Why does the output layer is simply zero at the end of the network? - tensorflow

Question

I am trying to train a model that takes a 15x15 image and classify each pixel into two classes (1/0).
This is my loss function:
smooth = 1
def tversky(y_true, y_pred):
y_true_pos = K.flatten(y_true)
y_pred_pos = K.flatten(y_pred)
true_pos = K.sum(y_true_pos * y_pred_pos)
false_neg = K.sum(y_true_pos * (1-y_pred_pos))
false_pos = K.sum((1-y_true_pos)*y_pred_pos)
alpha = 0.5
return (true_pos + smooth)/(true_pos + alpha*false_neg + (1-alpha)*false_pos + smooth)
def tversky_loss2(y_true, y_pred):
return 1 - tversky(y_true,y_pred)
This is the model:
input_image = layers.Input(shape=(size, size, 1))
b2 = layers.Conv2D(128, (3,3), padding='same', activation='relu')(input_image)
b2 = layers.Conv2D(128, (3,3), padding='same', activation='relu')(b2)
b2 = layers.Conv2D(128, (3,3), padding='same', activation='relu')(b2)
output = layers.Conv2D(1, (1,1), activation='sigmoid', padding='same')(b2)
model = models.Model(input_image, output)
model.compile(optimizer='adam', loss=tversky_loss2, metrics=['accuracy'])
The model left is the input and the label is the middle column and the prediction is always zero on the right column:
The training performs really poorly:
Epoch 1/10
100/100 [==============================] - 4s 38ms/step - loss: 0.9269 - acc: 0.1825
Epoch 2/10
100/100 [==============================] - 3s 29ms/step - loss: 0.9277 - acc: 0.0238
Epoch 3/10
100/100 [==============================] - 3s 29ms/step - loss: 0.9276 - acc: 0.0239
Epoch 4/10
100/100 [==============================] - 3s 29ms/step - loss: 0.9270 - acc: 0.0241
Epoch 5/10
100/100 [==============================] - 3s 30ms/step - loss: 0.9274 - acc: 0.0240
Epoch 6/10
100/100 [==============================] - 3s 29ms/step - loss: 0.9269 - acc: 0.0242
Epoch 7/10
100/100 [==============================] - 3s 29ms/step - loss: 0.9270 - acc: 0.0241
Epoch 8/10
100/100 [==============================] - 3s 29ms/step - loss: 0.9271 - acc: 0.0241
Epoch 9/10
100/100 [==============================] - 3s 29ms/step - loss: 0.9276 - acc: 0.0239
Epoch 10/10
100/100 [==============================] - 3s 29ms/step - loss: 0.9266 - acc: 0.0242

Answer 1

This sounds like a very imbalanced dataset with very tiny true regions. This might be hard to train indeed.
You may want to increase alpha to penalize more false negatives than false positives. Anyway, unless alpha is big enough, it's very normal that in the beginning your model first goes to all neg because it's definitely a great way to decrease the loss.
Now, there is a conceptual mistake regarding how Keras works in that loss. You need to keep the "samples" separate. Otherwise you are calculating a loss as if all images were one image. (Thus, it's probable that images with many positives have a reasoable result, while images with few positives don't, and this will be a good solution)
Fix the loss as:
def tversky(y_true, y_pred):
y_true_pos = K.batch_flatten(y_true) #keep the batch dimension
y_pred_pos = K.batch_flatten(y_pred)
true_pos = K.sum(y_true_pos * y_pred_pos, axis=-1) #don't sum over the batch dimension
false_neg = K.sum(y_true_pos * (1-y_pred_pos), axis=-1)
false_pos = K.sum((1-y_true_pos)*y_pred_pos, axis=-1)
alpha = 0.5
return (true_pos + smooth)/(true_pos + alpha*false_neg + (1-alpha)*false_pos + smooth)
This way you have an individual loss value for each image, so the exitence of images with many positives don't affect the results of images with few positives.