jarvis73 / deepcnn-tf2

Tensorflow 2.0 examples of baseline models and useful skills ......

Examples for Tensorflow v2

Tensorflow v2 examples of baseline networks.

0. Requirements

• python=3.7
• tensorflow=2.2.0
• numpy=1.18.8
• cudatoolkit=10.1.243
• cudnn=7.6.5
• opencv=3.4.2
• tqdm

1. Datasets

• mnist (Auto-download): 60000 training images with shape 28x28 and 10000 test images.
• Cifar (Auto-download): 60000 training

2. Training/Evaluating Model

2.1 Start with Tensorflow v2

Training a simple CNN model with the API fit() of the tf.Keras:

Usage: python xxx.py <mode> <gpu> [options]

python main_mini_cnn7layers.py train 0

Test model with test dataset:

python main_mini_cnn7layers.py test 0

2.2 Custom training/testing loop

python main_custom.py train

2.3 Start Tensorboard

tensorboard --logdir ./model_dir --port 16006
# Expore to the network
tensorboard --logdir ./model_dir --port 16006 --host 0.0.0.0

3. Visualization Data with Tensorboard Projector

python projector/mnist/main_cnn7layers.py
# With sprite image
python projector/mnist/main_cnn7layers.py -s

Open terminal in projector/mnist/dst_ckpt and open tensorboard by

tensorboard --logdir ./

Open browser with address: http://localhost:6006/#projector

Results in 2D (MNIST embedding):

4 LDA loss

Reference:

Dorfer M, Kelz R, Widmer G. Deep linear discriminant analysis[J]. arXiv preprint arXiv:1511.04707, 2015.

python main_lda.py train