rok commented on code in PR #8510: URL: https://github.com/apache/arrow/pull/8510#discussion_r1130837294
########## cpp/src/arrow/extension/fixed_shape_tensor.cc: ########## @@ -0,0 +1,299 @@ +// Licensed to the Apache Software Foundation (ASF) under one +// or more contributor license agreements. See the NOTICE file +// distributed with this work for additional information +// regarding copyright ownership. The ASF licenses this file +// to you under the Apache License, Version 2.0 (the +// "License"); you may not use this file except in compliance +// with the License. You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, +// software distributed under the License is distributed on an +// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +// KIND, either express or implied. See the License for the +// specific language governing permissions and limitations +// under the License. + +#include "arrow/extension/fixed_shape_tensor.h" + +#include "arrow/array/array_nested.h" +#include "arrow/array/array_primitive.h" +#include "arrow/json/rapidjson_defs.h" // IWYU pragma: keep +#include "arrow/tensor.h" +#include "arrow/util/logging.h" +#include "arrow/util/sort.h" + +#include <rapidjson/document.h> +#include <rapidjson/writer.h> + +namespace rj = arrow::rapidjson; + +namespace arrow { +namespace extension { + +bool FixedShapeTensorType::ExtensionEquals(const ExtensionType& other) const { + if (extension_name() != other.extension_name()) { + return false; + } + const auto& other_ext = static_cast<const FixedShapeTensorType&>(other); + + auto is_permutation_trivial = [](const std::vector<int64_t>& permutation) { + for (size_t i = 1; i < permutation.size(); ++i) { + if (permutation[i - 1] + 1 != permutation[i]) { + return false; + } + } + return true; + }; + const bool permutation_equivalent = + (permutation_ == other_ext.permutation()) || + ((permutation_.empty() && is_permutation_trivial(other_ext.permutation())) && + (is_permutation_trivial(permutation_) || other_ext.permutation().empty())); + + return storage_type()->Equals(other_ext.storage_type()) && + shape_ == other_ext.shape() && dim_names_ == other_ext.dim_names() && + permutation_equivalent; +} + +std::string FixedShapeTensorType::Serialize() const { + rj::Document document; + document.SetObject(); + rj::Document::AllocatorType& allocator = document.GetAllocator(); + + rj::Value shape(rj::kArrayType); + for (auto v : shape_) { + shape.PushBack(v, allocator); + } + document.AddMember(rj::Value("shape", allocator), shape, allocator); + + if (!permutation_.empty()) { + rj::Value permutation(rj::kArrayType); + for (auto v : permutation_) { + permutation.PushBack(v, allocator); + } + document.AddMember(rj::Value("permutation", allocator), permutation, allocator); + } + + if (!dim_names_.empty()) { + rj::Value dim_names(rj::kArrayType); + for (std::string v : dim_names_) { + dim_names.PushBack(rj::Value{}.SetString(v.c_str(), allocator), allocator); + } + document.AddMember(rj::Value("dim_names", allocator), dim_names, allocator); + } + + rj::StringBuffer buffer; + rj::Writer<rj::StringBuffer> writer(buffer); + document.Accept(writer); + return buffer.GetString(); +} + +Result<std::shared_ptr<DataType>> FixedShapeTensorType::Deserialize( + std::shared_ptr<DataType> storage_type, const std::string& serialized_data) const { + if (storage_type->id() != Type::FIXED_SIZE_LIST) { + return Status::Invalid("Expected FixedSizeList storage type, got ", + storage_type->ToString()); + } + auto value_type = + internal::checked_pointer_cast<FixedSizeListType>(storage_type)->value_type(); + rj::Document document; + if (document.Parse(serialized_data.data(), serialized_data.length()).HasParseError() || + !document.HasMember("shape") || !document["shape"].IsArray()) { + return Status::Invalid("Invalid serialized JSON data: ", serialized_data); + } + + std::vector<int64_t> shape; + for (auto& x : document["shape"].GetArray()) { + shape.emplace_back(x.GetInt64()); + } + std::vector<int64_t> permutation; + if (document.HasMember("permutation")) { + for (auto& x : document["permutation"].GetArray()) { + permutation.emplace_back(x.GetInt64()); + } + if (shape.size() != permutation.size()) { + return Status::Invalid("Invalid permutation"); + } + } + std::vector<std::string> dim_names; + if (document.HasMember("dim_names")) { + for (auto& x : document["dim_names"].GetArray()) { + dim_names.emplace_back(x.GetString()); + } + if (shape.size() != dim_names.size()) { + return Status::Invalid("Invalid dim_names"); + } + } + + return fixed_shape_tensor(value_type, shape, permutation, dim_names); +} + +std::shared_ptr<Array> FixedShapeTensorType::MakeArray( + std::shared_ptr<ArrayData> data) const { + return std::make_shared<ExtensionArray>(data); +} + +Result<std::shared_ptr<Array>> FixedShapeTensorType::MakeArray( + std::shared_ptr<Tensor> tensor) const { + auto permutation = internal::ArgSort(tensor->strides()); + std::reverse(permutation.begin(), permutation.end()); + if (permutation[0] != 0) { + return Status::Invalid( + "Only first-major tensors can be zero-copy converted to arrays"); + } + + auto cell_shape = tensor->shape(); + cell_shape.erase(cell_shape.begin()); + if (cell_shape != shape_) { + return Status::Invalid("Expected cell shape does not match input tensor shape"); + } + + permutation.erase(permutation.begin()); + for (auto& x : permutation) { + x--; + } + + auto ext_type = internal::checked_pointer_cast<ExtensionType>( + fixed_shape_tensor(tensor->type(), cell_shape, permutation, tensor->dim_names())); + + std::shared_ptr<FixedSizeListArray> arr; + std::shared_ptr<Array> value_array; + switch (tensor->type_id()) { + case Type::UINT8: { + value_array = std::make_shared<UInt8Array>(tensor->size(), tensor->data()); + break; + } + case Type::INT8: { + value_array = std::make_shared<Int8Array>(tensor->size(), tensor->data()); + break; + } + case Type::UINT16: { + value_array = std::make_shared<UInt16Array>(tensor->size(), tensor->data()); + break; + } + case Type::INT16: { + value_array = std::make_shared<Int16Array>(tensor->size(), tensor->data()); + break; + } + case Type::UINT32: { + value_array = std::make_shared<UInt32Array>(tensor->size(), tensor->data()); + break; + } + case Type::INT32: { + value_array = std::make_shared<Int32Array>(tensor->size(), tensor->data()); + break; + } + case Type::UINT64: { + value_array = std::make_shared<Int64Array>(tensor->size(), tensor->data()); + break; + } + case Type::INT64: { + value_array = std::make_shared<Int64Array>(tensor->size(), tensor->data()); + break; + } + case Type::HALF_FLOAT: { + value_array = std::make_shared<HalfFloatArray>(tensor->size(), tensor->data()); + break; + } + case Type::FLOAT: { + value_array = std::make_shared<FloatArray>(tensor->size(), tensor->data()); + break; + } + case Type::DOUBLE: { + value_array = std::make_shared<DoubleArray>(tensor->size(), tensor->data()); + break; + } + default: { + return Status::NotImplemented("Unsupported tensor type: ", + tensor->type()->ToString()); + } + } + arr = std::make_shared<FixedSizeListArray>(ext_type->storage_type(), tensor->shape()[0], + value_array); + auto ext_data = arr->data(); + ext_data->type = ext_type; + return MakeArray(ext_data); +} + +Result<std::shared_ptr<Tensor>> FixedShapeTensorType::ToTensor( + std::shared_ptr<Array> arr) { + // To convert an array of n dimensional tensors to a n+1 dimensional tensor we + // interpret the array's length as the first dimension the new tensor. Further, we + // define n+1 dimensional tensor's strides by front appending a new stride to the n + // dimensional tensor's strides. + + ARROW_CHECK(is_tensor_supported(this->value_type_->id())); + + ARROW_DCHECK_EQ(arr->null_count(), 0) << "Null values not supported in tensors."; + auto ext_arr = internal::checked_pointer_cast<FixedSizeListArray>( + internal::checked_pointer_cast<ExtensionArray>(arr)->storage()); + + std::vector<int64_t> shape = this->shape(); + shape.insert(shape.begin(), 1, arr->length()); + + std::vector<int64_t> tensor_strides = this->strides(); + tensor_strides.insert(tensor_strides.begin(), 1, arr->length() * tensor_strides[0]); + + std::shared_ptr<Buffer> buffer = ext_arr->values()->data()->buffers[1]; + return *Tensor::Make(ext_arr->value_type(), buffer, shape, tensor_strides, dim_names()); Review Comment: > Won't that give wrong results? (I am not super familiar with the Tensor implementation , though) The actual buffer that is being passed to Tensor::Make is row-major, and the doc comment of Tensor::Make says about the `strides` param that "if this is empty, the data assumed to be row-major". So if it is not empty and not a standard strides (but actually permuted strides), won't that give a wrong interpretation of the data? Tensor's strides are arbitrary as long they are valid ([CheckTensorStridesValidity](https://github.com/apache/arrow/blob/2ff4e3a2523bd0c58168d6ca4bcb14f45393ff2b/cpp/src/arrow/tensor.cc#L154-L191)). If we have elements with non-trivial strides we'd probably want to keep those (adding the 0-th stride to account for the "array length dimension") when calling `ToTensor`? If we want to avoid this we could also permute dim names so as to map to the physical layout as `Tensor` doesn't have permutation property (should we consider adding it?). > (also, if you edit strides, I would assume you also have to change the shape? And the dim_names) `ToTensor` currently amends `strides`, `shape` and `dim_names`. > Is there a test that covers this? I only see one ToTensor call in the tests (that is not checking an error) in `RoundtripTensor`, and I think that is not using custom strides? I'll add some. They will more or less just trigger error checks in `Tensor::Make` . > (personally, I would just leave any handling of strides/permutation to the user/application, and they can do that after getting the Tensor / numpy array) I think without handling strides in `ToTensor` we can't roundtrip numpy arrays with non-trivial arrays? -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
