mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 00:21:07 +01:00
f62bc01dfe
Summary: I got some tensor->variable conversion exceptions from `torch/csrc/autograd/variable.h`, which used the `TORCH_ASSERTM` macros instead of `AT_CHECK`, so they didn't have backtraces. This was such a substantial loss for debugability that I decided to update the whole codebase to use the backtrace-enabled ATen macros instead of `TORCH_ASSERT` and `JIT_ASSERT`, the latter having been an alias of the former. ezyang apaszke zdevito Pull Request resolved: https://github.com/pytorch/pytorch/pull/9575 Differential Revision: D8924566 Pulled By: goldsborough fbshipit-source-id: 7a4013b13eec9dbf024cef94cf49fca72f61d441
96 lines
3.1 KiB
C++
96 lines
3.1 KiB
C++
#include "torch/csrc/utils/tensor_flatten.h"
|
|
|
|
#include <unordered_map>
|
|
|
|
namespace torch { namespace utils {
|
|
|
|
using namespace at;
|
|
|
|
std::vector<TensorGroup> take_tensors(TensorList tensors, size_t size_limit) {
|
|
std::vector<TensorGroup> results;
|
|
results.reserve(tensors.size()); // an overapproximation, but at least we won't have to copy stuff around
|
|
std::unordered_map<at::Type*, TensorGroup> groups;
|
|
for (const auto & tensor : tensors) {
|
|
auto & type = tensor.type();
|
|
size_t tensor_size;
|
|
if (type.is_sparse()) {
|
|
const auto& indices = tensor._indices();
|
|
const auto& values = tensor._values();
|
|
tensor_size = indices.numel() * indices.type().elementSizeInBytes() +
|
|
values.numel() * indices.type().elementSizeInBytes();
|
|
} else {
|
|
tensor_size = tensor.numel() * type.elementSizeInBytes();
|
|
}
|
|
auto & type_group = groups[&type];
|
|
type_group.tensors.push_back(tensor);
|
|
type_group.size += tensor_size;
|
|
if (type_group.size + tensor_size >= size_limit) {
|
|
results.emplace_back();
|
|
std::swap(results.back(), type_group);
|
|
}
|
|
}
|
|
// End case. Look for any remaining groups and return them.
|
|
for (auto & entry : groups) {
|
|
auto & group = entry.second;
|
|
if (group.size > 0) {
|
|
results.emplace_back(std::move(group));
|
|
}
|
|
}
|
|
return results;
|
|
}
|
|
|
|
void reorder_tensors_like(std::vector<Tensor>& tensors, TensorList order) {
|
|
AT_ASSERT(tensors.size() == order.size());
|
|
std::unordered_map<at::Type*, std::vector<size_t>> type_indices;
|
|
for (size_t i = 0, num_tensors = tensors.size(); i < num_tensors; ++i)
|
|
type_indices[&tensors[i].type()].push_back(i);
|
|
|
|
std::unordered_map<at::Type*, size_t> type_used;
|
|
std::vector<Tensor> ordered_tensors;
|
|
ordered_tensors.reserve(tensors.size());
|
|
for (auto & tmpl_tensor : order) {
|
|
auto * type = &tmpl_tensor.type();
|
|
auto & indices = type_indices[type];
|
|
auto & used = type_used[type];
|
|
ordered_tensors.push_back(tensors[indices[used++]]);
|
|
}
|
|
std::swap(tensors, ordered_tensors);
|
|
}
|
|
|
|
namespace {
|
|
|
|
at::Tensor get_indices(const at::Tensor& t) {
|
|
return t._indices();
|
|
}
|
|
|
|
at::Tensor get_values(const at::Tensor& t) {
|
|
return t._values();
|
|
}
|
|
|
|
}
|
|
|
|
std::pair<at::Tensor, at::Tensor> flatten_sparse_tensors(at::TensorList tensors) {
|
|
auto flat_indices = flatten_dense_tensors(fmap(tensors, &get_indices));
|
|
auto flat_values = flatten_dense_tensors(fmap(tensors, &get_values));
|
|
return std::make_pair(flat_indices, flat_values);
|
|
}
|
|
|
|
std::vector<at::Tensor> unflatten_sparse_tensors(
|
|
const at::Tensor& flat_indices, const at::Tensor& flat_values,
|
|
at::TensorList tensors) {
|
|
if (tensors.size() == 0) return {};
|
|
|
|
auto indices = unflatten_dense_tensors(flat_indices, fmap(tensors, &get_indices));
|
|
auto values = unflatten_dense_tensors(flat_values, fmap(tensors, &get_values));
|
|
|
|
std::vector<at::Tensor> outputs;
|
|
outputs.reserve(tensors.size());
|
|
auto & type = tensors[0].type();
|
|
for (size_t i = 0, num_tensors = tensors.size(); i < num_tensors; ++i)
|
|
outputs.emplace_back(type._sparse_coo_tensor_unsafe(indices[i], values[i], tensors[i].sizes()));
|
|
return outputs;
|
|
}
|
|
|
|
|
|
}}
|