Skip to content

Commit 51f2259

Browse files
author
Niraj Dev Pandey
authored
Add files via upload
1 parent 4d0c1da commit 51f2259

File tree

7 files changed

+2346
-0
lines changed

7 files changed

+2346
-0
lines changed

‎SSD_model/SSD.ipynb

Lines changed: 326 additions & 0 deletions
Large diffs are not rendered by default.

‎SSD_model/ssd_alpha.ipynb

Lines changed: 309 additions & 0 deletions
Large diffs are not rendered by default.

‎SSD_model/ssd_layers.ipynb

Lines changed: 246 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,246 @@
1+
{
2+
"cells": [
3+
{
4+
"cell_type": "code",
5+
"execution_count": 1,
6+
"metadata": {},
7+
"outputs": [
8+
{
9+
"name": "stderr",
10+
"output_type": "stream",
11+
"text": [
12+
"Using TensorFlow backend.\n"
13+
]
14+
}
15+
],
16+
"source": [
17+
"\"\"\"Some special pupropse layers for SSD.\"\"\"\n",
18+
"\n",
19+
"import keras.backend as K\n",
20+
"from keras.engine.topology import InputSpec\n",
21+
"from keras.engine.topology import Layer\n",
22+
"import numpy as np\n",
23+
"import tensorflow as tf"
24+
]
25+
},
26+
{
27+
"cell_type": "markdown",
28+
"metadata": {},
29+
"source": [
30+
" Normalization layer as described in ParseNet paper.\n",
31+
"### Arguments\n",
32+
" scale: Default feature scale.\n",
33+
"### Input shape\n",
34+
" 4D tensor with shape:\n",
35+
" (samples, channels, rows, cols) if dim_ordering='th'\n",
36+
" or 4D tensor with shape:\n",
37+
" (samples, rows, cols, channels) if dim_ordering='tf'.\n",
38+
"### Output shape\n",
39+
" Same as input\n",
40+
"#### References\n",
41+
" http://cs.unc.edu/~wliu/papers/parsenet.pdf\n",
42+
"#### TODO\n",
43+
" Add possibility to have one scale for all features"
44+
]
45+
},
46+
{
47+
"cell_type": "code",
48+
"execution_count": 2,
49+
"metadata": {},
50+
"outputs": [],
51+
"source": [
52+
"class Normalize(Layer):\n",
53+
" def __init__(self, scale, **kwargs):\n",
54+
" if K.image_dim_ordering() == 'tf':\n",
55+
" self.axis = 3\n",
56+
" else:\n",
57+
" self.axis = 1\n",
58+
" self.scale = scale\n",
59+
" super(Normalize, self).__init__(**kwargs)\n",
60+
"\n",
61+
" def build(self, input_shape):\n",
62+
" self.input_spec = [InputSpec(shape=input_shape)]\n",
63+
" shape = (input_shape[self.axis],)\n",
64+
" init_gamma = self.scale * np.ones(shape)\n",
65+
" self.gamma = K.variable(init_gamma, name='{}_gamma'.format(self.name))\n",
66+
" self.trainable_weights = [self.gamma]\n",
67+
"\n",
68+
" def call(self, x, mask=None):\n",
69+
" output = K.l2_normalize(x, self.axis)\n",
70+
" output *= self.gamma\n",
71+
" return output"
72+
]
73+
},
74+
{
75+
"cell_type": "markdown",
76+
"metadata": {},
77+
"source": [
78+
" Generate the prior boxes of designated sizes and aspect ratios.\n",
79+
" ### Arguments\n",
80+
" img_size: Size of the input image as tuple (w, h).\n",
81+
" min_size: Minimum box size in pixels.\n",
82+
" max_size: Maximum box size in pixels.\n",
83+
" aspect_ratios: List of aspect ratios of boxes.\n",
84+
" flip: Whether to consider reverse aspect ratios.\n",
85+
" variances: List of variances for x, y, w, h.\n",
86+
" clip: Whether to clip the prior's coordinates\n",
87+
" such that they are within [0, 1].\n",
88+
"### Input shape\n",
89+
" 4D tensor with shape:\n",
90+
" (samples, channels, rows, cols) if dim_ordering='th'\n",
91+
" or 4D tensor with shape:\n",
92+
" (samples, rows, cols, channels) if dim_ordering='tf'\n",
93+
"### Output shape\n",
94+
" 3D tensor with shape:\n",
95+
" (samples, num_boxes, 8)\n",
96+
"### References\n",
97+
" https://arxiv.org/abs/1512.02325\n",
98+
"### TODO\n",
99+
" Add possibility not to have variances.\n",
100+
" Add Theano support"
101+
]
102+
},
103+
{
104+
"cell_type": "code",
105+
"execution_count": 3,
106+
"metadata": {},
107+
"outputs": [],
108+
"source": [
109+
"class PriorBox(Layer):\n",
110+
" def __init__(self, img_size, min_size, max_size=None, aspect_ratios=None,\n",
111+
" flip=True, variances=[0.1], clip=True, **kwargs):\n",
112+
" if K.image_dim_ordering() == 'tf':\n",
113+
" self.waxis = 2\n",
114+
" self.haxis = 1\n",
115+
" else:\n",
116+
" self.waxis = 3\n",
117+
" self.haxis = 2\n",
118+
" self.img_size = img_size\n",
119+
" if min_size <= 0:\n",
120+
" raise Exception('min_size must be positive.')\n",
121+
" self.min_size = min_size\n",
122+
" self.max_size = max_size\n",
123+
" self.aspect_ratios = [1.0]\n",
124+
" if max_size:\n",
125+
" if max_size < min_size:\n",
126+
" raise Exception('max_size must be greater than min_size.')\n",
127+
" self.aspect_ratios.append(1.0)\n",
128+
" if aspect_ratios:\n",
129+
" for ar in aspect_ratios:\n",
130+
" if ar in self.aspect_ratios:\n",
131+
" continue\n",
132+
" self.aspect_ratios.append(ar)\n",
133+
" if flip:\n",
134+
" self.aspect_ratios.append(1.0 / ar)\n",
135+
" self.variances = np.array(variances)\n",
136+
" self.clip = True\n",
137+
" super(PriorBox, self).__init__(**kwargs)\n",
138+
"\n",
139+
" def get_output_shape_for(self, input_shape):\n",
140+
" num_priors_ = len(self.aspect_ratios)\n",
141+
" layer_width = input_shape[self.waxis]\n",
142+
" layer_height = input_shape[self.haxis]\n",
143+
" num_boxes = num_priors_ * layer_width * layer_height\n",
144+
" return input_shape[0], num_boxes, 8\n",
145+
"\n",
146+
" # support for Keras 2.0\n",
147+
" def compute_output_shape(self, input_shape):\n",
148+
" return self.get_output_shape_for(input_shape)\n",
149+
"\n",
150+
" def call(self, x, mask=None):\n",
151+
" if hasattr(x, '_keras_shape'):\n",
152+
" input_shape = x._keras_shape\n",
153+
" elif hasattr(K, 'int_shape'):\n",
154+
" input_shape = K.int_shape(x)\n",
155+
" layer_width = input_shape[self.waxis]\n",
156+
" layer_height = input_shape[self.haxis]\n",
157+
" img_width = self.img_size[0]\n",
158+
" img_height = self.img_size[1]\n",
159+
" # define prior boxes shapes\n",
160+
" box_widths = []\n",
161+
" box_heights = []\n",
162+
" for ar in self.aspect_ratios:\n",
163+
" if ar == 1 and len(box_widths) == 0:\n",
164+
" box_widths.append(self.min_size)\n",
165+
" box_heights.append(self.min_size)\n",
166+
" elif ar == 1 and len(box_widths) > 0:\n",
167+
" box_widths.append(np.sqrt(self.min_size * self.max_size))\n",
168+
" box_heights.append(np.sqrt(self.min_size * self.max_size))\n",
169+
" elif ar != 1:\n",
170+
" box_widths.append(self.min_size * np.sqrt(ar))\n",
171+
" box_heights.append(self.min_size / np.sqrt(ar))\n",
172+
" box_widths = 0.5 * np.array(box_widths)\n",
173+
" box_heights = 0.5 * np.array(box_heights)\n",
174+
" # define centers of prior boxes\n",
175+
" step_x = img_width / layer_width\n",
176+
" step_y = img_height / layer_height\n",
177+
" linx = np.linspace(0.5 * step_x, img_width - 0.5 * step_x,\n",
178+
" layer_width)\n",
179+
" liny = np.linspace(0.5 * step_y, img_height - 0.5 * step_y,\n",
180+
" layer_height)\n",
181+
" centers_x, centers_y = np.meshgrid(linx, liny)\n",
182+
" centers_x = centers_x.reshape(-1, 1)\n",
183+
" centers_y = centers_y.reshape(-1, 1)\n",
184+
" # define xmin, ymin, xmax, ymax of prior boxes\n",
185+
" num_priors_ = len(self.aspect_ratios)\n",
186+
" prior_boxes = np.concatenate((centers_x, centers_y), axis=1)\n",
187+
" prior_boxes = np.tile(prior_boxes, (1, 2 * num_priors_))\n",
188+
" prior_boxes[:, ::4] -= box_widths\n",
189+
" prior_boxes[:, 1::4] -= box_heights\n",
190+
" prior_boxes[:, 2::4] += box_widths\n",
191+
" prior_boxes[:, 3::4] += box_heights\n",
192+
" prior_boxes[:, ::2] /= img_width\n",
193+
" prior_boxes[:, 1::2] /= img_height\n",
194+
" prior_boxes = prior_boxes.reshape(-1, 4)\n",
195+
" if self.clip:\n",
196+
" prior_boxes = np.minimum(np.maximum(prior_boxes, 0.0), 1.0)\n",
197+
" # define variances\n",
198+
" num_boxes = len(prior_boxes)\n",
199+
" if len(self.variances) == 1:\n",
200+
" variances = np.ones((num_boxes, 4)) * self.variances[0]\n",
201+
" elif len(self.variances) == 4:\n",
202+
" variances = np.tile(self.variances, (num_boxes, 1))\n",
203+
" else:\n",
204+
" raise Exception('Must provide one or four variances.')\n",
205+
" prior_boxes = np.concatenate((prior_boxes, variances), axis=1)\n",
206+
" prior_boxes_tensor = K.expand_dims(K.variable(prior_boxes), 0)\n",
207+
" if K.backend() == 'tensorflow':\n",
208+
" pattern = [tf.shape(x)[0], 1, 1]\n",
209+
" prior_boxes_tensor = tf.tile(prior_boxes_tensor, pattern)\n",
210+
" elif K.backend() == 'theano':\n",
211+
" #TODO\n",
212+
" pass\n",
213+
" return prior_boxes_tensor"
214+
]
215+
},
216+
{
217+
"cell_type": "code",
218+
"execution_count": null,
219+
"metadata": {},
220+
"outputs": [],
221+
"source": []
222+
}
223+
],
224+
"metadata": {
225+
"anaconda-cloud": {},
226+
"kernelspec": {
227+
"display_name": "Python [default]",
228+
"language": "python",
229+
"name": "python3"
230+
},
231+
"language_info": {
232+
"codemirror_mode": {
233+
"name": "ipython",
234+
"version": 3
235+
},
236+
"file_extension": ".py",
237+
"mimetype": "text/x-python",
238+
"name": "python",
239+
"nbconvert_exporter": "python",
240+
"pygments_lexer": "ipython3",
241+
"version": "3.5.4"
242+
}
243+
},
244+
"nbformat": 4,
245+
"nbformat_minor": 2
246+
}

0 commit comments

Comments
 (0)