/aifarm-classification

implemented with base_line code

Primary LanguagePython

aifarm (소달구지)

Implementation Details.

  • Implemented with base_line code
  • Pretrained models from timm github (https://github.com/rwightman/pytorch-image-models) 
    pip install timm
  • Used Modified Cross Etropy Loss
    • Cross Entropy Loss (ce)
    • Weight Cross Etropy Loss (w_ce) 
      if LOSS_FN == "w_ce":          
   # Weight Cross Entropy Loss
   # Use the calculated num of img in trainset for cross entropy weights.
   _, num_imgs_class  = train_dataset.data_loader()
   num_imgs_class = torch.FloatTensor(num_imgs_class)
   print("num of imgs for classes:", num_imgs_class)
   class_percentage = num_imgs_class / num_imgs_class.sum()
   class_weights = 1.0 /class_percentage
   class_weights = (class_weights / class_weights.sum())
   class_weights = torch.exp(class_weights)
   class_weights = class_weights.to(device)
   print("weights for classes:", class_weights)
   criterion = nn.CrossEntropyLoss(weight=class_weights)
    • Focal Loss(focal_loss) 
      class FocalLoss(nn.Module):
    def __init__(self, alpha=1, gamma=2, weight=1, logits=False, reduce=True):
        super(FocalLoss, self).__init__()
        self.alpha = alpha
        self.gamma = gamma
        self.logits = logits
        self.reduce = reduce
        self.weight = weight

   def forward(self, inputs, targets):
        ce_loss = nn.CrossEntropyLoss(weight=self.weight, reduction="none")(inputs, targets)

        pt = torch.exp(-ce_loss)
        F_loss = self.alpha * (1-pt) ** self.gamma * ce_loss

       if self.reduce:
           return torch.mean(F_loss)
       else:
           F_loss
    • Weight Focal Loss: Weighted Cross Entropy Loss + Focal Loss (w_focal_loss) <= Mainly Used 
      if LOSS_FN == "w_focal_loss":
   # Combine Focal Loss & Weight Cross Entropy Loss
   _, num_imgs_class  = train_dataset.data_loader()
   num_imgs_class = torch.FloatTensor(num_imgs_class)
   print("num of imgs for classes:", num_imgs_class)
   class_percentage = num_imgs_class / num_imgs_class.sum()
   class_weights = 1.0 /class_percentage
   class_weights = (class_weights / class_weights.sum())
   class_weights = torch.exp(class_weights)
   class_weights = class_weights.to(device)
   print("weights for classes:", class_weights)
   criterion = FocalLoss(alpha=1, gamma=2, weight=class_weights, reduce=True)
  • Top 5 Models Trained with corresponding Config Files.
    • tf_efficientnetv2_l_in21ft1k => train_config_best1
    • tf_efficientnetv2_m_in21ft1k => train_config_best2
    • tf_efficientnet_b6_ns => train_config4
    • tf_efficientnetv2_l_in21ft1k => train_config2
    • swin_large_patch4_window7_224 => train_config3

Top 2 Models

tf_efficientnetv2_m_in21ft1k

  • train: 0.999549062049062 val: 0.994590695997115
  • loss: w_focal_loss
  • sche: msl
  • optim: adam
  • Test
    • Private: 0.9903 image image

tf_efficientnetv2_l_in21ft1k

  • train: 0.9999098124098124 val: 0.997836278398846
  • loss: ce
  • sche: msl
  • optim: adam
  • Test
    • Private: 0.9922 image image

Best Score Public: 0.9934

Directory.

├── aifarm_baseline
│   ├── config
│   ├── models
│   ├── modules
│   ├── results
│   │   ├── train
│   │   └── pred
│   ├── command.py
│   ├── logger_20210712.py
│   ├── predict.py
│   ├── README.md
│   └── train.py
└── dataset
    └── Tomato
        ├── test
        └── train
            ├──Tomato_D01
            ├──Tomato_D01
            ...
            └──Tomato_R01

Train Process.

python command.py

<command.py>

...

command_list = [
###

## Best
"python train.py --yml train_config_best1",  # tf_efficientnetv2_m_in21ft1k, bathsize 32 => pub test: 99.03
"python train.py --yml train_config_best2",  # tf_efficientnetv2_l_in21ft1k, bathsize 16 => pub test: 99.22

###
]

...

Example Config File (for model: tf_efficientnetv2_m_in21ft1k)

SEED:
  random_seed: 42
DATALOADER:
  num_workers: 8
  shuffle: true
  pin_memory: true
  drop_last: false
TRAIN:
  input_shape: 224
  num_epochs: 50
  batch_size: 32
  learning_rate: 1e-4
  early_stopping_patience: 10
  model: tf_efficientnetv2_m_in21ft1k
  optimizer: adam
  scheduler: msl
  momentum: null
  weight_decay: 5.0e-05
  loss_function: w_focal_loss
  metric_function: null
PERFORMANCE_RECORD:
  column_list:
  - train_serial
  - train_timestamp
  - model_str
  - optimizer_str
  - loss_function_str
  - metric_function_str
  - early_stopping_patience
  - batch_size
  - epoch
  - learning_rate
  - momentum
  - random_seed
  - epoch_index
  - train_loss
  - validation_loss
  - train_score
  - validation_score
  - elapsed_time

Test Process.

modify <predict.py> for tf_efficientnetv2_l_in21ft1k

DIRECTORY:
  data: '../data/test'
  model: "tf_efficientnetv2_l_in21ft1k_20211007123012" 

MODEL: "tf_efficientnetv2_l_in21ft1k"

modify <predict.py> for tf_efficientnetv2_l_in21ft1k

DIRECTORY:
  data: '../data/test'
  model: "tf_efficientnetv2_m_in21ft1k_20211006215437" 

MODEL: "tf_efficientnetv2_m_in21ft1k"

Run Prediction (TEST)

python predict.py

Top 2 Trained Model Weight File Directory

./results/train/best/tf_efficientnetv2_l_in21ft1k_20211007123012/best.pt ./results/train/best/tf_efficientnetv2_m_in21ft1k_20211006215437/best.pt

Top 2 Trained Model Prediction File Directory

./results/pred/pred_tf_efficientnetv2_l_in21ft1k_20211007123012.csv ./results/pred/pred_tf_efficientnetv2_m_in21ft1k_20211006215437.csv