[Python] Regression Model Always Predicting Mean Value

I'm working on implementing two different models to predicting an optimal tension value to use for cleaving fiber optic cables based off images. I collected the images myself and trained a CNN model to identify good or bad cleaves that I labeled myself. This part seems to be working well. I then am trying to train a regression model to predict a tension value if the image is classified as a bad cleave. However, I the regression model keeps predicting the exact mean value of the tension ranges that I input and I'm not sure why. For reference, the CNN was trained on numerical feature data as well, (e.g. Tension, Angle, Misting, etc.) but I randomly zeroed out these features often to prevent overfitting on the numerical data and not focusing on the image itself. Does anyone have any ideas?

def _build_pretrained_model(self,
param_shape: Tuple[int, ...],
dense1: int,
dense2: int,
dense3: int,
dense_combined: int) -> tf.keras.Model:
"""
Build MLP model for tension prediction.

Args:
param_shape: Dimensions of numerical parameters
dense1: Dimensionality of first dense layer
dense2: Dimensionality of second dense layer
dense3: Dimenionality of third dense layer
dense_combined: Dimenionaly of combined image and features layer

Returns:
tf.keras.Model: Regression model for tension prediction
"""
self.cnn_model.trainable = False
x = Dense(dense1, name="first_dense_layer", activation="relu")(self.feature_output)
x = Dense(dense2, name="second_dense_layer", activation="relu")(x)
feature_input = Input(shape=param_shape, name="feature_input")
y = Dense(dense3, name="third_dense_layer", activation="relu")(feature_input)

combined = Concatenate()([x, y])
z = Dense(dense_combined, name="dense_combined", activation="relu")(combined)
output = Dense(1, name="tension_output")(z)
regression_model = Model(
inputs=[self.image_input, feature_input], outputs=output
)
regression_model.summary()
return regression_model


Above is the code I have for the regressor model. I load the pre-trained CNN classifier and then retrain from the feature output layer. I feel like the model isn't associating the tensions values with the images.

Continue reading...