Python infer api update and add unit test (#19353)

* python inference api supports numpy and add unit test, fix unit test fail in test_slim_int8_googlenet and test_slim_int8_mobilenet

paddle/fluid/pybind/inference_api.cc

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include "paddle/fluid/pybind/inference_api.h"
+#include <pybind11/numpy.h>
 #include <pybind11/stl.h>
 #include <cstring>
 #include <iostream>
@@ -20,6 +21,7 @@
 #include <memory>
 #include <string>
 #include <unordered_set>
+#include <utility>
 #include <vector>
 #include "paddle/fluid/inference/api/analysis_predictor.h"
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
@@ -51,6 +53,81 @@ void BindAnalysisPredictor(py::module *m);
 #ifdef PADDLE_WITH_MKLDNN
 void BindMkldnnQuantizerConfig(py::module *m);
 #endif
+
+template <typename T>
+PaddleBuf PaddleBufCreate(py::array_t<T> data) {
+  PaddleBuf buf(data.size() * sizeof(T));
+  std::copy_n(static_cast<T *>(data.mutable_data()), data.size(),
+              static_cast<T *>(buf.data()));
+  return buf;
+}
+
+template <typename T>
+void PaddleBufReset(PaddleBuf &buf, py::array_t<T> data) {  // NOLINT
+  buf.Resize(data.size() * sizeof(T));
+  std::copy_n(static_cast<T *>(data.mutable_data()), data.size(),
+              static_cast<T *>(buf.data()));
+}
+
+template <typename T>
+PaddleDType PaddleTensorGetDType();
+
+template <>
+PaddleDType PaddleTensorGetDType<int32_t>() {
+  return PaddleDType::INT32;
+}
+
+template <>
+PaddleDType PaddleTensorGetDType<int64_t>() {
+  return PaddleDType::INT64;
+}
+
+template <>
+PaddleDType PaddleTensorGetDType<float>() {
+  return PaddleDType::FLOAT32;
+}
+
+template <typename T>
+PaddleTensor PaddleTensorCreate(
+    py::array_t<T> data, const std::string name = "",
+    const std::vector<std::vector<size_t>> &lod = {}, bool copy = true) {
+  PaddleTensor tensor;
+
+  if (copy) {
+    PaddleBuf buf(data.size() * sizeof(T));
+    std::copy_n(static_cast<T *>(data.mutable_data()), data.size(),
+                static_cast<T *>(buf.data()));
+    tensor.data = std::move(buf);
+  } else {
+    tensor.data = PaddleBuf(data.mutable_data(), data.size() * sizeof(T));
+  }
+
+  tensor.dtype = PaddleTensorGetDType<T>();
+  tensor.name = name;
+  tensor.lod = lod;
+  tensor.shape.resize(data.ndim());
+  std::copy_n(data.shape(), data.ndim(), tensor.shape.begin());
+
+  return tensor;
+}
+
+py::array PaddleTensorGetData(PaddleTensor &tensor) {  // NOLINT
+  py::dtype dt;
+  switch (tensor.dtype) {
+    case PaddleDType::INT32:
+      dt = py::dtype::of<int32_t>();
+      break;
+    case PaddleDType::INT64:
+      dt = py::dtype::of<int64_t>();
+      break;
+    case PaddleDType::FLOAT32:
+      dt = py::dtype::of<float>();
+      break;
+    default:
+      LOG(FATAL) << "unsupported dtype";
+  }
+  return py::array(dt, {tensor.shape}, tensor.data.data());
+}
 }  // namespace
 
 void BindInferenceApi(py::module *m) {
@@ -89,23 +166,39 @@ void BindPaddleBuf(py::module *m) {
         std::memcpy(buf.data(), static_cast<void *>(data.data()), buf.length());
         return buf;
       }))
-      .def(py::init([](std::vector<int64_t> &data) {
+      .def(py::init(&PaddleBufCreate<int32_t>))
-        auto buf = PaddleBuf(data.size() * sizeof(int64_t));
+      .def(py::init(&PaddleBufCreate<int64_t>))
-        std::memcpy(buf.data(), static_cast<void *>(data.data()), buf.length());
+      .def(py::init(&PaddleBufCreate<float>))
-        return buf;
-      }))
       .def("resize", &PaddleBuf::Resize)
       .def("reset",
            [](PaddleBuf &self, std::vector<float> &data) {
              self.Resize(data.size() * sizeof(float));
              std::memcpy(self.data(), data.data(), self.length());
            })
-      .def("reset",
+      .def("reset", &PaddleBufReset<int32_t>)
-           [](PaddleBuf &self, std::vector<int64_t> &data) {
+      .def("reset", &PaddleBufReset<int64_t>)
-             self.Resize(data.size() * sizeof(int64_t));
+      .def("reset", &PaddleBufReset<float>)
-             std::memcpy(self.data(), data.data(), self.length());
-           })
       .def("empty", &PaddleBuf::empty)
+      .def("tolist",
+           [](PaddleBuf &self, const std::string &dtype) -> py::list {
+             py::list l;
+             if (dtype == "int32") {
+               auto *data = static_cast<int32_t *>(self.data());
+               auto size = self.length() / sizeof(int32_t);
+               l = py::cast(std::vector<int32_t>(data, data + size));
+             } else if (dtype == "int64") {
+               auto *data = static_cast<int64_t *>(self.data());
+               auto size = self.length() / sizeof(int64_t);
+               l = py::cast(std::vector<int64_t>(data, data + size));
+             } else if (dtype == "float32") {
+               auto *data = static_cast<float *>(self.data());
+               auto size = self.length() / sizeof(float);
+               l = py::cast(std::vector<float>(data, data + size));
+             } else {
+               LOG(FATAL) << "unsupported dtype";
+             }
+             return l;
+           })
       .def("float_data",
            [](PaddleBuf &self) -> std::vector<float> {
              auto *data = static_cast<float *>(self.data());
@@ -127,6 +220,19 @@ void BindPaddleBuf(py::module *m) {
 void BindPaddleTensor(py::module *m) {
   py::class_<PaddleTensor>(*m, "PaddleTensor")
       .def(py::init<>())
+      .def(py::init(&PaddleTensorCreate<int32_t>), py::arg("data"),
+           py::arg("name") = "",
+           py::arg("lod") = std::vector<std::vector<size_t>>(),
+           py::arg("copy") = true)
+      .def(py::init(&PaddleTensorCreate<int64_t>), py::arg("data"),
+           py::arg("name") = "",
+           py::arg("lod") = std::vector<std::vector<size_t>>(),
+           py::arg("copy") = true)
+      .def(py::init(&PaddleTensorCreate<float>), py::arg("data"),
+           py::arg("name") = "",
+           py::arg("lod") = std::vector<std::vector<size_t>>(),
+           py::arg("copy") = true)
+      .def("as_ndarray", &PaddleTensorGetData)
       .def_readwrite("name", &PaddleTensor::name)
       .def_readwrite("shape", &PaddleTensor::shape)
       .def_readwrite("data", &PaddleTensor::data)

python/paddle/fluid/contrib/slim/quantization/mkldnn_post_training_strategy.py

@@ -86,21 +86,16 @@ class MKLDNNPostTrainingQuantStrategy(Strategy):
         # TODO (Intel) Remove limits that MKLDNNPostTrainingQuantStrategy
         # only support image classification
         num_images = len(data)
-        images = core.PaddleTensor()
-        images.name = "x"
-        images.shape = [num_images, ] + list(data[0][0].shape)
-        images.dtype = core.PaddleDType.FLOAT32
         image_data = [img.tolist() for (img, _) in data]
-        image_data = np.array(image_data).astype("float32")
+        image_data = np.array(image_data).astype("float32").reshape(
+            [num_images, ] + list(data[0][0].shape))
         image_data = image_data.ravel()
-        images.data = core.PaddleBuf(image_data.tolist())
+        images = core.PaddleTensor(image_data, "x")
+        images.shape = [num_images, ] + list(data[0][0].shape)
 
-        labels = core.PaddleTensor()
-        labels.name = "y"
-        labels.shape = [num_images, 1]
-        labels.dtype = core.PaddleDType.INT64
         label_data = [label for (_, label) in data]
-        labels.data = core.PaddleBuf(label_data)
+        labels = core.PaddleTensor(
+            np.array(label_data).astype("int64").reshape([num_images, 1]), "y")
 
         warmup_data = [images, labels]
 

python/paddle/fluid/tests/unittests/test_inference_api.py

@@ -0,0 +1,74 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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.
+
+import os, shutil
+import unittest
+import numpy as np
+import paddle.fluid as fluid
+from paddle.fluid.core import PaddleTensor
+from paddle.fluid.core import PaddleDType
+
+
+class TestInferenceApi(unittest.TestCase):
+    def test_inference_api(self):
+        tensor32 = np.random.randint(10, 20, size=[20, 2]).astype('int32')
+        paddletensor32 = PaddleTensor(tensor32)
+        value32 = np.array(paddletensor32.data.int32_data()).reshape(*[20, 2])
+        dtype32 = paddletensor32.dtype
+        self.assertEqual(value32.all(), tensor32.all())
+        self.assertEqual(dtype32, PaddleDType.INT32)
+        self.assertEqual(
+            type(paddletensor32.data.tolist('int32')), type(tensor32.tolist()))
+        self.assertEqual(
+            paddletensor32.data.tolist('int32'), tensor32.ravel().tolist())
+        self.assertEqual(type(paddletensor32.as_ndarray()), type(tensor32))
+        paddletensor32.data.reset(tensor32)
+        self.assertEqual(paddletensor32.as_ndarray().all(), tensor32.all())
+
+        tensor64 = np.random.randint(10, 20, size=[20, 2]).astype('int64')
+        paddletensor64 = PaddleTensor(tensor64)
+        value64 = np.array(paddletensor64.data.int64_data()).reshape(*[20, 2])
+        dtype64 = paddletensor64.dtype
+        self.assertEqual(value64.all(), tensor64.all())
+        self.assertEqual(dtype64, PaddleDType.INT64)
+        self.assertEqual(
+            type(paddletensor64.data.tolist('int64')), type(tensor64.tolist()))
+        self.assertEqual(
+            paddletensor64.data.tolist('int64'), tensor64.ravel().tolist())
+        self.assertEqual(type(paddletensor64.as_ndarray()), type(tensor64))
+        paddletensor64.data.reset(tensor64)
+        self.assertEqual(paddletensor64.as_ndarray().all(), tensor64.all())
+
+        tensor_float = np.random.randn(20, 2).astype('float32')
+        paddletensor_float = PaddleTensor(tensor_float)
+        value_float = np.array(paddletensor_float.data.float_data()).reshape(
+            *[20, 2])
+        dtype_float = paddletensor_float.dtype
+        self.assertEqual(value_float.all(), tensor_float.all())
+        self.assertEqual(dtype_float, PaddleDType.FLOAT32)
+        self.assertEqual(
+            type(paddletensor_float.data.tolist('float32')),
+            type(tensor_float.tolist()))
+        self.assertEqual(
+            paddletensor_float.data.tolist('float32'),
+            tensor_float.ravel().tolist())
+        self.assertEqual(
+            type(paddletensor_float.as_ndarray()), type(tensor_float))
+        paddletensor_float.data.reset(tensor_float)
+        self.assertEqual(paddletensor_float.as_ndarray().all(),
+                         tensor_float.all())
+
+
+if __name__ == '__main__':
+    unittest.main()