0422'
img = image.load_img(os.path.join(args.rank_dir, jpg_file))
pred = predict_from_img(models[label_category], img)
pred_ind = pred.argmax(axis=-1)
print('prediction for {}\n\tis {}. '.format(
jpg_file.split('/')[-1], pred_ind + int(need_label_0)),
end='')
print('label:', label_lists[label_category][pred_ind])
label_pred = [0 for _ in range(int(need_label_0))]
label_pred.extend(pred)
label = ';'.join(
['{:.5f}'.format(prob) for prob in label_pred])
print('output:', label)
answer_writer.writerow([jpg_file, label_category, label])
def bottle_pred_func(img):
# target_size = (IM_WIDTH, IM_HEIGHT)
# img = image.load_img(file, target_size=target_size)
return predict_from_img(base_model, img)
def main(args):
"""Use transfer learning and fine-tuning to train a network on a new dataset"""
base_model = None
image_lists = load_training_data(args)
n_classes = len(image_lists.keys())
weights_file = os.path.join(args.model_dir, args.weights)
if not os.path.exists(weights_file):
print('weights_file: {} not found'.format(weights_file))
raise FileNotFoundError
bottleneck_dir = os.path.join(args.model_dir, 'distorted_bottlenecks/')
if args.make_new_model:
# use bottleneck, here the model must be identical to the original top layer
# print(base_model.output.shape)
target_size = (IM_WIDTH, IM_HEIGHT)
base_model = gen_base_model()
model = make_model(base_model, add_final_layer, weights_file,
n_classes)
for category in ['testing', 'validation', 'training']:
print('\ncategory:', category)
img_list = []
truths_list = []
for label_index, label_name in enumerate(image_lists.keys()):
for image_index, image_name in enumerate(
image_lists[label_name][category]):
image_name = get_image_path(image_lists, label_name,
image_index, args.image_dir,
category)
img_list.append(
image.load_img(image_name, target_size=target_size))
truths_list.append(label_index)
preds = predict_from_img(model, img_list)
pred_classes = preds.argmax(axis=-1)
ind = np.nonzero(np.array(pred_classes) != np.array(truths_list))
print('truths:', np.array(truths_list)[ind])
print('preds:', pred_classes[ind], preds[ind])
# img = image.load_img(file, target_size=target_size)
# model.save(os.path.join(args.model_dir, 'inceptionv3-ft.model'))
else:
retrain_input_tensor = Input(shape=(BOTTLENECK_DIM, ))
retrain_model = add_final_layer(retrain_input_tensor,
retrain_input_tensor, n_classes)
print('loading weights: {}'.format(weights_file))
retrain_model.load_weights(weights_file)
compile_retrain_model(retrain_model)
base_model = None
def bottle_pred_func(img):
# target_size = (IM_WIDTH, IM_HEIGHT)
# img = image.load_img(file, target_size=target_size)
return predict_from_img(base_model, img)
for category in ['testing', 'validation', 'training']:
print('\n{} data:'.format(category))
(test_x, test_y, files) = get_random_cached_bottlenecks(
image_lists, -1, category, bottleneck_dir, args.image_dir,
bottle_pred_func)
print(retrain_model.test_on_batch(test_x, test_y))
preds = retrain_model.predict_on_batch(np.array(test_x))
pred_classes = preds.argmax(axis=-1)
ind = np.nonzero(np.array(pred_classes) != np.array(test_y))
print('truths:', test_y[ind])
print('preds:', pred_classes[ind], preds[ind])