Would Edge Impulse like to share the dataset which is used for training the pre-trained keyword spotting model?

Help
#1

Question/Issue:
There are two keyword spotting models which are updated from MobileNetV1 and MobileNetV2 respectively, and their input shape is 99 x 40 x 1. Edge Impulse also kindly provides the pre-trained model weights for these two models. While I want to change the input shape to 120 x 40 x 1, or 160 x 40 x 1. Then l can use the new model to analyze more longer audio keywords. So the wieghts of pre-trained models should be changed too.

Project ID:
[Provide the project ID]

Context/Use case:
Developing keyword spotting.

Steps Taken:
None

Expected Outcome:
Would you like to share the train dataset for these two models? So I can use it train the model and use it in the following transferring learning. Thank you very much!
Or tell me the dataset name and version numer if it’s public. Thanks again.

Actual Outcome:
None.

Reproducibility:

  • [v] Always
  • [ ] Sometimes
  • [ ] Rarely

Environment:
None

Additional Information:
None

1 Like
#2

Hi @dobetter

Can you share the project links you are referring to? If you want to see our public datasets or export a projects data see below :point_down:

Public Datasets

Export any public project dataset

  • Public projects - Clone > Export Data (Download)

Hopefully this answers your question? If not please rephrase the question as I may have misunderstood.

Regards,

Eoin

#3

Regarding the second part of your question @dobetter

Reshaping the input is detailed in this tutorial:
.Classify multiple 2D input features - Edge Impulse Documentation

Best

Eoin

#4

@jorduino Thank you very much for your reply and suggestion! The project link is: Audio - Dashboard - Edge Impulse.
And the pre-trained model for transfer learning is MobileNet1.0 and MobileNetV2, which are provided by Edge Impluse. Both of them have input shape 99x40x1, and their weights should be trained by Edge Impluse. As following snapshot,

t
t1784Ă—914 145 KB

I want to know if I can get the dateset for pre-training these two models, because I want to change input shape from 99x40x1 to about 120x40x1. I saw the accuracy of transfer learning based these two models with 99x40x1 is good. So If can’t get the original dataset or its probability distribution, I can’t train a model having good accuracy in later transfer learning when changing input shape from 99x40x1 to 120x99x1.
Not sure the dataset for pre-training these two models is the ones “audio/keyword-spotting” and “audio/faucet-vs-noise” in the link you mentioned.
Please help me. Thank you!