Update PiT_MNIST_Colab_README.md

PiT_MNIST_Colab_README.md

@@ -1,10 +1,12 @@
 The following PiT_MNIST_V1.0.ipynb is a direct implementationi of the PiT pixel transformer described in the 2024 paper titled 
 An Image is Worth More Than 16 x 16 Patches: Exploring Transformers on Individual Pixels 
-at https://arxiv.org/html/2406.09415v1
-Which describes "directly treating each individual pixel as a token and achieve highly performant results"
+at https://arxiv.org/html/2406.09415v1 which describes "directly treating each individual pixel as a token and achieve highly performant results"
 This script simply applies this PiT model architecture without any modifications to the standard NMIST numeral-images-classification dataset that is provided in Google Colab sample_data folder.
 The script was ran for 25 epochs and obtained 92.30 Accuracy on the Validation set ( Train Loss: 0.2800 | Val Loss: 0.2441 | Val Acc: 92.30%) by epoch 15.
-Loss fell and Accuracy increased monontonically per each epoch.
+Loss fell and Accuracy increased (almost) monontonically per each epoch until Epoch 18.  (one minor dip in accuracy between Epoch 13 and 14, and again at Epoch 18-19, and 23-24 while Train Loss always continued to drop)
+Final Test Accuracy: 95.01% (25 Epochs)
+Final Test Loss: 0.1662
+
 
 # ==============================================================================
 # PiT_MNIST_V1.0.py   [in colab: PiT_MNIST_V1.0.ipynb]
@@ -305,6 +307,39 @@ Epoch 08/25 | Train Loss: 0.4682 | Val Loss: 0.3680 | Val Acc: 88.05%
   -> New best validation accuracy! Saving model state.
 Epoch 09/25 | Train Loss: 0.4264 | Val Loss: 0.3446 | Val Acc: 89.20%
   -> New best validation accuracy! Saving model state.
+Epoch 10/25 | Train Loss: 0.4038 | Val Loss: 0.3163 | Val Acc: 89.95%
+  -> New best validation accuracy! Saving model state.
+Epoch 11/25 | Train Loss: 0.3641 | Val Loss: 0.2941 | Val Acc: 90.80%
+  -> New best validation accuracy! Saving model state.
+Epoch 12/25 | Train Loss: 0.3447 | Val Loss: 0.2759 | Val Acc: 91.45%
+  -> New best validation accuracy! Saving model state.
+Epoch 13/25 | Train Loss: 0.3181 | Val Loss: 0.2603 | Val Acc: 92.05%
+  -> New best validation accuracy! Saving model state.
+Epoch 14/25 | Train Loss: 0.3023 | Val Loss: 0.2695 | Val Acc: 91.90%
+Epoch 15/25 | Train Loss: 0.2800 | Val Loss: 0.2441 | Val Acc: 92.30%
+  -> New best validation accuracy! Saving model state.
+Epoch 16/25 | Train Loss: 0.2677 | Val Loss: 0.2377 | Val Acc: 92.65%
+  -> New best validation accuracy! Saving model state.
+Epoch 17/25 | Train Loss: 0.2535 | Val Loss: 0.2143 | Val Acc: 93.80%
+  -> New best validation accuracy! Saving model state.
+Epoch 18/25 | Train Loss: 0.2395 | Val Loss: 0.2059 | Val Acc: 94.05%
+  -> New best validation accuracy! Saving model state.
+Epoch 19/25 | Train Loss: 0.2276 | Val Loss: 0.2126 | Val Acc: 93.75%
+Epoch 20/25 | Train Loss: 0.2189 | Val Loss: 0.1907 | Val Acc: 94.40%
+  -> New best validation accuracy! Saving model state.
+Epoch 21/25 | Train Loss: 0.2113 | Val Loss: 0.1892 | Val Acc: 94.35%
+Epoch 22/25 | Train Loss: 0.2004 | Val Loss: 0.1775 | Val Acc: 94.50%
+  -> New best validation accuracy! Saving model state.
+Epoch 23/25 | Train Loss: 0.1927 | Val Loss: 0.1912 | Val Acc: 94.15%
+Epoch 24/25 | Train Loss: 0.1836 | Val Loss: 0.1746 | Val Acc: 94.75%
+  -> New best validation accuracy! Saving model state.
+Epoch 25/25 | Train Loss: 0.1804 | Val Loss: 0.1642 | Val Acc: 94.75%
+--- Training Finished ---
+
+--- Evaluating on Test Set ---
+                                                                   Final Test Loss: 0.1662
+Final Test Accuracy: 95.01%
+----------------------------