Dataset Link: https://www.kaggle.com/datasets/tom99763/testtt
Kaggle Notebook: https://www.kaggle.com/code/shasan07/fer-pytorch
The images were transformed and normalized using torch, then we fed them to the pretrained DenseNet121 model for fine-Tuning. (Several other popular pre-trained models are also listed along with it).
A learning rate of 1e-4 was used, along with batch size of 8, and the training was conducted for 50 epochs. The best validation accuracy achieved by our model is 96.26%.
Training Summary:
Epoch [1/50] Train Loss: 1.1324 | Train Acc: 60.38% || Val Loss: 0.4929 | Val Acc: 81.80%
Saved new best model with val_acc: 81.80%
Epoch [2/50] Train Loss: 0.4077 | Train Acc: 87.49% || Val Loss: 0.3704 | Val Acc: 87.41%
Saved new best model with val_acc: 87.41%
Epoch [3/50] Train Loss: 0.2273 | Train Acc: 92.98% || Val Loss: 0.2418 | Val Acc: 92.01%
Saved new best model with val_acc: 92.01%
Epoch [4/50] Train Loss: 0.1222 | Train Acc: 96.89% || Val Loss: 0.2428 | Val Acc: 92.01%
Epoch [5/50] Train Loss: 0.1046 | Train Acc: 97.06% || Val Loss: 0.2372 | Val Acc: 91.67%
Epoch [6/50] Train Loss: 0.0743 | Train Acc: 98.13% || Val Loss: 0.1703 | Val Acc: 94.39%
Saved new best model with val_acc: 94.39%
Epoch [7/50] Train Loss: 0.0617 | Train Acc: 98.30% || Val Loss: 0.3382 | Val Acc: 88.44%
Epoch [8/50] Train Loss: 0.0567 | Train Acc: 98.47% || Val Loss: 0.1975 | Val Acc: 92.86%
Epoch [9/50] Train Loss: 0.0582 | Train Acc: 98.38% || Val Loss: 0.2311 | Val Acc: 92.52%
Epoch [10/50] Train Loss: 0.0417 | Train Acc: 98.64% || Val Loss: 0.2763 | Val Acc: 92.35%
Epoch [11/50] Train Loss: 0.0753 | Train Acc: 97.62% || Val Loss: 0.2098 | Val Acc: 92.86%
Epoch [12/50] Train Loss: 0.0429 | Train Acc: 98.77% || Val Loss: 0.2102 | Val Acc: 94.05%
Epoch [13/50] Train Loss: 0.0469 | Train Acc: 98.43% || Val Loss: 0.2433 | Val Acc: 91.67%
Epoch [14/50] Train Loss: 0.0195 | Train Acc: 99.45% || Val Loss: 0.2564 | Val Acc: 93.88%
Epoch [15/50] Train Loss: 0.0322 | Train Acc: 99.06% || Val Loss: 0.2419 | Val Acc: 92.35%
Epoch [16/50] Train Loss: 0.0456 | Train Acc: 98.60% || Val Loss: 0.2217 | Val Acc: 93.20%
Epoch [17/50] Train Loss: 0.0301 | Train Acc: 98.98% || Val Loss: 0.2320 | Val Acc: 93.03%
Epoch [18/50] Train Loss: 0.0312 | Train Acc: 99.15% || Val Loss: 0.1616 | Val Acc: 94.39%
Epoch [19/50] Train Loss: 0.0059 | Train Acc: 99.91% || Val Loss: 0.1668 | Val Acc: 94.73%
Saved new best model with val_acc: 94.73%
Epoch [20/50] Train Loss: 0.0031 | Train Acc: 99.96% || Val Loss: 0.1425 | Val Acc: 95.58%
Saved new best model with val_acc: 95.58%
Epoch [21/50] Train Loss: 0.0022 | Train Acc: 99.96% || Val Loss: 0.1634 | Val Acc: 94.90%
Epoch [22/50] Train Loss: 0.0251 | Train Acc: 99.28% || Val Loss: 0.2292 | Val Acc: 92.52%
Epoch [23/50] Train Loss: 0.0402 | Train Acc: 98.81% || Val Loss: 0.2504 | Val Acc: 92.86%
Epoch [24/50] Train Loss: 0.0508 | Train Acc: 98.38% || Val Loss: 0.1554 | Val Acc: 95.41%
Epoch [25/50] Train Loss: 0.0179 | Train Acc: 99.40% || Val Loss: 0.2003 | Val Acc: 95.24%
Epoch [26/50] Train Loss: 0.0129 | Train Acc: 99.70% || Val Loss: 0.1573 | Val Acc: 94.56%
Epoch [27/50] Train Loss: 0.0050 | Train Acc: 99.91% || Val Loss: 0.1797 | Val Acc: 95.24%
Epoch [28/50] Train Loss: 0.0026 | Train Acc: 99.96% || Val Loss: 0.1645 | Val Acc: 95.41%
Epoch [29/50] Train Loss: 0.0276 | Train Acc: 98.98% || Val Loss: 0.3311 | Val Acc: 90.65%
Epoch [30/50] Train Loss: 0.0572 | Train Acc: 98.26% || Val Loss: 0.2155 | Val Acc: 93.37%
Epoch [31/50] Train Loss: 0.0254 | Train Acc: 99.49% || Val Loss: 0.2288 | Val Acc: 93.88%
Epoch [32/50] Train Loss: 0.0124 | Train Acc: 99.70% || Val Loss: 0.2176 | Val Acc: 94.56%
Epoch [33/50] Train Loss: 0.0137 | Train Acc: 99.57% || Val Loss: 0.2769 | Val Acc: 91.84%
Epoch [34/50] Train Loss: 0.0111 | Train Acc: 99.70% || Val Loss: 0.2203 | Val Acc: 94.73%
Epoch [35/50] Train Loss: 0.0463 | Train Acc: 98.34% || Val Loss: 0.2478 | Val Acc: 92.18%
Epoch [36/50] Train Loss: 0.0212 | Train Acc: 99.28% || Val Loss: 0.2360 | Val Acc: 93.03%
Epoch [37/50] Train Loss: 0.0259 | Train Acc: 99.32% || Val Loss: 0.2064 | Val Acc: 92.86%
Epoch [38/50] Train Loss: 0.0039 | Train Acc: 99.96% || Val Loss: 0.1524 | Val Acc: 96.26%
Saved new best model with val_acc: 96.26%
Epoch [39/50] Train Loss: 0.0053 | Train Acc: 99.87% || Val Loss: 0.2094 | Val Acc: 95.58%
Epoch [40/50] Train Loss: 0.0265 | Train Acc: 99.23% || Val Loss: 0.4456 | Val Acc: 88.27%
Epoch [41/50] Train Loss: 0.0172 | Train Acc: 99.53% || Val Loss: 0.2251 | Val Acc: 94.22%
Epoch [42/50] Train Loss: 0.0023 | Train Acc: 99.96% || Val Loss: 0.1783 | Val Acc: 95.24%
Epoch [43/50] Train Loss: 0.0017 | Train Acc: 99.96% || Val Loss: 0.1831 | Val Acc: 95.58%
Epoch [44/50] Train Loss: 0.0016 | Train Acc: 99.96% || Val Loss: 0.1795 | Val Acc: 96.09%
Epoch [45/50] Train Loss: 0.0046 | Train Acc: 99.79% || Val Loss: 0.2901 | Val Acc: 92.86%
Epoch [46/50] Train Loss: 0.0731 | Train Acc: 97.70% || Val Loss: 0.2256 | Val Acc: 94.05%
Epoch [47/50] Train Loss: 0.0118 | Train Acc: 99.66% || Val Loss: 0.2706 | Val Acc: 93.88%
Epoch [48/50] Train Loss: 0.0114 | Train Acc: 99.74% || Val Loss: 0.2619 | Val Acc: 93.71%
Epoch [49/50] Train Loss: 0.0169 | Train Acc: 99.66% || Val Loss: 0.3609 | Val Acc: 93.20%
Epoch [50/50] Train Loss: 0.0147 | Train Acc: 99.57% || Val Loss: 0.2783 | Val Acc: 93.71%
Obtained Results:
Classification Report:
precision recall f1-score support
Anger 0.99 0.92 0.95 84
Disgust 1.00 0.95 0.98 84
Fear 0.92 0.94 0.93 84
Happy 0.98 0.99 0.98 84
Neutral 0.97 0.99 0.98 84
Sad 0.94 0.96 0.95 84
Surprise 0.95 0.99 0.97 84
accuracy 0.96 588
macro avg 0.96 0.96 0.96 588
weighted avg 0.96 0.96 0.96 588
