Training the Original Superpoint Network from scratch

[vanguard478]

Hi,
Thanks a lot for this amazing repo and also the blog post.
I am trying to train the original superpoint network SuperPoint_MagicLeap using the steps highlighted in this repo. I had to make some changes to deal with some issues in the first steps as below :

• In models/SuperPointNet_pretrained.py I changed this line to class SuperPointNet_pretrained(torch.nn.Module)
• Also in the Train_model_heatmap.py , I had to make changes to the variable receiving the return of the model as in the modified Superpoint Models, the return type is dict where as the original one returns a tuple:

           if self.config["model"]["name"] == "SuperPointNet_pretrained":
               semi, coarse_desc = outs[0], outs[1]
           else:
               semi, coarse_desc = outs["semi"], outs["desc"]

I am using the same config file magicpoint_shapes_pair.yaml as the one mentioned in Step 1 of the README with the following parameters changed :

front_end_model: 'Train_model_heatmap'
model:
    name: 'SuperPointNet_pretrained'
    params: {
    }
    detector_loss:
        loss_type: 'softmax'

    batch_size:  512 # 64
retrain: True # set true for new model
reset_iter: True
#pretrained: ##train from scratch

With these changes, I am able to start the training without any issues. I have also changed the batch size parameter to batch_size: 512 and increased the learning_rate: 0.008
The training loss from tensorboard is as below and am not sure if the training is going corretcly:

Is my current changes sufficient for the training of the MagicPoint network using the Original Superpoint Network?