ABX script for CPC

cpc/model.py

@@ -304,7 +304,6 @@ def __init__(self,
     def forward(self, batchData, label):
         cFeature, encodedData, label = self.cpc(batchData, label)
         cFeature = self.nullspace(cFeature)
-        encodedData = self.nullspace(encodedData)
         return cFeature, encodedData, label
 
 

finetune_nullspace.sh

@@ -1,30 +1,77 @@
-SAVE_DIR="/pio/scratch/1/i273233/linear_separability/cpc/gru_level2/cpc_official"
 SPEAKERS="speakers_factorized"
 PHONEMES="phonemes_nullspace"
 SPEAKERS_NULLSPACE="speakers_nullspace"
 
-DIM_INTER=$1
+DATASET_PATH=false
+TRAIN_SPLIT_FILE_PATH=false
+VALIDATION_SPLIT_FILE_PATH=false
+BASELINE_NO_CLUSTERING_CHECKPOINT_PATH=false
+SAVE_DIR=false
+DIM_INBETWEEN=false
 FROM_STEP=$SPEAKERS
-if [[ $# -ge 2 ]]; then
-    FROM_STEP=$2
+PHONEME_ALIGNMENTS_FILE=false
+
+print_usage() {
+  echo -e "Usage: ./finetune_nullspace.sh"
+  echo -e "\t-d DATASET_PATH (E.g. LIBRISPEECH_DATASET_PATH/train-clean-100)"
+  echo -e "\t-t TRAIN_SPLIT_FILE_PATH (E.g. LIBRISPEECH_TRAIN_CLEAN_100_TRAIN_SPLIT_FILE_PATH)"
+  echo -e "\t-v VALIDATION_SPLIT_FILE_PATH (E.g. LIBRISPEECH_TRAIN_CLEAN_100_TEST_SPLIT_FILE_PATH)"
+  echo -e "\t-c BASELINE_NO_CLUSTERING_CHECKPOINT_PATH"
+  echo -e "\t-o SAVE_DIR"
+  echo -e "\t-n DIM_INBETWEEN (Dimension of nullspace will be DIM_EMBEDDING - DIM_INBETWEEN)"
+  echo -e "\t-p PHONEME_ALIGNMENTS_FILE (Path to the file containing phonemes for the entire dataset)"
+  echo -e "OPTIONAL ARGS:"
+  echo -e "\t-f FROM_STEP (From which step do you want to start. Order: $SPEAKERS [default] -> $PHONEMES -> $SPEAKERS_NULLSPACE)"
+}
+
+while getopts 'd:t:v:c:o:n:f:p:' flag; do
+    case "${flag}" in
+        d) DATASET_PATH="${OPTARG}" ;;
+        t) TRAIN_SPLIT_FILE_PATH="${OPTARG}" ;;
+        v) VALIDATION_SPLIT_FILE_PATH="${OPTARG}" ;;
+        c) BASELINE_NO_CLUSTERING_CHECKPOINT_PATH="${OPTARG}" ;;
+        o) SAVE_DIR="${OPTARG}" ;;
+        n) DIM_INBETWEEN="${OPTARG}" ;;
+        f) FROM_STEP="${OPTARG}" ;;
+        p) PHONEME_ALIGNMENTS_FILE="${OPTARG}" ;;
+        *) print_usage
+           exit 1 ;;
+    esac
+done
+
+echo $DATASET_PATH $TRAIN_SPLIT_FILE_PATH $VALIDATION_SPLIT_FILE_PATH $BASELINE_NO_CLUSTERING_CHECKPOINT_PATH $SAVE_DIR $DIM_INBETWEEN $FROM_STEP $PHONEME_ALIGNMENTS_FILE
+
+if [[ $DATASET_PATH == false || $TRAIN_SPLIT_FILE_PATH == false || $VALIDATION_SPLIT_FILE_PATH == false || $BASELINE_NO_CLUSTERING_CHECKPOINT_PATH == false || $SAVE_DIR == false || $DIM_INBETWEEN == false || $PHONEME_ALIGNMENTS_FILE == false ]]
+then
+    echo "Either DATASET_PATH, TRAIN_SPLIT_FILE_PATH, VALIDATION_SPLIT_FILE_PATH, BASELINE_NO_CLUSTERING_CHECKPOINT_PATH, SAVE_DIR, DIM_INBETWEEN or PHONEME_ALIGNMENTS_FILE is not set."
+    print_usage
+    exit 1
 fi
 
+mkdir -p $SAVE_DIR
+
 case $FROM_STEP in
 $SPEAKERS)
     echo $SPEAKERS
-    mkdir -p ${SAVE_DIR}_${SPEAKERS}_${DIM_INTER} && python cpc/eval/linear_separability.py $zd/LibriSpeech/train-clean-100/ $zd/LibriSpeech/labels_split/train_split_100.txt $zd/LibriSpeech/labels_split/test_split_100.txt $zd/checkpoints/CPC-big-kmeans50/cpc_ll6k/checkpoint_32.pt --pathCheckpoint ${SAVE_DIR}_${SPEAKERS}_${DIM_INTER} --mode $SPEAKERS --max_size_loaded 40000000 --n_process_loader 2 --model cpc --dim_inter $DIM_INTER --gru_level 2 | tee ${SAVE_DIR}_${SPEAKERS}_${DIM_INTER}/log.txt
+    mkdir -p ${SAVE_DIR}/${SPEAKERS}_${DIM_INBETWEEN}
+    python cpc/eval/linear_separability.py $DATASET_PATH $TRAIN_SPLIT_FILE_PATH $VALIDATION_SPLIT_FILE_PATH $BASELINE_NO_CLUSTERING_CHECKPOINT_PATH --pathCheckpoint ${SAVE_DIR}/${SPEAKERS}_${DIM_INBETWEEN} --mode $SPEAKERS --max_size_loaded 40000000 --n_process_loader 2 --model cpc --dim_inter $DIM_INBETWEEN --gru_level 2
     ;&
 $PHONEMES)
     echo $PHONEMES
-    mkdir -p ${SAVE_DIR}_${PHONEMES}_${DIM_INTER} && python cpc/eval/linear_separability.py $zd/LibriSpeech/train-clean-100/ $zd/LibriSpeech/labels_split/train_split_100.txt $zd/LibriSpeech/labels_split/test_split_100.txt $zd/checkpoints/CPC-big-kmeans50/cpc_ll6k/checkpoint_32.pt --pathCheckpoint ${SAVE_DIR}_${PHONEMES}_${DIM_INTER} --mode $PHONEMES --max_size_loaded 40000000 --n_process_loader 2 --model cpc --pathPhone $zd/LibriSpeech/alignments2/converted_aligned_phones.txt --path_speakers_factorized ${SAVE_DIR}_${SPEAKERS}_${DIM_INTER}/checkpoint_9.pt --dim_inter $DIM_INTER --gru_level 2 | tee ${SAVE_DIR}_${PHONEMES}_${DIM_INTER}/log.txt
+    mkdir -p ${SAVE_DIR}/${PHONEMES}_${DIM_INBETWEEN}
+    python cpc/eval/linear_separability.py $DATASET_PATH $TRAIN_SPLIT_FILE_PATH $VALIDATION_SPLIT_FILE_PATH $BASELINE_NO_CLUSTERING_CHECKPOINT_PATH --pathCheckpoint ${SAVE_DIR}/${PHONEMES}_${DIM_INBETWEEN} --mode $PHONEMES --max_size_loaded 40000000 --n_process_loader 2 --model cpc --pathPhone $PHONEME_ALIGNMENTS_FILE --path_speakers_factorized ${SAVE_DIR}/${SPEAKERS}_${DIM_INBETWEEN}/checkpoint_9.pt --dim_inter $DIM_INBETWEEN --gru_level 2
     ;&
 $SPEAKERS_NULLSPACE)
     echo $SPEAKERS_NULLSPACE
-    mkdir -p ${SAVE_DIR}_${SPEAKERS_NULLSPACE}_${DIM_INTER} && python cpc/eval/linear_separability.py $zd/LibriSpeech/train-clean-100/ $zd/LibriSpeech/labels_split/train_split_100.txt $zd/LibriSpeech/labels_split/test_split_100.txt $zd/checkpoints/CPC-big-kmeans50/cpc_ll6k/checkpoint_32.pt --pathCheckpoint ${SAVE_DIR}_${SPEAKERS_NULLSPACE}_${DIM_INTER} --mode $SPEAKERS_NULLSPACE --max_size_loaded 40000000 --n_process_loader 2 --model cpc --path_speakers_factorized ${SAVE_DIR}_${SPEAKERS}_${DIM_INTER}/checkpoint_9.pt --dim_inter $DIM_INTER --gru_level 2 | tee ${SAVE_DIR}_${SPEAKERS_NULLSPACE}_${DIM_INTER}/log.txt
+    mkdir -p ${SAVE_DIR}/${SPEAKERS_NULLSPACE}_${DIM_INBETWEEN}
+    python cpc/eval/linear_separability.py $DATASET_PATH $TRAIN_SPLIT_FILE_PATH $VALIDATION_SPLIT_FILE_PATH $BASELINE_NO_CLUSTERING_CHECKPOINT_PATH --pathCheckpoint ${SAVE_DIR}/${SPEAKERS_NULLSPACE}_${DIM_INBETWEEN} --mode $SPEAKERS_NULLSPACE --max_size_loaded 40000000 --n_process_loader 2 --model cpc --path_speakers_factorized ${SAVE_DIR}/${SPEAKERS}_${DIM_INBETWEEN}/checkpoint_9.pt --dim_inter $DIM_INBETWEEN --gru_level 2
     ;;
 *)
     echo "Invalid from step: ${FROM_STEP} while it should be either ${SPEAKERS}, ${PHONEMES} or ${SPEAKERS_NULLSPACE}"
     ;;
 esac
 
+echo "Checkpoint with nullspace is located in ${SAVE_DIR}/${PHONEMES}_${DIM_INBETWEEN}/checkpoint_9.pt"
+echo "The results of all the experiments are located in ${SAVE_DIR}/DIRECTORY/checkpoint_logs.json"
+
 exit 0

scripts/create_ls_dataset_for_abx_eval.py

@@ -0,0 +1,56 @@
+import os
+import sys
+import shutil
+import argparse
+from pathlib import Path
+import numpy as np
+import soundfile as sf
+
+def parse_args():
+    # Run parameters
+    parser = argparse.ArgumentParser()
+    parser.add_argument("librispeech_path", type=str,
+                        help="Path to the root directory of LibriSpeech.")
+    parser.add_argument("zerospeech_dataset_path", type=str,
+                        help="Path to the ZeroSpeech dataset.")
+    parser.add_argument("target_path", type=str,
+                        help="Path to the output directory.")
+    parser.add_argument("--file_extension", type=str, default="flac",
+                          help="Extension of the audio files in the dataset (default: flac).")
+    return parser.parse_args()
+
+def main():
+    # Parse and print args
+    args = parse_args()
+    logger.info(args)
+
+    phonetic = "phonetic"
+    datasets = ["dev-clean", "dev-other", "test-clean", "test-other"]
+
+    for dataset in datasets:
+        print("> {}".format(dataset))
+        target_dirname = os.path.join(args.target_path, phonetic, dataset)
+        Path(target_dirname).mkdir(parents=True, exist_ok=True)
+
+        librispeech_dirname = os.path.join(args.librispeech_path, dataset)
+        files = [(filename, dirname) for dirname, _, files in os.walk(librispeech_dirname, followlinks=True) for filename in files if filename.endswith(args.file_extension)]
+        for i, (filename, dirname) in enumerate(files):
+            print("Progress {:2.1%}".format(i / len(files)), end="\r")
+            input_path = os.path.join(dirname, filename)
+            output_path = os.path.join(target_dirname, os.path.splitext(filename)[0] + ".wav")
+            data, sample_rate = sf.read(input_path)
+            sf.write(output_path, data, sample_rate)
+
+        if dataset.startswith("dev"):
+            source_item_path = os.path.join(args.zerospeech_dataset_path, phonetic, dataset, dataset + ".item")
+            target_item_path = os.path.join(target_dirname, dataset + ".item")
+            shutil.copy(source_item_path, target_item_path)
+
+
+if __name__ == "__main__":
+    #import ptvsd
+    #ptvsd.enable_attach(('0.0.0.0', 7310))
+    #print("Attach debugger now")
+    #ptvsd.wait_for_attach()
+    main()
+

scripts/embeddings_abx.py

@@ -0,0 +1,133 @@
+#!/usr/bin/env python3 -u
+
+import logging
+import os
+import sys
+import argparse
+from itertools import chain
+from pathlib import Path
+import time
+import copy
+import numpy as np
+import soundfile as sf
+
+from cpc.feature_loader import loadModel, FeatureModule
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+logging.basicConfig(
+    format="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
+    datefmt="%Y-%m-%d %H:%M:%S",
+    level=os.environ.get("LOGLEVEL", "INFO").upper(),
+    stream=sys.stdout,
+)
+logger = logging.getLogger("zerospeech2021 abx")
+
+def parse_args():
+    # Run parameters
+    parser = argparse.ArgumentParser()
+    parser.add_argument("path_checkpoint", type=str,
+                        help="Path to the trained fairseq wav2vec2.0 model.")
+    parser.add_argument("path_data", type=str,
+                        help="Path to the dataset that we want to compute ABX for.")
+    parser.add_argument("path_output_dir", type=str,
+                        help="Path to the output directory.")
+    parser.add_argument("--debug", action="store_true",
+                        help="Load only a very small amount of files for "
+                        "debugging purposes.")
+    parser.add_argument("--cpu", action="store_true",
+                        help="Run on a cpu machine.")
+    parser.add_argument("--file_extension", type=str, default="wav",
+                          help="Extension of the audio files in the dataset (default: wav).")
+    parser.add_argument("--no_test", action="store_true",
+                        help="Don't compute embeddings for test-* parts of dataset")
+    parser.add_argument('--gru_level', type=int, default=-1,
+                        help='Hidden level of the LSTM autoregressive model to be taken'
+                        '(default: -1, last layer).')
+    parser.add_argument('--nullspace', action='store_true',
+                        help="Additionally load nullspace")
+    return parser.parse_args()
+
+def main():
+    # Parse and print args
+    args = parse_args()
+    logger.info(args)
+
+    # Load the model
+    print("")
+    print(f"Loading model from {args.path_checkpoint}") 
+
+    if args.gru_level is not None and args.gru_level > 0:
+        updateConfig = argparse.Namespace(nLevelsGRU=args.gru_level)
+    else:
+        updateConfig = None
+
+    model = loadModel([args.path_checkpoint], load_nullspace=args.nullspace, updateConfig=updateConfig)[0]
+
+    if args.gru_level is not None and args.gru_level > 0:
+        # Keep hidden units at LSTM layers on sequential batches
+        if args.nullspace:
+            model.cpc.gAR.keepHidden = True
+        else:
+            model.gAR.keepHidden = True
+
+    device = "cuda" if torch.cuda.is_available() and not args.cpu else "cpu"
+
+    # Register the hooks
+    layer_outputs = {}
+    def get_layer_output(name):
+        def hook(model, input, output):
+            if type(output) is tuple:
+                layer_outputs[name] = output[0].detach().squeeze(1).cpu().numpy()
+            elif type(output) is dict:
+                layer_outputs[name] = output["x"].detach().squeeze(0).cpu().numpy()
+            else:
+                layer_outputs[name] = output.detach().squeeze(0).cpu().numpy()
+        return hook
+
+    layer_names = []
+    layer_name = os.path.basename(os.path.dirname(args.path_checkpoint))
+    layer_names.append(layer_name)
+    if not args.nullspace:
+        model.gAR.register_forward_hook(get_layer_output(layer_name))
+    else:
+        model.nullspace.register_forward_hook(get_layer_output(layer_name))
+
+    model = model.eval().to(device)  
+    print("Model loaded!")
+    print(model)
+
+    # Extract values from chosen layers and save them to files
+    phonetic = "phonetic"
+    datasets_path = os.path.join(args.path_data, phonetic)
+    datasets = os.listdir(datasets_path)
+    datasets = [dataset for dataset in datasets if not args.no_test or not dataset.startswith("test")]
+    print(datasets)
+
+    with torch.no_grad():     
+        for dataset in datasets:
+            print("> {}".format(dataset))
+            dataset_path = os.path.join(datasets_path, dataset)
+            files = [f for f in os.listdir(dataset_path) if f.endswith(args.file_extension)]
+            for i, f in enumerate(files):
+                print("Progress {:2.1%}".format(i / len(files)), end="\r")
+                input_f = os.path.join(dataset_path, f)
+                x, sample_rate = sf.read(input_f)
+                x = torch.tensor(x).float().reshape(1,1,-1).to(device)
+                output = model(x, None)[0]
+
+                for layer_name, value in layer_outputs.items():
+                    output_dir = os.path.join(args.path_output_dir, layer_name, phonetic, dataset)
+                    Path(output_dir).mkdir(parents=True, exist_ok=True)
+                    out_f = os.path.join(output_dir, os.path.splitext(f)[0] + ".txt")
+                    np.savetxt(out_f, value)
+
+if __name__ == "__main__":
+    #import ptvsd
+    #ptvsd.enable_attach(('0.0.0.0', 7310))
+    #print("Attach debugger now")
+    #ptvsd.wait_for_attach()
+    main()
+