Advertisement
Not a member of Pastebin yet?
Sign Up,
it unlocks many cool features!
- ## ToDo
- ### Report
- * write **Kaggle Competition: Diabetic Retinopathy Detection** section in your own words
- * include stats of the data set provided in terms of the number of examples per class
- * some of the **DL Frameworks** review parts are direct copy/paste - make sure it is all written in your own words.
- ### Pre-training second layer
- * make sure pre-training values are between 0.1 and 0.9 (actually look into existing datasets used for the pretraining and see what min and max values do they use)
- train first layer with unsupervised training
- * convolve input data with learned filters
- * replace inputs with feature maps (1 -> 64)
- * create 200k samples of 64x13x13 tensors use them to pre-train second layer filters (2048 -> 512). Should I keep the connection table with me (would not hurt I guess).
- ## Notes
- * Use Photoshop to mark symptom features for supervised learning. Save PSD files, where one layer is the original image and second layer is features marked with different color for each.
- * Extracting filenames in labels.csv:
- ```
- cat trainLabels.csv | grep ",1" | head -n20 | awk -F "," '{print "processed/train/"$1".jpeg"}' | xargs open
- ```
Advertisement
Add Comment
Please, Sign In to add comment
Advertisement