Skip to content

Latest commit

 

History

History
112 lines (106 loc) · 4.93 KB

models config.md

File metadata and controls

112 lines (106 loc) · 4.93 KB
Raw

Configurations for Different Models used in ensembling

No. of models used in ensembling - 5
Each of the five models were trained on different folds of training data and validated on different validation sets.

The following configurations were same for all the models:

  1. Image size - 300x300
  2. Multilabel target
  3. Augmentations -
    • Rotation (25 - 360 degree on different models)
    • vertical flip
    • horizontal flip
    • Zoom range - (0.25 - 0.35)
  4. optimizer - Adam
  5. Loss function - Binary Crossentropy
  6. output activation - Sigmoid
  7. callback - Early Stopping (patience = 5)

1. MODEL 1

  • Training on old competition data
    • Training on a total of about 35000 images divided into 9 batches, each batch trained for 10 epochs.
    • Total no. of epochs : 92 (divided into 9 batches of 10 epochs for training data), (also trained 2 epochs on the validation data).
    • Trained on validation data as well (only for 2 epochs)
    • augmentations - rotation_range = 25, zoom_range = 0.25, vertial and horizontal flip
    • preprocessing - only cropping
    • train-validation-split - 90% training and 10% validation
    • batch size : 32
  • Training on current competition data
    • Trained the above model on about 3600 images for 10 epochs
    • preprocessing - only cropping
    • No validation on new data
  • Added TTA on test data

Model 1 Performance :

  • Training Accuracy around 95%
  • Validation accuracy around 94%
  • Leaderboard score - Public : 0.797 Private : 0.910

2. MODEL 2

  • Training on old competition data :
    • Total no. of EPOCHS : 75 (8 epochs for 9 batches and 3 epochs while training on the validation set)
    • preprocessing - Cropping
    • augmentations - rotation_range = 25, zoom_range = 0.35, vertial and horizontal flip
    • batch size = 32
    • Train-validation-split - 90% training and 10% validation
  • Training on current competition data
    • Total no. of epochs : 12
    • Learning Rate : 1e-4 with ReduceLrOnPlateau(patience = 1, factor = 0.5, min_lr = 1e-6)
    • 90% training data and 10% validation data
    • preprocessing - Cropping (crop_from_gray)
  • Added TTA on test data

Model 2 Performance

  • Training Accuracy around 94%
  • validation accuracy around 94.5%
  • Leaderboard score - Public : 0.790 Private : 0.912

3. MODEL 3

  • Training on old competition data :
    • Total no. of EPOCHS : 72 (8 epochs for 9 batches)
    • augmentations - rotation_range = 360, zoom_range = 0.25, vertial and horizontal flip
    • learning rate = 1e-4 with ReduceLrOPlateau(patience = 1, factor = 0.5, min_lr = 5e-5)
    • EarlyStopping(patience = 5)
    • preprocessing - Cropping
    • batch size = 32
    • Train-validation-split - 90% training and 10% validation
  • Training on current competition data
    • Total no. of epochs : 8
    • Learning Rate : 5e-5
    • 90% training data and 10% validation data
    • preprocessing - Cropping (crop_from_gray)
  • Added TTA on test data

Model 3 Performance

  • Training Accuracy around 95%
  • validation accuracy around 96%
  • Leaderboard score - Public : 0.801 Private : 0.917

4. MODEL 4

  • Training on old competition data :
    • Total no. of EPOCHS : 77 (8 epochs for 9 batches and 5 epochs while training on the validation set)
    • augmentations - rotation_range = 360, zoom_range = 0.35, vertial and horizontal flip
    • No preprocessing
    • batch size = 32
    • Train-validation-split - 90% training and 10% validation
  • Training on current competition data
    • Total no. of epochs : 10
    • Learning Rate : 5e-5
    • 90% training data and 10% validation data
    • No preprocessing

Model 4 Performance

  • Training Accuracy around 94%
  • validation accuracy around 93%
  • Leaderboard score - Public : 0.797 Private : 0.914

5. MODEL 5

  • Training on old competition data :
    • Total no. of EPOCHS : 100
    • preprocessing - Cropping
    • Learning Rate : 5e-5
    • batch size = 32
    • no Validation
  • Training on current competition data
    • Total no. of epochs : 10
    • Learning Rate : 5e-5
    • 90% training data and 10% validation data
    • No preprocessing

Model 5 Performance

  • Training Accuracy around 96%
  • validation accuracy around 94%
  • Leaderboard score - Public : 0.794 Private : 0.915

6. Ensembling and TTA :

  • Each model gives predictions on 1. original image, 2. image rotated at 1-6 degrees, 3. image rotated at 7-12 degrees
  • stacking the predictions in such a way that for each type of tta we get one list of 5 elements (1 prediction from each model)
    • explanation - m1_p1 : model 1 predictions 1 on original image
      m1_p2 : model 1 prediction 2 on rotated image
  • this gives 3 lists of predictions such as : [m1_p1, m2_p1, m3_p1, m4_p1, m5_p1]
    [m1_p2, m2_p2, m3_p2, m4_p2, m5_p2]
    [m1_p3, m2_p3, m3_p3, m4_p3, m5_p3]
  • mode of each list is taken (this gives 3 values) and then mode of these 3 values is taken and this prediction is considered final.