Custom loss function with multiple inputs for validation

Question

I am creating a custom loss function following the instruction found here. When I add validation_data, I get an error message on ValueError. When I set validation_data=None, this error disappears. I found a similar question on Stackoverflow, but I think my issue is different because I am trying to use a custom loss function.

Here is my code:

from tensorflow.keras.layers import *
from tensorflow.keras.models import Model
import numpy as np
import tensorflow.keras.backend as K
from tensorflow.keras import regularizers

def loss_fcn(y_true, y_pred, w):
    loss = K.mean(K.square((y_true-y_pred)*w))
    return loss

# since tensor flow sets the batch_size default to 32.  The number of samples have to be a multiple of 32 when it is great than 32.
data_x = np.random.rand(32, 51)
data_w = np.random.rand(32, 5)
data_y = np.random.rand(32, 5)

val_x = np.random.rand(4, 51)
val_w = np.random.rand(4, 5)
val_y = np.random.rand(4, 5)

input_x = Input(shape=(51,), name="input")
y_true = Input(shape=(5,), name="true_y")
w = Input(shape=(5,), name="weights")

out = Dense(128, kernel_regularizer=regularizers.l2(0.001), name="HL1")(input_x)
y = Dense(5, name="HL2", activation="tanh")(out)

model = Model(inputs=[input_x, y_true, w], outputs=y)
model.add_loss(loss_fcn(y_true, y, w))

model.compile()
model.fit((data_x, data_y, data_w), validation_data=(val_x, val_y, val_w))

The error message:

ValueError: Error when checking model input: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 3 array(s), but instead got the following list of 1 arrays: [array([[0.74785946, 0.63599707, 0.45929641, 0.98855504, 0.84815295,
        0.28217452, 0.93502174, 0.23942027, 0.11885888, 0.32092279,
        0.47407394, 0.19737623, 0.85962504, 0.35906666, 0.22262...

score 0 · Answer 1 · answered May 26 '21 at 17:20

0

You model has 3 inputs and one output.

The arguments to model fit should be:

x = list (or tuple) of 3 tensors / arrays
y = output values
validation_data = tuple(list of 3 inputs, output values)

answered May 26 '21 at 17:20

Pedro Marques

2,642
1
10
10

The model is not expecting an output because I pass it as an input to be used in the custom loss function. For example, the model works fine if I set: model.fit(x=(data_x, data_y, data_w),y=None). A problem occurs when I pass a validation argument. – Andrew May 26 '21 at 17:36
On second thought, your comment turned out to be super helpful. – Andrew May 26 '21 at 23:19

score 0 · Accepted Answer · answered May 26 '21 at 17:29

0

Instead of tuples, make the training and validation data as list:

model.fit([data_x, data_y, data_w], validation_data=[val_x, val_y, val_w])

answered May 26 '21 at 17:29

Prakash Dahal

4,388
2
11
25

I copy and pasted your code. Unfortunately, it doesn't work as list. – Andrew May 26 '21 at 17:39
@Andrew I have checked the code in colab and its working fine – Prakash Dahal May 26 '21 at 17:43
I am using my university's supercomputing tf.2.0-gpu. Is it Google Colab you are using? Is it free? – Andrew May 26 '21 at 17:46
In any case, I will try Colab this afternoon. Seems interesting – Andrew May 26 '21 at 18:06

Custom loss function with multiple inputs for validation

2 Answers2