While we’ve extensively tested our Cloud Compiler to provide you with a smooth experience, we know that occasional hiccups can happen. Complex systems can face challenges from things like model differences (architecture, size, task type, input resolution), network conditions, or internet speeds.
We’re here to help you quickly overcome these issues and keep your projects moving. In this FAQ, you’ll find common compiler challenges, their possible causes, and helpful solutions to get you back on track.
If you don’t find your issue listed, please reply with a post below or create a new topic in the General category.
Can I use models from the Cloud Compiler offline?
Yes, you can use models you get from the Cloud Compiler offline. Download the model from the AI Hub, and you can use it offline right away.
How long does the Cloud Compiler take?
The Cloud Compiler typically takes between a few minutes up to 30 minutes. Contact us or post below if a Cloud Compiler task takes much longer than 30 minutes.
Where can I access my compiled models?
You may access compiled models in the AI Hub at the workspace/zoo specified in the Model Zoo URL field of the Cloud Compiler form.
For example, if your workspace is named hello, and it features a model zoo named world, then your model would be located in the model zoo at https://hub.degirum.com/hello/world.
Click here for more information about how to navigate models in the AI Hub.
How are compiled models named?
Compiled models are named in this format:
<model-name-prefix>–<image-width>x<image-height>_<float or quant>_<runtime>_<device>_<version>
The values for each segment of the name are based on the related fields in the Cloud Compiler form. For example, if you had entered the following into the Cloud Compiler form:
- Model Name Prefix:
foo - Model Version:
1 - Image Width:
640 - Image Height:
640 - Model Zoo URL:
hello/world - Runtime:
HAILORT - Device:
HAILO8L - Quant/Float:
Quant
Then your model would be named:
foo--640x640_quant_hailort_hailo8l_1
and your model would be accessible for AI inferences at:
https://hub.degirum.com/hello/world/foo--640x640_quant_hailort_hailo8l_1
Some models have support for many devices. Those models will be named with multidevice instead of the Device, such as in foo--640x640_quant_openvino_multidevice_1.
What are Float and Quant models?
The Cloud Compiler can compile floating-point (float) and quantized (quant) models. Compiling for float retains full precision, but the resulting model may be computationally demanding. Compiling for quant prioritizes speed and efficiency at the cost of some precision.
We recommend compiling for quant as opposed to float to benefit from the speed and efficiency. At DeGirum, we achieve very little loss of precision when compiling quant models. If you would like more information about how we achieve this, feel free to contact us.
If you select both Float and Quant options in the Cloud Compiler form, then your Cloud Compiler task will compile one quant model and one float model.
What is the custom dataset used for quantization?
The Cloud Compiler uses a dataset of images during the compilation process for quant models. You may submit a custom dataset for quantization through the Advanced Options dropdown menu.
We recommend using between 64 to 100 distinct images. If your dataset is very large, or your images aren’t distinct enough, then the Cloud Compiler will return an error.
If you do not submit a custom dataset for quantization, then a subset of the COCO dataset will be used during the quantization process.
Unsure how to do model height and width. Perhaps a better question is ‘How should I set image width and image height?’
Why can’t I find my Model Zoo?
If you cannot access your model zoo, it’s usually because of one of these:
- Workspace with access to the model zoo isn’t selected
- The Workspace doesn’t exist yet
To resolve this, first ensure you have selected the Workspace with the Model Zoo. For example, if I have four Workspaces and I want Model Zoos from my Workspace named Operator, I need to select it in the dropdown menu on the left.
In this first image, the Operator Workspace is not selected:
In this second image, the Operator Workspace is selected:
When you have selected a Workspace, there’s a few ways to find your model zoo:
- Navigate to model zoos and check model zoos with your Workspace:
- Go to Tasks, select the Cloud Compiler task associated with your model, and check the
cloud_zoo_urlparameter:
- Manually input the URL into your address bar. For example, for the Operator Workspace with the OperatorZoo model zoo, the URL would be
https://hub.degirum.com/Operator/OperatorZoo.
How do I create a model zoo?
For in-depth information on creating model zoos, visit our docs site. However, if your workspace has permission to create model zoos, follow these steps:
- Ensure a Workspace is selected.
- Click Private Models in the Navigation Bar on the left of the AI Hub. You will see tiles for Model Zoos and All Models. Both of these tiles are scoped to the selected Workspace, so you will see only Model Zoos and all models for the current Workspace.
- Click the Model Zoos tile.
- Click New Model Zoo.
- If enabled for your workspace, there will be a New Model Zoo button. Click the New Model Zoo button.
- Set up the model zoo using the window that appears.