tensorflow - can we show underfitting with mobilenet?

Question

Are there any examples which could prove/demonstrate that we can see underfitting while classifying images with mobilenet too?

I have tried transfer learning and feature extraction with mobilenet in ml5.js Since it's already trained on several thousand images, even when I add and train only 3 new images, I seem to get correct results.

I am looking for an example such that I can demonstrate to the user that underfitting is possible with mobilenet as well. It could be by changing a particular parameter while building the model or something close. Open to any tech stack (tensorflow.js / Ml5.js / keras).

For instance this is from keras's documentation:

application_mobilenet(
  input_shape = NULL,
  alpha = 1,
  depth_multiplier = 1,
  dropout = 0.001,
  include_top = TRUE,
  weights = "imagenet",
  input_tensor = NULL,
  pooling = NULL,
  classes = 1000
)

mobilenet_preprocess_input(x)

mobilenet_decode_predictions(preds, top = 5)

mobilenet_load_model_hdf5(filepath)

so is there a variable that the user could change and observe the difference/underfitting?

Additionally, here's a codelab link for doing image classification with mobilenet & tensorflow.js. Basically, I want to do something similar but just show the user that underfitting is also possible here. Is there any way I could modify this code?

https://codelabs.developers.google.com/codelabs/tensorflowjs-teachablemachine-codelab#0

Answer 1

Waitting for answers