Single Shot Multibox Detector Keras: how to train with own dataset

(I am using linux mint)
We have installed and tested if the SSD works in the last post. Now we will train a model for ssd_keras.

At first we need an dataset. I downloaded 120 pics (.jpg) of Thora Birch of Ghost World.
I made a folder like this:

And pictures are in the folder:

From these pictures, we will make a dataset.

$ git clone https://github.com/tzutalin/labelImg.git

$ cd labelImg

$ python3.5 labelImg.py

labelImg will open. Select "Open Dir".

Choose the thora folder.

Opened.

From the top of window, "Edit" -> "Create Rectbox". And create a rectangle box on her face. The category name is "thora".

Then save it.
Create rectbox on other 119 pictures of her also.

Finished.

Go to "PASCAL_VOC" of ssd_keras and create a folder "xml". Please make sure that there is get_data_from_xml.py file in this folder. We will use this file later.

Move all of xml file of thora we created into the xml folder.

We need to change get_data_from_XML.py a bit. Use this code:

import numpy as np
import os
from xml.etree import ElementTree

class XML_preprocessor(object):

    def __init__(self, data_path):
        self.path_prefix = data_path
        self.num_classes = 1 ##changed num of the classes
        self.data = dict()
        self._preprocess_XML()

    def _preprocess_XML(self):
        filenames = os.listdir(self.path_prefix)
        for filename in filenames:
            tree = ElementTree.parse(self.path_prefix + filename)
            root = tree.getroot()
            bounding_boxes = []
            one_hot_classes = []
            size_tree = root.find('size')
            width = float(size_tree.find('width').text)
            height = float(size_tree.find('height').text)
            for object_tree in root.findall('object'):
                for bounding_box in object_tree.iter('bndbox'):
                    xmin = float(bounding_box.find('xmin').text)/width
                    ymin = float(bounding_box.find('ymin').text)/height
                    xmax = float(bounding_box.find('xmax').text)/width
                    ymax = float(bounding_box.find('ymax').text)/height
                bounding_box = [xmin,ymin,xmax,ymax]
                bounding_boxes.append(bounding_box)
                class_name = object_tree.find('name').text
                one_hot_class = self._to_one_hot(class_name)
                one_hot_classes.append(one_hot_class)
            image_name = root.find('filename').text
            bounding_boxes = np.asarray(bounding_boxes)
            one_hot_classes = np.asarray(one_hot_classes)
            image_data = np.hstack((bounding_boxes, one_hot_classes))
            self.data[image_name] = image_data

## Changed to one category.
    def _to_one_hot(self,name):
        one_hot_vector = [0] * self.num_classes
        if name == 'thora':
            one_hot_vector[0] = 1
        else:
            print('unknown label: %s' %name)

        return one_hot_vector

## example on how to use it
import pickle
data = XML_preprocessor('./xml/').data
pickle.dump(data,open('thora.pkl','wb'))

It isn't so big change. I just changed the number of categories to 1 (thora) and made it to generate thoara.pkl. Delete the original code of get_data_from_XML.py and copy and paste this code into the file.

And run it:

$ python3.5 python3.5 get_data_from_XML.py

You will get thora.pkl.

Move it to ssd_keras folder.

Now you are ready to use the .pkl file and its dataset. But we need to configure the pyrhon file to use the .pkl file and the dataset.

Go to the directory of SSD on console.

$ cd your/directory/to/ssd_keras
$ jupyter notebook

And you will see this:

Open SSD_training.ipynb.

You will see this:

Change the number of classes to "2" and the pkl file to load to "thora.pkl".
2 Classes: thora class and others class. Others class is always needed.

Also change the path_prefix and the path passed to Generator. Please note that the number 16 passed in Generator is a batch size (which means how many pictures you load at once for training). If the number of picture samples are not enough in the dataset, decrease it to smaller number. The bigger the batch size is, the better result you tend to have. But you need strong GPU if the batch size is too big.

Before running the program, make sure that you made "checkpoints" folder in ssd_keras. Learned weight will be saved in this folder (every epoch).

Run the program.

The result:

You can transform .ipynb file to .py file this way:

$ jupyter nbconvert --to python SSD_training.ipynb  

My code transformed for your reference (.py):

# coding: utf-8

# In[24]:


import cv2
import keras
from keras.applications.imagenet_utils import preprocess_input
from keras.backend.tensorflow_backend import set_session
from keras.models import Model
from keras.preprocessing import image
import matplotlib.pyplot as plt
import numpy as np
import pickle
from random import shuffle
from scipy.misc import imread
from scipy.misc import imresize
import tensorflow as tf

from ssd import SSD300
from ssd_training import MultiboxLoss
from ssd_utils import BBoxUtility

#get_ipython().run_line_magic('matplotlib', 'inline')
plt.rcParams['figure.figsize'] = (8, 8)
plt.rcParams['image.interpolation'] = 'nearest'

np.set_printoptions(suppress=True)

# config = tf.ConfigProto()
# config.gpu_options.per_process_gpu_memory_fraction = 0.9
# set_session(tf.Session(config=config))


# In[25]:


# some constants
NUM_CLASSES = 2
input_shape = (300, 300, 3)


# In[26]:


priors = pickle.load(open('prior_boxes_ssd300.pkl', 'rb'))
bbox_util = BBoxUtility(NUM_CLASSES, priors)


# In[27]:


gt = pickle.load(open('thora.pkl', 'rb'))
keys = sorted(gt.keys())
num_train = int(round(0.9 * len(keys)))
train_keys = keys[:num_train]
val_keys = keys[num_train:]
num_val = len(val_keys)


# In[28]:


class Generator(object):
    def __init__(self, gt, bbox_util,
                 batch_size, path_prefix,
                 train_keys, val_keys, image_size,
                 saturation_var=0.5,
                 brightness_var=0.5,
                 contrast_var=0.5,
                 lighting_std=0.5,
                 hflip_prob=0.5,
                 vflip_prob=0.5,
                 do_crop=True,
                 crop_area_range=[0.75, 1.0],
                 aspect_ratio_range=[3./4., 4./3.]):
        self.gt = gt
        self.bbox_util = bbox_util
        self.batch_size = batch_size
        self.path_prefix = path_prefix
        self.train_keys = train_keys
        self.val_keys = val_keys
        self.train_batches = len(train_keys)
        self.val_batches = len(val_keys)
        self.image_size = image_size
        self.color_jitter = []
        if saturation_var:
            self.saturation_var = saturation_var
            self.color_jitter.append(self.saturation)
        if brightness_var:
            self.brightness_var = brightness_var
            self.color_jitter.append(self.brightness)
        if contrast_var:
            self.contrast_var = contrast_var
            self.color_jitter.append(self.contrast)
        self.lighting_std = lighting_std
        self.hflip_prob = hflip_prob
        self.vflip_prob = vflip_prob
        self.do_crop = do_crop
        self.crop_area_range = crop_area_range
        self.aspect_ratio_range = aspect_ratio_range
     
    def grayscale(self, rgb):
        return rgb.dot([0.299, 0.587, 0.114])

    def saturation(self, rgb):
        gs = self.grayscale(rgb)
        alpha = 2 * np.random.random() * self.saturation_var
        alpha += 1 - self.saturation_var
        rgb = rgb * alpha + (1 - alpha) * gs[:, :, None]
        return np.clip(rgb, 0, 255)

    def brightness(self, rgb):
        alpha = 2 * np.random.random() * self.brightness_var
        alpha += 1 - self.saturation_var
        rgb = rgb * alpha
        return np.clip(rgb, 0, 255)

    def contrast(self, rgb):
        gs = self.grayscale(rgb).mean() * np.ones_like(rgb)
        alpha = 2 * np.random.random() * self.contrast_var
        alpha += 1 - self.contrast_var
        rgb = rgb * alpha + (1 - alpha) * gs
        return np.clip(rgb, 0, 255)

    def lighting(self, img):
        cov = np.cov(img.reshape(-1, 3) / 255.0, rowvar=False)
        eigval, eigvec = np.linalg.eigh(cov)
        noise = np.random.randn(3) * self.lighting_std
        noise = eigvec.dot(eigval * noise) * 255
        img += noise
        return np.clip(img, 0, 255)
 
    def horizontal_flip(self, img, y):
        if np.random.random() < self.hflip_prob:
            img = img[:, ::-1]
            y[:, [0, 2]] = 1 - y[:, [2, 0]]
        return img, y
 
    def vertical_flip(self, img, y):
        if np.random.random() < self.vflip_prob:
            img = img[::-1]
            y[:, [1, 3]] = 1 - y[:, [3, 1]]
        return img, y
 
    def random_sized_crop(self, img, targets):
        img_w = img.shape[1]
        img_h = img.shape[0]
        img_area = img_w * img_h
        random_scale = np.random.random()
        random_scale *= (self.crop_area_range[1] -
                         self.crop_area_range[0])
        random_scale += self.crop_area_range[0]
        target_area = random_scale * img_area
        random_ratio = np.random.random()
        random_ratio *= (self.aspect_ratio_range[1] -
                         self.aspect_ratio_range[0])
        random_ratio += self.aspect_ratio_range[0]
        w = np.round(np.sqrt(target_area * random_ratio))   
        h = np.round(np.sqrt(target_area / random_ratio))
        if np.random.random() < 0.5:
            w, h = h, w
        w = min(w, img_w)
        w_rel = w / img_w
        w = int(w)
        h = min(h, img_h)
        h_rel = h / img_h
        h = int(h)
        x = np.random.random() * (img_w - w)
        x_rel = x / img_w
        x = int(x)
        y = np.random.random() * (img_h - h)
        y_rel = y / img_h
        y = int(y)
        img = img[y:y+h, x:x+w]
        new_targets = []
        for box in targets:
            cx = 0.5 * (box[0] + box[2])
            cy = 0.5 * (box[1] + box[3])
            if (x_rel < cx < x_rel + w_rel and
                y_rel < cy < y_rel + h_rel):
                xmin = (box[0] - x_rel) / w_rel
                ymin = (box[1] - y_rel) / h_rel
                xmax = (box[2] - x_rel) / w_rel
                ymax = (box[3] - y_rel) / h_rel
                xmin = max(0, xmin)
                ymin = max(0, ymin)
                xmax = min(1, xmax)
                ymax = min(1, ymax)
                box[:4] = [xmin, ymin, xmax, ymax]
                new_targets.append(box)
        new_targets = np.asarray(new_targets).reshape(-1, targets.shape[1])
        return img, new_targets
 
    def generate(self, train=True):
        while True:
            if train:
                shuffle(self.train_keys)
                keys = self.train_keys
            else:
                shuffle(self.val_keys)
                keys = self.val_keys
            inputs = []
            targets = []
            for key in keys:         
                img_path = self.path_prefix + key
                img = imread(img_path).astype('float32')
                y = self.gt[key].copy()
                if train and self.do_crop:
                    img, y = self.random_sized_crop(img, y)
                img = imresize(img, self.image_size).astype('float32')
                if train:
                    shuffle(self.color_jitter)
                    for jitter in self.color_jitter:
                        img = jitter(img)
                    if self.lighting_std:
                        img = self.lighting(img)
                    if self.hflip_prob > 0:
                        img, y = self.horizontal_flip(img, y)
                    if self.vflip_prob > 0:
                        img, y = self.vertical_flip(img, y)
                y = self.bbox_util.assign_boxes(y)
                inputs.append(img)             
                targets.append(y)
                if len(targets) == self.batch_size:
                    tmp_inp = np.array(inputs)
                    tmp_targets = np.array(targets)
                    inputs = []
                    targets = []
                    yield preprocess_input(tmp_inp), tmp_targets


# In[29]:


path_prefix = 'thora/'
gen = Generator(gt, bbox_util, 1, 'thora/',
                train_keys, val_keys,
                (input_shape[0], input_shape[1]), do_crop=False)


# In[30]:


model = SSD300(input_shape, num_classes=NUM_CLASSES)
#model.load_weights('weights_SSD300.hdf5', by_name=True)


# In[31]:


freeze = ['input_1', 'conv1_1', 'conv1_2', 'pool1',
          'conv2_1', 'conv2_2', 'pool2',
          'conv3_1', 'conv3_2', 'conv3_3', 'pool3']#,
#           'conv4_1', 'conv4_2', 'conv4_3', 'pool4']

for L in model.layers:
    if L.name in freeze:
        L.trainable = False


# In[32]:


def schedule(epoch, decay=0.9):
    return base_lr * decay**(epoch)

callbacks = [keras.callbacks.ModelCheckpoint('./checkpoints/weights.{epoch:02d}-{val_loss:.2f}.hdf5',
                                             verbose=1,
                                             save_weights_only=True),
             keras.callbacks.LearningRateScheduler(schedule)]


# In[33]:


base_lr = 3e-4
optim = keras.optimizers.Adam(lr=base_lr)
# optim = keras.optimizers.RMSprop(lr=base_lr)
# optim = keras.optimizers.SGD(lr=base_lr, momentum=0.9, decay=decay, nesterov=True)
model.compile(optimizer=optim,
              loss=MultiboxLoss(NUM_CLASSES, neg_pos_ratio=2.0).compute_loss)


# In[34]:


nb_epoch = 10
history = model.fit_generator(gen.generate(True), gen.train_batches,
                              nb_epoch, verbose=1,
                              callbacks=callbacks,
                              validation_data=gen.generate(False),
                              nb_val_samples=gen.val_batches,
                              nb_worker=1)


# In[ ]:


inputs = []
images = []
img_path = path_prefix + sorted(val_keys)[0]
img = image.load_img(img_path, target_size=(300, 300))
img = image.img_to_array(img)
images.append(imread(img_path))
inputs.append(img.copy())
inputs = preprocess_input(np.array(inputs))


# In[35]:


preds = model.predict(inputs, batch_size=1, verbose=1)
results = bbox_util.detection_out(preds)


# In[ ]:


for i, img in enumerate(images):
    # Parse the outputs.
    det_label = results[i][:, 0]
    det_conf = results[i][:, 1]
    det_xmin = results[i][:, 2]
    det_ymin = results[i][:, 3]
    det_xmax = results[i][:, 4]
    det_ymax = results[i][:, 5]

    # Get detections with confidence higher than 0.6.
    top_indices = [i for i, conf in enumerate(det_conf) if conf >= 0.6]

    top_conf = det_conf[top_indices]
    top_label_indices = det_label[top_indices].tolist()
    top_xmin = det_xmin[top_indices]
    top_ymin = det_ymin[top_indices]
    top_xmax = det_xmax[top_indices]
    top_ymax = det_ymax[top_indices]

    colors = plt.cm.hsv(np.linspace(0, 1, 4)).tolist()

    plt.imshow(img / 255.)
    currentAxis = plt.gca()

    for i in range(top_conf.shape[0]):
        xmin = int(round(top_xmin[i] * img.shape[1]))
        ymin = int(round(top_ymin[i] * img.shape[0]))
        xmax = int(round(top_xmax[i] * img.shape[1]))
        ymax = int(round(top_ymax[i] * img.shape[0]))
        score = top_conf[i]
        label = int(top_label_indices[i])
#         label_name = voc_classes[label - 1]
        display_txt = '{:0.2f}, {}'.format(score, label)
        coords = (xmin, ymin), xmax-xmin+1, ymax-ymin+1
        color = colors[label]
        currentAxis.add_patch(plt.Rectangle(*coords, fill=False, edgecolor=color, linewidth=2))
        currentAxis.text(xmin, ymin, display_txt, bbox={'facecolor':color, 'alpha':0.5})
 
    plt.show(block=True)

Video

If you want to try the trained model on video, at first, download the video.
And use "videotest_example.py" that is in "testing_utils" folder after writing necessary information in the file (like path to the video, path to the learned weight, categories in the model and so on).
But maybe you need to add +1 to num_classes like:

model = SSD(input_shape, num_classes=NUM_CLASSES+1)

Contents

How to use SSD: Single Shot MultiBox Detector
Single Shot Multibox Detector: how to train with own dataset

Opencv: how to install

Install opencv:

$ sudo apt-get install build-essential
$ sudo apt-get install cmake git libgtk2.0-dev pkg-config libavcodec-dev libavformat-dev libswscale-dev
$ sudo apt-get install python-dev python-numpy libtbb2 libtbb-dev libjpeg-dev libpng-dev libtiff-dev libjasper-dev libdc1394-22-dev
$ sudo -H pip2 install -U pip numpy
$ sudo -H pip3 install -U pip numpy
$ git clone https://github.com/opencv/opencv.git
$ cd opencv
$ mkdir build

$ cd build

$ cmake -D CMAKE_BUILD_TYPE=RELEASE       -D CMAKE_INSTALL_PREFIX=/usr/local ..
$ make
$ sudo make install
$ sudo ldconfig

Or more simply, this might be enough:

$ sudo apt-get install python-opencv

Or:

$ conda install -c menpo opencv3

Django: how to install (python3, apache)

Install essentials

I will explain how to install Django. I'm using Linux mint.
At first, we need to make a local virtual server and install necessary things.
See here.

If you want to use nginx instead of apache, see here.

Then we will install git.

$ sudo yum install -y git

Check if python3 and pip3 are installed.

$ python3.5 -m pip --version
pip 9.0.1 from /usr/lib/python3.5/site-packages (python 3.5)

We use pyenv. Why do we need pyenv/virtualenv?

$ cd /vagrant
$ pyvenv-3.5 djangoapp

If this doesn't work and can't create the folder, it's maybe because of the permission. Stop virtual machine (vagrant halt) and run CMD and PuTTy or Teraterm as an administrator. Then try these commands again.

Start the virtual environment of pyenv:

$ source ./djangoapp/bin/activate

If you want to stop the virtual environment:

$ deactivate

But you don't need to stop it yet.

Update pip of the virtual environment (pyenv) if necessary:

$ python3.5 -m pip install --upgrade pip

Because we are in the virtual environment (pyenv), the python and pip of virtual environement will be used:

$ python --version
Python 3.5.4
$ which python
/vagrant/djangoapp/bin/python
$ pip --version
pip 9.0.1 from /vagrant/djangoapp/lib64/python3.5/site-packages (python 3.5)

and we can  use/install/update/downgrade python and python's modules in this environment without affecting the system's python libraries. (If you need to care system's and django's python/python's module versions simultaneously... it would be boring but hard task. So we simply use pyenv and separate the environments.)

Sign in to mysql and create a table:

$ mysql -u root -proot
mysql> CREATE DATABASE testdb CHARACTER SET utf8;
mysql> exit

Install Django

In the pyenv virtual environment, install django.

$ python3.5 -m pip install django

You can check the version of django.

$ python3.5 -m django --version

Install uwsgi (WSGI: Web Server Gateway Interface):

$ python3.5 -m pip install uwsgi

Create a project of Django:

$ cd /vagrant/djangoapp
$ django-admin startproject mypj

After doing these commands, you will get a project folder of Django in this directory: /vagrant/djangoapp. The name of the project folder is "mypj". And in the mypj folder, you will find that there is another mypj folder there.

In the folder....

Open the file "setting.py" that is in the folder "mypj".

Add "localhost" and "192.168.33.10" to "Allowed_Hosts" like this:

ALLOWED_HOSTS = ['localhost','192.168.33.10']

(Assuming that your virtual server's IP address is "192.168.33.10")

Run this command and start the built-in server of Django:

$ python3.5 mypj/manage.py runserver 0.0.0:8000

Go to the URL: http://192.168.33.10:8000/

If this is displayed, it means you successfully installed Django :)

Hello World

Now we stop the built-in server with ctrl+c.

Go to the mypj directory on the console.

$ cd ./mypj

There is manage,py in the top directory of django

mypj project automatically created in the "mypj" directory is a submodule of "mypj". We will create a top-level module in the directory where manage.py is.

$ python3.5 manage.py startapp mywebsite

mywebsite directory is created

In the mywebsite directory, you can see that these files are there:

Open the view.py and write as follows:

from django.shortcuts import render
# Create your views here.
from django.http import HttpResponse

def index(request):
    return HttpResponse("Hello, world Django!!!!")

Create a file called "urls.py".

Write as follows in urls.py:

from django.urls import path
from . import views

urlpatterns = [
    path('', views.index, name='index'),
]

Now go back to djangoapp/mypj/mypj (the submodule of mypj) and open urls.py.

Import "include" and add this path.

from django.contrib import admin
from django.urls import include, path

urlpatterns = [
    path('mywebsite/', include('mywebsite.urls')),
    path('admin/', admin.site.urls),

]

Make sure you at the top directory of Django.

Run the server.

$ python3.5 manage.py runserver 0.0.0:8000

Go to http://192.168.33.10:8000/mywebsite/ and you will see the message you defined in index function.

Please note, when you deploy your django project to production, you can't use "runserver". Google "how to deploy django apache".

Numba makes python greatly faster

Numba is a module developed by Anaconda and that makes python greatly faster using Just In Time Compiler (JIT compiler).

How to install

You can install Numba with pip.

$ pip3 install numba

I installed it this way:

$ python3.5 -m pip install --user numba

How to use

You can use numba with a "jit" decorator.

from numba import jit

@jit
def example():
    for ....blabla
    return something

You can specify the type of returned value (A) and the type of the arguments (B1, B2, B3, ...).

from numba import jit

@jit(A(B1,B2,...))
def example(B1, B2, ...):
    for ....blabla
    return something

Examples

Examples of Numba.

from numba import jit

@jit
def sum_jit(x):
    sum_num = 0
    for i in range(x):
        sum_num += i
    return sum_num

@jit('int64(int64)')
def sum_jit_type_specified(x):
    sum_num = 0
    for i in range(x):
        sum_num += i
    return sum_num


print(sum_jit(10000000))
print(sum_jit_type_specified(10000000))

It was incredibly fast!!

$ python3.5 test.py
49999995000000
49999995000000

Want to make .exe with Python

Maybe use Cython.
Maybe optimization like C/C++ is also possible for Cython? (like -Ofast, O2, O3 option..)

How to use Tensorflow Object Detection API

1. How to use Tensorflow Object Detection API
2. How to train for Tensorflow Object Detection API
3. How to use Tensorboard
4. How to use a trained model of TF Detect in Android

I am using Linux Mint.

At first, you need tensorflow:

$ python3.6 -m pip install tensorflow

If your pc has GPU, install tensorflow-gpu:

$ python3.6 -m pip install tensorflow-gpu

(These instructions might help you if you have problem:
Install Keras on Linux
Install Keras on windows)

Then you need to install tensorflow object detection:

$ cd ~/Documents
$ git clone https://github.com/tensorflow/models.git 

Then you can read the official start guide from
/models/research/object_detection/g3doc/installation.md

Install dependencies:

$ sudo apt-get install protobuf-compiler python-pil python-lxml python-tk
$ python3.6 -m pip install jupyter
$ python3.6 -m pip install matplotlib

$ python3.6 -m pip install pillow
$ python3.6 -m pip install lxml

Make sure you have the latest cython:

$ python3.6 -m pip install --upgrade --user cython

COCO API installation:

git clone https://github.com/cocodataset/cocoapi.git
cd cocoapi/PythonAPI

We must rewrite the "PythonAPI/Makefile".

all:
    # install pycocotools locally
python3.6 setup.py build_ext --inplace #Change the python version!
rm -rf build

install:
# install pycocotools to the Python site-packages #Change the python version!
python3.6 setup.py build_ext install
rm -rf build

Then do the make command.

make
cp -r pycocotools ~/Documents/models/research/

Protobuf Compilation:

$ cd ~/Documents/models/research/
$ protoc object_detection/protos/*.proto --python_out=.
$ export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim

Note: you need to export the PYTHONPATH every time you open the console. It's just temporarily adding it to the path.

Check if you successfully installed everything:

$ cd ~/Documents/models/research/
$ python3.6 object_detection/builders/model_builder_test.py

If it says it's OK, finally you can use the object detection API now!
Go to the research directory and start jupyter.

$ cd ~/Documents/models/research/object_detection/
$ jupyter notebook

If vscode opens instead of the browser, try "export BROWSER=/usr/bin/firefox" and "jupyter notebook".

You will see this:

Click on "object_detection_tutorial.ipynb":

Click on "Run". Every time you click on it, the program execution will proceed one by one.

If you want to close the notebook, hit "ctrl + c" on your console.

Report: the detector of tensorflow on android app

1. You can use Tensorflow Object Detector API, YOLO, Single Shot MultiBox Detector on android demo app of tensforflow.

2. But only Keras 1 is available for SSD (Single Shot MultiBox Detector) though seems like you can customize it to enable Keras 2. Maybe I will use Tensorflow Object Detector API.
EDIT: It was too difficult to make and train models for tensorflow od api. Maybe SSD or Yolo is better!

3. Tutorial of Tensorflow OD API: https://medium.com/@WuStangDan/step-by-step-tensorflow-object-detection-api-tutorial-part-1-selecting-a-model-a02b6aabe39e

Integral and Integration

What's integral

According to Weinsstein, "An integral is a mathematical object that can be interpreted as an area or a generalization of area."[1] Integration is a process of computing an integral.

How to integrate a function

Generally, this is the expression of integral:

where f(x) is a function for x. This function of x is surounded by integral's symbol and dx.

This can be solved this way:

where

If we suppose the function is x²+5, a is 2, b is 1, the integral can be solved by:

We know the value of a and b of the integral before calculation, so this called definite integral.

What if we don't know the value of a and b?

where C is a Constant number like 1, 4, 7, 20, 100, 1.5, 1/3... whatever number as long as it is constant.

This is called indefinite integral.

[1] WolframMathworld.  http://mathworld.wolfram.com/Integral.html 21 March 2018 visited.

How to use SSD: Single Shot MultiBox Detector

I'm using Linux mint OS and keras 2.1.4. The github's page of SSD is here.

Copy the project to your local.

$ git clone https://github.com/rykov8/ssd_keras.git
$ cd ssd_keras

Maybe you need

$ sudo apt-get install python-opencv

Download the learned weights from here.
And move the downloaded weight to the "ssd_keras" folder.

Open the jupyter notebook.

$ jupyter notebook

If you get "AttributeError: type object 'IOLoop' has no attribute 'initialized'.", maybe you should uninstall tornade5 and install tornade4 instead.

$ pip uninstall tornado
$ pip install tornado==4.5.3

Click on "SSD.ipynb".

And you can run the sample.
But if you want to use Keras 2, you need to customize a bit.

Open ssd_layers.py:

And change "def get_output_shape_for(self, input_shape):"

to "def compute_output_shape(self, input_shape):"

(Just copy and paste and change the function name.)


Also, on SSD.ipynb (on jupyter notebook), we need to add dummy compile to here:

model.compile('sgd','mse') is a dummy compile.

Then run the example.

References

https://github.com/rykov8/ssd_keras/issues/60
https://github.com/keras-team/keras/issues/9394

Contents

How to use SSD: Single Shot MultiBox Detector
Single Shot Multibox Detector: how to train with own dataset