computervision-recipes
computervision-recipes copied to clipboard
microsoft
Best Practices, code samples, and documentation for Computer Vision.
### Description Here https://github.com/microsoft/ComputerVision/blob/3e0631e0dc7d5ddbfc6283b1e89b3ce51f0bd449/utils_cv/common/azureml.py#L52 you are catching a general exception which could hide the true cause of failure and make it really hard to debug. If it fails to get...
### Description In e.g. the 01_training_introduction.ipynb notebook in the object detection folder, after each epoch is complete, output from pycocotools such as the following is displayed: ``` Average Precision (AP)...
### Description In the 01_training_introduction.ipynb file in the object detection folder, I ran a code cell with the following code/output: ``` detector = DetectionLearner(data, im_size=IM_SIZE) > Downloading: "https://download.pytorch.org/models/fasterrcnn_resnet50_fpn_coco-258fb6c6.pth" to C:\Users\mawah/.cache\torch\checkpoints\fasterrcnn_resnet50_fpn_coco-258fb6c6.pth...
### Description The code cell to download the data occurs before a description of the dataset being downloaded and how many images it contains. Moving the description will help the...
### Description In https://github.com/microsoft/ComputerVision/blob/master/scenarios/detection/11_exploring_hyperparameters_on_azureml.ipynb you copy the whole directory in order to make use of the utils_cv This is a bit cumbersome and unecesarily copies things around. You can create...
### Description If the utils are pip installed then there is no need to add the relative path to utils_cv While developing simply install using pip install -e and any...
### Description In the setup guide, users may need more guidance on what SKU to select. From experience I know which VMs have GPUs, and how to request more quota...
- Explain in notebook/FAQ what non-maxima suppression is what values to set (threshold on IoU) - Explain and provide code how to pick a good score threshold (reuse Patrick's plot...
### Description Add evaluation how many images are correct: 
Currently a negative image is added by creating a tiny pseudo bbox with id 0 if the image does not have any objects-of-interest in it. Otherwise torchvision's training script throws...