如何在 Keras 中正确制作自定义损失函数

Question

i am making a mode that the prediction is a metrix from a conv layer. my loss function is

def custom_loss(y_true, y_pred):
    print("in loss...")
    final_loss = float(0)
    print(y_pred.shape)
    print(y_true.shape)
    for i in range(7):
        for j in range(14):
            tl = float(0)
            gt = y_true[i,j]
            gp = y_pred[i,j]
            if gt[0] == 0:
                tl = K.square(gp[0] - gt[0])
            else:
                for l in range(5):
                    tl = tl + K.square(gp[l] - gt[l])/5
            final_loss = final_loss + tl/98
    return final_loss

the shapes that printed out from the arguments are

(?, 7, 14, 5)

(?, ?, ?, ?)

the labels are in the shape of 7x14x5.

it seems like the loss function gets called for a list of predictions instead of one prediction at a time. I am relatively new to Keras and don't really understand how these things work.

this is my model

model = Sequential()
input_shape=(360, 640, 1)

model.add(Conv2D(24, (5, 5), strides=(1, 1), input_shape=input_shape))
model.add(MaxPooling2D((2,4), strides=(2, 2)))

model.add(Conv2D(48, (5, 5), padding="valid"))
model.add(MaxPooling2D((2,4), strides=(2, 2)))

model.add(Conv2D(48, (5, 5), padding="valid"))
model.add(MaxPooling2D((2,4), strides=(2, 2)))

model.add(Conv2D(24, (5, 5), padding="valid"))
model.add(MaxPooling2D((2,4), strides=(2, 2)))

model.add(Conv2D(5, (5, 5), padding="valid"))
model.add(MaxPooling2D((2,4), strides=(2, 2)))


model.compile(
    optimizer="Adam",
    loss=custom_loss,
    metrics=['accuracy'])

print(model.summary())

I am getting an error like

ValueError: slice index 7 of dimension 1 out of bounds. for 'loss/max_pooling2d_5_loss/custom_loss/strided_slice_92' (op: 'StridedSlice') with input shapes: [?,7,14,5], [2], [2], [2] and with computed input tensors: input[1] = <0 7>, input[2] = <1 8>, input[3] = <1 1>.

I think I know this is because of the arguments to the loss function is given in many predictions at a time with 4D.

how can I fix? is the problem in the way I assign the loss function or in the loss function. for now, the output of the loss function is a float. but what is it supposed to be.

Answer 1

To answer some of your concerns,

I don't see anyone use loops in the loss function

Usually it's a pretty bad practice. Deep nets train on millions of samples usually. Having loops instead of using vectorized operations therefore, will really bring down your model performance.

Implementing without loops.

I'm not sure if I've exactly captured what you wanted in your loss function. But I'm quite sure it's very close (if not this is what you needed). I could have compared your loss against mine with fixed random seeds to see if I get exactly the result given by your loss function. However, since your loss is not working, I can't do that.

def custom_loss_v2(y_true, y_pred):
  # We create MSE loss that captures what's in the else condition -> shape [batch_size, height, width]
  mse = tf.reduce_mean((y_true-y_pred)**2, axis=-1)

  # We create pred_first_ch tensor that captures what's in the if condition -> shape [batch, height, width]
  pred_first_ch = tf.gather(tf.transpose(y_pred**2, [3,0,1,2]),0)

  # We create this to get a boolean array that satisfy the conditions in the if else statement
  true_first_zero_mask = tf.equal(tf.gather(tf.transpose(y_true, [3,0,1,2]),0), 0)

  # Then we use tf.where with reduce_mean to get the final loss
  res = tf.where(true_first_zero_mask, pred_first_ch, mse)
  return tf.reduce_mean(res)