Open coldwarrl opened 7 years ago
Agreed,
A rough outline of an all inclusive doc on ETL for images would be best solution.
I think the classes/features that should be mentioned would be.
Path Label Generator shuffling transform pipelines including shuffling of transforms such as skew and crop scaling and standardizing
dead link in https://deeplearning4j.org/model-zoo:
Transfer Learning
Pretrained models are perfect for transfer learning! You can read more about transfer learning using DL4J here.
Thanks, fixed the link.
The documentation, explaining the ETL process for images (https://deeplearning4j.org/simple-image-load-transforml) could be improved regarding describing the transformation possibilities:
"The pipeline can also be randomly shuffled with each transform in order to increase the available dataset. If you set shuffle to true, you can get an image pipeline that executes each transformation at random."
The randomization is not directly related to increasing the available dataset. At least me has confused this.
It would be also helpful to point to the related example: (https://github.com/deeplearning4j/dl4j-examples/blob/master/dl4j-examples/src/main/java/org/deeplearning4j/examples/convolution/AnimalsClassification.java)