def train(batchNum = 500, batchSize = 200000, learningRate = 0.001,
ImagePatchWidth = 20, #layers = [500, 1000, 500],
ImagePatchStep = 4, labelOptionNum = 100,
labelMode = 'NUM'):
trainDS = ds.read_data_sets(ImagePatchWidth, ImagePatchStep,
labelOptionNum, 'train', labelMode)
classifier = skflow.TensorFlowEstimator(
model_fn = conv_model,
n_classes = labelOptionNum,
batch_size = batchSize,
steps = batchNum,
learning_rate = learningRate)
classifier.fit(trainDS.images, np.argmax(trainDS.labels, axis = 1),
logdir = gv.__DIR__ + gv.tensorflow_log_dir)
return classifier
def test(classifier, ImagePatchWidth = 20, ImagePatchStep = 4,
labelOptionNum = 100, labelMode = 'PRO'):
image_files, bubble_num, bubble_regions = getinfo()
result_filename = gv.cnn__result_filename
accuracy_filename = gv.cnn__accuracy_filename
result = np.zeros((len(image_files),1))
accuracy = np.zeros((len(image_files),4))
index = -1
start_time = time.time()
PROGRESS = progress.progress(0, len(image_files), prefix_info = 'Labeling ')
for i, image_file in enumerate(image_files):
testDS = ds.read_data_sets(ImagePatchWidth, ImagePatchStep,
labelOptionNum, 'test', labelMode,
imageName = image_file)
y = classifier.predict(testDS.images)
index = index + 1
result[index] = np.sum(y)
# saving labeled result as image
io.imsave(gv.__DIR__ + gv.cnn__image_dir + image_file,
np.reshape(y, (testDS.ylength, testDS.xlength)))
_y = np.argmax(testDS.labels, axis = 1)
# total accuracy
accuracy[index, 0] = np.true_divide(np.sum(y == _y), _y.size)
# accuracy of negative labeled instances
accuracy[index, 1] = np.true_divide(np.sum(np.all(
[y == _y, _y == 0], axis = 0)), np.sum(_y == 0))
# accuracy of positive labeled instances
accuracy[index, 2] = np.true_divide(np.sum(np.all(
[y == _y, _y > 0], axis = 0)), np.sum(_y > 0))
# average difference sum
accuracy[index, 3] = np.true_divide(
np.sum(np.absolute(np.subtract(y, _y))), _y.size)
PROGRESS.setCurrentIteration(i+1)
PROGRESS.setInfo(suffix_info = image_file)
PROGRESS.printProgress()
accuracy.tofile(accuracy_filename, sep =" ")
return [result, accuracy]