- keras
- pandas
- numpy
- matplotlib.plot
$ python3 main.py --help
usage: main.py [-h] {training,prediction,splitting,plotting} ...
Image Processing Project
positional arguments:
{training,prediction,splitting,plotting}
optional arguments:
-h, --help show this help message and exit$ python3 main.py splitting$ python3 main.py training $ python3 main.py training --help
Using TensorFlow backend.
usage: main.py training [-h] [--epochs EPOCHS] [--batch-size BATCH_SIZE]
[--steps-per-epoch STEPS_PER_EPOCH]
optional arguments:
-h, --help show this help message and exit
--epochs EPOCHS
--batch-size BATCH_SIZE
--steps-per-epoch STEPS_PER_EPOCH
You can change the shown parameters like this:
$ python3 main.py training --epochs 32You have to put your images to Prediction folder. If there is no folder like this. The program will warn you to create it.
$ python3 main.py prediction
Result should be like this for 1 image:
Prediction/example_pizza.jpeg
Answer: 9
Predicted: pizza
pizzaIn this part of application, images in folder Dataset are splitting into parts in rate of desired value. In default it's 80% - training set, 20% test set.
Directory hierarchy is like this after splitting
--->training
||
||---->butterfly
||
||--->image001.jpg, image002.jpg,....
--->test
||
||---->butterfly
||
||--->image001.jpg, image002.jpg,....
CNN Model in this part we are using like this. You can change it in your desires: # conv => max pool => dropout => conv => max pool => dropout => fully connected (2 layer)
Keras libraries are used to build this model.
Dropout: Dropout is a technique where randomly selected neurons are ignored during training
Training step is can take a lot of time according to dataset size. So model and info about model is being saved to seperated files for future uses(like to predict images with created model)
Assume that you have a dataset of 10 examples (or samples). You have a batch size of 2, and you've specified you want the algorithm to run for 3 epochs. Therefore, in each epoch, you have 5 batches (10/2 = 5). Each batch gets passed through the algorithm, therefore you have 5 iterations per epoch. reference : https://stackoverflow.com/questions/4752626/epoch-vs-iteration-when-training-neural-networks
To avoid overfitting problem, we need to expand artificially our dataset. Alter the training data with small transformations to reproduce the variations of images. For example, the car is not centered The scale is not the same. The image is rotated. ImageDataGenerator's parameters are hyperparameters. It can be set according to data and the results.
train_datagen = ImageDataGenerator(rescale = 1.0 / 255 , shear_range = 0.2 , zoom_range = 0.2 , horizontal_flip = True)
You can see the accuracy and loss values when training action performed.
Epoch 1/16
32/32 [=======*] - 56s 2s/step - loss: 4.4368 - acc: 0.4382 - val_loss: 1.8855 - val_acc: 0.4292
Epoch 2/16
32/32 [=======*] - 56s 2s/step - loss: 1.3432 - acc: 0.6016 - val_loss: 1.6124 - val_acc: 0.5685
Epoch 3/16
32/32 [=======*] - 57s 2s/step - loss: 1.0990 - acc: 0.6582 - val_loss: 1.1177 - val_acc: 0.7374
Epoch 4/16
32/32 [=======*] - 56s 2s/step - loss: 0.9275 - acc: 0.7126 - val_loss: 0.9164 - val_acc: 0.7603
Epoch 5/16
32/32 [=======*] - 55s 2s/step - loss: 0.8592 - acc: 0.7508 - val_loss: 0.7677 - val_acc: 0.7854
Epoch 6/16
32/32 [=======*] - 56s 2s/step - loss: 0.7570 - acc: 0.7612 - val_loss: 0.6741 - val_acc: 0.8402
Epoch 7/16
32/32 [=======*] - 57s 2s/step - loss: 0.6795 - acc: 0.7910 - val_loss: 0.6298 - val_acc: 0.8082
Epoch 8/16
32/32 [=======*] - 57s 2s/step - loss: 0.6260 - acc: 0.8231 - val_loss: 0.5743 - val_acc: 0.8425
Epoch 9/16
32/32 [=======*] - 70s 2s/step - loss: 0.6303 - acc: 0.8105 - val_loss: 0.6314 - val_acc: 0.8699
Epoch 10/16
32/32 [=======*] - 68s 2s/step - loss: 0.5699 - acc: 0.8247 - val_loss: 0.5304 - val_acc: 0.8493
Epoch 11/16
32/32 [=======*] - 56s 2s/step - loss: 0.5148 - acc: 0.8436 - val_loss: 0.5142 - val_acc: 0.8516
Epoch 12/16
32/32 [=======*] - 55s 2s/step - loss: 0.5040 - acc: 0.8516 - val_loss: 0.5033 - val_acc: 0.8539
Epoch 13/16
32/32 [=======*] - 57s 2s/step - loss: 0.4961 - acc: 0.8485 - val_loss: 0.4333 - val_acc: 0.8790
Epoch 14/16
32/32 [=======*] - 57s 2s/step - loss: 0.4464 - acc: 0.8477 - val_loss: 0.3994 - val_acc: 0.8904
Epoch 15/16
32/32 [=======*] - 56s 2s/step - loss: 0.4749 - acc: 0.8475 - val_loss: 0.3994 - val_acc: 0.9018
Epoch 16/16
32/32 [=======*] - 56s 2s/step - loss: 0.4431 - acc: 0.8691 - val_loss: 0.4150 - val_acc: 0.8881
Execution Time: 15.470373169581096 minutes
In this part you can test the images different than the dataset images. It uses saved model in training step
You can plot the model accuracy and model loss with the model info saved in training step.
