6_Understand_CNN.md

VI. Understand the accuracy of the model

1. Check the accuracy

Plot the accuracy of model against size of epoch (train and val)
Plot the loss of model against size of epoch (train and val)

...
axs[0].plot(m.history['accuracy'])
axs[0].plot(m.history['val_accuracy'])
axs[1].plot(m.history['loss'])
axs[1].plot(m.history['val_loss'])

-->>>>>>>>>>>>>>>>>>graph

• accuracy keeps increasing: probably providing more epochs can improve a model
• there is a some overfitting: even though train and val accuracy are pretty close to each other, sometimes val_loss 'jumps'

2. Score the model

Function evaluate_generator evaluates the model on a data generator.
In this case score is a list of scalars (loss and accuracy).

loss:0.8221790790557861
accuracy:0.6345810890197754

So the accuracy for a model from scratch ~ 65%. Any suggestions on improving a model are taking up 😊

3. Confusion matrix

Confusion matrix can be pretty useful when evaluating multiclass classifications. Because of great amount of classes just plot misclassified gemstones by model. numpy.argmax() function returns the indices of maximum elements along the specific axis inside the array (axis = 1 - 'horizontally').

Create a list of misclassified indexes which will be substitued into validation set X_val.
Plot misclassified gemstones.

x=(y_pre_test-y_val!=0).tolist()
x=[i for i,l in enumerate(x) if l!=False]

Don't judge poor model. Just look at Almandine, Garnet Red, Hessonite, Pyrope and Rhodolite. Can you distinguish between them? -->>>> IMAGES

4. Save the model

• Save weights to reuse them instead of training again. Keras function save creates h5 file with weights. Use new_model.load_weights('model_gemstones.h5') to reuse it in other models.

model.save('model_gemstones.h5')