Skip to main content

Pokemon Classification by Siamese Network and One Shot Learning

Some of my recent clients are interested in image classification using limited learning data. A major use case is in detecting defective products on the manufacturing line. Defective products are normally caught by human operators watching the production line. It requires constant concentration and effort so if this can be automated it will bring a lot of productivity boost to manufacturers.

One method to overcome this issue is by creating a network to compare input images with the training dataset. For defective products, the images will look different from the passing products.

Experiments

For this experiment, I used a Siamese Network to generate embeddings for images of Pokemon. The Pokemon has 4 classes, Pikachu, Squirtle, Bulbasaur and Charmander. The images are scraped from the internet using the Bing API.


The images in the training dataset is trained to find the set of weights that will clearly separate the 4 classes. If a new image which is not included in the 4 classes is used, the resulting distance (difference) will be large enough to separate it with the 4 available classes.

The Siamese network trains triplets of Anchor, Positive and Negative images, and I generated some hard batches to improve the training accuracy. For details look into the original article in Credits section.

Results

The best result I got was AUC 84%.


And the image tests with this set of weights are as follows




Discussions

As we can see in the second row, the input image of Bulbasaur is wrongly classified as Squirtle. This error can happen when the input image of Bulbasaur has the same color tone as Squirtle. Therefore, we can conclude that although One-Shot learning can be useful for classification with minimal data because it only measures the distance between images, it can also produce errors when the images are similar to each other. For all use cases, it will depend on the data available, and if the images have enough distance (sufficiently different). 

Credit

This experiment was created by modifying the original pipeline found here:
https://medium.com/@crimy/one-shot-learning-siamese-networks-and-triplet-loss-with-keras-2885ed022352










Comments

Post a Comment

Popular posts from this blog

Object detection with Google Colab and Tensorflow

This is just a memo of the challenges I faced when running a model training on Google Colab, while following a great tutorial here . Mind the versions Tensorflow is currently at version 2.2.0 but most tutorials are still using the contrib package, and there is no known easy way to update the code to remove dependency on contrib. So my best bet is to downgrade the tensorflow version to 1.x. Since Google Colab only gives the options of either 1.x or 2.x and we cannot specify the exact version, I ended up with version 1.15.2. Even with the command :  %tensorflow_version  1.15.0 I ended up with : 1.15.2 Another pitfall was the version of numpy. Installing numpy gives us the version 1.18.3 but for some reason this generates the error : TypeError: 'numpy.float64' object cannot be interpreted as an integer Downgrading numpy to version 1.17.4 solved this for me. It seems we don't need ngrok for tensorboard With the command :  %load_ext tensorboard W...
Post a Comment
Read more

Installing a custom ROM on Android (on the GT-N8013)

It's been a while since my last entry and since it is a new start in 2019, I thought I'd write something about "gone with the old and in with the new". I've had my Samsung Galaxy Note 10.1 (pnotewifi) since 2014, and it's one of the early Galaxy Note tablet series. It has served me well all this years but now it just sits there collecting dust. My old Samsung GT-N8013 I've known a long time about custom Android ROMs like CyanogenMod but has never had the motivation to try them out, until now ! Overview of the process For beginners like me, I didn't have an understanding of the installation process and so it looked complicated and it was one of the reasons I was put off in trying the custom ROM. I just want to say, it's not complicated at all!   Basically you will need to Prepare an SD card and install Android SDK (you need adb ). Install a custom boot loader ( TWRP is the de facto tool at the moment). Use adb to copy custom...
15 comments
Read more

Using FCM with the new HTTP v1 API and NodeJS

When trying to send FCM notifications I found out that Google has changed their API specifications. The legacy API still works but if you want to use the latest v1 API you need to make several changes. The list of changes is listed on their site so I won't be repeating them again but I'll just mention some of the things that caused some trial and error on my project. The official guide from Google is here : Official Migration Guide to v1 . The request must have a Body with the JSON containing the message data. Most importantly it needs to have "message" field which must contain the target of the notification. Usually this is a Topic, or Device IDs. Since my previous project was using GAS, my request had a field called "payload" instead of "body". Using the request from my previous project, my request in Node JS was as follows: request ({ url: 'https://fcm.googleapis.com/v1/projects/safe-door-278108/messages:send' , method: ...
Post a Comment
Read more