I am following along and I have found some really great tools to help with the learning process. Along with the IBM Keras Neural Network course I am about to finish

First off I wanted to see the shape of the data at different points in the Sequential model I had created. And I realized it's actually really easy to extend Keras Layers and create your. So I made quick debug layer to print out the first input it sees. This allows me to see the shape of the data as it passes through different layers:

So in my Jupyter notebook, right where I am creating my Keras model I just create my own layer with:

import functools
import operator
import numpy as np
from keras.src.engine.base_layer import Layer
from keras.src.engine.input_spec import InputSpec
from keras.src.utils import conv_utils
# isort: off
from tensorflow.python.util.tf_export import keras_export
# @keras_export("keras.layers.PrintInputs")
class PrintInputs(Layer):

    def __init__(self, data_format=None, **kwargs):
        super().__init__(**kwargs)
        self.count = 0

    def call(self, inputs):
        if self.count == 0:
            print(inputs)
            self.count += 1
        return inputs

This is super helpful and I can place it at any point in the model to see what shape is coming in and out of layers. Ex use:

def convolutional_model():
    
    # create model
    model = Sequential()
    model.add(Conv2D(16, (5, 5), activation='relu', input_shape=(28, 28, 1)))
    model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
    
    model.add(Conv2D(8, (2, 2), activation='relu'))
    model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
    
    
    model.add(PrintInputs())  ############### RIGHT HERE 
    model.add(Flatten())
    model.add(PrintInputs())  ############### RIGHT HERE 
    model.add(Dense(100, activation='relu'))
    model.add(Dense(num_classes, activation='softmax'))
    
    # Compile model
    model.compile(optimizer='adam', loss='categorical_crossentropy',  metrics=['accuracy'])
    return model

I can see Flatten does indeed make flatten down to a single dimension from (None, 5, 5, 8) to (None, 200)

Now I have always wasted way too much time scripting away tasks with tools like fzf.. But now I can just ask ChatGPT to do it for me. I just Tell it what function I want to write and it then shows me the function it came up with, and I can choose whether or not I want to have it written to a file that is loaded automatically. It actually works surprisingly well. It's important that you set the System prompt correctly though. I use a tool sgpt or shell gpt I think just to make it easy. Its a quick python install. I am not going to spend the time to document clearly how to set this up, but you should be able to figure it out from below:

#sgpt role created with:
#sgpt --create-role zsh
#Enter role description: Response only with a valid zsh function or zsh alias that can be added to a zshrc file to complete my request. Do not surround this script with any markdow
n. The response needs to be zsh script I can pipe into zsh itself.
#Role "zsh" already exists, overwrite it? [y/N]: y


 function gptzsh() {
   local prompt response confirm
   read "prompt?Enter your prompt for ChatGPT: "

   # Simulate sending the prompt to ChatGPT and getting a response
   response=$(echo "$prompt" | sgpt --no-md --role zsh | grep -v "\`")

   echo "Response: $response"
   read "confirm?Do you want to write this response to $HOME/zsh/gpt.zsh? (y/n): "

   if [[ "$confirm" == "y" ]]; then
     echo "\n\n$response" >> "$HOME/zsh/gpt.zsh"
     echo "Response written to $HOME/zsh/gpt.zsh"
   else
     echo "Response not written to file."
   fi
 }

Some of this function was even written by ChatGPT..