Improving validation accuracy of an image sentiment identification system

AI Discussions
,
1

Greetings!!

I am trying to develop an image sentiment identification (to start with, only 2 sentiments - happy or sad) system through transfer learning by using InceptionResNetV2 as the base model. I have given the code below. (I got this approach for transfer learning from the book Deep Learning with Python by Chollet.)

I found InceptionResNetV2 to have the best validation accuracy for image classification based on the approach described by Marie Stephen Leo (How to Choose the Best Keras Pre-Trained Model for Image Classification | by Marie Stephen Leo | Towards Data Science).

However, the validation accuracy for sentiment identification is stuck around 81%. I have tried multiple things such as adding dropout, increasing the number of epochs, increasing/decreasing number of nodes in the last layer between 64 and 512, curating the training/validation dataset, etc.

Please advise how I can solve this. Thank you!

conv_base = tf.keras.applications.InceptionResNetV2(
    include_top=False,
    weights='imagenet',
    input_shape=(IMAGE_HEIGHT,IMAGE_WIDTH,3),
)
conv_base.trainable = False


inputs = tf.keras.Input(shape=(IMAGE_HEIGHT,IMAGE_WIDTH,3))
x = data_augmentation(inputs)
x = tf.keras.applications.inception_resnet_v2.preprocess_input(x)
x = conv_base(x)
x = tf.keras.layers.Flatten()(x)
x = tf.keras.layers.Dense(64)(x)
x = tf.keras.layers.Dropout(0.5)(x)
outputs = tf.keras.layers.Dense(1, activation='sigmoid')(x)

model = tf.keras.Model(inputs, outputs)
2

How does the validation error compare with the training error?

You ask “how can I solve this”? But you have not established whether 81% is a good or bad result.

Maybe 81% is as good as you can do with that dataset.

3

Excellent points. The client requires 99% validation accuracy.
I have posted the plots for accuracies and losses.
accuracies
losses

I am unable to find any large labeled datasets for image sentiment identification. I could find only one dataset with about ~400 images. I have done data augmentation as well. Can I generate training data or any other datasets available? kindly let me know. thank you!

4

That’s going to be extremely difficult - especially given the lack of data.

“Augmenting the dataset” is the only valid way to create new training data. You can’t really create new examples from scratch.

I’m not aware of any data sources for this task. Maybe someone from the community will have a better idea.

5

Yeah, I am unable to find good datasets. If anyone can recommend, it will be very helpful. Will I be able to generate images using text-to-image models such as DALL-E or Imagen? I know Imagen is not widely available.