Transformer impv

.gitignore

@@ -0,0 +1,4 @@
+*.csv
+*.pyc
+*.png
+*.json

README.md

@@ -1,2 +1 @@
 # denoising
-Spectra denoising

configs/config.yml

@@ -23,4 +23,9 @@ split:
   train_fraction: 0.70
   val_fraction: 0.15
   test_fraction: 0.15
-  random_seed: 42
+  random_seed: 42
+  stratify_by_recali: true # preserve groups while matching the recali True/False ratio in every split
+  length_weight: true # if true, multiply the FDR weight by a clipped per-spectrum true/false ratio
+  length_weight_eps: 1.0 # smoothing so spectra with zero true or zero false labels stay finite
+  length_weight_min: 0.5 # lower clip for the true/false ratio factor
+  length_weight_max: 2.0 # upper clip for the true/false ratio factor

configs/infer_transformer_imp.yml

@@ -0,0 +1,48 @@
+input:
+  mgf_path: "/mnt/c/Users/yynk1/Downloads/denoising/B190717_12_HF_LS_IN_130_ECLP_DSSO_01_SCX23_hSAX05_rep1.mgf"
+  peak_list_file_name: "B190717_12_HF_LS_IN_130_ECLP_DSSO_01_SCX23_hSAX05_rep1.mgf" #PeakListFileName
+  recali: null # set true/false when the checkpoint was trained with use_recali_label: true
+
+model:
+  checkpoint_path: "/mnt/c/Users/yynk1/Downloads/denoising/transformer_peak_classifier_imp_1pct_filtered_imp_03_threshold/best_model.pt"
+
+features:
+  peak_feature_columns:
+    - "peak_feature_log_intensity"
+    - "peak_feature_relative_intensity"
+    - "peak_feature_mz_over_precursor"
+    - "peak_feature_delta_to_precursor"
+    - "peak_feature_delta_prev"
+    - "peak_feature_delta_next"
+  spectrum_feature_columns:
+    - "Charge"
+    - "exp m/z"
+    - "num_peaks"
+    - "tic"
+  use_instrument_label: false
+  use_recali_label: false
+  recali_label_source_column: "recali"
+  use_raw_peak_mz: true
+  raw_peak_mz_column: "mz_arr"
+  use_raw_peak_intensity: false
+  raw_peak_intensity_column: "int_arr"
+  sort_raw_peak_inputs_by_mz: true
+  broadcast_spectrum_features_to_peaks: false
+  normalize_peak_features: true
+  normalize_spectrum_features: true
+
+evaluation:
+  use_checkpoint_threshold: true
+  threshold_for_binary_metrics: 0.3
+
+runtime:
+  device: "cuda"
+  batch_size: 2
+
+output:
+  output_dir: "/mnt/c/Users/yynk1/Downloads/denoising/inference_transformer_peak_classifier_imp"
+  write_filtered_mgf: true
+  write_dropped_peaks_report: true
+  dropped_peaks_report_name: "dropped_peaks.csv"
+  #default: drop low-probability peaks instead of writing zero-intensity peaks.
+  filtered_mgf_mode: "drop_below_threshold"

configs/train_autogluon.yml

@@ -0,0 +1,85 @@
+data:
+  train_path: "../data/splits_5pct_filtered_new/train.parquet"
+  val_path: "../data/splits_5pct_filtered_new/val.parquet"
+  test_path: "../data/splits_5pct_filtered_new/test.parquet"
+
+  target_column: "annotation_mask" # per-peak binary labels after flattening each spectrum into peak rows
+  use_training_weights: true # optionally repeat a spectrum weight onto each peak row for AutoGluon
+  weight_column: "weight"
+
+output:
+  #output_dir: "../outputs/splits_6pct_filtered"
+  output_dir: "../outputs/autogluon_imp_5pct_recali_instrument"
+  predictor_subdir: "autogluon_predictor"
+  unique_predictor_subdir: true # create a new predictor directory each run to avoid "Learner is already fit"
+  save_metrics_summary: true
+  save_confusion_matrices: true
+  save_predictions: true
+  save_flattened_tables: false # can become very large; enable only when you need to inspect the flat peak table
+  save_fit_summary: true
+  enable_file_logging: true
+  log_file_name: "training.log"
+
+features:
+  peak_feature_columns:
+    - "peak_feature_mz"
+    - "peak_feature_log_intensity"
+    - "peak_feature_relative_intensity"
+    - "peak_feature_mz_over_precursor"
+    - "peak_feature_delta_to_precursor"
+    - "peak_feature_delta_prev"
+    - "peak_feature_delta_next"
+
+  spectrum_feature_columns:
+    - "Charge"
+    - "exp m/z"
+    - "num_peaks"
+    - "tic"
+
+  use_raw_peak_mz: true
+  raw_peak_mz_column: "mz_arr"
+  use_raw_peak_intensity: true
+  raw_peak_intensity_column: "int_arr"
+  sort_raw_peak_inputs_by_mz: true
+  broadcast_spectrum_features_to_peaks: true
+  normalize_peak_features: true
+  normalize_spectrum_features: true
+  normalization_fit_split: "train"
+
+autogluon:
+  label: "annotation_mask"
+  problem_type: "binary"
+  eval_metric: "average_precision" # good fit for imbalanced signal-vs-noise prediction
+  presets: "high_quality" # strong baseline without the very large cost of best_quality
+  time_limit: 14400 # 4h default budget for a serious baseline on a large table
+  verbosity: 2
+  positive_class: 1
+  sample_weight_column: null
+  weight_evaluation: false
+  dynamic_stacking: false # avoid expensive DyStack / Ray side work unless you really need it
+  num_stack_levels: 0
+  use_bag_holdout: false # valid when tuning_data is passed alongside bagging
+  fit_weighted_ensemble: true
+  save_bag_folds: true
+  save_leaderboard: true
+  save_feature_importance: true
+  feature_importance_subsample_size: 50000
+  feature_importance_num_shuffle_sets: 5
+  feature_importance_time_limit: 600
+
+evaluation:
+  primary_metric: "pr_auc"
+  threshold_for_binary_metrics: 0.3
+
+  report_metrics:
+    - "pr_auc"
+    - "roc_auc"
+    - "f1"
+    - "mcc"
+    - "precision"
+    - "recall"
+
+  retained_peak_fractions:
+    - 0.3
+    - 0.5
+    - 0.7

configs/train_mlp.yml

@@ -0,0 +1,109 @@
+data:
+  train_path: "../data/splits/train.parquet"
+  val_path: "../data/splits/val.parquet"
+  test_path: "../data/splits/test.parquet"
+
+  target_column: "annotation_mask" # peak-level labels where True means signal and False means noise
+  use_training_weights: false # stage 1 MLP ignores FDR-derived weights (this will be used only in the Loss)
+  weight_column: "weight" # column name for training weights used in later stages
+
+output:
+  output_dir: "outputs/mlp_baseline" # directory where training artifacts will be saved
+  save_best_model: true # save the best checkpoint for later inference
+  save_epoch_history: true # save epoch-by-epoch metrics as a table
+  save_metrics_summary: true # save final validation and test metrics
+  save_plots: true # save training curves such as loss and PR-AUC
+  save_confusion_matrices: true # save confusion matrices for validation and test
+  device_verbose: true # print device infos! important for GPU/used-CPUs :)
+  enable_file_logging: true # mirror all console training output into a log file
+  log_file_name: "training.log" # written inside the model output directory
+
+features:
+  peak_feature_columns:
+    - "peak_feature_mz"
+    - "peak_feature_log_intensity"
+    - "peak_feature_relative_intensity"
+    - "peak_feature_mz_over_precursor"
+    - "peak_feature_delta_to_precursor" # precursor m/z minus peak m/z
+    - "peak_feature_delta_prev" # m/z difference to the previous peak after sorting
+    - "peak_feature_delta_next" # m/z difference to the next peak after sorting
+
+  spectrum_feature_columns:
+    - "Charge" # precursor charge state
+    - "exp m/z" # precursor experimental m/z
+    - "num_peaks" # number of peaks in the processed spectrum
+    - "tic" # total ion current computed as the sum of intensities
+
+  use_raw_peak_mz: false # optional raw per-peak m/z input from mz_arr
+  raw_peak_mz_column: "mz_arr"
+  use_raw_peak_intensity: false # optional raw per-peak intensity input from int_arr
+  raw_peak_intensity_column: "int_arr"
+  sort_raw_peak_inputs_by_mz: true # keep raw arrays aligned with sorted engineered peak features
+  broadcast_spectrum_features_to_peaks: true # repeat spectrum-level features for every peak before concatenation
+  normalize_peak_features: true # apply feature normalization to peak-level inputs
+  normalize_spectrum_features: true # apply feature normalization to spectrum-level inputs
+  normalization_fit_split: "train" # fit normalization statistics only on the training split
+
+model:
+  name: "mlp_peak_classifier" 
+
+  mlp:
+    hidden_dims: [128, 64] # hidden layer sizes of the MLP
+    dropout: 0.1 # dropout probability used between hidden layers
+    activation: "gelu" # non-linear activation function used in the MLP
+    use_layer_norm: true # apply layer normalization in hidden layers
+    output_dim: 1 # one output logit per peak
+
+training:
+  seed: 42 # random seed for reproducible training behavior
+  batch_size: 256 # GPU-oriented default; reduce only if you hit CUDA OOM
+  num_workers: 16 # enough workers to keep the GPU fed without excessive loader overhead
+  max_epochs: 50 # maximum number of training epochs
+  cpu_num_threads: 8 # leave most CPU capacity available for dataloader workers and system overhead
+  cpu_num_interop_threads: 2 # low interop setting usually works better for single-GPU training
+  dataloader_prefetch_factor: 4 # let each worker prepare several batches ahead
+  dataloader_persistent_workers: true # keep workers alive across epochs
+  dataloader_pin_memory: true # force pinned host memory for faster CUDA transfers
+  enable_amp: true # use mixed precision on CUDA
+  compile_model: true # use torch.compile when supported by the local PyTorch build
+  cache_dataset_in_memory: true # preprocess arrays once instead of rebuilding them per sample
+
+  optimizer:
+    name: "adamw" # optimizer used for parameter updates
+    learning_rate: 0.001 # optimizer learning rate
+    weight_decay: 0.0001 # L2-style regularization strength
+
+  scheduler:
+    enabled: false # no learning-rate scheduler is used in the first baseline
+
+  gradient_clip_norm: 1.0 # clip gradients to avoid unstable large updates
+
+loss:
+  name: "bce_with_logits" # binary cross-entropy loss applied to raw logits
+  use_pos_weight: true # compensate for class imbalance using positive-class weighting
+  pos_weight: null # if null, estimate the positive-class weight from the training split
+  reduction: "none" # keep unreduced loss values so masking and weighting stay flexible
+
+early_stopping:
+  enabled: true # stop training early when validation performance stops improving
+  monitor: "val_pr_auc" # validation metric used to decide early stopping
+  mode: "max" # higher values of the monitored metric are better
+  patience: 7 # number of bad epochs allowed before stopping
+  min_delta: 0.0001 # minimum improvement required to count as progress
+
+evaluation:
+  primary_metric: "pr_auc" # main metric used to compare models
+  threshold_for_binary_metrics: 0.5 # probability threshold for binary predictions
+
+  report_metrics:
+    - "pr_auc" # precision-recall area under the curve
+    - "roc_auc" # ROC area under the curve
+    - "f1" # harmonic mean of precision and recall
+    - "mcc" # Matthews correlation coefficient for imbalanced binary classification
+    - "precision" # fraction of predicted positives that are correct
+    - "recall" # fraction of true positives that are recovered
+
+  retained_peak_fractions:
+    - 0.3 # evaluate how well the model keeps signal when retaining the top 30% of peaks
+    - 0.5 # evaluate how well the model keeps signal when retaining the top 50% of peaks
+    - 0.7 # evaluate how well the model keeps signal when retaining the top 70% of peaks

configs/train_transformer.yml

@@ -0,0 +1,129 @@
+data:
+  train_path: "../data/splits_1pct_filtered/train.parquet"
+  val_path: "../data/splits_1pct_filtered/val.parquet"
+  test_path: "../data/splits_1pct_filtered/test.parquet"
+  # train_path: "../data/splits_6pct_filtered/train.parquet"
+  # val_path: "../data/splits_6pct_filtered/val.parquet"
+  # test_path: "../data/splits_6pct_filtered/test.parquet"
+  #train_path: "../data/demo/train_ten_rows.parquet"
+  #val_path: "../data/demo/val_ten_rows.parquet"
+  #test_path: "../data/demo/test_ten_rows.parquet"
+
+  target_column: "annotation_mask" # Column containing the per-peak binary targets.
+  #loss_matrix = loss_matrix * weights otherwise: 
+  #FDR weights are already constrained to a safe range in src/features/fdr_weights.py
+  #the caution about “very large weights” does not really apply to the used-setup, we can keep it! ;)
+  use_training_weights: true # Enables per-example weighting in the loss if available.
+  weight_column: "weight" # Column name to read sample weights from when weighting is enabled.
+
+output:
+  output_dir: "../outputs/transformer_peak_classifier_1pct_filtered" # Directory where all training artifacts are written.
+  #output_dir: "../outputs/transformer_peak_classifier_6pct_filtered" 
+  save_best_model: true # Writes the best validation checkpoint to disk.
+  save_epoch_history: true # Saves per-epoch metrics to a CSV file.
+  save_metrics_summary: true # Saves the final validation and test metrics summary as JSON.
+  save_plots: true # Generates training curve plots after training finishes.
+  save_confusion_matrices: true # Saves confusion-matrix artifacts for validation and test sets.
+  device_verbose: true # Prints detailed device and accelerator information at startup.
+  enable_file_logging: true # Mirrors console logs into a file in the output directory.
+  log_file_name: "training.log" # File name used for the persisted training log.
+
+features:
+  peak_feature_columns:
+    - "peak_feature_mz" # Peak m/z-based engineered feature.
+    - "peak_feature_log_intensity" # Log-scaled peak intensity feature.
+    - "peak_feature_relative_intensity" # Intensity normalized relative to the spectrum.
+    - "peak_feature_mz_over_precursor" # Ratio between peak m/z and precursor m/z.
+    - "peak_feature_delta_to_precursor" # Difference between peak m/z and precursor m/z.
+    - "peak_feature_delta_prev" # Distance from the previous peak in the ordered spectrum.
+    - "peak_feature_delta_next" # Distance to the next peak in the ordered spectrum.
+
+  spectrum_feature_columns:
+    - "Charge" # Precursor charge state feature.
+    - "exp m/z" # Experimental precursor m/z feature.
+    - "num_peaks" # Total number of peaks in the spectrum.
+    - "tic" # Total ion current summary feature.
+
+  use_raw_peak_mz: true # Appends the raw m/z sequence as an additional peak input column.
+  raw_peak_mz_column: "mz_arr" # Source column holding the raw m/z array.
+  use_raw_peak_intensity: true # Appends the raw intensity sequence as an additional peak input column.
+  raw_peak_intensity_column: "int_arr" # Source column holding the raw intensity array.
+  sort_raw_peak_inputs_by_mz: true # Sorts raw peak arrays by m/z before building tensors.
+  broadcast_spectrum_features_to_peaks: false # Kept for config compatibility but unused by the transformer path.
+  normalize_peak_features: true # Standardizes peak-level input features using train-set statistics.
+  normalize_spectrum_features: true # Standardizes spectrum-level input features using train-set statistics.
+  normalization_fit_split: "train" # Split used to fit feature normalization statistics.
+
+model:
+  name: "transformer_peak_classifier" # Identifier for this model family.
+
+  transformer:
+    d_model: 96 # Hidden width used for transformer token embeddings.
+    num_heads: 4 # Number of attention heads in each encoder block.
+    num_layers: 3 # Number of stacked transformer encoder blocks.
+    ff_multiplier: 4.0 # Expansion factor for the feed-forward hidden layer.
+    #It prevents the model from relying too much on a few specific peak-to-peak relationships.
+    dropout: 0.1 # Dropout probability applied inside the transformer.
+    activation: "gelu" # Nonlinearity used in the feed-forward network and head.
+    use_layer_norm: true # Enables layer normalization inside the encoder blocks.
+    output_dim: 1 # Number of logits produced per peak.
+    use_spectrum_context_gating: true # Gates peak tokens with spectrum-level context.
+    use_peak_positional_projection: true # Projects the first two peak features as a lightweight positional cue.
+
+training:
+  seed: 42 # Global random seed for reproducible training behavior.
+  batch_size: 8 # Number of spectra per batch before peak padding is applied.
+  num_workers: 10 # Number of dataloader worker processes.
+  max_epochs: 50 # Maximum number of training epochs to run.
+  cpu_num_threads: 8 # CPU compute threads used by PyTorch kernels where applicable.
+  cpu_num_interop_threads: 2 # CPU inter-op threads used to schedule parallel work.
+  dataloader_prefetch_factor: 4 # Number of batches each worker preloads ahead of time.
+  dataloader_persistent_workers: true # Keeps dataloader workers alive across epochs.
+  dataloader_pin_memory: true # Pins host memory to speed CPU-to-GPU transfers.
+  enable_amp: true # Enables automatic mixed precision on CUDA.
+  #Mixed precision tries to match each op to its appropriate datatype. see https://docs.pytorch.org/docs/stable/amp.html
+  compile_model: false # Disables torch.compile to reduce runtime memory risk.
+  cache_dataset_in_memory: true # Precomputes dataset samples and keeps them in RAM.
+  save_latest_checkpoint: true # Saves a resumable checkpoint after each epoch.
+  resume_from_checkpoint: null # Set to "auto" or a checkpoint path to continue after interruption.
+
+  optimizer:
+    name: "adamw" # Optimizer used to update model parameters.
+    learning_rate: 0.0003 # Base learning rate passed to the optimizer.
+    weight_decay: 0.0001 # L2-style parameter decay applied by the optimizer.
+
+  scheduler:
+    # sofar not important, can be added afterwards
+    enabled: false # Placeholder switch for learning-rate scheduling support.
+
+  gradient_clip_norm: 1.0 # Max gradient norm used for clipping before optimizer step.
+
+loss: #pos_weight is important, because we have class-imbalanced data! 
+  name: "bce_with_logits" # Per-peak binary classification loss function.
+  use_pos_weight: true # Uses a positive-class weight estimated from the training split.
+  pos_weight: null # Manual positive-class weight override; null means auto-compute.
+  reduction: "none" # Keeps unreduced loss values so padding can be masked out manually.
+
+early_stopping:
+  enabled: true # Stops training early when the monitored metric stops improving.
+  monitor: "val_pr_auc" # Validation metric used to select the best epoch.
+  mode: "max" # Improvement direction for the monitored metric.
+  patience: 7 # Number of unimproved epochs tolerated before stopping.
+  min_delta: 0.0001 # Minimum metric change required to count as an improvement.
+
+evaluation:
+  primary_metric: "pr_auc" # Main evaluation metric emphasized in reports.
+  threshold_for_binary_metrics: 0.3 # Probability cutoff used for discrete binary metrics.
+
+  report_metrics:
+    - "pr_auc" # Precision-recall area under the curve.
+    - "roc_auc" # Receiver operating characteristic area under the curve.
+    - "f1" # Harmonic mean of precision and recall.
+    - "mcc" # Matthews correlation coefficient.
+    - "precision" # Positive predictive value at the chosen threshold.
+    - "recall" # True positive rate at the chosen threshold.
+
+  retained_peak_fractions:
+    - 0.3 # Fraction used for retained-peak style analysis.
+    - 0.5 # Fraction used for retained-peak style analysis.
+    - 0.7 # Fraction used for retained-peak style analysis.