Deep learning 10-Let us create a semantic segmentation model(LinkNet) by PyTorch

  Deep learning, in recent years this technique take over many difficult tasks of computer vision, semantic segmentation is one of them. The first segmentation net I implement is LinkNet, it is a fast and accurate segmentation network. 

Introduction

Q : What is LinkNet?

A :  LinkNet is a convolution neural network designed for semantic segmentation. This network is 10 times faster than SegNet and more accurate.

Q : What is semantic segmentation? Any difference with segmentation?

A :  Of course they are difference. Segmentation partition image into several "similar" parts, but you do not know what are those parts presents. On the other hand, semantic segmentation partition the image into different pre-determined labels. Those labels are present as color as the end results. For example, checkout the following images(from camvid).

Q : Semantic segmentation sounds like object detection, are they the same thing?

A : No, they are not, although you may achieve the same goal by both of them.
From the aspect of tech, they use different approach. From the view of end results, semantic segmentation tell you what are those pixels are, but they do not tell you how many instance in your images, object detection show you how many instance in your images by minimal bounding box, but it do not give you delienation of objects. For example, checkout below images(from yolo).

Network architectures

  LinkNet paper describe their network architecture with excellent graphs and simple descriptions, following are the figures copy shameless from the paper.

  LinkNet adopt encoder-decoder architecture, according to the paper, LinkNet performance or come from adding the output of encoder to the decoder, this help the decoder easier to recover the information. If you want to know the details, please study section 3 of the paper, it is nice writing, very easy to understand.

Q : The paper is easy to read, but they do not explain what is full convolution, could you tell me what that means?

A :  Full convolution indicates that the neural network is composed of convolution layers and activation only, without any full connection or pooling layers. 

Q : How do they perform down-sampling without pooling layers?

A : Make the stride of convolution as 2 x 2 and do zero padding, if you cannot figure it out why this work, I suggest you create an excel file, write down some data and do some experiment.

Q : Which optimizer work best?

A : According to the paper, rmsprop is the winner, my experiments told me the same thing too, in case you are interesting, below are the graph of training loss. From left to right is rmsprop, adam, sgd. Hyper parameters are

Initial learning rate : adam and rmsprop are 5e-4, sgd is 1e-3
Augmentation : random crop(480,320) and horizontal flip
Normalize : subtract mean(based on imagenet mean value) and divided by 255
Batch size : 16
Epoch : 800
Training examples : 368

  The results of adam and rmsprop are very close. Loss of sgd steadily decrease, but it converge very slow even with higher learning rate, maybe higher learning rate would work better for SGD.

Data pre-processing

  Almost every computer vision task need you to pre-process your data, segmentation is not an exception, following are my steps.

1 : Convert the color do not exist in the category into void(0, 0, 0)
2 : Convert the color into integer
3 : Zero mean(mean value come from imagenet)

Experiment on camvid

  Enough of Q&A, let us have some benchmark and pictures😊.

Performance

  Model 1,2,3 all train with same parameters, pre-processing but with different input size when training, they are (128,128), (256,256), (512, 512). When testing, the size of the images are (960,720).

  Following are some examples, from left to right is original image, ground truth and predicted image.

  Results looks quite good and  IoU is much better than the paper, possible reasons are

1 : I augment the data by random crop and horizontal flip, the paper may use another methods or do not perform augmentation at all(?).

2 : My pre-processing are different with the paper

3 : I did not omit void when training

4 : My measurement on IoU is wrong

5 : My model is more complicated than the paper(wrong implementation)

6 : It is overfit

7 : Random shuffle training and testing data create data leakage because many images of camvid
are very similar to each other

Trained models and codes

1 : As usual, located at github.
2 :  Model trained with 368 images, 12 labels(include void), random crop (128x128),800 epoch
3 :  Model trained with 368 images, 12 labels(include void), random crop (480x320),800 epoch
4 :  Model trained with 368 images, 12 labels(include void), random crop (512x512),800 epoch

Miscellaneous

Q : Is it possible to create portable model by PyTorch?

A : It is possible, but not easy. you could check out ONNX and caffe2 if you want to try it. Someone manage to convert pytorch model to caffe model and loaded by opencv dnn. Right now opencv dnn do not support PyTorch but PyTorch. Thanks god opencv dnn can import model trained by torch  at ease(right now opencv dnn do not support nngraph).

Q : What are IoU and iIoU in the paper refer to?

A : This page give good definition, although I still can't figure out how to calculate iIoU.


  If you liked this article, please help others find it by clicking the little g+ icon below. Thanks a lot!

Deep learning 09-Performance of perceptual losses for super resolution

    Have you ever scratch your head when upscaling low resolution images? I do, because we all know the quality of the images after upscaling degrade. Thanks to the rise of machine learning in recent years, we are able to upscale single image with better results compare with traditional solutions(ex : bilinear, bicubic. You do not need to know what they are except they are apply widely in many products), we call this technique super resolution.

    This sound great, but how could we do it?I did not know it either until I study the tutorials of part2 of the marvelous Practical Deep learning for Coders, this course is fantastic to get your feet wet on deep learning.

    I will try my best to explain everything with minimal prerequisite knowledge on machine learning and computer vision, however, some knowledge of convolution neural network(cnn) is needed. The course of part1 is excellent if you want to learn cnn in depth. If you are in a hurry, pyimagesearch and medium has a short tutorial about cnn.

What is super resolution and how does it work

Q : What is super resolution

A :  Super resolution is a class of technique to enhance the resolution of images or videos.

Q : There are many softwares could help us upscale images, why do we need super resolution?

A : Traditional solutions of upscaling image apply interpolation algorithm on one image only(ex: bilinear or bicubic). In the contrast, super resolution exploit info from another source, either from contiguous frames, from the model trained by machine learning or different scale from one image.

Q : How does super resolution work

A : Super resolution I want to introduce today is based on Perceptual losses for Real-Time style Transfer and Super-Resolution.(please consult wiki if you want to study another type of super resolution).  The most interesting part of this solution is it treat super resolution as an image transformation problem(it is a process where an input image is transformed into an output image). This mean we may use the same technique to solve colorization, denoising, depth estimation, semantic segmentation and another tasks(It is not a problem if you do not know what they are).

Q : How do we transformed low resolution image to high resolution image?

A : A picture worth a thousand words.

    This network is composed by two components, image transformation network and a loss network. Image transformation network transform low resolution image into high resolution image, while loss network measuring the difference between predicted high resolution image and the true high resolution image

Q : What is the loss network anyway?Why do we use it to measure the loss?

A : Loss network is an image classification network train on imagenet (ex : vgg16, resnet, densenet). We use it to measure the loss because we want our network to better measure perceptual and semantic difference between images. The paper call the loss measure by this loss network perceptual loss.

Q : What makes the loss network able to generate better loss?

A : The loss network can generate better loss because the convolutional neural network trained for image classification have already learned to encode the perceptual and semantic information we want.

Q : The color of the image is different after upscale, how could I fixed it?

A : You could apply histogram matching as the paper mentioned, this should be able to deal with most of the cases.

Q : Any draw back of this algorithm?

A : Of course, nothing is perfect.

1 : Not all of the image work, they may look very ugly after upscale.
2 : The image maybe ice cream to your eyes, but it is not reconstructing the photo exactly but create details based on its training from example images.It is impossible to reconstruct the image with perfect results, because we have no way to retrieve the information did not exist from the beginning.
3 : Color of part of the images change after upscale, even histogram matching cannot fix it.

Q : What is histogram matching?

A : It is a way to make the color distribution of image A looks like image B.

Experiment

    All of the experiments use same network architecture and train on 80000 images from imagenet, 2 epoch. From left to right are original image, image upscale 4x by bicubic, image upscale by super resolution by 4x.

    The results are not perfect, but this is not the end, super resolution is a hot research topic, every paper is a stepping stone for next algorithm, we will see more and more better, advance techniques pop out in the future.

Sharing trained model and codes

1 : Notebook to transform the imagenet data to training data
2 : Notebook to train and use the super resolution model
3 : Network model with transformation network and loss network, trained on 80000 images

    If you liked this article, please help others find it by clicking the little g+ icon below. Thanks a lot!

Deep learning 08--Neural style by Keras

    Today I want to write down how to implement the neural style of the paper A Neural Algorithm of Artistic Style by Keras learn from fast.ai course. You can find the codes located at github.

    Before I begin to explain how to do it, I want to mentioned that generate artistic style by deep neural network is different with image classification, we need to learn new concepts and add them into our tool boxes, if you find it hard to understand at the first time you saw it, do not fear, I have the same feeling too. You can ask me the questions or go to fast ai forum.

    The paper present an algorithm to generate artistic style image by combine two image together using convolution neural network. Here are examples combine source images(bird, dog, building) with style images like starry , alice and tes_teach. From left to right is style image, source image, image combined by convolution neural network.

    Let us begin our journey of the implementation of the algorithm(I assume you know how to install Keras, tensorflow, numpy, cuda and other tools, I recommend using ubuntu16.04.x as your os, this could save you tons of headache when setup your deep learning toolbox).

Step 1 : Import file and modules

 
from PIL import Image

import os

import keras.backend as K
import vgg16_avg

from keras.models import Model
from keras.layers import *
from keras import metrics

from scipy.optimize import fmin_l_bfgs_b
from scipy.misc import imsave

Step 2 : Preprocess our input image

 
#the value of rn_mean is come from image net data set
rn_mean = np.array([123.68, 116.779, 103.939], dtype=np.float32)

#create image close to zero mean and convert rgb channel to bgr channel 
#since the vgg model need bgr channel. ::-1 invert the order of axis 0
preproc = lambda x: (x - rn_mean)[:,:,:,::-1]
#We need to undo the preprocessing before we save it to our hard disk
deproc = lambda x: x[:,:,:,::-1] + rn_mean

Step 3 : Read the source image and style image

    Source image is the image you want to apply style on it. Style image is the style you want to apply on the source image.

dpath= os.getcwd() + "/"

#I make the size of content image, style image, generated img
#have the same shape, but this is not mandatory
#since we do not use any full connection layer
def read_img(im_name, shp):
    style_img = Image.open(im_name)
    if len(shp) > 0:
        style_img = style_img.resize((shp[2], shp[1]))
    style_arr = np.array(style_img)    
    #The image read by PIL is three dimensions, but the model
    #need a four dimensions tensor(first dim is batch size)
    style_arr = np.expand_dims(style_arr, 0)
    
    return preproc(style_arr)

content_img_name = "dog"
content_img_arr = read_img(dpath + "img/{}.png".format(content_img_name), [])
content_shp = content_img_arr.shape
style_img_arr = read_img(dpath + "img/starry.png", content_shp)

Step 4 : Load vgg16_avg

    Unlike doing image classification with pure sequential api of Keras, to build a neural style network, we need to use backend api of Keras.

content_base = K.variable(content_img_arr)
style_base = K.variable(style_img_arr)
gen_img = K.placeholder(content_shp)
batch = K.concatenate([content_base, style_base, gen_img], 0)

#Feed the batch into the vgg model, every time we call the model/layer to
#generate output, it will generate output of content_base, style_base,
#gen_img. Unlike content_base and style_base, gen_img is a placeholder,
#that means we will need to provide data to this placeholder later on
model = vgg16_avg.VGG16_Avg(input_tensor = batch, include_top=False)

#build a dict of model layers
outputs = {l.name:l.output for l in model.layers}
#I prefer these 1~3 layers hierarchy as my style_layers, 
#you can try it out with different range
style_layers = [outputs['block{}_conv1'.format(i)] for i in range(1,4)]
content_layer = outputs['block4_conv2']

    If you find K.variable, K.placeholder very confuse, please check the document of TensorFlow and Keras backend api.

Step 5 : Create function to find loss and gradient

#gram matrix is a matrix collect the correlation of all of the vectors
#in a set. Check wiki(https://en.wikipedia.org/wiki/Gramian_matrix) 
#for more details
def gram_matrix(x):
    #change height,width,depth to depth, height, width, it could be 2,1,0 too
    #maybe 2,0,1 is more efficient due to underlying memory layout
    features = K.permute_dimensions(x, (2,0,1))
    #batch flatten make features become 2D array
    features = K.batch_flatten(features)
    return K.dot(features, K.transpose(features)) / x.get_shape().num_elements()    

def style_loss(x, targ):
    return metrics.mse(gram_matrix(x), gram_matrix(targ))
    
content_loss = lambda base, gen: metrics.mse(gen, base)    

#l[1] is the output(activation) of style_base, l[2] is the
#output of gen_img loss of style image and gen_img. As the
#paper suggest, we add the loss of all convolution layers
loss = sum([style_loss(l[1], l[2]) for l in style_layers]) 

#content_layer[0] is the output of content_base,
#content_layer[2] is the output of gen_img
#loss of content image and gen_img
loss += content_loss(content_layer[0], content_layer[2]) / 10. 

#The loss need two variables but we only pass in one,
#because we only got one placeholder in the graph,
#the other variable already determine by K.variable
grad = K.gradients(loss, gen_img)
#We cannot call loss and grad directly, we need
#to create a function(convert it to symbolic definition)
#before we can feed it into the solver
fn = K.function([gen_img], [loss] + grad)

    You can adjust the weight of style loss and content loss by yourself until you think the image looks good enough. The function at the end only tells you that the concatenated list of loss and grads is the output that you want to - eventually - minimize. So, when you feed it to the solver bfgs, it will try to minimize the loss and will stop when the gradients are also zero (a minimum, hopefully not just a local one).

Step 6 : Create a helper class to separate loss and gradient

#fn will return loss and grad, but fmin_l_bfgs need to seperate them
#that is why we need a class to separate loss and gradient and store them
class Evaluator:
    def __init__(self, fn_, shp_):
        self.fn = fn_
        self.shp = shp_
        
    def loss(self, x):
        loss_, grads_ = self.fn([x.reshape(self.shp)])
        self.grads = grads_.flatten().astype(np.float64)
        
        return loss_.astype(np.float64)
    
    def grad(self, x):
        return np.copy(self.grads)
    
evaluator = Evaluator(fn, content_shp)

Step 7 : Generate a random noise image(white noise image mentioned by the paper)

#This is the real value of the placeholder--gen_img
rand_img = lambda shape: np.random.uniform(-2.5, 2.5, shape)/100

Step 8 : Minimize the loss of rand_img with the source image and style image

def solve_img(evalu, niter, x):
    for i in range(0, niter):
        x, min_val, info = fmin_l_bfgs_b(evalu.loss, x.flatten(), 
                                         fprime=evalu.grad, maxfun = 20)
        #value of PIL lie within -127 and 127
        x = np.clip(x, -127, 127)
        print(i, ',Current loss value:', min_val)
        x = x.reshape(content_shp)
        simg = deproc(x.copy())
        img_name = '{}_{}_neural_style_img_{}.png'.
                    format(dpath + "gen_img/", content_img_name, i)
        imsave(img_name, simg[0])
    return x

solve_img(evaluator, 10, rand_img(content_shp)/10.)

    You may ask, why using fmin_l_bfgs_b but not stochastic gradient descent? The answer is we can, but we have a better choice. Unlike image classification, we do not have a lot of batch to run, right now we only need to figure out the loss and gradient between three inputs, they are source image, style image and the random image, using fmin_l_bfgs_b is more than enough.

 

Deep learning 07-Challenge dog vs cat fun competition of kaggle by dlib and mxnet

  Today I want to record down the experiences I learned from the dog vs cat fun competition of kaggle which I spend about two weeks on it(this is the first competition I took). My best rank in this competition is 67, this rank is close to top 5%(there are 1314 team).

  The first tool I give it a try is dlib, although this library lack a lot of features compare with another deep learning toolbox, I still like it very much, especially the fact that dlib can work as zero dependency library.

What I have learned

1 : Remember to record down the parameters you used

At first I write down the records in my header file, this make my codes become harder to read as times go on, I should save those records in excel like format from the beginning. The other thing I learn is, I should record the parameters even I am running out of times, I find out without the records, I become more panic when the dead line was closer.

2 : Feed pseudo labels into the mini-batch with naive way do not work

 I should split the data of mini-batch with some sort of ratio, like 2/3 truth labels, 1/3 pseudo labels.

3 : Leverage pretrained model is much easier to get good results

I do not use pre-trained model but train the network from scratch, this do not give me great results, especially when I cannot afford to train the image with bigger size since my gpu only got 2GB of rams, my score was 0.27468 with brand new model. To speed things up, I treat resnet34 of dlib as feature extractor, save the features extracted by resnet34 and train on new network, this push my score to 0.09627, a big improve.

4 : Ensemble and k-cross validation

To improve my score, I split the data set into 5 cross data set and ensemble the results by average them, this push my score to 0.06266.  I do not apply stacking because I learn this technique after the competition finished. Maybe I can get better results if I know this technique earlier.

5 : How to use dlib, keras and mxnet

I put the codes of dlib and mxnet on github, I removed all of the non-work solutions, that is why you do not see any codes related to keras. keras did not help me improve my score but mxnet did, what I have done with mxnet was finetune all of the resnet pretrained models and ensemble them with the results trained by dlib. This improve my score to 0.05051.

6 : Read the post at forums, it may give you useful info

I learned that the data set got some "errors"  in it, I removed those false images from the training data set.

7 : Fast ai course is awesome, I should view them earlier

If I have watched the videos before I take this competition, I believe I could perform better in this competition. The forum of this course is very helpful too, it is royal free and open.

8 : X-Crop validation may help you improve your score

I found this technique from PyImageSearch and dlib, but I do not have enough of times to try this technique out.

9 : Save settings in JSON format

Rather than hard code the parameters, save them in JSON file is better, because

a : Do not need to change the source codes frequently, this save compile time
b : Every models, experiments can have their own settings record, easier to reproduce
training result