topic Re: Training with eIQ portal in i.MX Processors

Training with eIQ portal

ejmfrancisco — Mon, 13 Mar 2023 06:08:32 GMT

I am trying to train a face detection model using the eIQ portal. I am trying to import dataset from tensorflow dataset (tfds) specifically coco/2017 dataset. However, I only want to import wider_face. But when I try to do that, Importer error occurs which is shown in the image below. Any help will do. Thanks!

Re: Training with eIQ portal

joanxie — Mon, 13 Mar 2023 10:49:14 GMT

what bsp version do you use? pls check this

https://tfhub.dev/tensorflow/collections/object_detection/1

Re: Training with eIQ portal

ejmfrancisco — Tue, 14 Mar 2023 03:47:58 GMT

The running bsp is based on i.MX Yocto Kirkstone BSP 5.15.71_2.2.0 release.
eIQ toolkit is downloaded from https://www.nxp.com/design/software/development-software/eiq-ml-development-environment/eiq-toolkit-for-end-to-end-model-development-and-deployment:EIQ-TOOLKIT
The eIQ installer is the eIQ toolkit 1.6.9 Ubuntu 20.04 Installer.

Also, the link you provide is already a pre-trained model. I am trying to train my model from the scratch by importing from tensorflow datasets using the eIQ portal.

Target dataset: Coco2017/wider_face

But it seems that there are errors downloading the dataset.

Re: Training with eIQ portal

joanxie — Wed, 15 Mar 2023 13:16:59 GMT

did you train other dataset? is the issue only for Coco2017/wider_face? I send to you an example for how to use eIQ toolkit, maybe you can refer to that

Re: Training with eIQ portal

ejmfrancisco — Wed, 15 Mar 2023 21:31:11 GMT

I already have do that process before but Im referring to the wider_face dataset from tfds since I only wanted to focus on face detection.

Steps are:

1. Import dataset

2. Select tensorflow datasets

3. Select object detection

4. Go to advance options and choose wider_face dataset

5. Split dataset to 80%/20%

The error happens from step 1. I tried to download the whole coco2017 dataset aswell but I just want to use the face detection instead thats why I select wider_face dataset.