search

numericalalgorithmsgroup / AzureML_Best_Practice

Reference Implementations of AzureML Workloads with Cost/Performance Best Practice
3 stars 1 forks source link

Where is maskrcnn_benchmark file ? #1

Open j-sieger opened 3 years ago

j-sieger commented 3 years ago

I am trying to solution for one of problem. After looking at read me file i felt that i can follow ur approach to solve my issue. I want to check make_data_loader function to find how to mounted azure dataset during training

ptooley commented 3 years ago

Hi there, thanks for your interest.

The maskrcnn_benchmark package is published by NVIDIA and can be found here: https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Segmentation/MaskRCNN/pytorch we install the package directly into our Docker image which is why it is not forked into this repo. (Note that this is itself a fork of the Facebook detectron2 repo).

The way that the dataset is located is actually a bit tricky because it is managed by a "catalog" interface that is part of the MaskR-CNN implementation. We provide a custom catalog file in train/dataset_catalog.py that points to /data inside the Docker container.

We mount the dataset to this location within the training script by picking up the dataset that was passed in by the AzureML runtime and mounting it to the right location (see train/train_net_mount.py)

Hopefully this helps!