Bug Neuropod doesn't deal correctly with non-contiguous numpy arra

Put up a fix in <a class="issue-link js-issue-link" data-error-text="Failed to load ti

Thanks for the bug report and the repro ! I'm pretty sure this

Wrong results with non-contiguous arrays about neuropod HOT 3 CLOSED

uber commented on May 18, 2024

Wrong results with non-contiguous arrays

from neuropod.

Comments (3)

VivekPanyam commented on May 18, 2024 1

Put up a fix in #399. Thanks again for the detailed bug report and repro!

from neuropod.

VivekPanyam commented on May 18, 2024

Thanks for the bug report and the repro script!

I'm pretty sure this behavior is because of this TODO:

neuropod/source/neuropod/bindings/python_bindings.cc

Lines 102 to 130 in b3a5291

 std::shared_ptr<NeuropodTensor> tensor_from_numpy(NeuropodTensorAllocator &allocator, py::array array) 

 { 

 // TODO(vip): Make sure it's contiguous and aligned 

 auto ndims = array.ndim(); 

 auto dims = array.shape(); 

 auto dtype = get_array_type(array); 

 auto data = array.mutable_data(); 

 // Capture the array in our deleter so it doesn't get deallocated 

 // until we're done 

 auto to_delete = std::make_shared<py::array>(array); 

 auto deleter = [to_delete](void *unused) mutable { 

 py::gil_scoped_acquire gil; 

 to_delete.reset(); 

 }; 

 // Create a vector with the shape info 

 std::vector<int64_t> shape(&dims[0], &dims[ndims]); 

 if (dtype != STRING_TENSOR) 

 { 

 // Wrap the data from the numpy array 

 return allocator.tensor_from_memory(shape, dtype, data, deleter); 

 } 

 else 

 { 

 return tensor_from_string_numpy(allocator, array, shape); 

 } 

 }

In addition to fixing that TODO (ensure contiguous and aligned), I also want to look into correctly propagating the strides through into Neuropod.

Unfortunately, there are many cases where we depend (or need to depend) on data being contiguous so I'm not sure whether propagating the stride information will be worth it (if we always need to convert into a contiguous tensor later for example). I'll create another issue about this to describe some of the pros and cons.

Thanks again for the detailed bug report!

from neuropod.

dsuess commented on May 18, 2024

That TODO does look pretty much on point :D

Maybe it's worth just putting in a error for now if the data is not contiguous. It took me a while to figure out because the testcase run in create_python_neuropod passes. If a user is alerted when the neuropod is created, it's pretty easy to fix.

from neuropod.

Recommend Projects

Wrong results with non-contiguous arrays about neuropod HOT 3 CLOSED

Comments (3)

Related Issues (20)

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent

	std::shared_ptr<NeuropodTensor> tensor_from_numpy(NeuropodTensorAllocator &allocator, py::array array)
	{
	// TODO(vip): Make sure it's contiguous and aligned
	auto ndims = array.ndim();
	auto dims = array.shape();
	auto dtype = get_array_type(array);
	auto data = array.mutable_data();

	// Capture the array in our deleter so it doesn't get deallocated
	// until we're done
	auto to_delete = std::make_shared<py::array>(array);
	auto deleter = [to_delete](void *unused) mutable {
	py::gil_scoped_acquire gil;
	to_delete.reset();
	};

	// Create a vector with the shape info
	std::vector<int64_t> shape(&dims[0], &dims[ndims]);

	if (dtype != STRING_TENSOR)
	{
	// Wrap the data from the numpy array
	return allocator.tensor_from_memory(shape, dtype, data, deleter);
	}
	else
	{
	return tensor_from_string_numpy(allocator, array, shape);
	}
	}