Repository for tensorflow examples and experiments.
- 1_cnn_using_keras_api: Train a custom convolutional neural network using the tf.keras API and ImageDataGenerator.
- 2_data_augmentation_using_keras_api.py: Train a custom convolutional neural network tf.keras API and ImageDataGenerator with data augmentation.
- 3_transfer_learning_using_keras_applications: Train and fine-tune a neural network using a pre-trained model from the keras.applications API.
- 4_cnn_visualize_class_activation_heatmaps: Visualize class activation heatmaps from a convolutional neural network, to determine where in the image the network is focusing.
Use Visual Studio Installer -> select Modify.
Workload:
- Install Data science and analytical applications workload
Components:
- Install Python language support
- Install Python 3 64-bit (e.g. 3.6.3)
Install required packages:
- Start Visual Studio, open Python Environments
- Select Python 3.6 (64-bit)
- On the Overview combo box select Install from requirements.txt
If using the gpu version of tensorflow, make sure to check which version of cuda and cuDNN is supported. Then follow the instructions on cudnn-install-windows
The cats and dogs data set used in the examples, can be downloaded from kaggle cats and dogs.
The examples are based on splitting the training set into 3 smaller data sets:
- Train: the first 2000 cat and the first 2000 dog images.
- Validation: The next 1000 cat and the next 1000 dog images.
- Test: The next 1000 cat and the next 1000 dog images.
The 3 sets should each be placed in a separate folder, with sub folders for each class (cat and dog):
- Train: Data/Train/Cat, Data/Train/Dog.
- Validation: Data/Validation/Cat, Data/Validation/Dog.
- Train: Data/Test/Cat, Data/Test/Dog.
This structure will work with the keras ImageDataGenerator.flow_from_directory method.