Question/Issue:
While trying to train, I’ve got the following:
ValueError: Attempt to convert a value ({‘sampleId’: 123684157, ‘boundingBoxes’: [{‘label’: 1, ‘x’: 12, ‘y’: 38, ‘w’: 10, ‘h’: 29}, {‘label’: 2, ‘x’: 19, ‘y’: 48, ‘w’: 8, ‘h’: 21}]}) with an unsupported type (<class ‘dict’>) to a Tensor.
Application exited with code 1
I checked that I labels only 2 classes and this is what i see in the details. Not sure what that messages says and how to solve/train?
Thanks
Project ID:
127972
Context/Use case:
Train for object detection of 2 classes. I’m using a tiny dataset of less than ~200 images, just to test this feature before the actual large training process.
It’s interesting as your dataset is setup for object detection but you have a classification block in your learning model. I switched your block to object detection and it’s all good now (FYI you can also use MobileNet SSD model with 320x320px images if your target is a CPU/GPU type).
By chance could you confirm if you first created a classification model, and then switched to object detection? It’s something we should caught in the UI.
Thanks @aurel !
Yes, I can confirm that I started with a Classification project and only later switched to Detection.
What I missed? Where is the place that I should change that block?
In addition, when i’m trying to check the EON Tuner feature, I get an error that maybe related to the same issue:
“An error occurred during the last EON Tuner run: Labelling method ‘object detection’ not currently supported”
Is that related?
Thanks
It’s something we should detect to prevent the error, I’ll open a ticket on our side.
After switching to object detection, the Transfer Learning block should have been removed and then replaced by the Object Detection one.
Regarding the EON Tuner, it is actually not supporting object detection.
I am doing an object detection project and I have a few questions related to Tamir’s here.
Object Detection vs Object Classification:
You mention above the project Tamir started on was a classification project and then it turned into a detection project. It’s my understanding the difference is detection is pixel-based location and classification is color-based identification. Is that the case here? My goal is to identify when/where an object is within an image. Would that mean it’s a detection / classification project you think? Does classification of an object within an image still yield the location within the image (boxed output)?
EON Tuner Qs:
I also ran into the labelling method not supported for object detection. Is it supported for object classification? Is there an ETA on when it will be supported? Why is it not supported (difference from detection and other labelling methods; maybe classification if it works)? I ask this given I’m continuing a project & my predecessor alleges she used the EON Tuner for the aforementioned project description.
My end goal is to determine if the object exists within the image at minimum, and the object’s location within the image for live-inferencing with those red boxes is the full goal.