def predict_views(batch_size, batch_threads, dataset_factory, estimator,
image_size, output_directory, dataset_part):
print('Starting views evaluation...')
dataset = dataset_factory.get_dataset(dataset_part)
output_directory = os.path.join(output_directory, dataset_part)
if os.path.exists(output_directory):
shutil.rmtree(output_directory)
os.makedirs(output_directory)
print('\n\nRunning Prediction for %s' % dataset_part)
input_function = get_input_function(dataset, batch_size, batch_threads,
False, image_size)
predicted = estimator.predict(input_fn=input_function)
num_samples = dataset.get_number_of_samples()
for sample, prediction in enumerate(predicted):
target_path = os.path.join(output_directory,
str(prediction['views_classifications']),
prediction['file_names'].decode('UTF-8'))
original_path = prediction['paths'].decode('UTF-8')
directory = os.path.dirname(target_path)
if not os.path.exists(directory):
os.makedirs(directory)
shutil.copy(original_path, target_path)
if (sample + 1) % batch_size == 0:
sys.stdout.write('\r>> Processed %d samples of %d' %
(sample + 1, num_samples))
sys.stdout.flush()
print('Finished views prediction.')
def predict_views(batch_size, batch_threads, dataset_factory, estimator, image_size, dataset_part):
print('Starting views evaluation...')
dataset = dataset_factory.get_dataset(dataset_part)
print('\n\nRunning Prediction for %s' % dataset_part)
input_function = get_input_function(dataset, batch_size, batch_threads, False, image_size)
predicted = estimator.predict(input_fn=input_function)
num_samples = dataset.get_number_of_samples()
counts = [0, 0, 0]
correct = [0, 0, 0]
for sample, prediction in enumerate(predicted):
expected = prediction['views']
actual = prediction['views_classifications']
counts[expected] += 1
if expected == actual:
correct[expected] += 1
if (sample + 1) % batch_size == 0:
sys.stdout.write('\r>> Processed %d samples of %d' % (sample + 1, num_samples))
sys.stdout.flush()
print('Finished views prediction.')
print('counts')
print(counts)
print('correct')
print(correct)
accuracies = [correct[0] / counts[0], correct[1] / counts[1], correct[2] / counts[2]]
print('accuracies')
print(accuracies)
def run_validation(dataset_factory, batch_size, batch_threads, estimator,
image_size):
val_input_function = get_input_function(dataset_factory.get_dataset('val'),
batch_size, batch_threads, False,
image_size)
evaluation_result = estimator.evaluate(input_fn=val_input_function)
print('Finished Validation: ')
print(evaluation_result)
def run_training(dataset, batch_size, batch_threads, epoch, estimator,
num_epochs, image_size):
print('\n\nRunning training of epoch %d of %d:\n' %
(epoch + 1, num_epochs))
train_input_function = get_input_function(dataset, batch_size,
batch_threads, True, image_size)
estimator.train(input_fn=train_input_function)
print('\nFinished Training epoch %d' % (epoch + 1))
def run_evaluation(batch_size, batch_threads, dataset_factory, estimator,
image_size):
print('\n\nRunning test set evaluation:')
evaluation_result = estimator.evaluate(
input_fn=get_input_function(dataset_factory.get_dataset(
'test'), batch_size, batch_threads, False, image_size))
print('Finished test evaluation: ')
print(evaluation_result)
def run_prediction_and_store_features(dataset_factory, batch_size,
batch_threads, estimator,
output_directory, dataset_part,
image_size):
dataset = dataset_factory.get_dataset(dataset_part)
output_directory = os.path.join(output_directory, dataset_part)
if os.path.exists(output_directory):
shutil.rmtree(output_directory)
os.makedirs(output_directory)
print('\n\nRunning Prediction for %s' % dataset_part)
input_function = get_input_function(dataset, batch_size, batch_threads,
False, image_size)
predicted = estimator.predict(input_fn=input_function)
num_samples = dataset.get_number_of_samples()
with open(output_directory + '/features.csv', 'w', newline='') as features_file, open(output_directory + '/labels.csv', 'w', newline='') as labels_file, \
open(output_directory + '/cameras.csv', 'w', newline='') as cameras_file, open(output_directory + '/names.csv', 'w', newline='') as file_names_file:
features_writer = csv.writer(features_file, delimiter=',')
labels_writer = csv.writer(labels_file, delimiter=',')
cameras_writer = csv.writer(cameras_file, delimiter=',')
file_names_writer = csv.writer(file_names_file, delimiter=',')
for sample, prediction in enumerate(predicted):
if (sample + 1) % batch_size == 0:
sys.stdout.write('\r>> Processed %d samples of %d' %
(sample + 1, num_samples))
sys.stdout.flush()
pre_logits = prediction['pre_logits']
features_writer.writerow(np.squeeze(pre_logits))
actual_labels = prediction['actual_labels']
labels_writer.writerow([actual_labels])
camera = prediction['cameras']
cameras_writer.writerow([camera])
file_names = prediction['file_names'].decode('UTF-8')
file_names_writer.writerow([file_names])
print('\nFinished Prediction %s' % dataset_part)
def run_prediction_and_store_images(dataset_factory, batch_size, batch_threads,
estimator, output_directory, dataset_part,
image_size, result_dir):
dataset = dataset_factory.get_dataset(dataset_part)
print('\n\nRunning Prediction for %s' % dataset_part)
input_function = get_input_function(dataset, batch_size, batch_threads,
False, image_size)
predicted = estimator.predict(input_fn=input_function)
num_samples = dataset.get_number_of_samples()
if not os.path.exists(result_dir):
os.makedirs(result_dir)
if not os.path.exists(os.path.join(result_dir, 'input')):
os.makedirs(os.path.join(result_dir, 'input'))
if not os.path.exists(os.path.join(result_dir, 'prediction')):
os.makedirs(os.path.join(result_dir, 'prediction'))
if not os.path.exists(os.path.join(result_dir, 'target')):
os.makedirs(os.path.join(result_dir, 'target'))
for sample, prediction in enumerate(predicted):
if (sample + 1) % batch_size == 0:
sys.stdout.write('\r>> Processed %d samples of %d' %
(sample + 1, num_samples))
sys.stdout.flush()
out_im = (1 * prediction['output']) #.astype(np.uint8)
in_im = (1 * prediction['input']) #.astype(np.uint8)
tar_im = (1 * prediction['target']) #.astype(np.uint8)
imageio.imwrite(
os.path.join(result_dir, 'input',
prediction['file_names'].decode('UTF-8')), in_im)
imageio.imwrite(
os.path.join(result_dir, 'prediction',
prediction['file_names'].decode('UTF-8')), out_im)
imageio.imwrite(
os.path.join(result_dir, 'target',
prediction['file_names'].decode('UTF-8')), tar_im)
print('\nFinished Prediction %s' % dataset_part)
def predict_views(batch_size, batch_threads, dataset_factory, estimator,
image_size, output_directory, dataset_part):
print('Starting views evaluation...')
dataset = dataset_factory.get_dataset(dataset_part)
output_directory = os.path.join(output_directory, dataset_part)
if os.path.exists(output_directory):
shutil.rmtree(output_directory)
os.makedirs(output_directory)
print('\n\nRunning Prediction for %s' % dataset_part)
input_function = get_input_function(dataset, batch_size, batch_threads,
False, image_size)
predicted = estimator.predict(input_fn=input_function)
num_samples = dataset.get_number_of_samples()
sum_images = np.zeros([3, 128, 64, 3], dtype=np.longlong)
counters = np.zeros([3])
for sample, prediction in enumerate(predicted):
original_path = prediction['paths'].decode('UTF-8')
image = cv2.imread(original_path)
image = cv2.resize(image, (64, 128))
predicted_view = prediction['views_classifications']
sum_images[predicted_view, :, :, :] += image
counters[predicted_view] += 1
if (sample + 1) % batch_size == 0:
sys.stdout.write('\r>> Processed %d samples of %d' %
(sample + 1, num_samples))
sys.stdout.flush()
for view in range(3):
mean_image = sum_images[view] / counters[view]
cv2.imwrite(os.path.join(output_directory,
str(view) + '.png'), mean_image)
print('\n\nFinished views prediction.')