Creating a custom piecewise loss function in tf.keras with three variables - tensorflow

Question

I am using the following code to try and train a model using a custom piecewise loss function that incorporates three variables but I am unable to get it to work. I am new to tensorflow so if anyone has any suggestions that would be helpful.
I want to incorporate a third variable "p" into the loss function where "p" varies with each y_true/y_pred pair. "p" represents one column from the original dataframe. For this problem "p" is crucial to determining if the model is correct or not. If the model is correct I assign a loss of zero and if its incorrect I assign a loss of one. I sum the loss values and divide by the batch size to determine the loss value for that batch. Is what I am trying to do even possible? And if not, what is an alternative way I could achieve my intended outcome.
import tensorflow as tf
import pandas as pd
from tensorflow.keras import layers
# Read in statistics and outcomes dataframe
df = pd.read_csv(r'gs.csv')
df = df.drop(['prediction_ou'], axis=1)
# Change categorical columns to numeric
df['date'] = pd.Categorical(df['date'])
df['date'] = df.date.cat.codes
df['away_team'] = pd.Categorical(df['away_team'])
df['away_team'] = df.away_team.cat.codes
df['away_conf'] = pd.Categorical(df['away_conf'])
df['away_conf'] = df.away_conf.cat.codes
df['home_team'] = pd.Categorical(df['home_team'])
df['home_team'] = df.home_team.cat.codes
df['home_conf'] = pd.Categorical(df['home_conf'])
df['home_conf'] = df.home_conf.cat.codes
# Create target data
target = df.pop('actual_spread')
# Create tensorflow dataset
dataset = tf.data.Dataset.from_tensor_slices((df.values, target.values))
# Shuffle and batch
train_dataset = dataset.shuffle(len(df)).batch(32)
# Model
model = tf.keras.Sequential([
layers.Dense(128, activation='relu'),
layers.Dense(128, activation='relu'),
layers.Dense(1)
])
# Custom loss function
def cbb_loss_higher(p):
def cbb_loss(y_true,y_pred):
c=0
for i in range(len(y_true)):
if ((y_true[i]>p[i]) and (y_pred[i]<p[i])) or ((y_true[i]<p[i]) and (y_pred[i]>p[i])):
c+=1
elif ((y_true[i]>p[i]) and (y_pred[i]>p[i])) or ((y_true[i]<p[i]) and (y_pred[i]<p[i])):
c+=0
else:
c+=0.5
cbb_loss = c/len(y_true)
return cbb_loss
model.compile(optimizer='adam',
loss=cbb_loss_higher(p = df.prediction_spread),
metrics=['accuracy'])
model.fit(train_dataset,
epochs=10)
When the code is run as is I receive the following error:
File "cbb_ml.py", line 129, in <module>
epochs=10)
...
ValueError: No gradients provided for any variable: ['dense/kernel:0', 'dense/bias:0', 'dense_1/kernel:0', 'dense_1/bias:0', 'dense_2/kernel:0', 'dense_2/bias:0'].

Answer 1

Loss function is a part of computation graph built by keras. You can not use python len() function within it. This function don't support backpropagation of gradient. Replace it by tf.shape().

Answer 2

I have just re-arranged your cost function a little bit. Since I did not have your csv file, I mocked some inputs and a simple model.
p = tf.random.normal(shape=[10])
# Custom loss function
def cbb_loss(y_true,y_pred):
c=0.0
for i in range(len(y_true)):
if ((y_true[i]>p[i]) and (y_pred[i]<p[i])) or ((y_true[i]<p[i]) and (y_pred[i]>p[i])):
c+=1.0
elif ((y_true[i]>p[i]) and (y_pred[i]>p[i])) or ((y_true[i]<p[i]) and (y_pred[i]<p[i])):
c+=0.0
else:
c+=0.5
cbb_loss = c/tf.cast(len(y_true),dtype=tf.float32)
return cbb_loss
x = tf.random.normal(shape=(10,10))
y = tf.random.normal(shape=(10,1))
model = tf.keras.Sequential([
layers.Dense(units=1)
])
model.compile(optimizer='adam',
loss=cbb_loss,
metrics=['accuracy'])
model.fit(x=x, y=y, epochs=100,verbose=1)