tduffy000 / deep-learning-group Goto Github PK

From this confusion matrix,

There are some confusions between classes 7 and 8. Almost all the other classes, the error seems to be randomly distributed.

We should update the top-level README to reflect the final structure and contents of this repo.

• Create a 'subset' dataset with separate train & test folders (10 classes)
• Establish a working pipeline on Google Collab
• Train, validate, and evaluate a simple CNN bare model on Google Collab
• Testing additional model architecture
• Since I have access to Google Collab Pro, testing models with bigger dataset
• Summarize initial results with 4 models
• Review Tom's transfer learning models
• Review Sam's journals
• Organize results into team's final presentation

Now that we have a couple of working models from using transfer learning from VGG & Densenet, we can bump the number of classes from 17 to 50 and then the full dataset.

We, currently, don't have any notebooks or other code fully dedicated to exploratory data analysis, which is a crucial task. We should take one of the notebooks and modify it such that it simply serves as exploring the source dataset and providing the reader with examples and distributions.

May 7 Slides

Here's the link to our final presentation that I'll be working on:

Final presentation

Here are the tasks that I am aiming to complete before Wednesday:

• Select "some" labels from dataset to train for the project instead of using them all
• Train, validate, and evaluate a simple CNN bare model. Plot the val_accuracy and accuracy
• Incorporate callbacks checkpoints to terminate early and to save best model and training history
• Do the same thing, but with VGG16 imagenet using transfer learning.
• Compare the result between the two simple model and optimize the hyper-parameters so they can be used as baseline.
• Create checkpoint to save and load the best model

• Repo clean-up
• Benchmark model (e.g. random choice)
• Class prior probability analysis

Hi guys, here are some issues I ran into and am working on fixing right now

• The original model.fit function is too slow to train the model. I'm looking for an alternative. Using this function takes an abysmal amount of time per epoch.
• The default value for steps_per_epoch is too low which causes longer training time. I'm searching for optimal value.
• Should we limit the number of validation steps per epoch using validation_steps? Right now, the validation_split is 0.2

Hi guys,

Here is the dropbox link to the trimmed dataset. I select the 17 classes that have the most images. Feel free to suggest me to add or remove any classes. I just want to make that we all use the same trimmed dataset.