RF-DETR not detecting more than 300 objects

[emerson-w]

I am installing RFDETR directly from the source using:
pip install git+https://github.com/roboflow/rf-detr.git

And training like so:
model = RFDETRBase()
model.train(
dataset_dir='...',
num_queries = 700,
num_select = 700
)

I'm getting this runtime error.

RuntimeError Traceback (most recent call last)
in <cell line: 0>()
18 model = RFDETRBase()
19
---> 20 model.train(
21 dataset_dir='/content/drive/MyDrive/RMMC_CAN_ds_v3/images',
22 num_queries = 700,

7 frames
/usr/local/lib/python3.11/dist-packages/rfdetr/models/lwdetr.py in forward(self, outputs, target_sizes)
552
553 prob = out_logits.sigmoid()
--> 554 topk_values, topk_indexes = torch.topk(prob.view(out_logits.shape[0], -1), self.num_select, dim=1)
555 scores = topk_values
556 topk_boxes = topk_indexes // out_logits.shape[2]

RuntimeError: selected index k out of range

Is it possible to train RF-DETR to detect more than 300 images? How can I do this?

[isaacrob-roboflow]

unfortunately, using the pretrained checkpoint you won't be able to detect more than 300 objects in a given image. you'd have to train from scratch.

can you say a bit more about your use case?

[emerson-w]

I am now training from scratch like so:
model = RFDETRBase(
pretrain_weights=None,
num_classes=1,
num_queries =700,
num_select = 700
)

model.train(
dataset_dir= '...',
num_queries = 700,
num_select = 700,
epochs=100,
...
)

detections = model.predict(image, threshold=0.1, num_select = 700, num_queries = 700)

But there are still only 300 detections. Is it possible for RF-DETR to predict more than 300 objects?

My use case is counting the number of small objects in an image. These objects are all similar (num_classes = 1) and the training data supports this large number. I would like to be able to identify more than 300 objects without splitting up the image.

[isaacrob-roboflow]

that is likely a bug :) probably the num_select argument isn't getting propagated correctly .. are you comfortable submitting a PR?