No Mask output for Instance segmentation task

[raviakash]

Hello all,

I am trying to setup this model for Multi class instance segmentation task on custom dataset. If I closely monitor the results out of 15 classes only 5 classes are showing the very low output AP result and the rest all are showing the value 0.0AP. It seems the model has not learned anything and ran the model for around 450000 iterations with batch size of 16. All the dataset were properly labeled and double checked the configs
here are some of the configs
MODEL.SEM_HEAD: FastInstHead
MODEL.SEM_HEAD.NUM_CLASSES: 15
MODEL.FASTINST.TRANSFORMER_DECODER_NAME: FastInstDecoder
MODEL.FASTINST.TEST.INSTANCE_ON: True

When I load the model and few test images for inference I am only getting box predictions and there are no output mask predictions. If I execute this command after loading the model I get No masks found as the result.

# Check for masks
if instances.has("pred_masks"):
    print("Masks found:", instances.pred_masks.shape)
else:
    print("No masks found.")
    

can you guys point me out where the error is stemming out or where I am going wrong?

[akang-ai]

你好，我也是使用自制的数据集训练的，我在训练的时候也遇到了类似的问题，似乎在鱼类实例上没有学习到任何东西，请问你的问题解决了吗？

[akang-ai]

并且推理的时候速度很慢，并没有像作者论文里说的那样块，可能是我使用的方法不对吧，在output文件夹里还有一个inference文件夹里面也偶遇一个模型（不知道是干嘛的）

[akang-ai]

我的是只有掩码预测，没有框预测

[raviakash]

Hello, I also used a self-made dataset for training. I encountered a similar problem during training. It seems that I did not learn anything on the fish examples. Did you solve your problem?

Hi, No I was not able to solve the problem and also the model did not learn anything related to the various classes

[raviakash]

I only have mask prediction, no box prediction

How did you get mask prediction? could you guide me in setting up the model for generating mask predictions?