def inference(saved_model_filepath, image_folder, output_filepath,
image_format):
img_filepath_list = [
os.path.join(image_folder, fn) for fn in os.listdir(image_folder)
if fn.endswith('.{}'.format(image_format))
]
model = tf.saved_model.load(saved_model_filepath)
with open(output_filepath, 'w') as fh:
print('Starting inference of file list')
for i in range(len(img_filepath_list)):
img_filepath = img_filepath_list[i]
_, img_name = os.path.split(img_filepath)
print('{}/{} : {}'.format(i, len(img_filepath_list), img_name))
img = imagereader.imread(img_filepath)
img = img.astype(np.float32)
# normalize with whole image stats
img = imagereader.zscore_normalize(img)
pred = _inference(img, model)
fh.write('{}, {}\n'.format(img_name, float(pred)))
print(' Regression Value: {} '.format(pred))
def inference(checkpoint_filepath, image_folder, output_folder, number_classes, number_channels, image_format):
print('Arguments:')
print('checkpoint_filepath = {}'.format(checkpoint_filepath))
print('image_folder = {}'.format(image_folder))
print('output_folder = {}'.format(output_folder))
print('number_classes = {}'.format(number_classes))
print('number_channels = {}'.format(number_channels))
print('image_format = {}'.format(image_format))
# create output filepath
if not os.path.exists(output_folder):
os.mkdir(output_folder)
img_filepath_list = [os.path.join(image_folder, fn) for fn in os.listdir(image_folder) if fn.endswith('.{}'.format(image_format))]
fcd_model = model.FCDensenet(number_classes, 1, number_channels, 1e-4)
fcd_model.load_checkpoint(checkpoint_filepath)
print('Starting inference of file list')
for i in range(len(img_filepath_list)):
img_filepath = img_filepath_list[i]
_, slide_name = os.path.split(img_filepath)
print('{}/{} : {}'.format(i, len(img_filepath_list), slide_name))
print('Loading image: {}'.format(img_filepath))
img = imagereader.imread(img_filepath)
img = img.astype(np.float32)
# normalize with whole image stats
img = imagereader.zscore_normalize(img)
print(' img.shape={}'.format(img.shape))
if img.shape[0] > INF_TILE_SIZE or img.shape[1] > INF_TILE_SIZE:
tile_size = INF_TILE_SIZE
segmented_mask = _inference_tiling(img, fcd_model, tile_size)
else:
segmented_mask = _inference(img, fcd_model)
if 0 <= np.max(segmented_mask) <= 255:
segmented_mask = segmented_mask.astype(np.uint8)
if 255 < np.max(segmented_mask) < 65536:
segmented_mask = segmented_mask.astype(np.uint16)
if np.max(segmented_mask) > 65536:
segmented_mask = segmented_mask.astype(np.int32)
if 'tif' in image_format:
skimage.io.imsave(os.path.join(output_folder, slide_name), segmented_mask, compress=6, bigtiff=True, tile=(1024,1024))
else:
try:
skimage.io.imsave(os.path.join(output_folder, slide_name), segmented_mask, compress=6)
except TypeError: # compress option not valid
skimage.io.imsave(os.path.join(output_folder, slide_name), segmented_mask)
def inference(image_folder, image_format, saved_model_filepath, output_folder,
min_box_size):
# create output filepath
if not os.path.exists(output_folder):
os.mkdir(output_folder)
if image_format.startswith('.'):
image_format = image_format[1:]
img_filepath_list = [
os.path.join(image_folder, fn) for fn in os.listdir(image_folder)
if fn.endswith('.{}'.format(image_format))
]
# load the saved model
yolo_model = tf.saved_model.load(saved_model_filepath)
print('Starting inference of file list')
for i in range(len(img_filepath_list)):
img_filepath = img_filepath_list[i]
_, file_name = os.path.split(img_filepath)
print('{}/{} : {}'.format(i, len(img_filepath_list), file_name))
print('Loading image: {}'.format(img_filepath))
img = imagereader.imread(img_filepath)
height, width, channels = img.shape
img = img.astype(np.float32)
# normalize with whole image stats
img = imagereader.zscore_normalize(img)
print(' img.shape={}'.format(img.shape))
# convert HWC to CHW
batch_data = img.transpose((2, 0, 1))
# convert CHW to NCHW
batch_data = batch_data.reshape(
(1, batch_data.shape[0], batch_data.shape[1], batch_data.shape[2]))
batch_data = tf.convert_to_tensor(batch_data)
boxes = yolo_model(batch_data, training=False)
# boxes are [x1,y1,x2,y2,c]
# constrain boxes to image coordinates
boxes[:, 0] = np.clip(boxes[:, 0], 0, width)
boxes[:, 1] = np.clip(boxes[:, 1], 0, height)
boxes[:, 2] = np.clip(boxes[:, 2], 0, width)
boxes[:, 3] = np.clip(boxes[:, 3], 0, height)
# strip out batch_size which is fixed at one
boxes = np.array(boxes)
boxes = boxes[0, ]
boxes = bbox_utils.filter_small_boxes(boxes, min_box_size)
objectness = boxes[:, 4:5]
class_probs = boxes[:, 5:]
boxes = boxes[:, 0:4]
# nms boxes per tile being passed through the network
boxes, scores, class_label = bbox_utils.per_class_nms(
boxes, objectness, class_probs)
# convert [x1, y1, x2, y2] to [x, y, w, h]
boxes[:, 2] = boxes[:, 2] - boxes[:, 0]
boxes[:, 3] = boxes[:, 3] - boxes[:, 1]
class_label = np.reshape(class_label, (-1, 1))
boxes = np.concatenate((boxes, class_label), axis=-1)
boxes = boxes.astype(np.int32)
# # draw boxes on the images and save
# import skimage.io
# ofp = './inference-results'
# img = img - np.min(img)
# img = img / np.max(img)
# img = np.asarray(img * 255, dtype=np.uint8)
# fn = '{:08d}.png'.format(i)
# print(boxes)
# skimage.io.imsave(os.path.join(ofp, fn), bbox_utils.draw_boxes(img.copy(), boxes))
# write merged rois
print('Found: {} rois'.format(boxes.shape[0]))
output_csv_file = os.path.join(output_folder,
file_name.replace(image_format, 'csv'))
bbox_utils.write_boxes_from_xywhc(boxes, output_csv_file)