• Adding presentation and official documentation
• Splitting notebook per sections
• Adding additional comments to the code

• Adding ViT-B/16/S model on 25 epochs with constant learning rate
• Checking on training and validation accuracy/loss parameters according to the training log
• Display results on standalone plots

• Adding evaluation for ResNet18
• Debugging on SAM model evaluation
• Improvment Training Plot support curves on N < 5 lines
• Model adaptation during loading on evaluation (standalone) with adapting on backbones

• Adding ResNet 18 (11M parameters)
• Upload history for loss and accuracy
• Upload epoch 20 dump
• Upload final model checkpoint

• Adding ViT-B/16/SG
• Gradual learning rate every 10 epochs
• SGD optimization
• Adding loss and accuracy histories

• Adding VIT-B/16 model checkpoint using customized learning rate scheduler
• Adding SAM to the model as a optimization algorithm to smooth the loss landscape
• Adding history for training and validation loss
• Adding history for training and validation accuracy

• Configuration of resnet18 with gradual learning rate
• Starting learning rate at 0.01
• Epochs 50 with plateau at 25
• Loading training and validation accuracy histories

• Adding SAM optimization for VIT-B/16
• Defining closure for sharpness-aware minimization efficiency
• Debugging model loader for the checkpoints recovery

• Upload optimal model on AffectNet
• Defines evaluation plots on accuracy and loss values

• Adding gradual learning rate
• Modify dataset with AffectNet in validation and testing set
• Adding scheduler for learning rate adjustment

• Extends data analysis for the AffectNet, CK+48 and FER-2013

• Creation of AVFER with the following features

  • Splitting initial dataset in training and testing set with ratio 80/20
  • Splitting validation and training set with ratio 90/10
  • Testing and validation set contains only samples from AffectNet (RGB and high quality images)

Drive of AVFER: https://drive.google.com/drive/folders/1-8WG_CNrU3chL_OHpkM8EYx3Bm129cnE?usp=sharing

• Adjust train and test splitting
• Balancing augmentation over 150.000 samples
• Removing augmentation on validation to increment variability
• Loading of vfer for 5, 15 and 25 epochs of training on the result dataset
• Loading history for training and validation accuracy/loss

• Adding integration with partial training during the transformer weights improvements (best-fit)
• Updating of the VFER model on 5/50 training epochs with 62% accuracy (state-of-art of AffectNet visual transformer)
• Integrating with fluid system for face detection in the cropping phase

• Adjust normalization parameters from [0.48, 0.28] to 0.5
• Balancing dataset with not augment element in validation
• Resize the training set on double capacity for less epochs on training phase
• Adding featuring and inference on video capture tools in OpenCV for models applications

• Model dump for batch 50 on 12 epochs for the VFER transformer, accuracy of 69%
• Model dump for batch 60 on 24 epochs for the VFER transformer, accuracy of 70%
• Model dump for batch 60 on 50 epochs for the VFER transformer, accuracy of 71%
• Debugging notebook for the loss evaluation
• Adding every section until the evaluation
• Integration of the dataset available here