def test(self):
target = []
predict = []
predict_without_threshold = []
self.model.eval()
losssum = 0
for (in_data, out_target) in tqdm(self.test_data):
### transform data ###
in_data_1 = transform_torch_vars(in_data[0], self.is_cuda)
in_data_2 = transform_torch_vars(in_data[1], self.is_cuda)
out_data = transform_torch_vars(out_target, self.is_cuda)
# cal loss
output = self.model.forward_encoder_vector(in_data_1, in_data_2)
out_data = torch.unsqueeze(out_data,1)
temp_loss = self.evaluate_metric(output, out_data)
losssum += temp_loss.item()
out_pred = self._postprocess_output(output)
predict.extend(self._postprocess_label(out_pred))
predict_without_threshold.extend(self._postprocess_label_without_threshold(out_pred))
target.extend(self._postprocess_label_without_threshold(out_target))
print(losssum/len(self.test_data))
FileTool.writePickle("predict_score.pkl", predict_without_threshold)
self._print_out_result(target, predict)
def __init__(self, img_dir, file_info, transform):
print("read pickle files")
self.img_dir = img_dir
self.imgIds = sorted(FileTool.readPickle(file_info["ids"]))
self.file_info = FileTool.readPickle(file_info["file_info"])
self.annotations = FileTool.readPickle(file_info["annotations"])
self.transform = transform
def _print_out_result(self, target, pred):
"""
change code here to display result
:param pred:
:param target:
:return:
"""
FileTool.writePickle("target.pkl", target)
FileTool.writePickle("predict.pkl", pred)
print(precision_recall_fscore_support(target, pred))
print(accuracy_score(target, pred))
def encode(self):
encoder_file = []
self.model.eval()
for (in_data, out_target) in tqdm(self.test_data):
### transform data ###
in_data = transform_torch_vars(in_data, self.is_cuda)
out_data = transform_torch_vars(out_target, self.is_cuda)
# cal loss
output = self.model.forward_encode(in_data[0])
predict = list(output.detach().cpu().numpy())
target = self._postprocess_label_without_threshold(out_data)
encoder_file.extend([[data, target[id]] for id, data in enumerate(predict)])
FileTool.writePickle("features_siamese.pkl", encoder_file)
def get_img_annotation(self, ind):
img_path = self.imgIds[ind]
image = cv2.imread(img_path)
annos = json.loads(
FileTool.read_text_file(img_path.replace(".jpg", ".json"))[0])
handpoints = np.array(annos["hand_pts"])
hand_center = np.array(annos["hand_box_center"])
max_x, min_x = np.max(handpoints[:, 0]), np.min(handpoints[:, 0])
max_y, min_y = np.max(handpoints[:, 1]), np.min(handpoints[:, 1])
size = int(max_x - min_x) * 3 if (max_x - min_x) > (max_y - min_y) \
else int(max_y - min_y) * 3
image = cv2.copyMakeBorder(image, size // 2, size // 2, size // 2,
size // 2, cv2.BORDER_REPLICATE)
handpoints[:, 0:2] = handpoints[:, 0:2] + size // 2
hand_center = hand_center + size // 2
image = image[int(hand_center[1] - size / 2):int(hand_center[1] +
size / 2),
int(hand_center[0] - size / 2):int(hand_center[0] +
size / 2), :]
handpoints[:, 0] = handpoints[:, 0] - int(hand_center[0] - size / 2)
handpoints[:, 1] = handpoints[:, 1] - int(hand_center[1] - size / 2)
return image, handpoints
if ele["image_id"] not in ids_add:
ids_add.append(ele["image_id"])
if ele["image_id"] in annotations:
annotations[ele["image_id"]].append({
"image_id": ele["image_id"],
"is_crowd": ele["iscrowd"],
"area": ele["area"],
"keypoints": ele["keypoints"],
"num_keypoints": ele["num_keypoints"]
})
else:
annotations[ele["image_id"]] = [{
"image_id": ele["image_id"],
"is_crowd": ele["iscrowd"],
"area": ele["area"],
"keypoints": ele["keypoints"],
"num_keypoints": ele["num_keypoints"]
}]
import os
print("write output")
# FileTool.writePickle(os.path.join(output_file_path, "val_ids.pkl"), ids_add)
# FileTool.writePickle(os.path.join(output_file_path, "val_file_infos.pkl"), file_infos)
# FileTool.writePickle(os.path.join(output_file_path, "val_annotation_ids.pkl"), annotations)
FileTool.writePickle(os.path.join(output_file_path, "ids.pkl"), ids_add)
FileTool.writePickle(os.path.join(output_file_path, "file_infos.pkl"), file_infos)
FileTool.writePickle(os.path.join(output_file_path, "annotation_ids.pkl"), annotations)