Create example for SSD detection. (#997)

* Create example for SSD detection.

* Create script to convert PyTorch model to TinyDNN.

* Fix some bugs.

* Add README.

examples/CMakeLists.txt

@@ -36,7 +36,11 @@ if(USE_SERIALIZER)
     target_link_libraries(example_deconv_train
         ${project_library_target_name} ${REQUIRED_LIBRARIES})
 
-    cotire(example_mnist_train example_mnist_test example_mnist_quantized_train example_deconv_train)
+    add_executable(example_ssd_test ssd_detection/test.cpp ${tiny_dnn_headers})
+    target_link_libraries(example_ssd_test
+            ${project_library_target_name} ${REQUIRED_LIBRARIES})
+
+    cotire(example_mnist_train example_mnist_test example_mnist_quantized_train example_deconv_train example_ssd_test)
 endif()
 
 add_executable(example_deconv_visual deconv/visual.cpp ${tiny_dnn_headers})

examples/ssd_detection/README.md

@@ -0,0 +1,47 @@
+# Single Shot MultiBox Detector (SSD)
+
+A TinyDNN implementation of [Single Shot MultiBox Detector](http://arxiv.org/abs/1512.02325) from the 2016 paper by Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang, and Alexander C. Berg.  The official and original Caffe code can be found [here](https://github.com/weiliu89/caffe/tree/ssd).
+
+## Prerequisites for this example
+
+- Download PyTorch pretrained models from [here](https://s3.amazonaws.com/amdegroot-models/ssd300_mAP_77.43_v2.pth)
+- Convert PyTorch models to TinyDNN with following commands
+
+```
+mkdir models
+python convert_models.py /path/to/model models
+```
+
+## Use SSD in Object Detection
+
+```
+./example_ssd_test /folder/to/models/ /path/to/image
+```
+
+You will get output similar to following:
+
+```
+Bounding box coordinates:
+x_min = 19.6578, x_max = 66.2639, y_min = 240.801, y_max = 270.333, class = 1, score = 0.989524
+x_min = -3.73388, x_max = 302.883, y_min = 43.5821, y_max = 202.043, class = 1, score = 0.900133
+x_min = 17.771, x_max = 54.9812, y_min = 237.505, y_max = 259.987, class = 1, score = 0.660543
+```
+
+If you see the following message:
+
+```
+Failed to load weights from models/01.weights
+Failed to load weights from models/02.weights
+...
+Failed to load weights from models/18.weights
+```
+
+Please make sure the path of weight files.
+
+## Detection Results
+
+Here's an example of object detection results produced by SSD.
+
+![Airplane](https://user-images.githubusercontent.com/1730504/47263759-67bc1900-d53a-11e8-91cd-4bb4648668b7.png)
+
+![Sofa](https://user-images.githubusercontent.com/1730504/47264055-d9976100-d540-11e8-98a5-0af7871374fd.png)

examples/ssd_detection/convert_models.py

@@ -0,0 +1,77 @@
+#!/usr/bin/python3
+# -*- coding: utf-8 -*-
+
+import os
+import sys
+import torch
+
+
+nets = {
+    1: ["vgg.0", "vgg.2", "vgg.5", "vgg.7", "vgg.10", "vgg.12", "vgg.14", "vgg.17", "vgg.19", "vgg.21"],
+    2: ["vgg.24", "vgg.26", "vgg.28", "vgg.31", "vgg.33"],
+    3: ["extras.0", "extras.1"],
+    4: ["extras.2", "extras.3"],
+    5: ["extras.4", "extras.5"],
+    6: ["extras.6", "extras.7"],
+    7: ["loc.0"],
+    8: ["loc.1"],
+    9: ["loc.2"],
+    10: ["loc.3"],
+    11: ["loc.4"],
+    12: ["loc.5"],
+    13: ["conf.0"],
+    14: ["conf.1"],
+    15: ["conf.2"],
+    16: ["conf.3"],
+    17: ["conf.4"],
+    18: ["conf.5"],
+}
+
+
+def dump_layer_weights(f, weight, bias):
+    for i in range(weight.shape[0]):
+        for j in range(weight.shape[1]):
+            for k in range(weight.shape[2]):
+                for m in range(weight.shape[3]):
+                    f.write('%.24f ' % weight[i][j][k][m])
+    _ = f.write('\n')
+    for i in range(bias.shape[0]):
+        f.write('%.24f ' % bias[i])
+    _ = f.write('\n')
+
+
+
+def dump_net_weights(model_path, output_folder):
+    ckpt = torch.load(model_path)
+
+    for net_id in nets:
+        layers = nets[net_id]
+        output_file_path = os.path.join(output_folder, '%02d.weights' % net_id)
+
+        print('Saving weights to %s' % output_file_path)
+        with open(output_file_path, 'w') as f:
+            for layer in layers:
+                weight = ckpt['%s.weight' % layer]
+                bias = ckpt['%s.bias' % layer]
+                dump_layer_weights(f, weight, bias)
+
+
+def main():
+    if not len(sys.argv) == 3:
+        print('python convert_models.py model_path output_folder')
+        sys.exit(1)
+
+    model_path = sys.argv[1]
+    output_folder = sys.argv[2]
+    if not os.path.exists(model_path) or os.path.isdir(model_path):
+        print('[ERROR] Model file not exists!')
+        sys.exit(2)
+    if not os.path.exists(output_folder) or not os.path.isdir(output_folder):
+        print('[ERROR] Output folder not exists!')
+        sys.exit(2)
+
+    dump_net_weights(model_path, output_folder)
+
+
+if __name__ == '__main__':
+    main()