Implement check to skip value range transform #198
Conversation
Yeah, I was wondering about that. Any suggestion @qlzh727? We’d have to inspect the graph, right? |
|
@qlzh727 How is this working between the preprocessing layers chain? Are we going to have some important casting overhead as we have profiled at #165 (comment)? |
|
if the value range for all the image are (0, 255) for the input image, then it will benefit from this change (since it skip the conversion). When the input range are (0, 1), then any chain of the layer will keep converting them between (0, 1) and (0, 255). We might want to add some warning to user if this method has been hit several times in the pipeline, and suggest user to pre-convert their value to (0, 255). Anyway, this method is always good to have since it always skip the unnecessary conversion. |
Can we find a way to communicate/propagate between KLP layers that the range is converted? |
|
As I suppose that a chain of transformations is a quite common use case and it could be the O() of the casting overhead. |
* Implement check to skip value range transform * add simple test case * Run format
* Implement check to skip value range transform * add simple test case * Run format
int64 is not a supported type in jax bye default. Trying to use it gets the following error. UserWarning: Explicitly requested dtype int64 requested in array is not available, and will be truncated to dtype int32. To enable more dtypes, set the jax_enable_x64 configuration option or the JAX_ENABLE_X64 shell environment variable. See https://github.com/google/jax#current-gotchas for more. target = jnp.array(target, dtype="int64") We can stick to int32 (which is what will be used anyway).
No description provided.