Visualization of batches

[mistaleee]

Hi there,

is there a possibility to visualize the batches from the input layer to check if the batches are korrekt?

Thanks for any hints.

[apatsekin]

I use this data augmentation class to dump augmented images:

class DumpSamples(DetectionAugmentation):
    def __init__(self, debug_dir):
        super().__init__()
        self.debug_dir = debug_dir

    def apply(self, input_record):
        image = input_record["image"]
        gt_bbox = input_record["gt_bbox"]
        img = image.astype(np.uint8).transpose((1, 2, 0)).copy()
        for box in gt_bbox:
            cv2.rectangle(img, tuple(box[:2]), tuple(box[2:4]), color=(255, 0, 0),thickness=3)
        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
        print("DEBUG: writing file {}".format(os.path.join(self.debug_dir,"{}.jpg".format(input_record['im_id']))))
        if not os.path.exists(self.debug_dir):
            os.makedirs(self.debug_dir)
        cv2.imwrite(os.path.join(self.debug_dir,"{}.jpg".format(input_record['im_id'])),img)

Make sure you have debug/data_aug directory. And then just append DumpSamples("debug/") in configuration file to the end of transform list. Also make sure you stop execution when you get enough examples.

[mistaleee]

Awesome, thanks for your code!

I put the class in the config file tridentnet_r101v2c4_c5_2x.py and append the DumpSamples("debug/") at the end of

metric_list = [rpn_acc_metric, rpn_l1_metric, box_acc_metric, box_l1_metric] return General, KvstoreParam, RpnParam, RoiParam, BboxParam, DatasetParam, \ ModelParam, OptimizeParam, TestParam, \ transform, data_name, label_name, metric_list, DumpSamples("debug/")

But I tells me that 'NameError: name 'DetectionAugmentation' is not defined'.

Is there anything I missed?

[apatsekin]

yeah, the class definition goes to core.detection_input file OR should import DetectionAugmentation from that file. If you put into core.detection_input don't forget to import it later in configuration file (as others data augmentations classes: from core.detection_input import DumpSamples

[mistaleee]

Thanks, I'll try it and come back to you.

[mistaleee]

yes, it worked. Thanks!