Tensorflow-Image-Classification

Easy and updated tensorflow image classificator

See this proyect being used in a webpage:

https://github.com/AxelAli/Tensorflow-Image-Classifier-Web-Demo

BASED ON / Thanks To:

https://codelabs.developers.google.com/codelabs/tensorflow-for-poets/ https://github.com/llSourcell/tensorflow_image_classifier

Requirements

Tensorflow

Python

Pip

Getting Started!

Clone This Project (In Terminal or Desktop App)

git clone https://github.com/AxelAli/Tensorflow-Image-Classification.git

Start Docker! (in project folder)

docker run -it -v $HOME/Tensorflow-Image-Classification:/Tensorflow-Image-Classification  gcr.io/tensorflow/tensorflow:latest-devel

Use the following command to install latest Tensorflow package

sudo pip install tensorflow --upgrade

Now. (Just in case)

cd ..
cd ..
cd /tensorflow
git pull

Then

cd ..
cd ..
cd /Tensorflow-Image-Classification
git pull

Optional (Getting a dataset)

Great!

Install Dependencies for The Image Downloader (Optional, Makes it easier to download) (In Terminal)

sudo pip install BeautifulSoup bs4 requests image

We have 2 Python scripts to download images from google.

With "user interface"

python GetImagesfromgoogle.py

We get the following Question:

What do i search for?

Here you can write anything you want, (Ex. Dog,Car,Flower).

You can also add extra words to be more specific or to get a wider dataset (Ex. White Dog,Ferrari Car, Sun Flower).

The terminal will start to output each image downloaded

there are total 100 images
1
2
3
4
5
[...]

With console commands

python GetImagesfromgoogleCommand.py

How do i use it? Easy, just write after GetImagesfromgoogleCommand.py you query

python GetImagesfromgoogleCommand.py QUERY

Ex.

python GetImagesfromgoogleCommand.py Dog

You can also "Batch download"

python GetImagesfromgoogleCommand.py DOG
python GetImagesfromgoogleCommand.py CAT
python GetImagesfromgoogleCommand.py PARROT
python GetImagesfromgoogleCommand.py RAT

NOTE!

If you want to search something like "Mashed Potatoes" or "Sports Car" replace all spaces with '+' !

Example:

python GetImagesfromgoogleCommand.py Cute+Dog
python GetImagesfromgoogleCommand.py Red+Car
python GetImagesfromgoogleCommand.py Coke+Bottle

Now you can let it download overnight

Your folder will look like this:

NOTE:All downloads are in the data folder, there is also a folder for each query.

Let's Train!

Just Run the following commands

python image_retraining/retrain.py --bottleneck_dir=./processed-data/bottlenecks --how_many_training_steps 1000  --model_dir=./processed-data/inception --output_graph=./processed-data/retrained_graph.pb --output_labels=./processed-data/retrained_labels.txt --image_dir ./data/

What are these arguments?

--bottleneck_dir : Bottleneck directory, you should not change this unless you want to change things later (/PATH/TO/FOLDER)

**--how_many_training_steps : **Training Steps, these make your accuracy better (XXXX Number)

[NOTE: if you have <1000 images just place 1000] If you arent getting enough Precision , increase images and training steps

--model_dir: Model directory, you should not change this unless you want to change things later (/PATH/TO/FOLDER)

--output_graph: Output Graph file, you should not change this unless you want to change things later (/PATH/TO/FOLDER/FILE.pb)

--output_labels: Output Labels files, you should not change this unless you want to change things later (/PATH/TO/FOLDER/FILE.txt)

--img_dir: Image Directory , you should not change this unless you want to change things later (/PATH/TO/FOLDER)

[NOTE: DO NOT , i repeat , DO NOT place folders inside folders

You should get:

>> Downloading inception-XXXX-XX-XX.tgz %

Wait till downloaded (It's Around 200 mb)

You Might get this Warning, dont worry

W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use XXX instructions, but these are available on your machine and could speed up CPU computations.

Then (This searches the folders)

Looking for images in 'apple+pie'
Looking for images in 'cheesecake'

Eventually you will get this:

Creating bottleneck at /data/bottlenecks/XXX/*_XX.jpg.txt

YOU MIGHT GET THE FOLLOWING ERROR: (Usually 1 in every 50 images)

Not a JPEG file: starts with 0x89 0x50

UPDATE: April/2017

YOU MIGHT NEED TO INSTALL ImageMagick

SEE:

http://www.imagemagick.org/script/download.php

Run the following commands (inside each folder [example: data/dog])!

mogrify -format jpg *.ashx
rm *.ashx
mogrify -format jpg *.png
rm *.png
mogrify -format jpg *.jpeg
rm *.jpeg
mogrify -format jpg *.jpg

**SOLUTION**
```
Creating bottleneck at /data/bottlenecks/apple+pie/*_47.jpg.txt.
Not a JPEG file: starts with 0x89 0x50
```
Delete the image that it got stuck with. (in this case *_47.jpg inside data/apple+pie/)

**Why does this happen?!** [is a renamed file (png>jpg) , your web browser can show it , but the library does not]

After this, repeat 
**Let's Train!**

~~~~~~~~ NO LONGER NEEDED AS OF APRIL/17 ~~~~~~~~

After a couple minutes you should start seeing 
```
2017-02-19 08:55:55.615936: Step XXXX: Cross entropy = 0.008041
2017-02-19 08:55:55.660845: Step XXXX: Validation accuracy = 100.0%
```
And when finished:
```
Final test accuracy = XX.X%
```
Great! We Trained it successfully!

### Guessing the images!

Now we just test an image!

Just in case:
```
pip install image
```

Run the following:
```
python label_image.py  data/cheesecake/*_9.jpg
```
##### What are these arguments?
The only argument is the path to the file:
(PATH/TO/FILE.jpg)

You will get :
```
cheesecake (score = 0.90928)
carrot cake (score = 0.06337)
apple pie (score = 0.02339)
smashed potatoes (score = 0.00396)
```
(Multiply result by 10 ,[ Ex. score = 0.90928 x 10 = 90.9% ])This is the likelihood that the image is of a cheesecake

If you want to add another item just start again from **Let ' s Train!**


### Known Bugs / Errors / Improvements
~~>We should have a system that checks every image and checks if it is a renamed png~~

Fixed April/17