-1

I am trying to teach image classification model to define a number characteristic from an image. I am sure that SparseCategoricalCrossentropy loss function doesn't work for me, as for training I need to penalize big differences more than small ones. Ideally I would like to use Mean Squared Error loss function.

I use TensorFlow tutorial to prepare the model - https://www.tensorflow.org/tutorials/images/classification.

Class names are numbers for me, I tried the following options:

  • ['00', '01', '02', '03', '04', '05', '06', '07', '08', '09', '10', '11', '12']
  • ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9', '10', '11', '12']

The only change I made against tutorial (except the dataset) is exchanging SparseCategoricalCrossentropy loss function to 'mean_squared_error'.

But the loss function clearly doesn't work for me. It returns values, that gets smaller with training, but accuracy is never more than 5%, and it even goes down as loss value becomes smaller. Results also do not make sense. The data is fine, I can easily achieve 95% accuracy with SparseCategoricalCrossentropy loss function. What am I missing?

UPDATE: I think what I really need is a way to define regression problem in TensorFlow using images labeled with numbers.

kharit
  • 1
  • 2
  • MSE is **not** a classification loss. Please see [What function defines accuracy in Keras when the loss is mean squared error (MSE)?](https://stackoverflow.com/questions/48775305/what-function-defines-accuracy-in-keras-when-the-loss-is-mean-squared-error-mse/48788577#48788577) – desertnaut Sep 02 '20 at 10:34
  • I saw it. It seems, I don't need classification loss. What do I need then? Is it possible to use regression loss on image models? – kharit Sep 02 '20 at 10:39
  • What do you mean "image models"? Loss is defined by the kind of *problem* you are trying to solve; if it is a *classification* problem, you cannot use MSE or other losses appropriate for *regression*. – desertnaut Sep 02 '20 at 10:41
  • I think what I need is a way to define regression problem in TensorFlow using images labeled with numbers. – kharit Sep 02 '20 at 10:56
  • This way, the error between a "9" and a "2" would be greater than between "9" and "8". If this is what you want, sounds indeed like a regression problem. – desertnaut Sep 02 '20 at 11:01

1 Answers1

0

Turns out it is quite easy to turn image classification problem into a regression problem. Against tutorial referenced in question I had to make the following changes:

  1. Different dataset with numbers as 'classes' (folder names).

  2. Changed loss function to Mean Squared Error or other loss function suitable for regression.

  3. Made the last layer for model with just 1 neurone instead of number of classes (and without softmax):

    ...
layers.Dense(128, activation='relu'),
layers.Dense(1)    # changed from num_classes to 1

  4. Changed interpretation of prediction results:

     ...
 predictions = model.predict(img_array)
 # score = tf.nn.softmax(predictions[0])    # correct for classification, but not regression
 score = predictions.flatten()[0]    # correct result for regression
 ...
kharit
  • 1
  • 2
  • But your basic premise by which you begin your question, "*I am trying to teach image **classification** model*", does not hold anymore, does it? – desertnaut Sep 02 '20 at 14:50
  • You are right. If I knew this stuff, I wouldn't ask here, — I would just figure out myself, as it is faster. I don't want to update the question, as I think it is clear and I saw a number of people looking for an answer for a similar question and not receiving it. I hope it will help someone. – kharit Sep 02 '20 at 15:30