add local user data conversion into full_pascalvoc_test_preprocess.py (#19283)

* add local user data conversion into full_pascalvoc_test_preprocess.py
test=develop

* change PADDLE_ENFORCE to PADDLE_ENFORCE_GE
test=develop

* change according to reviews
test=develop

paddle/fluid/inference/api/helper.h

@@ -317,7 +317,7 @@ static void PrintTime(int batch_size, int repeat, int num_threads, int tid,
                       double batch_latency, int epoch = 1,
                       const framework::proto::VarType::Type data_type =
                           framework::proto::VarType::FP32) {
-  PADDLE_ENFORCE(batch_size > 0, "Non-positive batch size.");
+  PADDLE_ENFORCE_GT(batch_size, 0, "Non-positive batch size.");
   double sample_latency = batch_latency / batch_size;
   LOG(INFO) << "====== threads: " << num_threads << ", thread id: " << tid
             << " ======";

paddle/fluid/inference/tests/api/CMakeLists.txt

@@ -250,6 +250,9 @@ if(WITH_MKLDNN)
   # download dataset if necessary
   download_int8_data(${INT8_DATA_DIR} "pascalvoc_val_head_300.tar.gz")
 
+  # download small demo set of pascalvoc for testing local userdata preprocessing
+  download_int8_data(${INT8_DATA_DIR} "pascalvoc_small.tar.gz")
+
   # build test binary to be used in subsequent tests
   inference_analysis_api_int8_test_build(${INT8_OBJ_DETECT_TEST_APP} ${INT8_OBJ_DETECT_TEST_APP_SRC})
 

paddle/fluid/inference/tests/api/analyzer_int8_object_detection_tester.cc

@@ -144,8 +144,8 @@ std::shared_ptr<std::vector<PaddleTensor>> GetWarmupData(
     int32_t num_images = FLAGS_warmup_batch_size) {
   int test_data_batch_size = test_data[0][0].shape[0];
   auto iterations = test_data.size();
-  PADDLE_ENFORCE(
-      static_cast<size_t>(num_images) <= iterations * test_data_batch_size,
+  PADDLE_ENFORCE_LE(
+      static_cast<size_t>(num_images), iterations * test_data_batch_size,
       "The requested quantization warmup data size " +
           std::to_string(num_images) + " is bigger than all test data size.");
 
@@ -235,8 +235,8 @@ std::shared_ptr<std::vector<PaddleTensor>> GetWarmupData(
                 static_cast<int64_t *>(difficult.data.data()) + objects_accum);
     objects_accum = objects_accum + objects_remain;
   }
-  PADDLE_ENFORCE(
-      static_cast<size_t>(num_objects) == static_cast<size_t>(objects_accum),
+  PADDLE_ENFORCE_EQ(
+      static_cast<size_t>(num_objects), static_cast<size_t>(objects_accum),
       "The requested num of objects " + std::to_string(num_objects) +
           " is the same as objects_accum.");
 
@@ -274,7 +274,7 @@ TEST(Analyzer_int8_mobilenet_ssd, quantization) {
   q_cfg.mkldnn_quantizer_config()->SetWarmupData(warmup_data);
   q_cfg.mkldnn_quantizer_config()->SetWarmupBatchSize(FLAGS_warmup_batch_size);
 
-  // 0 is avg_cose, 1 is top1_acc, 2 is top5_acc or mAP
+  // 0 is avg_cost, 1 is top1_acc, 2 is top5_acc or mAP
   CompareQuantizedAndAnalysis(&cfg, &q_cfg, input_slots_all, 2);
 }
 

paddle/fluid/inference/tests/api/full_pascalvoc_test_preprocess.py

@@ -11,7 +11,8 @@
 # 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 xml.etree.ElementTree as ET
+
+import xml.etree.ElementTree
 from PIL import Image
 import numpy as np
 import os
@@ -21,6 +22,7 @@ import tarfile
 import StringIO
 import hashlib
 import tarfile
+import argparse
 
 DATA_URL = "http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar"
 DATA_DIR = os.path.expanduser("~/.cache/paddle/dataset/pascalvoc/")
@@ -28,8 +30,8 @@ TAR_FILE = "VOCtest_06-Nov-2007.tar"
 TAR_PATH = os.path.join(DATA_DIR, TAR_FILE)
 RESIZE_H = 300
 RESIZE_W = 300
-mean_value = [127.5, 127.5, 127.5]
-ap_version = '11point'
+MEAN_VALUE = [127.5, 127.5, 127.5]
+AP_VERSION = '11point'
 DATA_OUT = 'pascalvoc_full.bin'
 DATA_OUT_PATH = os.path.join(DATA_DIR, DATA_OUT)
 BIN_TARGETHASH = "f6546cadc42f5ff13178b84ed29b740b"
@@ -40,10 +42,8 @@ BIN_FULLSIZE = 5348678856
 
 def preprocess(img):
     img_width, img_height = img.size
-
     img = img.resize((RESIZE_W, RESIZE_H), Image.ANTIALIAS)
     img = np.array(img)
-
     # HWC to CHW
     if len(img.shape) == 3:
         img = np.swapaxes(img, 1, 2)
@@ -51,12 +51,92 @@ def preprocess(img):
     # RBG to BGR
     img = img[[2, 1, 0], :, :]
     img = img.astype('float32')
-    img_mean = np.array(mean_value)[:, np.newaxis, np.newaxis].astype('float32')
+    img_mean = np.array(MEAN_VALUE)[:, np.newaxis, np.newaxis].astype('float32')
     img -= img_mean
     img = img * 0.007843
     return img
 
 
+def convert_pascalvoc_local2bin(args):
+    data_dir = os.path.expanduser(args.data_dir)
+    label_fpath = os.path.join(data_dir, args.label_file)
+    flabel = open(label_fpath)
+    label_list = [line.strip() for line in flabel]
+
+    img_annotation_list_path = os.path.join(data_dir, args.img_annotation_list)
+    flist = open(img_annotation_list_path)
+    lines = [line.strip() for line in flist]
+
+    output_file_path = os.path.join(data_dir, args.output_file)
+    f1 = open(output_file_path, "w+b")
+    f1.seek(0)
+    image_nums = len(lines)
+    f1.write(np.array(image_nums).astype('int64').tobytes())
+
+    boxes = []
+    lbls = []
+    difficults = []
+    object_nums = []
+
+    for line in lines:
+        image_path, label_path = line.split()
+        image_path = os.path.join(data_dir, image_path)
+        label_path = os.path.join(data_dir, label_path)
+
+        im = Image.open(image_path)
+        if im.mode == 'L':
+            im = im.convert('RGB')
+        im_width, im_height = im.size
+
+        im = preprocess(im)
+        np_im = np.array(im)
+        f1.write(np_im.astype('float32').tobytes())
+
+        # layout: label | xmin | ymin | xmax | ymax | difficult
+        bbox_labels = []
+        root = xml.etree.ElementTree.parse(label_path).getroot()
+
+        objects = root.findall('object')
+        objects_size = len(objects)
+        object_nums.append(objects_size)
+
+        for object in objects:
+            bbox_sample = []
+            # start from 1
+            bbox_sample.append(
+                float(label_list.index(object.find('name').text)))
+            bbox = object.find('bndbox')
+            difficult = float(object.find('difficult').text)
+            bbox_sample.append(float(bbox.find('xmin').text) / im_width)
+            bbox_sample.append(float(bbox.find('ymin').text) / im_height)
+            bbox_sample.append(float(bbox.find('xmax').text) / im_width)
+            bbox_sample.append(float(bbox.find('ymax').text) / im_height)
+            bbox_sample.append(difficult)
+            bbox_labels.append(bbox_sample)
+
+        bbox_labels = np.array(bbox_labels)
+        if len(bbox_labels) == 0: continue
+
+        lbls.extend(bbox_labels[:, 0])
+        boxes.extend(bbox_labels[:, 1:5])
+        difficults.extend(bbox_labels[:, -1])
+
+    f1.write(np.array(object_nums).astype('uint64').tobytes())
+    f1.write(np.array(lbls).astype('int64').tobytes())
+    f1.write(np.array(boxes).astype('float32').tobytes())
+    f1.write(np.array(difficults).astype('int64').tobytes())
+    f1.close()
+
+    object_nums_sum = sum(object_nums)
+    target_size = 8 + image_nums * 3 * args.resize_h * args.resize_h * 4 + image_nums * 8 + object_nums_sum * (
+        8 + 4 * 4 + 8)
+    if (os.path.getsize(output_file_path) == target_size):
+        print("Success! \nThe output binary file can be found at: ",
+              output_file_path)
+    else:
+        print("Conversion failed!")
+
+
 def print_processbar(done_percentage):
     done_filled = done_percentage * '='
     empty_filled = (100 - done_percentage) * ' '
@@ -65,7 +145,7 @@ def print_processbar(done_percentage):
     sys.stdout.flush()
 
 
-def convert_pascalvoc(tar_path, data_out_path):
+def convert_pascalvoc_tar2bin(tar_path, data_out_path):
     print("Start converting ...\n")
     images = {}
     gt_labels = {}
@@ -87,12 +167,12 @@ def convert_pascalvoc(tar_path, data_out_path):
     f_test = tar.extractfile(TEST_LIST_KEY).read()
     lines = f_test.split('\n')
     del lines[-1]
-    line_len = len(lines)
-    per_percentage = line_len / 100
+    image_nums = len(lines)
+    per_percentage = image_nums / 100
 
     f1 = open(data_out_path, "w+b")
     f1.seek(0)
-    f1.write(np.array(line_len).astype('int64').tobytes())
+    f1.write(np.array(image_nums).astype('int64').tobytes())
     for tarInfo in tar:
         if tarInfo.isfile():
             tmp_filename = tarInfo.name
@@ -115,7 +195,7 @@ def convert_pascalvoc(tar_path, data_out_path):
 
         # layout: label | xmin | ymin | xmax | ymax | difficult
         bbox_labels = []
-        root = ET.fromstring(gt_labels[name_prefix])
+        root = xml.etree.ElementTree.fromstring(gt_labels[name_prefix])
 
         objects = root.findall('object')
         objects_size = len(objects)
@@ -179,9 +259,48 @@ def run_convert():
             retry = retry + 1
         else:
             download_pascalvoc(DATA_URL, DATA_DIR, TAR_TARGETHASH, TAR_PATH)
-            convert_pascalvoc(TAR_PATH, DATA_OUT_PATH)
-    print("Success! \nThe binary file can be found at %s\n" % DATA_OUT_PATH)
+            convert_pascalvoc_tar2bin(TAR_PATH, DATA_OUT_PATH)
+    print("Success!\nThe binary file can be found at %s\n" % DATA_OUT_PATH)
+
+
+def main_pascalvoc_preprocess(args):
+    parser = argparse.ArgumentParser(
+        description="Convert the full pascalvoc val set or local data to binary file."
+    )
+    parser.add_argument(
+        '--choice', choices=['local', 'VOC_test_2007'], required=True)
+    parser.add_argument(
+        "--data_dir",
+        default="/home/li/AIPG-Paddle/paddle/build/third_party/inference_demo/int8v2/pascalvoc_small",
+        type=str,
+        help="Dataset root directory")
+    parser.add_argument(
+        "--img_annotation_list",
+        type=str,
+        default="test_100.txt",
+        help="A file containing the image file path and relevant annotation file path"
+    )
+    parser.add_argument(
+        "--label_file",
+        type=str,
+        default="label_list",
+        help="List the labels in the same sequence as denoted in the annotation file"
+    )
+    parser.add_argument(
+        "--output_file",
+        type=str,
+        default="pascalvoc_small.bin",
+        help="File path of the output binary file")
+    parser.add_argument("--resize_h", type=int, default=RESIZE_H)
+    parser.add_argument("--resize_w", type=int, default=RESIZE_W)
+    parser.add_argument("--mean_value", type=str, default=MEAN_VALUE)
+    parser.add_argument("--ap_version", type=str, default=AP_VERSION)
+    args = parser.parse_args()
+    if args.choice == 'local':
+        convert_pascalvoc_local2bin(args)
+    elif args.choice == 'VOC_test_2007':
+        run_convert()
 
 
 if __name__ == "__main__":
-    run_convert()
+    main_pascalvoc_preprocess(sys.argv)

paddle/fluid/inference/tests/api/test_detection_dataset_preprocess.py

@@ -0,0 +1,35 @@
+# 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.
+
+from __future__ import print_function
+
+from full_pascalvoc_test_preprocess import main_pascalvoc_preprocess
+import numpy as np
+import paddle.fluid.core as core
+import paddle.fluid as fluid
+import unittest
+import os
+
+
+class Test_Preprocess(unittest.TestCase):
+    def test_local_convert(self):
+        os.system("python full_pascalvoc_test_preprocess.py --choice=local")
+
+    def test_online_convert(self):
+        os.system(
+            "python full_pascalvoc_test_preprocess.py --choice=VOC_test_2007")
+
+
+if __name__ == '__main__':
+    unittest.main()

paddle/fluid/inference/tests/api/tester_helper.h

@@ -445,9 +445,12 @@ void TestPrediction(const PaddlePredictor::Config *config,
 
 void SummarizeAccuracy(float avg_acc_fp32, float avg_acc_int8,
                        int compared_idx) {
-  PADDLE_ENFORCE(compared_idx <= 2 && compared_idx >= 1,
-                 "Compare either top1 accuracy either mAP(top5), the "
-                 "compared_idx is out of range");
+  PADDLE_ENFORCE_LE(compared_idx, 2,
+                    "Compare either top1 accuracy or mAP (top5), the "
+                    "compared_idx is out of range");
+  PADDLE_ENFORCE_GE(compared_idx, 1,
+                    "Compare either top1 accuracy or mAP (top5), the "
+                    "compared_idx is out of range");
   std::string prefix = (compared_idx == 1) ? "top1_accuracy " : "mAP ";
   LOG(INFO) << "--- Accuracy summary --- ";
   LOG(INFO) << "Accepted " << prefix
@@ -485,8 +488,23 @@ void CompareAccuracy(
   float total_accs_quant{0};
   float total_accs_ref{0};
   for (size_t i = 0; i < output_slots_quant.size(); ++i) {
-    PADDLE_ENFORCE(output_slots_quant[i].size() >= 2UL);
-    PADDLE_ENFORCE(output_slots_ref[i].size() >= 2UL);
+    if (compared_idx == 1) {
+      PADDLE_ENFORCE_GE(
+          output_slots_quant[i].size(), 2UL,
+          "To achieve top 1 accuracy, output_slots_quant[i].size()>=2");
+      PADDLE_ENFORCE_GE(
+          output_slots_ref[i].size(), 2UL,
+          "To achieve top 1 accuracy, output_slots_ref[i].size()>=2");
+    } else if (compared_idx == 2) {
+      PADDLE_ENFORCE_GE(output_slots_quant[i].size(), 3UL,
+                        "To achieve mAP, output_slots_quant[i].size()>=3");
+      PADDLE_ENFORCE_GE(output_slots_ref[i].size(), 3UL,
+                        "To achieve mAP, output_slots_ref[i].size()>=3");
+    } else {
+      throw std::invalid_argument(
+          "CompareAccuracy: compared_idx is out of range.");
+    }
+
     if (output_slots_quant[i][compared_idx].lod.size() > 0 ||
         output_slots_ref[i][compared_idx].lod.size() > 0)
       throw std::invalid_argument("CompareAccuracy: output has nonempty LoD.");
@@ -535,8 +553,8 @@ void CompareNativeAndAnalysis(
   std::vector<std::vector<PaddleTensor>> native_outputs, analysis_outputs;
   TestOneThreadPrediction(config, inputs, &native_outputs, false);
   TestOneThreadPrediction(config, inputs, &analysis_outputs, true);
-  PADDLE_ENFORCE(native_outputs.size() > 0, "Native output is empty.");
-  PADDLE_ENFORCE(analysis_outputs.size() > 0, "Analysis output is empty.");
+  PADDLE_ENFORCE_GT(native_outputs.size(), 0, "Native output is empty.");
+  PADDLE_ENFORCE_GT(analysis_outputs.size(), 0, "Analysis output is empty.");
   CompareResult(analysis_outputs.back(), native_outputs.back());
 }