The last modifications of this post were around 3 years ago, some information may be outdated!

This is my note for the first course of TensorFlow in Practice Specialization given by deeplearning.ai and taught by Laurence Moroney on Coursera.

ðŸ‘‰ Check the codes on my Github.
ðŸ‘‰ Official notebooks on Github.

ðŸ‘‰ Go to course 2 - CNN in Tensorflow.
ðŸ‘‰ Go to course 3 - NLP in Tensorflow.
ðŸ‘‰ Go to course 4 - Sequences, Time Series and Prediction.

## Basic DL on MNIST

import tensorflow as tf# stop the training with conditionclass myCallback(tf.keras.callbacks.Callback):    def on_epoch_end(self, epoch, logs={}): # compare at the end of each epoch        if(logs.get('accuracy') > 0.99):            self.model.stop_training = Truemnist = tf.keras.datasets.mnist(x_train, y_train),(x_test, y_test) = mnist.load_data()x_train, x_test = x_train / 255.0, x_test / 255.0 # normalizecallbacks = myCallback() # define the callbackmodel = tf.keras.models.Sequential([    tf.keras.layers.Flatten(input_shape=(28, 28)), # Takes that square and                                                   # turns it into a 1 dim    tf.keras.layers.Dense(512, activation=tf.nn.relu),    tf.keras.layers.Dense(10, activation=tf.nn.softmax) # 10 outputs])model.compile(optimizer='adam',              loss='sparse_categorical_crossentropy',              metrics=['accuracy'])model.fit(x_train, y_train, epochs=10, callbacks=[callbacks])

1. Adding more Neurons we have to do more calculations, slowing down the process, but get more accurate.
2. The first layer in your network should be the same shape as your data.
3. The number of neurons in the last layer should match the number of classes you are classifying for.
4. Extra layers are often necessary.
5. Flatten as the name implies, converts your multidimensional matrices (Batch.Size x Img.W x Img.H x Kernel.Size) to a nice single 2-dimensional matrix: (Batch.Size x (Img.W x Img.H x Kernel.Size)). During backpropagation it also converts back your delta of size (Batch.Size x (Img.W x Img.H x Kernel.Size)) to the original (Batch.Size x Img.W x Img.H x Kernel.Size).
6. Dense layer is of course the standard fully connected layer.

CNN layers, cource of image.

## Basic DL on Fashion-MNIST

# the same as in MINST# different at below line of loading datamnist = tf.keras.datasets.fashion_mnist

## Basic CNN on Fashion-MNIST

import tensorflow as tfmnist = tf.keras.datasets.fashion_mnistclass myCallback(tf.keras.callbacks.Callback):    def on_epoch_end(self, epochs, logs={}) :        if(logs.get('accuracy') is not None and logs.get('accuracy') >= 0.998) :            print('\nReached 99.8% accuracy so cancelling training!')            self.model.stop_training = True(training_images, training_labels), (test_images, test_labels) = mnist.load_data()# Why reshape?# The first convolution expects a single tensor containing everything,# so instead of 60000 28x28x1 items in a list, we have a single 4D list# that is 60000x28x28x1## training_images' shape (before reshape): (60000, 28, 28)# training_images' shape (after reshape): (60000, 28, 28, 1)# trainaing_labels' shape: (60000,)training_images=training_images.reshape(60000, 28, 28, 1)training_images=training_images / 255.0test_images = test_images.reshape(10000, 28, 28, 1)test_images=test_images/255.0model = tf.keras.models.Sequential([  tf.keras.layers.Conv2D(64, (3,3), activation='relu', input_shape=(28, 28, 1)),  tf.keras.layers.MaxPooling2D(2, 2),  tf.keras.layers.Conv2D(64, (3,3), activation='relu'),  tf.keras.layers.MaxPooling2D(2,2),  tf.keras.layers.Flatten(),  tf.keras.layers.Dense(128, activation='relu'),  tf.keras.layers.Dense(10, activation='softmax')])model.compile(optimizer='adam',              loss='sparse_categorical_crossentropy',              metrics=['accuracy'])callbacks = myCallback()model.fit(training_images, training_labels, epochs=5, callbacks=[callbacks])test_loss = model.evaluate(test_images, test_labels)
model.summary() # model detail
Model: "sequential_1"_________________________________________________________________Layer (type)                 Output Shape              Param=================================================================conv2d (Conv2D)              (None, 26, 26, 64)        640                             # for every image, 64 convolution has been tried                             # 26 (=28-2) because we use 3x3 filter and we can't                             # count on edges, so the picture is 2 smaller on x and y.                             # if 5x5 filter => 4 smaller on x and y._________________________________________________________________max_pooling2d (MaxPooling2D) (None, 13, 13, 64)        0_________________________________________________________________conv2d_1 (Conv2D)            (None, 11, 11, 64)        36928_________________________________________________________________max_pooling2d_1 (MaxPooling2 (None, 5, 5, 64)          0_________________________________________________________________flatten_1 (Flatten)          (None, 1600)              0_________________________________________________________________dense_2 (Dense)              (None, 128)               204928_________________________________________________________________dense_3 (Dense)              (None, 10)                1290=================================================================Total params: 243,786Trainable params: 243,786Non-trainable params: 0

Refs:

1. Kernel in image processing: examples with images.
2. Pooling layer: non-linear down-sampling.

More?

1. Image Filtering -- Lode's Computer Graphics Tutorial

2. Applying Convolutions on top of our Deep neural network will make training => It depends on many factors. It might make your training faster or slower, and a poorly designed Convolutional layer may even be less efficient than a plain DNN!

3. What is a Convolution? => A technique to isolate features in images

4. What is a Pooling? => A technique to reduce the information in an image while maintaining features

5. How do Convolutions improve image recognition? => They isolate features in images

6. After passing a 3x3 conv filter over a 28x28 image, how big will the output be? => 26x26

7x7 to 5x5 (source)

7. After max pooling a 26x26 image with a 2x2 filter, how big will the output be? => 13x13

(source)

## Visualizing the Convolutions and Pooling

Using layer API, something like below, check more in the notebook.

import matplotlib.pyplot as pltf, axarr = plt.subplots()from tensorflow.keras import modelslayer_outputs = [layer.output for layer in model.layers]activation_model = tf.keras.models.Model(inputs = model.input, outputs = layer_outputs)for x in range(0,4):    f1 = activation_model.predict(test_images[FIRST_IMAGE].reshape(1, 28, 28, 1))[x]    axarr[0,x].imshow(f1[0, : , :, CONVOLUTION_NUMBER], cmap='inferno')    axarr[0,x].grid(False)    ...

## Using real-world images

An example of classifying horses and humans!

### ImageGenerator

# make images more used for training# (focus on object, split cleary objects, label images,...)# also help to augmenting data (rotate, skew, flip,...)from tensorflow.keras.preprocessing.image import ImageDataGeneratortrain_datagen = ImageDataGenerator(rescale=1./255)    # normalize -> No need to convert images and then put in the training    # do the scaling on the flytrain_generator = train_datagen.flow_from_directory(    train_dir, # dir contains the dir containing your images               # -> be careful!    target_size=(300, 300), # images will be resized when loaded, genial!                            # because NN always needs that!                            # -> experimenting with diff sizes without                            # impacting your source data    batch_size=128,    class_mode="binary"     # 2 diff things)test_datagen = ImageDataGenerator(rescale=1./255) # normalizevalidation_generator = test_datagen.flow_from_directory(    validation_dir, # dir contains the dir containing your images    target_size=(300, 300),    batch_size=32,    class_mode="binary")

More docs: