Open BennoStaub opened 6 years ago
Everything needed to train should be there. I'll be publishing an article on how to use Matterport's Mask RCNN implementation on the shapes dataset next week, which should hopefully answer most of your questions.
I am not sure whether my problem is the same one in this topic.
When I tried to run visualize_coco.ipynb
, it said there's no file named instances_shape_train2018.json
. Then I found it via your article in which I downloaded shapes_train_dataset
to replace the one given by the Jupyter notebook example.
Maybe it's a good idea to update the examples/shapes/train
content with corresponding json file, or indicate the download link in README.md?
It is not in GitHub, I think you can download from https://patrickwasp.com/wp-content/uploads/2018/04/shapes_train_dataset.zip
You need to comment out the line in coco.py where it tries to load that (it's trying to append an additional set of data for training). Also change the import minival line to just import val. The reason for this is that the model authors used a bigger training set and smaller validation set than the split of the original data.
Trying to train Mask R CNN, I get the error that above file is missing. Did you try to use your created dataset for a model like Mask R CNN? Maybe there are more files missing. Eager to hear from you and thx.