Analytics Vidya CV Practice Problem
- I used 4 different pretrained models.
(a) Resnet50 - 0.66109-0.73312
(b) Densenet121 - 0.6675-0.8292
(c) Efficientnet_b0 - 0.8725-0.9082
(d) Efficientnet_b1 - o.9082-0.91305
-
I have used 40-50 epochs for each model. I have used Adam and SGD optimizer together, first 20 with Adam to finetune the model(i.e. to train the last layer and the last 20-30 epochs with SGD to train full model.
-
Finetuning is to be used for resnet and densenet only as they have more parameters than efficientnet, which can be trained without finetuning for 30-40 epochs.
-
For the first 20 epochs I have used learning rate= 1e-3 and for the next 20-30 epochs used 1e-4.
-
I have also used ReduceOnPlateau scheduler with 3-4 patiece.
- Straightified K-fold
- Test Time Augmentation
- Ensembling
- Pseudolabeling