Solve LSTM Issue where model.predict losses do not correspond to model.fit -- 3
Budget: $250 – $750 USD
I am training an LSTM for a time series prediction task.
Environment: Google Vertex AI VM with JupyterLab - Google pre-configured Tensorflow 2.9.1 environment
I have successfully trained it to an accuracy that is sufficient for my needs. However, when using it to predict, it works very poorly. This is not a model issue, but something with Keras/TF because the predictions are done on the SAME dataset as the training. So they should be fairly close to the last saved state of the model after training.
The issue is documented here. Please READ this carefully so you understand it, and let me know if you have any questions
https://datascience.stackexchange.com/questions/113627/lstm-sequential-val-loss-train-loss-on-same-dataset
You will have access to a Google Cloud Vertex AI VM with all the code and data so you can work on it (i.e you don't have to pay for your own VM)
Expected deliverables: to complete the project, you MUST resolve this issue so that we can get
1) model.evaluate(X_train) losses is very close to model.fit(X_train) losses. It doesn't matter how the model performs on other data etc, we just need to resolve this particular problem
2) You can make any changes to the code that don't require re-training the model. So the model has to be-re loadable. If you find this is not possible, we will agree on an alternative compensation
Environment: Google Vertex AI VM with JupyterLab - Google pre-configured Tensorflow 2.9.1 environment
I have successfully trained it to an accuracy that is sufficient for my needs. However, when using it to predict, it works very poorly. This is not a model issue, but something with Keras/TF because the predictions are done on the SAME dataset as the training. So they should be fairly close to the last saved state of the model after training.
The issue is documented here. Please READ this carefully so you understand it, and let me know if you have any questions
https://datascience.stackexchange.com/questions/113627/lstm-sequential-val-loss-train-loss-on-same-dataset
You will have access to a Google Cloud Vertex AI VM with all the code and data so you can work on it (i.e you don't have to pay for your own VM)
Expected deliverables: to complete the project, you MUST resolve this issue so that we can get
1) model.evaluate(X_train) losses is very close to model.fit(X_train) losses. It doesn't matter how the model performs on other data etc, we just need to resolve this particular problem
2) You can make any changes to the code that don't require re-training the model. So the model has to be-re loadable. If you find this is not possible, we will agree on an alternative compensation