RML2016.10A - Transformer (6 encoders, no convolution) - QPSK step-by-step flow (Kaggle)
Model: /kaggle/input/datasets/tiodomm/rml2016-noconv-transformer/RML2016_10a_4Models/experiments/4models/TransformerNoConv/best_model.pt  (epoch 90)
==============================================================================

Input 2x128 I/Q -> 8 patches (8x32) -> Linear 32->128 -> class token + position (9x128)
-> 6 pre-norm encoders (4 heads, FFN 512, GELU) -> final LayerNorm -> class token -> Linear 128->11

QPSK signal: QPSK, 18 dB, sample 18 (dataset index 199018, test split)
Prediction : 9 = QPSK   probability = 0.921864
Stages     : 18 PASS / 0 FAIL (max error 6.75e-06)

TEST SET (44,000 signals)        CSV-weight NumPy   PyTorch (training run)
Test accuracy (all SNR)  :       62.76%           62.76%
Mean accuracy SNR >= 0 dB:       92.60%           92.60%
Accuracy at +18 dB       :       93.36%           93.36%
(differences of a few signals = float64 CPU vs float32 GPU rounding on borderline signals)

FOLDERS
01_training/            best_model.pt, training history, results, scripts, split + normalization
02_reference_outputs/   RML2016.10A_reference_outputs: stage outputs of the QPSK signal
                        02_preprocessed, 04_patches, 05_embeddings, 06_encoder_outputs,
                        07_classifier, 08_predictions, 10_verification
03_weight_export/       MRTransformer_RML2016: best_model.pt exported to CSV + PyTorch reference outputs
04_verification/        verification_output: python_calculated + report
05_test_results/        accuracy_vs_snr, confusion_matrix_all_snr, confusion_matrix_plus18dB
06_function_sheet/      RML2016_Transformer_Functions.xlsx
