CNTK Examples: Image/Classification/GoogLeNet/InceptionV3

BrainScript

InceptionV3.cntk

The Inception-V3 model is implemented according to the model described in Rethinking the Inception Architecture for Computer Vision with changes from TensorFlow implementation.

This implementation achieves 78.122% Top-1 accuracy and 94.028% Top-5 accuracy in validation set, which matches with TensorFlow implementation.

And we're using common SGD with Nesterov momentum in this implementation.

This example with a 256 batch-size should be trained with 8 GPUs.

You could run the example from the current folder using:

mpiexec -n 8 cntk configFile=InceptionV3.cntk

If you would like to run this example with a single card. You could divide the minibatchSize and learningRatesPerMB with a ratio 8 simultaneously.

And run this example using:

cntk configFile=InceptionV3.cntk