The purpose of this copy of the MNIST small dataset [mnist_test.csv (20,000 samples) and mnist_train_small.csv (10,000 samples)] copied from sample_data folder in Google Colab is simply to illustrate how WEIRD and totally deformed/unrecognizable are the 1% to 2% test samples that are difficult for a competent Vision model to correctly classify. See for yourself (up to 4 misclassified test samples shown per training epoch)
Vision_model_V2.1.py
--- Hyperparameters ---
[INFO] Loading datasets... [AUGMENT] Creating augmented training data... [AUGMENT] Created augmentation pipeline: [RandomAffine(degrees=[-8.0, 8.0], translate=(0.07142857142857142, 0), shear=[0.0, 0.0])] [AUGMENT] Original train size: 20000. New size: 40000 [INFO] Calculating class distribution and entropy...
--- Dataset Information --- Name: MNIST (Small) Source: Included in Google Colab's /sample_data directory Original Train Samples 20000 Total Train Samples (w/ Aug) 40000 Test Samples 10000 Image Dimensions 1x28x28 Classes [np.int64(0), np.int64(1), np.int64(2), np.int64(3), np.int64(4), np.int64(5), np.int64(6), np.int64(7), np.int64(8), np.int64(9)] Class Entropy Contrib: Class 0: 0.3286 Class 1: 0.3540 Class 2: 0.3312 Class 3: 0.3342 Class 4: 0.3249 Class 5: 0.3087 Class 6: 0.3358 Class 7: 0.3438 Class 8: 0.3238 Class 9: 0.3343 Total Label Entropy: 3.3192 (Max: 3.3219)
--- Starting Training ---
[MODEL] New best accuracy: 97.83%. Saving model to output/best_model.pt
Epoch [01/20] | Train Loss: 0.1843, Train Acc: 94.28% | Test Loss: 0.0667, Test Acc: 97.83% | LR: 1.00e-03
[ANALYSIS] Displaying up to 4 failed test samples for Epoch 1...
[PLOT] Saved failed samples plot to output/failed_samples_epoch_1.png
[MODEL] New best accuracy: 98.10%. Saving model to output/best_model.pt
Epoch [02/20] | Train Loss: 0.0830, Train Acc: 97.45% | Test Loss: 0.0570, Test Acc: 98.10% | LR: 1.00e-03
[ANALYSIS] Displaying up to 4 failed test samples for Epoch 2...
[PLOT] Saved failed samples plot to output/failed_samples_epoch_2.png
[MODEL] New best accuracy: 98.29%. Saving model to output/best_model.pt
Epoch [03/20] | Train Loss: 0.0589, Train Acc: 98.16% | Test Loss: 0.0529, Test Acc: 98.29% | LR: 1.00e-03
[ANALYSIS] Displaying up to 4 failed test samples for Epoch 3...
[PLOT] Saved failed samples plot to output/failed_samples_epoch_3.png
[MODEL] New best accuracy: 98.72%. Saving model to output/best_model.pt
Epoch [04/20] | Train Loss: 0.0473, Train Acc: 98.51% | Test Loss: 0.0426, Test Acc: 98.72% | LR: 1.00e-03
[ANALYSIS] Displaying up to 4 failed test samples for Epoch 4...
[PLOT] Saved failed samples plot to output/failed_samples_epoch_4.png
Epoch [05/20] | Train Loss: 0.0398, Train Acc: 98.71% | Test Loss: 0.0470, Test Acc: 98.54% | LR: 1.00e-03
[ANALYSIS] Displaying up to 4 failed test samples for Epoch 5...
[PLOT] Saved failed samples plot to output/failed_samples_epoch_5.png
[MODEL] New best accuracy: 98.75%. Saving model to output/best_model.pt
Epoch [06/20] | Train Loss: 0.0374, Train Acc: 98.79% | Test Loss: 0.0400, Test Acc: 98.75% | LR: 1.00e-03
[ANALYSIS] Displaying up to 4 failed test samples for Epoch 6...
[PLOT] Saved failed samples plot to output/failed_samples_epoch_6.png
Confusion Matrix:
Training Telemetry Plot:
