Add zipformer recipe for audio tagging

.pre-commit-config.yaml

@@ -26,7 +26,7 @@ repos:
       # E121,E123,E126,E226,E24,E704,W503,W504
 
   - repo: https://github.com/pycqa/isort
-    rev: 5.10.1
+    rev: 5.12.0
     hooks:
       - id: isort
         args: ["--profile=black"]

egs/audioset/AT/README.md

@@ -0,0 +1,12 @@
+# Introduction
+
+This is an audio tagging recipe. It aims at predicting the sound events of an audio clip.
+
+[./RESULTS.md](./RESULTS.md) contains the latest results.
+
+
+# Zipformer
+
+| Encoder | Feature type |
+| --------| -------------|
+| Zipformer | Frame level fbank|

egs/audioset/AT/RESULTS.md

@@ -0,0 +1,44 @@
+## Results
+
+### zipformer
+See <https://github.com/k2-fsa/icefall/pull/1421> for more details
+
+[zipformer](./zipformer)
+
+You can find a pretrained model, training logs, decoding logs, and decoding results at:
+<https://huggingface.co/marcoyang/icefall-audio-tagging-audioset-zipformer-2024-03-12#/>
+
+The model achieves the following mean averaged precision on AudioSet:
+
+| Model | mAP |
+| ------ | ------- |
+| Zipformer-AT | 45.1 |
+
+The training command is:
+
+```bash
+export CUDA_VISIBLE_DEVICES="4,5,6,7"
+subset=full
+
+python zipformer/train.py \
+    --world-size 4 \
+    --num-epochs 50 \
+    --exp-dir zipformer/exp_at_as_${subset} \
+    --start-epoch 1 \
+    --use-fp16 1 \
+    --num-events 527 \
+    --audioset-subset $subset \
+    --max-duration 1000 \
+    --enable-musan True \
+    --master-port 13455
+```
+
+The evaluation command is:
+
+```bash
+python zipformer/evaluate.py \
+    --epoch 32 \
+    --avg 8 \
+    --exp-dir zipformer/exp_at_as_full \
+    --max-duration 500
+```

egs/audioset/AT/local/compute_fbank_musan.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py

egs/audioset/AT/local/generate_audioset_manifest.py

@@ -0,0 +1,177 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file generates the manifest and computes the fbank features for AudioSet
+dataset. The generated manifests and features are stored in data/fbank.
+"""
+
+import argparse
+import csv
+import glob
+import logging
+import os
+from typing import Dict
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.audio import Recording
+from lhotse.cut import MonoCut
+from lhotse.supervision import SupervisionSegment
+
+from icefall.utils import get_executor
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_ID_mapping(csv_file):
+    # get a mapping between class ID and class name
+    mapping = {}
+    with open(csv_file, "r") as fin:
+        reader = csv.reader(fin, delimiter=",")
+        for i, row in enumerate(reader):
+            if i == 0:
+                continue
+            mapping[row[1]] = row[0]
+    return mapping
+
+
+def parse_csv(csv_file: str, id_mapping: Dict):
+    # The content of the csv file shoud be something like this
+    # ------------------------------------------------------
+    # filename  label
+    # dataset/AudioSet/balanced/xxxx.wav 0;451
+    # dataset/AudioSet/balanced/xxxy.wav 375
+    # ------------------------------------------------------
+
+    def name2id(names):
+        ids = [id_mapping[name] for name in names.split(",")]
+        return ";".join(ids)
+
+    mapping = {}
+    with open(csv_file, "r") as fin:
+        reader = csv.reader(fin, delimiter=" ")
+        for i, row in enumerate(reader):
+            if i <= 2:
+                continue
+            key = row[0].replace(",", "")
+            mapping[key] = name2id(row[-1])
+    return mapping
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument("--dataset-dir", type=str, default="downloads/audioset")
+
+    parser.add_argument(
+        "--split",
+        type=str,
+        default="balanced",
+        choices=["balanced", "unbalanced", "eval"],
+    )
+
+    parser.add_argument(
+        "--feat-output-dir",
+        type=str,
+        default="data/fbank",
+    )
+
+    return parser
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    dataset_dir = args.dataset_dir
+    split = args.split
+    feat_output_dir = args.feat_output_dir
+
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+
+    if split in ["balanced", "unbalanced"]:
+        csv_file = f"{dataset_dir}/{split}_train_segments.csv"
+    elif split == "eval":
+        csv_file = f"{dataset_dir}/eval_segments.csv"
+    else:
+        raise ValueError()
+
+    class_indices_csv = f"{dataset_dir}/class_labels_indices.csv"
+    id_mapping = get_ID_mapping(class_indices_csv)
+    labels = parse_csv(csv_file, id_mapping)
+
+    audio_files = glob.glob(f"{dataset_dir}/{split}/*.wav")
+
+    new_cuts = []
+    for i, audio in enumerate(audio_files):
+        cut_id = audio.split("/")[-1].split("_")[0]
+        recording = Recording.from_file(audio, cut_id)
+        cut = MonoCut(
+            id=cut_id,
+            start=0.0,
+            duration=recording.duration,
+            channel=0,
+            recording=recording,
+        )
+        supervision = SupervisionSegment(
+            id=cut_id,
+            recording_id=cut.recording.id,
+            start=0.0,
+            channel=0,
+            duration=cut.duration,
+        )
+        try:
+            supervision.audio_event = labels[cut_id]
+        except KeyError:
+            logging.info(f"No labels found for {cut_id}.")
+            continue
+        cut.supervisions = [supervision]
+        new_cuts.append(cut)
+
+        if i % 100 == 0 and i:
+            logging.info(f"Processed {i} cuts until now.")
+
+    cuts = CutSet.from_cuts(new_cuts)
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    logging.info(f"Computing fbank features for {split}")
+    with get_executor() as ex:
+        cuts = cuts.compute_and_store_features(
+            extractor=extractor,
+            storage_path=f"{feat_output_dir}/{split}_feats",
+            num_jobs=num_jobs if ex is None else 80,
+            executor=ex,
+            storage_type=LilcomChunkyWriter,
+        )
+
+    manifest_output_dir = feat_output_dir + "/" + f"cuts_audioset_{split}.jsonl.gz"
+
+    logging.info(f"Storing the manifest to {manifest_output_dir}")
+    cuts.to_jsonl(manifest_output_dir)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()

egs/audioset/AT/prepare.sh

@@ -0,0 +1,104 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+# run step 0 to step 5 by default
+stage=-1
+stop_stage=4
+
+dl_dir=$PWD/download
+
+# we assume that you have your downloaded the AudioSet and placed
+# it under $dl_dir/audioset, the folder structure should look like
+# this:
+# - $dl_dir/audioset
+#       - balanced
+#       - eval
+#       - unbalanced
+# If you haven't downloaded the AudioSet, please refer to
+# https://github.com/RicherMans/SAT/blob/main/datasets/audioset/1_download_audioset.sh.
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "Running prepare.sh"
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage 0: Download the necessary csv files"
+  if [ ! -e $dl_dir/audioset/.csv.done]; then
+    wget --continue "http://storage.googleapis.com/us_audioset/youtube_corpus/v1/csv/class_labels_indices.csv" -O "${dl_dir}/audioset/class_labels_indices.csv"
+    wget --continue http://storage.googleapis.com/us_audioset/youtube_corpus/v1/csv/balanced_train_segments.csv -O "${dl_dir}/audioset/balanced_train_segments.csv"
+    wget --continue http://storage.googleapis.com/us_audioset/youtube_corpus/v1/csv/eval_segments.csv -O "${dl_dir}/audioset/eval_segments.csv"
+    touch $dl_dir/audioset/.csv.done
+  fi
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Construct the audioset manifest and compute the fbank features for balanced set"
+  fbank_dir=data/fbank
+  if [! -e $fbank_dir/.balanced.done]; then
+    python local/generate_audioset_manifest.py \
+      --dataset-dir $dl_dir/audioset \
+      --split balanced \
+      --feat-output-dir $fbank_dir
+    touch $fbank_dir/.balanced.done
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Construct the audioset manifest and compute the fbank features for unbalanced set"
+  fbank_dir=data/fbank
+  if [! -e $fbank_dir/.unbalanced.done]; then
+    python local/generate_audioset_manifest.py \
+      --dataset-dir $dl_dir/audioset \
+      --split unbalanced \
+      --feat-output-dir $fbank_dir
+    touch $fbank_dir/.unbalanced.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Construct the audioset manifest and compute the fbank features for eval set"
+  fbank_dir=data/fbank
+  if [! -e $fbank_dir/.eval.done]; then
+    python local/generate_audioset_manifest.py \
+      --dataset-dir $dl_dir/audioset \
+      --split eval \
+      --feat-output-dir $fbank_dir
+    touch $fbank_dir/.eval.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to $dl_dir/musan
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.musan.done ]; then
+    lhotse prepare musan $dl_dir/musan data/manifests
+    touch data/manifests/.musan.done
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  mkdir -p data/fbank
+  if [ ! -e data/fbank/.musan.done ]; then
+    ./local/compute_fbank_musan.py
+    touch data/fbank/.musan.done
+  fi
+fi

egs/audioset/AT/shared

@@ -0,0 +1 @@
+../../../icefall/shared