Skip to content

notes.md

Latest commit

 

History

History
142 lines (111 loc) · 5.73 KB

notes.md

File metadata and controls

142 lines (111 loc) · 5.73 KB

Task list

rms, pitch, pos, rms_reg, word_onset, sentence_onset, pretraining, spec_target_pretraining

$REPO_DIR = "C:\Users\badri\OneDrive\Documents\EE 675 Neural Learning\Baseline Replication\PopulationTransformer" $BRAIN_REPO = "C:\Users\badri\OneDrive\Documents\EE 675 Neural Learning\Baseline Replication\Brain-Treebank" $BRAINTREEBANK_DIR = "$BRAIN_REPO\braintreebank_data"

$STFT_CKPT = "$BRAIN_REPO\pretrained_weights\brainbert_pretrained_weights\stft_large_pretrained.pth" $PRETRAINED_POPT = "$BRAIN_REPO\pretrained_weights\popt_pretrained_weights\pretrained_popt_brainbert_stft.pth"

$RESULTS_DIR = "$REPO_DIR\outputs\sub_1_word_onset_top1_popt_brainbert_stft"

$TASK = "word_onset" $TASK = "sentence_onset"

Restrict to sub_1, trials 000 and 002

$trialJson = @" { "sub_1": ["trial000", "trial002"] } "@

New-Item -ItemType Directory -Force -Path "$REPO_DIR\trial_selections" | Out-Null $trialJson | Out-File -Encoding utf8 "$REPO_DIR\trial_selections\speech_sub1_000_002.json"

Build the supervised dataset for word_onset

$CACHED_ALIGNS = "$BRAIN_REPO\semantics\saved_aligns" $CACHED_ARRAYS = "$BRAIN_REPO\cached_data_arrays" $TRANSCRIPTS_DIR = "$BRAINTREEBANK_DIR\transcripts"

write_multi_subject_multi_channel SCRIPT

'''' (dl_gpu_env) PS C:\Users\badri\OneDrive\Documents\EE 675 Neural Learning\Baseline Replication\PopulationTransformer> $REPO_DIR = "C:\Users\badri\OneDrive\Documents\EE 675 Neural Learning\Baseline Replication\PopulationTransformer" (dl_gpu_env) PS C:\Users\badri\OneDrive\Documents\EE 675 Neural Learning\Baseline Replication\PopulationTransformer> python -c "import json, pathlib; p = pathlib.Path(r'$REPO_DIR\trial_selections\speech_sub1_000_002.json'); p.write_text(json.dumps({'sub_1':['trial000','trial002']}), encoding='utf-8')" (dl_gpu_env) PS C:\Users\badri\OneDrive\Documents\EE 675 Neural Learning\Baseline Replication\PopulationTransformer> Get-Content "$REPO_DIR\trial_selections\speech_sub1_000_002.json" {"sub_1": ["trial000", "trial002"]}

'''

$TASK = "sentence_onset" $IDX = "idx05"

python -m data.write_multi_subject_multi_channel "+data_prep=pretrain_multi_subj_multi_chan_template" "++data_prep.task_name=$TASK" "++data_prep.brain_runs=$REPO_DIR/trial_selections/speech_sub1_000_002.json" "++data_prep.electrodes=$REPO_DIR/electrode_selections/clean_laplacian.json" "++data_prep.output_directory=$REPO_DIR/saved_examples/all_test_${TASK}_$IDX" "++data_prep.index_subsample=1.0" "++data_prep.separation_interval=0.5" "+preprocessor=multi_elec_spec_pretrained" "++preprocessor.upstream_ckpt=$STFT_CKPT" "+data=subject_data_template" "++data.cached_transcript_aligns=$CACHED_ALIGNS" "++data.cached_data_array=$CACHED_ARRAYS" "++data.raw_brain_data_dir=$BRAINTREEBANK_DIR" "++data.movie_transcripts_dir=$TRANSCRIPTS_DIR" "++data.rereference=None" "++data.duration=1.0" ` "++data.interval_duration=1.0"

$TASK = "word_onset" $IDX = "idx06" and "idx07,speech_sub1_2_word_onset"

python -m data.write_multi_subject_multi_channel "+data_prep=pretrain_multi_subj_multi_chan_template" "++data_prep.task_name=$TASK" "++data_prep.brain_runs=$REPO_DIR/trial_selections/speech_sub1_2_word_onset.json" "++data_prep.electrodes=$REPO_DIR/electrode_selections/clean_laplacian.json" "++data_prep.output_directory=$REPO_DIR/saved_examples/all_test_${TASK}_$IDX" "++data_prep.index_subsample=1.0" "++data_prep.separation_interval=0.5" "+preprocessor=multi_elec_spec_pretrained" "++preprocessor.upstream_ckpt=$STFT_CKPT" "+data=subject_data_template" "++data.cached_transcript_aligns=$CACHED_ALIGNS" "++data.cached_data_array=$CACHED_ARRAYS" "++data.raw_brain_data_dir=$BRAINTREEBANK_DIR" "++data.movie_transcripts_dir=$TRANSCRIPTS_DIR" "++data.rereference=None" "++data.duration=1.0" ` "++data.interval_duration=1.0"

solution

Turn off rereference ✅ (already did, rereference=None) Shrink the window duration (5s → 1s) Subsample the number of intervals (e.g., use ~10% of them

What this does: ++data.duration=1.0 → windows 1 second long instead of 5s ++data_prep.index_subsample=0.1 → use ~10% of intervals

Keeps the same word-onset logic (same labels, same pipeline). Uses shorter 1s windows → 5× less memory per window. Uses 10% of all intervals → 10× fewer windows. Combined memory reduction ≈ 50× vs the original 51 GB monster → into the ~1 GB range, which is realistic.

MANIFEST

SUB 1

python -m data.make_subject_specific_manifest "+data_prep=subject_specific_manifest" "++data_prep.data_path=$REPO_DIR/saved_examples/all_test_${TASK}$IDX" "++data_prep.subj=sub_1" "++data_prep.out_path=$REPO_DIR/saved_examples/sub_1${TASK}cr$IDX"

FINE TUNING

$TASK = "sentence_onset" $IDX = "idx05" $SUBJECT = "sub_1" $N = 1 $NAME = "popt_brainbert_stft_${TASK}$IDX" $PRETRAINED_POPT = "$BRAIN_REPO\pretrained_weights\popt_pretrained_weights\pretrained_popt_brainbert_stft.pth" $RESULTS_DIR = "$REPO_DIR\outputs${SUBJECT}${TASK}top${N}${NAME}"

python run_train.py "+exp=multi_elec_feature_extract" "++exp.runner.results_dir=$RESULTS_DIR" "++exp.runner.save_checkpoints=False" "++model.frozen_upstream=True" "+task=pt_feature_extract_coords" "+criterion=pt_feature_extract_coords_criterion" "+preprocessor=empty_preprocessor" "+data=pt_supervised_task_coords" "++data.data_path=$REPO_DIR/saved_examples/sub_1_${TASK}_cr_$IDX" "++data.saved_data_split=$REPO_DIR/saved_data_splits/sub_1_${TASK}fine_tuning$IDX" "++data.sub_sample_electrodes=$REPO_DIR/electrode_selections/clean_laplacian.json" "+model=pt_downstream_model" ` "++model.upstream_path=$PRETRAINED_POPT"

METIRCS

Get-Content "$RESULTS_DIR\results.json"