Week 2

Apr 04/14/23

1. Meeting with Vannary
   1. What is the mapping for PACP and wRc labels to grouped labels?

Metric calculation

Which labels are used for the metric calculation?

Labels which are in the test set and the trained model (intersection) as well as the labels which are predicted by the model.

Some points about labels

1. Label X is not in the training set but is a part of the test set, hence the model can't predict that. Hence, it is dropped from the metric calculation.
2. DSF and XP are predicted by the model, but are not present in the test set. They are misclassifications made by the model. Hence, will be included in the metric calculation.
3. OBB is present in both train and test set, but is not predicted by the model. It is a missclassification made by the model. Hence, even this label will be used in the metric calculation.

Newer runs to make

1. Lower threshold
   1. Training dataset: SD1 VB ML
   2. Training dataset: All Utility VB ML

Apr 04/11/23

Design to copy files from VS Research to Hydrotrek GDrive

1. Start a Colab Notebook
2. Mount Google Drive
3. Using the Linux cp command, copy files from VS_Research folder to the desired location in the Hydrotrek GDrive.

Ensemble methods in computer-vision

info

Prompt to ChatGPT: how will I use ensemble methods in computer vision for a multi-label classification problem?

Ensemble methods can be a powerful technique for improving the performance of multi-label classification models in computer vision. Here's a general approach you can take:

1. Data preprocessing: First, you'll need to preprocess your data. This typically involves resizing and normalizing images, as well as creating training and validation sets.
2. Base models: Next, you'll need to select several base models to use in your ensemble. These could be different types of models (e.g., CNNs, RNNs, SVMs), or different variants of the same type of model (e.g., with different hyperparameters). Ideally, the base models should have low correlation with each other, so that their predictions are diverse.
3. Training: You'll then train each base model on your training set. You may want to use data augmentation techniques to increase the size of your training set.
4. Predictions: Once you have trained your base models, you can use them to make predictions on your validation set. This will give you a set of predictions for each image in your validation set.
5. Ensemble: Finally, you'll need to combine the predictions from your base models to create a final prediction for each image. There are many ways to do this, including:
   1. Simple averaging: Take the average of the predicted probabilities for each label across all base models.
   2. Weighted averaging: Give more weight to the predictions of certain base models based on their performance on the validation set.
   3. Stacking: Train a meta-model (e.g., logistic regression or neural network) on the predictions of your base models, and use the meta-model to make the final prediction.
6. Evaluation: Once you have created your ensemble model, you can evaluate its performance on a separate test set.

Keep in mind that ensemble methods can be computationally expensive, especially if you're using a large number of base models. You may want to experiment with different numbers and types of base models to find the right balance between performance and computational efficiency.

Apr 04/10/23

Empty