retinanet

[vismayaps]

Hi, First of all thank you for the awesome content. But I was wondering how to test the model on custom images once the training is done. I've completed the training part and it is performing detection on the example images given. I would like to test the retina net model on my image data set. Any help would be appreciated. Thanks in Advance

[mj1156]

Hello vismayaps, Cna you please tell me how you created customized tf.dataset for retinaNet because tfds.load is for ready to use datasets such coco2017. I am also working with custom images but while training I am getting loss=000+e0 for all epochs It would be great if you can help me out here .

[srihari-humbarwadi]

@vismayaps this should help you run inference with your own images

image_path = ...  # path to your image file
image = tf.image.decode_image(tf.io.read_file(image_path), channels=3)
image = tf.cast(image, dtype=tf.float32)
input_image, ratio = prepare_image(image)
detections = inference_model.predict(input_image)
num_detections = detections.valid_detections[0]
class_names = [
    int2str(int(x)) for x in detections.nmsed_classes[0][:num_detections]
]
visualize_detections(
    image,
    detections.nmsed_boxes[0][:num_detections] / ratio,
    class_names,
    detections.nmsed_scores[0][:num_detections],
)

[mj1156]

Hello vismayaps, Thank you for helping me out

[oattao]

I don't understand the box_variance in classes LabelEncoder and DecoderPredictions. In the LabelEncoder, I saw (box_target = box_target / self._box_variance). In the DecoderPrediction, I saw (boxes = box_predictions * self._box_variance). Could you please explain why we need box_variance?

[gabrielpeixoto-cvai]

Hello @mj1156,

I am having the same issue as you. I could make it work with my custom dataset, but I am getting loss=000+e0. Did you solve this issue? Any ideas @vismayaps ?

[mj1156]

Hallo @gaburiero
I think, I got loss=000+e0 because of two reasons : 1. Dataset labelling is missing in my input pipeline and 2. the predefined anchor boxes values according to COCO dataset.

[gabrielpeixoto-cvai]

Hello @mj1156 ,

Thanks for the feedback, I could make it work (after some time of effort), It turns out that my dataset was in VOC format and I used some converter script to COCO format. However, the conversion was not compatible with the script, so I had to modify the preprocess method in order for it to accept my dataset, and after some tries, I could make it work.

My problem is a single class detection, and the confidence is pretty low (around 0.1), is there any way to improve it? I just trained for 10 epochs.

[adavradou]

Hi @mj1156, I also get a loss: 0.0000e+00 .

How did you manage to change the predefined anchor boxes values? I just changed the num_classes from 80 to 1 (since I have only one class).

Did you or @gabrielpeixoto-cvai manage to train the model successfully for your custom data?

[gabrielpeixoto-cvai]

Hello @adavradou,

Yes, I could make it work. However, I am not sure if my anchor box change attempt was successful (I think so though, but I down know how to prove it). For the anchor boxes, I changed the constructor of the AnchorBox class:

def __init__(self):
        self.aspect_ratios = [0.5, 1.0, 2.0]
        self.scales = [2 ** x for x in [0, 1 / 3, 2 / 3]]
        self._num_anchors = len(self.aspect_ratios) * len(self.scales)
        self._strides = [2 ** i for i in range(3, 8)]
        self._areas = [x ** 2 for x in [32.0, 64.0, 128.0, 256.0, 512.0]]
        self._anchor_dims = self._compute_dims()

and played with areas, scales, and aspect ratios. You can print the result anchor boxes and sizes from get_anchors method.

However, the confidence is still low, which is not a problem for my application, but I tried everything I read online (increase number of samples, changing anchor boxes, etc...) and I got no success. Let me know if you have any success in the future.

[nikeshdevkota]

@mj1156 can you provide details on how you set up a pipeline to load data for a custom dataset?