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 for the AdamW optimizer, along with batch size of 8, and the training was conducted for 50 epochs. The best validation accuracy achieved by our model is 94.49%.
Training Summary:
Epoch [1/50] Train Loss: 1.1639 | Train Acc: 57.76% || Val Loss: 0.5370 | Val Acc: 82.76%
Saved new best model with val_acc: 82.76%
Epoch [2/50] Train Loss: 0.4200 | Train Acc: 85.83% || Val Loss: 0.3926 | Val Acc: 86.73%
Saved new best model with val_acc: 86.73%
Epoch [3/50] Train Loss: 0.2386 | Train Acc: 92.34% || Val Loss: 0.3453 | Val Acc: 89.59%
Saved new best model with val_acc: 89.59%
Epoch [4/50] Train Loss: 0.1351 | Train Acc: 95.89% || Val Loss: 0.3586 | Val Acc: 89.90%
Saved new best model with val_acc: 89.90%
Epoch [5/50] Train Loss: 0.1163 | Train Acc: 96.38% || Val Loss: 0.2585 | Val Acc: 91.73%
Saved new best model with val_acc: 91.73%
Epoch [6/50] Train Loss: 0.0898 | Train Acc: 97.14% || Val Loss: 0.4773 | Val Acc: 92.55%
Saved new best model with val_acc: 92.55%
Epoch [7/50] Train Loss: 0.0843 | Train Acc: 97.14% || Val Loss: 0.2090 | Val Acc: 93.67%
Saved new best model with val_acc: 93.67%
Epoch [8/50] Train Loss: 0.0682 | Train Acc: 97.88% || Val Loss: 0.3692 | Val Acc: 89.39%
Epoch [9/50] Train Loss: 0.0634 | Train Acc: 98.03% || Val Loss: 0.3282 | Val Acc: 91.73%
Epoch [10/50] Train Loss: 0.0483 | Train Acc: 98.67% || Val Loss: 0.2505 | Val Acc: 92.35%
Epoch [11/50] Train Loss: 0.0515 | Train Acc: 98.37% || Val Loss: 0.3511 | Val Acc: 91.94%
Epoch [12/50] Train Loss: 0.0399 | Train Acc: 98.88% || Val Loss: 0.4205 | Val Acc: 89.59%
Epoch [13/50] Train Loss: 0.0360 | Train Acc: 99.00% || Val Loss: 0.2864 | Val Acc: 92.55%
Epoch [14/50] Train Loss: 0.0369 | Train Acc: 98.88% || Val Loss: 0.3243 | Val Acc: 92.24%
Epoch [15/50] Train Loss: 0.0299 | Train Acc: 99.16% || Val Loss: 0.4572 | Val Acc: 90.51%
Epoch [16/50] Train Loss: 0.0427 | Train Acc: 98.88% || Val Loss: 0.3331 | Val Acc: 90.92%
Epoch [17/50] Train Loss: 0.0337 | Train Acc: 99.13% || Val Loss: 0.3805 | Val Acc: 91.22%
Epoch [18/50] Train Loss: 0.0354 | Train Acc: 99.03% || Val Loss: 0.4215 | Val Acc: 89.80%
Epoch [19/50] Train Loss: 0.0449 | Train Acc: 98.70% || Val Loss: 0.2631 | Val Acc: 92.65%
Epoch [20/50] Train Loss: 0.0225 | Train Acc: 99.26% || Val Loss: 0.2593 | Val Acc: 93.88%
Saved new best model with val_acc: 93.88%
Epoch [21/50] Train Loss: 0.0298 | Train Acc: 99.08% || Val Loss: 0.3229 | Val Acc: 92.65%
Epoch [22/50] Train Loss: 0.0189 | Train Acc: 99.34% || Val Loss: 0.3232 | Val Acc: 92.96%
Epoch [23/50] Train Loss: 0.0320 | Train Acc: 99.23% || Val Loss: 0.4447 | Val Acc: 92.76%
Epoch [24/50] Train Loss: 0.0253 | Train Acc: 99.26% || Val Loss: 0.3163 | Val Acc: 92.35%
Epoch [25/50] Train Loss: 0.0300 | Train Acc: 99.23% || Val Loss: 0.2847 | Val Acc: 93.57%
Epoch [26/50] Train Loss: 0.0167 | Train Acc: 99.46% || Val Loss: 0.3133 | Val Acc: 92.96%
Epoch [27/50] Train Loss: 0.0194 | Train Acc: 99.57% || Val Loss: 0.2613 | Val Acc: 93.78%
Epoch [28/50] Train Loss: 0.0299 | Train Acc: 99.13% || Val Loss: 0.3292 | Val Acc: 93.06%
Epoch [29/50] Train Loss: 0.0210 | Train Acc: 99.34% || Val Loss: 0.3586 | Val Acc: 93.06%
Epoch [30/50] Train Loss: 0.0211 | Train Acc: 99.36% || Val Loss: 0.3061 | Val Acc: 93.78%
Epoch [31/50] Train Loss: 0.0211 | Train Acc: 99.41% || Val Loss: 0.3750 | Val Acc: 92.76%
Epoch [32/50] Train Loss: 0.0169 | Train Acc: 99.46% || Val Loss: 0.2669 | Val Acc: 93.98%
Saved new best model with val_acc: 93.98%
Epoch [33/50] Train Loss: 0.0238 | Train Acc: 99.11% || Val Loss: 0.3924 | Val Acc: 92.14%
Epoch [34/50] Train Loss: 0.0265 | Train Acc: 99.11% || Val Loss: 0.3600 | Val Acc: 92.45%
Epoch [35/50] Train Loss: 0.0269 | Train Acc: 99.16% || Val Loss: 0.2951 | Val Acc: 93.88%
Epoch [36/50] Train Loss: 0.0089 | Train Acc: 99.77% || Val Loss: 0.2502 | Val Acc: 93.47%
Epoch [37/50] Train Loss: 0.0158 | Train Acc: 99.59% || Val Loss: 0.3338 | Val Acc: 92.86%
Epoch [38/50] Train Loss: 0.0248 | Train Acc: 99.46% || Val Loss: 0.3596 | Val Acc: 92.96%
Epoch [39/50] Train Loss: 0.0290 | Train Acc: 99.23% || Val Loss: 0.3590 | Val Acc: 92.24%
Epoch [40/50] Train Loss: 0.0287 | Train Acc: 99.16% || Val Loss: 0.2972 | Val Acc: 93.78%
Epoch [41/50] Train Loss: 0.0138 | Train Acc: 99.59% || Val Loss: 0.3068 | Val Acc: 92.76%
Epoch [42/50] Train Loss: 0.0257 | Train Acc: 99.39% || Val Loss: 0.2569 | Val Acc: 93.57%
Epoch [43/50] Train Loss: 0.0149 | Train Acc: 99.67% || Val Loss: 0.2555 | Val Acc: 94.18%
Saved new best model with val_acc: 94.18%
Epoch [44/50] Train Loss: 0.0145 | Train Acc: 99.49% || Val Loss: 0.3501 | Val Acc: 92.24%
Epoch [45/50] Train Loss: 0.0093 | Train Acc: 99.74% || Val Loss: 0.3772 | Val Acc: 93.06%
Epoch [46/50] Train Loss: 0.0210 | Train Acc: 99.49% || Val Loss: 0.3676 | Val Acc: 92.76%
Epoch [47/50] Train Loss: 0.0131 | Train Acc: 99.74% || Val Loss: 0.2568 | Val Acc: 94.49%
Saved new best model with val_acc: 94.49%
Epoch [48/50] Train Loss: 0.0081 | Train Acc: 99.72% || Val Loss: 0.3309 | Val Acc: 93.88%
Epoch [49/50] Train Loss: 0.0169 | Train Acc: 99.57% || Val Loss: 0.2729 | Val Acc: 92.96%
Epoch [50/50] Train Loss: 0.0118 | Train Acc: 99.67% || Val Loss: 0.3195 | Val Acc: 93.57%
Obtained Results:
Classification Report:
precision recall f1-score support
Anger 0.98 0.91 0.95 140
Disgust 0.90 0.97 0.93 140
Fear 0.93 0.89 0.91 140
Happy 0.99 0.99 0.99 140
Neutral 0.96 0.98 0.97 140
Sad 0.93 0.92 0.92 140
Surprise 0.93 0.96 0.94 140
accuracy 0.94 980
macro avg 0.95 0.94 0.94 980
weighted avg 0.95 0.94 0.94 980
Confusion Matrix:
[[128 6 2 0 1 3 0]
[ 1 136 0 1 0 2 0]
[ 1 0 124 1 3 3 8]
[ 0 0 1 138 0 0 1]
[ 0 1 0 0 137 2 0]
[ 0 7 2 0 1 129 1]
[ 0 1 5 0 0 0 134]]
