NVIDIA
[DOCS] wp.array auto-conversion using wp.array()
Category
- [ ] Report an error in the documentation.
- [x] Request for something to be documented.
- [ ] Suggestion to improve the documentation.
- [x] Other: Warning / error message
Description
The documentation recommends constructing warp arrays from torch (I guess the same holds for numpy) using from_torch. However, it is also possible to do the following:
a = torch.arange(10)
wp1 = wp.array(a, dtype=wp.int32)
wp2 = wp.array(a, dtype=wp.int64)
While this does not seem to be recommended anywhere, I suggest raising a warning or even an error if this is done and it triggers a data type conversion.
The above example will result in:
wp1 -> [0 0 1 0 2 0 3 0 4 0]
wp2 -> [0 1 2 3 4 5 6 7 8 9]
without any warning, which can lead to hard-to-detect bugs.
Would it help if we put the warning at a higher verbosity level, e.g. if you set wp.config.verbose = True? I don't want to presume that everyone doing a non-standard type conversion would like to be warned by default. I don't know if setting wp.config.verbose = True is part of your debugging workflow, though.
Sure - it would have already helped me if it was mentioned in the Docs, e.g. here. Printing only if wp.config.verbose = True seems like a good solution, too.
Can you please clarify the code that produced the output in the original comment?
Input:
import torch
import warp as wp
wp.init()
a = torch.arange(10)
wp1 = wp.from_torch(a, dtype=wp.int32)
wp2 = wp.from_torch(a, dtype=wp.int64)
print(wp1)
print(wp2)
RuntimeError: Cannot convert Torch type torch.int64 to Warp type <class 'warp.types.int32'>
I'm sorry, apparently I made a mistake when copying my own code. The original issue is not clear in what I wanted to communicate. Here is an example with code that actually reproduces the issue I wanted talked about: Using from_torch works as intended. Using wp.array directly does not always give the expected results but also doesn't fail (as from_torch does).
import torch
import warp as wp
wp.init()
t_cpu = torch.arange(10, device='cpu')
t_gpu = torch.arange(10, device='cuda')
wp_1_cpu = wp.array(t_cpu, dtype=wp.int32)
wp_1_gpu = wp.array(t_gpu, dtype=wp.int32)
wp_2_cpu = wp.array(t_cpu, dtype=wp.int64)
wp_2_gpu = wp.array(t_gpu, dtype=wp.int64)
# int32 will throw an exception
wp_2_cpu_from_torch = wp.from_torch(t_cpu, dtype=wp.int64)
wp_2_gpu_from_torch = wp.from_torch(t_gpu, dtype=wp.int64)
print("Conversion as intended")
print('int64, from torch, cpu:', wp_2_cpu_from_torch)
print('int64, from torch, gpu:', wp_2_gpu_from_torch)
print("Direct conversion to array:")
print('int32, cpu:', wp_1_cpu)
print('int32, gpu:', wp_1_gpu)
print('int64, cpu:', wp_2_cpu)
print('int64, gpu:', wp_2_gpu)
Which results in
Conversion as intended
int64, from torch, cpu: [0 1 2 3 4 5 6 7 8 9]
int64, from torch, gpu: [0 1 2 3 4 5 6 7 8 9]
Direct conversion to array:
int32, cpu: [0 1 2 3 4 5 6 7 8 9]
int32, gpu: [0 0 1 0 2 0 3 0 4 0]
int64, cpu: [0 1 2 3 4 5 6 7 8 9]
int64, gpu: [0 1 2 3 4 5 6 7 8 9]
Here's what I have in a new branch (no need to flip on a verbose flag, this is the default behavior):
Warp UserWarning: The input data type int64 does not appear to be compatible with the requested dtype <class 'warp.types.int32'>
Conversion as intended
int64, from torch, cpu: [0 1 2 3 4 5 6 7 8 9]
int64, from torch, gpu: [0 1 2 3 4 5 6 7 8 9]
Direct conversion to array:
int32, cpu: [0 1 2 3 4 5 6 7 8 9]
int32, gpu: [0 0 1 0 2 0 3 0 4 0]
int64, cpu: [0 1 2 3 4 5 6 7 8 9]
int64, gpu: [0 1 2 3 4 5 6 7 8 9]
Let me know if this seems good to you.
Yes, I was looking for something like that when I opened the issues :)
This change went in with 5e32b5f0b419834761860a4ef3a7600dbea97bb0
Eventually, we might raise an exception for certain conversions, e.g. requesting a 32-bit type when the Torch array is a 64-bit type.