ImageNet for 8bit quantization

[bell__sefi]

The standard Imagenet dataset is divided into many folders where each folder represents the label of the images in that folder.

Normally there is no text file named 'val.txt'.

Do you support this data structure? what should I put in the conf json for this to work?

[Iffa_Intel · Accepted solution]

Greetings,

If you are trying to call certain values inside the val.txt, it should be doable as long as you declared/call it properly.

You should declare a string variable and allocate it with your values file & call it.

Example:

val = "C:/Desktop/OpenVino/val.txt"

Or if you are using python, you can use something like subprocess.Popen([r'C:/locations'])

You can refer here for more: https://docs.python.org/3/library/subprocess.html

For further details about quantization you can refer here:

1. https://docs.openvinotoolkit.org/latest/_docs_Workbench_DG_Int_8_Quantization.html
2. https://docs.openvinotoolkit.org/latest/_docs_Workbench_DG_Dataset_Types.html#imagenet

Sincerely,

Iffa

[bell__sefi]

We didn't understand each-other.

Imagenet is organized in 1000 folders (see attached image \ sketched example). Where each folder contains many images of the same class (the folder name represents the image it self).

However, the 8bit calibration expects to receive a .txt file.

Do you support this type of dataset hierarchy?

If not, how should the .txt file be prepared in this case? 

(Image would not upload, this is the structure:

> Imagenet

    > train

    > val

            > n01440764

                     > .........JPEG

                     > .........JPEG

            > n01443537

            > n01494475

[Iffa_Intel]

By default, the imagenet should come together with the val.txt file as you can see in here https://docs.openvinotoolkit.org/latest/workbench_docs_Workbench_DG_Download_and_Cut_Datasets.html

Hence, this is the common structure

|-- imagenet.zip

|-- val.txt

|-- 0001.jpg

|-- 0002.jpg

|...

|-- n.jpg

However, if yours does not have val.txt, you can convert them to unnotated as in here: https://docs.openvinotoolkit.org/latest/workbench_docs_Workbench_DG_Dataset_Types.html

Which forms the archive as below (no val.txt file):

|-- 0001.jpg

|-- 0002.jpg

|...

|-- n.jpg

If your process take too much time, you can also cut the dataset:

https://docs.openvinotoolkit.org/latest/workbench_docs_Workbench_DG_Download_and_Cut_Datasets.html

Sincerely,

Iffa