This tensorflow implementation code is based on Densely Connected Convolutional Networks.
- Using tf.Dataset and tf.estimator which has better computational efficiency and more concise code
- Using the tf.keras.layers(not include dropout and batch_normalization) which has a good visualization in Tensorboard
- Add the Compute time Graph
- Adding a compute graph which inclding time and memory
- Its accuracy rate can reach 94.33% using data augmentation in cifar10.
you only run experiment.py
to get the result.
python experiment.py
Notice:
Maybe it will have a error like "Couldn't open CUDA library libcupti.so.8.0 "
.You need to add /usr/local/cuda/extras/CUPTI/lib64/
to your LD_LIBRARY_PATH
. For detail, you can find there