Testbild

Yes I can confirm, same issue here also.

``` history = model.fit_generator( tuple(train_dataloader), steps_per_epoch=len(train_dataloader), epochs=EPOCHS, callbacks=callbacks, validation_data=tuple(valid_dataloader), validation_steps=len(valid_dataloader), ) ``` ``` /usr/local/lib/python3.7/dist-packages/tensorflow/python/keras/engine/training.py:1940: UserWarning: `Model.fit_generator` is deprecated and will be removed in a future version. Please use `Model.fit`, which...

https://github.com/qubvel/segmentation_models/issues/412 This issue explains it. Appareantly inside the dataloader it has to be made a tuple already and not in the model.fit()

Hi, have you tried to increase the loss_weights? ``` LOSS_WEIGHTS = { "rpn_class_loss": 1., "rpn_bbox_loss": 1., "mrcnn_class_loss": 1., "mrcnn_bbox_loss": 1., "mrcnn_mask_loss": 1. } ```

The weight to my understanding increases the importance of the loss for your problem. Hence if this is important to you and you want to "punish" the model more for...

As a strong guess I would say your image dimensions that you feed to the model and the ones the model expects are 3 pixels off (112 to 115). Do...

@jmayank23 You could use the roLabelImg that you linked and my pull request from the voc2coco repo https://github.com/Tony607/voc2coco/pull/12#issue-1023510998 There I added the rotation conversion, so you can convert the PASCAL...

Ok I see. I think it is still correct to say `import skimage` but the pip install has to be scikit-image. You do not have a docker container by any...

Do you by any chance remember the TF version that you used? I would assume TF1.x if I had to guess. Would be really nice, as we do have a...

Hi, sorry for the late reply. Do you have an E-Mail, where I can contact you? I mainly do this for educational purposes, but maybe there is a chance to...