Accuracy is coming constant for hybrid transfer model

PennyLane Development
1

[(‘Basic Variational Circuit’, QuantumIncepLayer(
(pre_net): Linear(in_features=512, out_features=4, bias=True)
(post_net): Linear(in_features=4, out_features=2, bias=True)
), ‘BVC’)]
Training Model for Basic Variational Circuit

Phase: train Epoch: 1/25 Loss: 0.7116 Acc: 0.5123
Phase: validation Epoch: 1/25 Loss: 0.6997 Acc: 0.4575
Phase: train Epoch: 2/25 Loss: 0.7038 Acc: 0.4385
Phase: validation Epoch: 2/25 Loss: 0.6921 Acc: 0.5425
Phase: train Epoch: 3/25 Loss: 0.6972 Acc: 0.4631
Phase: validation Epoch: 3/25 Loss: 0.6953 Acc: 0.4575
Phase: train Epoch: 4/25 Loss: 0.6972 Acc: 0.3811
Phase: validation Epoch: 4/25 Loss: 0.6910 Acc: 0.5425
Phase: train Epoch: 5/25 Loss: 0.6957 Acc: 0.4877
Phase: validation Epoch: 5/25 Loss: 0.6933 Acc: 0.4575
Phase: train Epoch: 6/25 Loss: 0.7009 Acc: 0.5123
Phase: validation Epoch: 6/25 Loss: 0.7162 Acc: 0.4575
Phase: train Epoch: 7/25 Loss: 0.6962 Acc: 0.5369
Phase: validation Epoch: 7/25 Loss: 0.6896 Acc: 0.5425
Phase: train Epoch: 8/25 Loss: 0.6981 Acc: 0.4959
Phase: validation Epoch: 8/25 Loss: 0.6896 Acc: 0.5425
Phase: train Epoch: 9/25 Loss: 0.6957 Acc: 0.4959
Phase: validation Epoch: 9/25 Loss: 0.6903 Acc: 0.5425
Phase: train Epoch: 10/25 Loss: 0.6947 Acc: 0.4959
Phase: validation Epoch: 10/25 Loss: 0.6911 Acc: 0.5425
Phase: train Epoch: 11/25 Loss: 0.6939 Acc: 0.4959
Phase: validation Epoch: 11/25 Loss: 0.6916 Acc: 0.5425
Phase: train Epoch: 12/25 Loss: 0.6940 Acc: 0.4959
Phase: validation Epoch: 12/25 Loss: 0.6919 Acc: 0.5425
Phase: train Epoch: 13/25 Loss: 0.6937 Acc: 0.4959
Phase: validation Epoch: 13/25 Loss: 0.6924 Acc: 0.5425

Can anyone help?
I am building a transfer learning model for object detection using hybrid Quantum approach as mentioned in the documentation of pennylane.

Accuracy is coming constant after some epochs and not improving from there?
Am I doing anything wrong or need to set some parameters?

2

Hi @Jay_Timbadia,

Thank you for your question!

In general, it is possible for a model to “plateau” during training, meaning that the cost function (or any other relevant figure of merit) fails to improve with subsequent optimization steps. This can be due to a variety of reasons, although typically it can be a signal that the model is either stuck in local minima, or is not expressive enough for the task at hand. I would recommend revisiting your choice of model to determine whether it is a good approach for the object detection problem you are interested in.

I may be able to provide more help if you send us more details about your code and the problem you are tackling.

Best,

Juan Miguel