Imagenet (synthetic) data generation methodology?

tensorflow / benchmarks

A benchmark framework for Tensorflow

Apache License 2.0

1.15k stars 632 forks source link

Imagenet (synthetic) data generation methodology? #494

Closed spatodia closed 4 years ago

spatodia commented 4 years ago

Hello folks,

Beginners question.. how does the Imagenet synthetic dataset used for training in the default case get generated? I see the4 training pipeline uses 1281167 images + 50000 validation images.

reedwm commented 4 years ago

It's randomly generated using tf.truncated_normal. You can see the code here:

https://github.com/tensorflow/benchmarks/blob/c55b4aac7861b3c519803d207045c045c9440c4d/scripts/tf_cnn_benchmarks/models/model.py#L243-L261