TensorFlow

  • Returns two tuples (two immutable lists in math parlance).

  • Split arguments into train and test, separating the two groups for different use.

  • Train tuple has 6000, test has 1000.

  • Shape is 28x28.

  • Use a Python library to visualize data: import matplotlib.pyplot as plt.

  • Images are identified by a number to map label → class. For fashion MNIST there are 10 classes: [T-shirt/top, Trouser, Pullover, Dress, Coat, Sandal, Shirt, Sneaker, Bag, Ankle boot].

    ![[RAtsYYyT 4.png|RAtsYYyT 4.png]]

  • Input shape defines the image size; Flatten makes them one-dimensional.

  • Before the dense layer:

    ![[QzC1J1S2 4.png|QzC1J1S2 4.png]]

Normal For a fully connected layer (everyone talks to everyone): ![[-eHHLsbk 4.png|-eHHLsbk 4.png]] Dense (fully connected) layer What I’m doing: ![[L2v4a_Jg 4.png|L2v4a_Jg 4.png]] To:

  1. Flatten 28x28 images.
  2. Apply ReLU; with few layers you can add more capacity; some only call it “deep” if more layers.
  3. Dropout of 0.2 (20%) helps reduce overfit.
  4. Softmax to output probabilities across classes; total must sum to 1.

Model example:

keras.Sequential([
  keras.layers.Flatten(input_shape=(28, 28)),
  keras.layers.Dense(256, activation=tensorflow.nn.relu),
  keras.layers.Dropout(0.2),
  keras.layers.Dense(10, activation=tensorflow.nn.softmax)
])

![[6IQgGSnE 4.png|6IQgGSnE 4.png]]

The exact number of neurons in Dense is flexible; 256 is a common choice. ReLU enables non-linear learning: ![[CGtpMIcU 4.png|CGtpMIcU 4.png]] It basically enables non-linear separation ![[Pu6OCEDg 4.png|Pu6OCEDg 4.png]]

![[zK48A3JU 4.png|zK48A3JU 4.png]]

Then for the output layer:

keras.layers.Dense(10, activation=tensorflow.nn.softmax)

Before training (fit) you must compile:

model.compile(
  optimizer='adam',
  loss='sparse_categorical_crossentropy',
  metrics=['accuracy']
)
  1. Adam is commonly used for optimization, especially with multiclass.
  2. For loss, we use multiclass cross-entropy.
  3. The metric is accuracy (simple and intuitive). Reducing loss and increasing accuracy tends to indicate better learning.

Then train (fit) the model. How much to train depends on the problem; you only know by testing:

  1. Compare train vs validation.
  2. 5 epochs is a decent starting point; more may overfit depending on the dataset.
  3. Split validation from training to observe generalization, e.g. validation_split=0.2.
history = model.fit(
  train_images, train_labels,
  epochs=5,
  validation_split=0.2
)

Finally, use the visualization library to inspect predictions and see if they match the expected class.