Changes to cnn defaults for task image classification #1096
Conversation
|
Looks solid! Ready to review once the PR is complete! |
|
Ok, this is ready! I've cleaned up and adjusted the examples a bit! In addition to the updated examples pictured below, I'll be releasing a video tutorial demonstrating this example: https://editor.p5js.org/codingtrain/sketches/ogxO8har_
How the examples look now:
|
|
Ack, apologies, I just realized my vscode settings made a lot of trivial formatting changes (like double -> single quotes). Hope that doesn't pollute this pull request to much, I am happy to clean up and re-submit if necessary. I really need to learn how to have my code editor follow the linting rules of the project I'm working with! (TBH I thought I have it set up this way but I guess I do not!) |
| }; | ||
|
|
||
| // construct the neural network | ||
| nn = ml5.neuralNetwork(options); | ||
| nn.loadData("daytime_nightime.json", train); | ||
| // nn.loadData('daytime_nightime.json', train); |
There was a problem hiding this comment.
Do we need to keep this comment in?
There was a problem hiding this comment.
I'll add a comment that explains it further, this is useful in case you don't want to collect your own data and just want to quickly train a model with sample data and see the classification process.
Apologies for the brief note here I am in the middle of recording a video tutorial about working with CNN in ml5.js. I will circle back and provide more details later today as well as update the examples. But just wanted to note for @bomanimc that this change is incoming and should if possible be a part of 0.6.0.
At the moment, the changes here are to the default configuration for the CNN architecture following this MNIST example from tf.js. The examples were not working with the previous library defaults.