↰ Return to documentation for file (morpheus/_lib/src/stages/preprocess_fil.cpp)
/*
* SPDX-FileCopyrightText: Copyright (c) 2021-2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* SPDX-License-Identifier: Apache-2.0
*
* Licensed 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 "morpheus/stages/preprocess_fil.hpp"
#include "mrc/segment/object.hpp" // for Object
#include "morpheus/messages/control.hpp" // for ControlMessage
#include "morpheus/messages/memory/inference_memory_fil.hpp" // for InferenceMemoryFIL
#include "morpheus/messages/memory/tensor_memory.hpp" // for TensorMemory
#include "morpheus/messages/meta.hpp" // for MessageMeta
#include "morpheus/messages/multi.hpp" // for MultiMessage
#include "morpheus/messages/multi_inference.hpp" // for MultiInferenceMessage
#include "morpheus/objects/dev_mem_info.hpp" // for DevMemInfo
#include "morpheus/objects/dtype.hpp" // for DType, TypeId
#include "morpheus/objects/table_info.hpp" // for TableInfo, MutableTableInfo
#include "morpheus/objects/tensor.hpp" // for Tensor
#include "morpheus/objects/tensor_object.hpp" // for TensorObject
#include "morpheus/types.hpp" // for TensorIndex
#include "morpheus/utilities/matx_util.hpp" // for MatxUtil
#include <cuda_runtime.h> // for cudaMemcpy, cudaMemcpyKind
#include <cudf/column/column.hpp> // for column
#include <cudf/column/column_view.hpp> // for column_view
#include <cudf/types.hpp> // for type_id, data_type
#include <cudf/unary.hpp> // for cast
#include <glog/logging.h> // for COMPACT_GOOGLE_LOG_ERROR, LOG, LogMessage
#include <mrc/cuda/common.hpp> // for __check_cuda_errors, MRC_CHECK_CUDA
#include <mrc/segment/builder.hpp> // for Builder
#include <pybind11/gil.h> // for gil_scoped_acquire
#include <pybind11/pybind11.h> // for object_api::operator(), operator""_a, arg
#include <pybind11/pytypes.h> // for object, str, object_api, generic_item, literals
#include <rmm/cuda_stream_view.hpp> // for cuda_stream_per_thread
#include <rmm/device_buffer.hpp> // for device_buffer
#include <algorithm> // for find
#include <cstddef> // for size_t
#include <memory> // for shared_ptr, __shared_ptr_access, allocator, mak...
#include <stdexcept> // for runtime_error
#include <type_traits> // for is_same_v
#include <typeinfo> // for type_info
#include <utility> // for move
namespace morpheus {
// Component public implementations
// ************ PreprocessFILStage ************************* //
template <typename InputT, typename OutputT>
PreprocessFILStage<InputT, OutputT>::PreprocessFILStage(const std::vector<std::string>& features) :
base_t(rxcpp::operators::map([this](sink_type_t x) {
return this->on_data(std::move(x));
})),
m_fea_cols(std::move(features))
{}
template <typename InputT, typename OutputT>
void PreprocessFILStage<InputT, OutputT>::transform_bad_columns(std::vector<std::string>& fea_cols,
morpheus::MutableTableInfo& mutable_info)
{
auto df_meta_col_names = mutable_info.get_column_names();
std::vector<std::string> bad_cols;
// Only check the feature columns. Leave the rest unchanged
for (auto& fea_col : fea_cols)
{
// Find the index of the column in the dataframe
auto col_idx =
std::find(df_meta_col_names.begin(), df_meta_col_names.end(), fea_col) - df_meta_col_names.begin();
if (col_idx == df_meta_col_names.size())
{
// This feature was not found. Ignore it.
continue;
}
if (mutable_info.get_column(col_idx).type().id() == cudf::type_id::STRING)
{
bad_cols.push_back(fea_col);
}
}
// Exit early if there is nothing to do
if (!bad_cols.empty())
{
// Need to ensure all string columns have been converted to numbers. This requires running a
// regex which is too difficult to do from C++ at this time. So grab the GIL, make the
// conversions, and release. This is horribly inefficient, but so is the JSON lines format for
// this workflow
using namespace pybind11::literals;
pybind11::gil_scoped_acquire gil;
// pybind11::object df = x->meta->get_py_table();
auto pdf = mutable_info.checkout_obj();
auto& df = *pdf;
std::string regex = R"((\d+))";
for (auto c : bad_cols)
{
df[pybind11::str(c)] = df[pybind11::str(c)]
.attr("str")
.attr("extract")(pybind11::str(regex), "expand"_a = true)
.attr("astype")(pybind11::str("float32"));
}
mutable_info.return_obj(std::move(pdf));
}
}
template <typename InputT, typename OutputT>
TableInfo PreprocessFILStage<InputT, OutputT>::fix_bad_columns(sink_type_t x)
{
if constexpr (std::is_same_v<sink_type_t, std::shared_ptr<MultiMessage>>)
{
{
// Get the mutable info for the entire meta object so we only do this once per dataframe
auto mutable_info = x->meta->get_mutable_info();
transform_bad_columns(this->m_fea_cols, mutable_info);
}
// Now re-get the meta
return x->get_meta(m_fea_cols);
}
else if constexpr (std::is_same_v<sink_type_t, std::shared_ptr<ControlMessage>>)
{
{
// Get the mutable info for the entire meta object so we only do this once per dataframe
auto mutable_info = x->payload()->get_mutable_info();
transform_bad_columns(this->m_fea_cols, mutable_info);
}
// Now re-get the meta
return x->payload()->get_info(m_fea_cols);
}
// sink_type_t not supported
else
{
std::string error_msg{"PreProcessFILStage receives unsupported input type: " + std::string(typeid(x).name())};
LOG(ERROR) << error_msg;
throw std::runtime_error(error_msg);
}
}
template <typename InputT, typename OutputT>
PreprocessFILStage<InputT, OutputT>::source_type_t PreprocessFILStage<InputT, OutputT>::on_data(sink_type_t x)
{
if constexpr (std::is_same_v<sink_type_t, std::shared_ptr<MultiMessage>>)
{
return on_multi_message(x);
}
else if constexpr (std::is_same_v<sink_type_t, std::shared_ptr<ControlMessage>>)
{
return on_control_message(x);
}
// sink_type_t not supported
else
{
std::string error_msg{"PreProcessFILStage receives unsupported input type: " + std::string(typeid(x).name())};
LOG(ERROR) << error_msg;
throw std::runtime_error(error_msg);
}
}
template <>
std::shared_ptr<MultiInferenceMessage> PreprocessFILStage<MultiMessage, MultiInferenceMessage>::on_multi_message(
std::shared_ptr<MultiMessage> x)
{
auto packed_data = std::make_shared<rmm::device_buffer>(m_fea_cols.size() * x->mess_count * sizeof(float),
rmm::cuda_stream_per_thread);
auto df_meta = this->fix_bad_columns(x);
for (size_t i = 0; i < df_meta.num_columns(); ++i)
{
auto curr_col = df_meta.get_column(i);
auto curr_ptr = static_cast<float*>(packed_data->data()) + i * df_meta.num_rows();
// Check if we are something other than float
if (curr_col.type().id() != cudf::type_id::FLOAT32)
{
auto float_data = cudf::cast(curr_col, cudf::data_type(cudf::type_id::FLOAT32))->release();
// Do the copy here before it goes out of scope
MRC_CHECK_CUDA(cudaMemcpy(
curr_ptr, float_data.data->data(), df_meta.num_rows() * sizeof(float), cudaMemcpyDeviceToDevice));
}
else
{
MRC_CHECK_CUDA(cudaMemcpy(curr_ptr,
curr_col.template data<float>(),
df_meta.num_rows() * sizeof(float),
cudaMemcpyDeviceToDevice));
}
}
// Need to convert from row major to column major
// Easiest way to do this is to transpose the data from [fea_len, row_count] to [row_count, fea_len]
auto transposed_data = MatxUtil::transpose(DevMemInfo{packed_data,
TypeId::FLOAT32,
{static_cast<TensorIndex>(m_fea_cols.size()), x->mess_count},
{x->mess_count, 1}});
// Create the tensor which will be row-major and size [row_count, fea_len]
auto input__0 = Tensor::create(
transposed_data, DType::create<float>(), {x->mess_count, static_cast<TensorIndex>(m_fea_cols.size())}, {}, 0);
auto seq_id_dtype = DType::create<TensorIndex>();
auto seq_ids = Tensor::create(
MatxUtil::create_seq_ids(
x->mess_count, m_fea_cols.size(), seq_id_dtype.type_id(), input__0.get_memory(), x->mess_offset),
seq_id_dtype,
{x->mess_count, 3},
{},
0);
// Build the results
auto memory = std::make_shared<InferenceMemoryFIL>(x->mess_count, std::move(input__0), std::move(seq_ids));
auto next = std::make_shared<MultiInferenceMessage>(
x->meta, x->mess_offset, x->mess_count, std::move(memory), 0, memory->count);
return next;
}
template <>
std::shared_ptr<ControlMessage> PreprocessFILStage<ControlMessage, ControlMessage>::on_control_message(
std::shared_ptr<ControlMessage> x)
{
auto num_rows = x->payload()->get_info().num_rows();
auto packed_data =
std::make_shared<rmm::device_buffer>(m_fea_cols.size() * num_rows * sizeof(float), rmm::cuda_stream_per_thread);
auto df_meta = this->fix_bad_columns(x);
for (size_t i = 0; i < df_meta.num_columns(); ++i)
{
auto curr_col = df_meta.get_column(i);
auto curr_ptr = static_cast<float*>(packed_data->data()) + i * df_meta.num_rows();
// Check if we are something other than float
if (curr_col.type().id() != cudf::type_id::FLOAT32)
{
auto float_data = cudf::cast(curr_col, cudf::data_type(cudf::type_id::FLOAT32))->release();
// Do the copy here before it goes out of scope
MRC_CHECK_CUDA(cudaMemcpy(
curr_ptr, float_data.data->data(), df_meta.num_rows() * sizeof(float), cudaMemcpyDeviceToDevice));
}
else
{
MRC_CHECK_CUDA(cudaMemcpy(curr_ptr,
curr_col.template data<float>(),
df_meta.num_rows() * sizeof(float),
cudaMemcpyDeviceToDevice));
}
}
// Need to convert from row major to column major
// Easiest way to do this is to transpose the data from [fea_len, row_count] to [row_count, fea_len]
auto transposed_data = MatxUtil::transpose(DevMemInfo{
packed_data, TypeId::FLOAT32, {static_cast<TensorIndex>(m_fea_cols.size()), num_rows}, {num_rows, 1}});
// Create the tensor which will be row-major and size [row_count, fea_len]
auto input__0 = Tensor::create(
transposed_data, DType::create<float>(), {num_rows, static_cast<TensorIndex>(m_fea_cols.size())}, {}, 0);
auto seq_id_dtype = DType::create<TensorIndex>();
auto seq_ids = Tensor::create(
MatxUtil::create_seq_ids(num_rows, m_fea_cols.size(), seq_id_dtype.type_id(), input__0.get_memory(), 0),
seq_id_dtype,
{num_rows, 3},
{},
0);
// Build the results
auto memory = std::make_shared<TensorMemory>(num_rows);
memory->set_tensor("input__0", std::move(input__0));
memory->set_tensor("seq_ids", std::move(seq_ids));
auto next = x;
next->tensors(memory);
return next;
}
template class PreprocessFILStage<MultiMessage, MultiInferenceMessage>;
template class PreprocessFILStage<ControlMessage, ControlMessage>;
// ************ PreprocessFILStageInterfaceProxy *********** //
std::shared_ptr<mrc::segment::Object<PreprocessFILStageMM>> PreprocessFILStageInterfaceProxy::init_multi(
mrc::segment::Builder& builder, const std::string& name, const std::vector<std::string>& features)
{
auto stage = builder.construct_object<PreprocessFILStageMM>(name, features);
return stage;
}
std::shared_ptr<mrc::segment::Object<PreprocessFILStageCM>> PreprocessFILStageInterfaceProxy::init_cm(
mrc::segment::Builder& builder, const std::string& name, const std::vector<std::string>& features)
{
auto stage = builder.construct_object<PreprocessFILStageCM>(name, features);
return stage;
}
} // namespace morpheus