stephanakkerman / chart-recognizer Goto Github PK

Some images are RGBA, this causes the issue that all images must be converted to RGB first before applying it to a model

Maybe try using the default settings for this first to see if there is any clear improvement

If the images on the hub get updated the local images should change as well

• Get complete dataset from HF
• Compare image ids with the images that are currently saved
  • If an image with an id exist skip it
  • If an image with an id does not yet exist, create it
  • If there an image with an id that does not exist in the HF dataset, delete it

We use all data now for validation and training, we should save some data so we can do analysis to see what images are hard to classify

We track the results with prior parameters so we can see the changes in performance when changing parameters. Maybe we can also add a config.json for changing the settings of the model.

The parameters we should track are:

• Timm model name

• Batch size

• Datablock parameters

• Datasets used

• Number of images used

• Validation percentage

• Metrics results on val and test set

• Loss function

• Callbacks

• Epoch

• Learning rate

• Confusion matrix results (TP, TN, FP, FN)

• Add config.json for changing parameters

• Write model output + used parameters to results.json

Add functionality to find the optimal parameters over a specified range

Should be easily doable using PIL

So we know which images are difficult to classify, add a notebook that covers:

• The images from the test set that failed
• Others?

Now it is based on nothing, using a hash will make it easier to perform #8
This hash can then also be used to find duplicates