Hluboké učení

Convolutional networks for computer vision

S. Lazebnik - Principles of convolution networks (2024)
S. Lazebnik - Architectures of convolution networks (2024)
Záznam přednášky 2021 na Sharepoint - část 1 (česky)
Záznam přednášky 2021 na Sharepoint - část 2 (česky)
Fooling neural networks (extension topic)
Dogs & Cats (8 points)

Fine-tune the network inception_V3 to classify the dataset Dogs vs. Cats available at Kaggle. Use the methodology in Chapter 8 of Deep Learning with Python. Please find the detailed instructions bellow.

Submit a sharing link to your Colab notebook containing your accuracy plot.

Instruction to submit Dogs & Cats
Instead of VGG16, use:

inception_model = tf.keras.applications.InceptionV3(input_shape=(180,180,3), include_top=False)

  • Print the model: inception_model.summary()
  • Unfreeze the last Inception module (starting at layer mixed9)
  • Start with an input layer 
  • Then add a rescalling layer: tf.keras.layers.Rescaling(1./255)
  • Then add the inception model and then classifying layers
  • Train for 30 or more epochs
  • Experiment with hyperparameters (augmentation, size of classification layers, ...)
  • You can also unfreeze another module starting at layer mixed8
  • You should obtain validation accuracy better than with VGG16 (i.e., about 0.99)

Předchozí
Následující