{ "cells": [ { "cell_type": "code", "execution_count": 19, "id": "c666bc86", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "compute-permanent-node-36\r\n" ] } ], "source": [ "!echo $HOSTNAME" ] }, { "cell_type": "code", "execution_count": 18, "id": "08218f19", "metadata": {}, "outputs": [], "source": [ "import os\n", "os.environ[\"CUDA_VISIBLE_DEVICES\"] = \"\"" ] }, { "cell_type": "code", "execution_count": 2, "id": "76fc0b2c", "metadata": {}, "outputs": [], "source": [ "# vocab: \n", "# 0-60_000 text\n", "# 1x0-3999 semantic\n", "# 8x0-4095 coarse\n", "\n", "# 4000 semantic pad token\n", "# 4001 semantic infer token\n", "# 2048 coarse pad token\n", "# 2049 coarse infer token\n", "\n", "# Memmaps:\n", "# Nx9x3584 for audio tokens\n", "# Jsons:\n", "# N*Dict with meta keys \n", "# \"dataset\"\n", "# \"original_id\", \"original_duration_s\",\n", "# \"start_s\", \"end_s\", \n", "# \"text_segments\", \"private_text_segments\",\n", "# \"text\", \"private_text\",\n", "# \"tags\", \"private_tags\",\n", "# \"views\",\n", "# Dict with meta keys {\"dataset\": [\"idx_list\"]}\n", "\n", "# Bundles (mert_v2_2x1k & dac_2c_25_8):\n", "# s3://suno-data/datasets/bundles/\n", "# v1/youtube_music\n", "# v1/genius_hq\n", "# v1/jamendo\n", "# v1/imslp\n", "# v2/pond5\n", "# v2/deezer\n", "# v2/ytm_tagged\n", "# v2/ytm_mb" ] }, { "cell_type": "code", "execution_count": 3, "id": "ce43e8ad", "metadata": {}, "outputs": [], "source": [ "%matplotlib inline\n", "from matplotlib import pyplot as plt" ] }, { "cell_type": "code", "execution_count": 4, "id": "5164dde2", "metadata": {}, "outputs": [], "source": [ "# TODO: add original_filepath on s3\n", "# TODO: exclude super short segment?" ] }, { "cell_type": "code", "execution_count": 5, "id": "0e33ff3d", "metadata": {}, "outputs": [], "source": [ "import math\n", "import numpy as np\n", "import tqdm\n", "import torch\n", "import funcy\n", "import json\n", "import gc\n", "import re\n", "import random\n", "import tempfile\n", "import collections\n", "from collections import defaultdict\n", "from joblib import Parallel, delayed\n", "from transformers import BertTokenizer\n", "\n", "from suno_utils.utils.text import write_jsonl, read_jsonl, write_json, read_json, normalize_whitespace\n", "from suno_utils.utils.s3 import read_from_s3, check_s3_file_exists\n", "\n", "TEXT_CODEBOOK_SIZE = 60_001\n", "TEXT_PAD_TOKEN = TEXT_CODEBOOK_SIZE\n", "TEXT_VOCAB_SIZE = 60_032\n", "\n", "SEMANTIC_CODEBOOK_SIZE = 4000\n", "SEMANTIC_N_CODEBOOKS = 1\n", "SEMANTIC_PAD_TOKEN = SEMANTIC_CODEBOOK_SIZE\n", "SEMANTIC_INFER_TOKEN = SEMANTIC_CODEBOOK_SIZE + 1\n", "SEMANTIC_VOCAB_SIZE = 4032\n", "SEMANTIC_RATE_HZ = 25\n", "SEMANTIC_SHIFT_FACTOR = 50\n", "assert(SEMANTIC_VOCAB_SIZE == (np.floor(SEMANTIC_CODEBOOK_SIZE // 64) + 1) * 64)\n", "\n", "COARSE_CODEBOOK_SIZE = 4096\n", "COARSE_N_CODEBOOKS = 12\n", "COARSE_PAD_TOKEN = COARSE_CODEBOOK_SIZE\n", "COARSE_INFER_TOKEN = COARSE_CODEBOOK_SIZE + 1\n", "COARSE_VOCAB_SIZE = 4160\n", "COARSE_RATE_HZ = 25\n", "COARSE_SHIFT_FACTOR = 5\n", "assert(COARSE_CODEBOOK_SIZE + 3 < COARSE_VOCAB_SIZE)\n", "assert(COARSE_VOCAB_SIZE % 64 == 0)\n", "\n", "assert(SEMANTIC_RATE_HZ == COARSE_RATE_HZ)\n", "\n", "BLOCK_SIZE = 4288\n", "N_TOKENS_TEXT = 1152\n", "N_TOKENS_AUDIO = 3008 # max 120s of audio\n", "# make sure we have enough space for shift 10\n", "assert(\n", " BLOCK_SIZE >= (\n", " N_TOKENS_TEXT + N_TOKENS_AUDIO + \n", " SEMANTIC_N_CODEBOOKS * SEMANTIC_SHIFT_FACTOR + \n", " (COARSE_N_CODEBOOKS - 1) * COARSE_SHIFT_FACTOR\n", " )\n", ")\n", "\n", "SEMANTIC_EMBED_DIR = \"mert_25_2x4k\"\n", "CODEC_EMBED_DIR = \"dac_2c_25_12\"\n", "\n", "METAS_DIR = \"/app/suno/data/chirp_v3/metadata/\"\n", "OUT_DATA_DIR = \"/mnt/localdisk/data\"\n", "os.makedirs(OUT_DATA_DIR, exist_ok=True)" ] }, { "cell_type": "code", "execution_count": 6, "id": "d188ba34", "metadata": {}, "outputs": [], "source": [ "# aws s3 cp s3://suno-data/georg/data/chirp_v2_5/filtered_metas/genius_hq.jsonl /app/suno/data/chirp_v3/metadata/\n", "# aws s3 cp s3://suno-data/georg/data/chirp_v2_5/filtered_metas/youtube_music.jsonl /app/suno/data/chirp_v3/metadata/\n", "# aws s3 cp s3://suno-data/georg/data/chirp_v2_5/filtered_metas/jamendo.jsonl /app/suno/data/chirp_v3/metadata/\n", "# aws s3 cp s3://suno-data/georg/data/chirp_v2_5/filtered_metas/imslp.jsonl /app/suno/data/chirp_v3/metadata/\n", "# aws s3 cp s3://suno-data/georg/data/chirp_v2_5/filtered_metas/pond5_music.jsonl /app/suno/data/chirp_v3/metadata/\n", "# aws s3 cp s3://suno-data/georg/data/chirp_v2_5/filtered_metas/deezer.jsonl /app/suno/data/chirp_v3/metadata/\n", "# aws s3 cp s3://suno-data/georg/data/chirp_v2_5/filtered_metas/ytm_tagged.jsonl /app/suno/data/chirp_v3/metadata/\n", "# aws s3 cp s3://suno-data/georg/data/chirp_v2_5/filtered_metas/musescore.jsonl /app/suno/data/chirp_v3/metadata/\n", "# aws s3 cp s3://suno-data/georg/data/chirp_v2_5/filtered_metas/ytm_mb.jsonl /app/suno/data/chirp_v3/metadata/" ] }, { "cell_type": "code", "execution_count": 7, "id": "b30bfb77", "metadata": {}, "outputs": [], "source": [ "# load manifests of IDs and text and tags etc\n", "meta_info_map = {\n", " \"genius_hq\": {m[\"id\"]: m for m in read_jsonl(os.path.join(METAS_DIR, \"genius_hq.jsonl\"))},\n", " \"youtube_music\": {m[\"id\"]: m for m in read_jsonl(os.path.join(METAS_DIR, \"youtube_music.jsonl\"))},\n", " \"jamendo\": {m[\"id\"]: m for m in read_jsonl(os.path.join(METAS_DIR, \"jamendo.jsonl\"))},\n", " \"imslp\": {m[\"id\"]: m for m in read_jsonl(os.path.join(METAS_DIR, \"imslp.jsonl\"))},\n", " \"pond5_music\": {m[\"id\"]: m for m in read_jsonl(os.path.join(METAS_DIR, \"pond5_music.jsonl\"))},\n", " \"deezer\": {m[\"id\"]: m for m in read_jsonl(os.path.join(METAS_DIR, \"deezer.jsonl\"))},\n", " \"ytm_tagged\": {m[\"id\"]: m for m in read_jsonl(os.path.join(METAS_DIR, \"ytm_tagged.jsonl\"))},\n", " \"musescore\": {m[\"id\"]: m for m in read_jsonl(os.path.join(METAS_DIR, \"musescore.jsonl\"))},\n", " \"ytm_mb\": {m[\"id\"]: m for m in read_jsonl(os.path.join(METAS_DIR, \"ytm_mb.jsonl\"))},\n", "}" ] }, { "cell_type": "code", "execution_count": 25, "id": "9c043b02", "metadata": {}, "outputs": [], "source": [ "def _verify_stuff(dset_name, meta_info):\n", " if dset_name == \"genius_hq\":\n", " assert(\n", " \"private_text_segments\" in meta_info or \n", " \"private_text\" in meta_info or\n", " \"text_segments\" in meta_info or \n", " \"text\" in meta_info\n", " )\n", "\n", "def _trim_to_common(arr_1, arr_2):\n", " common_len = min(len(arr_1), len(arr_2))\n", " arr_1 = arr_1[:common_len]\n", " arr_2 = arr_2[:common_len]\n", " return arr_1, arr_2\n", "\n", "def _parse_arrays(dset_name, meta_info, semantic_arr, coarse_arr):\n", " _verify_stuff(dset_name, meta_info)\n", " # prep segment metas (use semantic for timekeeping)\n", " segments_info = []\n", " # first check if we have known segments\n", " if \"text_segments\" in meta_info:\n", " for m in meta_info[\"text_segments\"]:\n", " segments_info.append((\n", " int(round(m[\"start_s\"]*SEMANTIC_RATE_HZ)), \n", " min(len(semantic_arr), int(round(m[\"end_s\"]*SEMANTIC_RATE_HZ))),\n", " m[\"text\"],\n", " m.get(\"private_text\"),\n", " True,\n", " ))\n", " else:\n", " # randomize offset to not get only multiples if no text available\n", " offs = 0\n", " if \"text\" not in meta_info and random.random() > 0.5:\n", " # don't randomize offsets if we have lyrics\n", " offs = random.randint(1, N_TOKENS_AUDIO-1)\n", " segments_info.append((0, min(len(semantic_arr), offs), None, None, False))\n", " for n in range(int(np.ceil((len(semantic_arr)-offs)/N_TOKENS_AUDIO))):\n", " start_idx = offs + n*N_TOKENS_AUDIO\n", " end_idx = min(len(semantic_arr), offs+(n+1)*N_TOKENS_AUDIO)\n", " if end_idx - start_idx < SEMANTIC_RATE_HZ:\n", " # might as well skip mini ones\n", " continue\n", " segments_info.append((\n", " start_idx, \n", " end_idx,\n", " meta_info.get(\"text\") if n == 0 else None, # add text only to first piece\n", " meta_info.get(\"private_text\") if n == 0 else None, # add text only to first piece\n", " False,\n", " ))\n", " if dset_name == \"musescore\":\n", " # break after first piece only text-audio pairs useful here\n", " # TODO: lang here is a hack\n", " meta_info[\"lang\"] = \"en\"\n", " break\n", " arr_list = []\n", " for sem_start_idx, sem_end_idx, text, private_text, is_aligned in segments_info:\n", " if (\n", " sem_end_idx - sem_start_idx > N_TOKENS_AUDIO or \n", " sem_end_idx - sem_start_idx < SEMANTIC_RATE_HZ # arbitrary\n", " ):\n", " continue\n", " coarse_start_idx = int(round(sem_start_idx * COARSE_RATE_HZ / SEMANTIC_RATE_HZ))\n", " coarse_end_idx = int(round(sem_end_idx * COARSE_RATE_HZ / SEMANTIC_RATE_HZ))\n", " assert(sem_end_idx >= 0 and coarse_start_idx >= 0)\n", " if sem_end_idx > len(semantic_arr) or coarse_end_idx > len(coarse_arr):\n", " continue\n", " # get array segments\n", " arr_s = semantic_arr[sem_start_idx:sem_end_idx,:SEMANTIC_N_CODEBOOKS].copy()\n", " arr_c = coarse_arr[coarse_start_idx:coarse_end_idx,:COARSE_N_CODEBOOKS].copy()\n", " # fix any alignment mistakes\n", " arr_s, arr_c = _trim_to_common(arr_s, arr_c)\n", " assert(len(arr_s) == len(arr_c))\n", " # concat and stack\n", " if len(arr_c) < N_TOKENS_AUDIO:\n", " arr_c = np.pad(\n", " arr_c, \n", " ((0, N_TOKENS_AUDIO-len(arr_c)), (0, 0)),\n", " constant_values=COARSE_PAD_TOKEN, \n", " mode=\"constant\",\n", " )\n", " arr_s = np.pad(\n", " arr_s, \n", " ((0, N_TOKENS_AUDIO-len(arr_s)), (0, 0)),\n", " constant_values=SEMANTIC_PAD_TOKEN, \n", " mode=\"constant\",\n", " )\n", " arr = np.concatenate([arr_s, arr_c], axis=-1)\n", " arr = arr.astype(np.uint16)\n", " assert(arr.shape == (N_TOKENS_AUDIO, SEMANTIC_N_CODEBOOKS + COARSE_N_CODEBOOKS))\n", " new_meta = {\n", " \"id\": meta_info[\"id\"],\n", " \"start_s\": round(sem_start_idx / SEMANTIC_RATE_HZ, 2),\n", " \"end_s\": round(sem_end_idx / SEMANTIC_RATE_HZ, 2),\n", " \"original_duration_s\": round(len(semantic_arr) / SEMANTIC_RATE_HZ, 2),\n", " }\n", " if text is not None:\n", " new_meta[\"text\"] = text\n", " if private_text is not None and private_text != private_text:\n", " new_meta[\"text_private\"] = private_text\n", " new_meta[\"text_lang\"] = meta_info[\"lang\"]\n", " new_meta[\"text_aligned\"] = is_aligned\n", " new_meta[\"dset_suffix\"] = \"lyrics\" if meta_info[\"lang\"] == \"en\" else \"lyrics_foreign\" \n", " if \"tags\" in meta_info:\n", " new_meta[\"tags\"] = meta_info[\"tags\"]\n", " if \"private_tags\" in meta_info:\n", " if \"tags\" in meta_info:\n", " # verify that superset\n", " assert(len(set(meta_info[\"tags\"]) - set(meta_info[\"private_tags\"])) == 0)\n", " if \"tags\" not in meta_info or meta_info[\"private_tags\"] != meta_info[\"tags\"]:\n", " new_meta[\"tags_private\"] = meta_info[\"private_tags\"]\n", " if \"original_id\" in meta_info:\n", " new_meta[\"original_id\"] = meta_info[\"original_id\"]\n", " if \"views\" in meta_info:\n", " new_meta[\"views\"] = meta_info[\"views\"]\n", " arr_list.append((arr, new_meta))\n", " del arr_s, arr_c\n", " return arr_list\n", "\n", "def _process_archives(\n", " dset_name,\n", " s3_semantic_archive_filepaths, \n", " s3_coarse_archive_filepaths,\n", " relevant_metas,\n", "):\n", " semantic_archive = {}\n", " for s3_semantic_archive_filepath in s3_semantic_archive_filepaths:\n", " if not check_s3_file_exists(s3_semantic_archive_filepath):\n", " continue\n", " for k, v in read_from_s3(s3_semantic_archive_filepath, read_f=np.load).items():\n", " semantic_archive[k] = v\n", " coarse_archive = {}\n", " for s3_coarse_archive_filepath in s3_coarse_archive_filepaths:\n", " if not check_s3_file_exists(s3_coarse_archive_filepath):\n", " continue\n", " for k, v in read_from_s3(s3_coarse_archive_filepath, read_f=np.load).items():\n", " coarse_archive[k] = v\n", " semantic_uids, coarse_uids = set(semantic_archive.keys()), set(coarse_archive.keys())\n", " # removing this for now just incase (eg imslp)\n", "# assert(\n", "# (len(semantic_uids) < 10 and len(coarse_uids) < 10) or\n", "# (len(semantic_uids & coarse_uids) / (len(semantic_uids) + len(coarse_uids)) > 0.1)\n", "# )\n", "# assert(len(semantic_uids & coarse_uids) > 0)\n", " arr_list = []\n", " for uid in semantic_uids & coarse_uids:\n", " if uid not in relevant_metas:\n", " continue\n", " semantic_arr = semantic_archive[uid]\n", " coarse_arr = coarse_archive[uid]\n", " if np.abs(len(coarse_arr) / COARSE_RATE_HZ - len(semantic_arr) / SEMANTIC_RATE_HZ) > 0.1:\n", " # skip if embeddings not roughly the same duration\n", " continue\n", " semantic_arr, coarse_arr = _trim_to_common(semantic_arr, coarse_arr)\n", " assert(len(coarse_arr) == len(semantic_arr) * COARSE_RATE_HZ / SEMANTIC_RATE_HZ)\n", " arr_list.extend(_parse_arrays(dset_name, relevant_metas[uid], semantic_arr, coarse_arr))\n", " del semantic_archive, coarse_archive\n", " gc.collect()\n", " return arr_list\n", "\n", "def _collect_uids(\n", " s3_semantic_metas_filepaths, \n", " s3_coarse_metas_filepaths,\n", "):\n", " semantic_uids = []\n", " for fp in s3_semantic_metas_filepaths:\n", " semantic_uids.extend([m[\"id\"] for m in read_from_s3(fp, read_f=read_jsonl)])\n", " coarse_uids = []\n", " for fp in s3_coarse_metas_filepaths:\n", " coarse_uids.extend([m[\"id\"] for m in read_from_s3(fp, read_f=read_jsonl)])\n", " return set(semantic_uids) & set(coarse_uids)\n", "\n", "def _prep_data(\n", " dataset,\n", " njobs=5,\n", " chunksize=10,\n", " is_val=False,\n", " n_offs=0,\n", "):\n", " dset_name, dset_version, (start_idx, end_idx), n_sem, n_coarse = dataset\n", " dset_type = \"val\" if is_val else \"tr\"\n", " out_mm_filepath = os.path.join(OUT_DATA_DIR, f\"data_{dset_type}.bin\")\n", " out_metas_filepath = os.path.join(OUT_DATA_DIR, f\"metas_{dset_type}.jsonl\")\n", " tot_duration_dict = defaultdict(int)\n", " n_chunks = int(np.ceil((end_idx - start_idx) / chunksize))\n", " for idx_chunk in tqdm.tqdm(funcy.chunks(chunksize, list(range(start_idx, end_idx))), total=n_chunks):\n", " n_jobs = np.min([njobs, chunksize, len(idx_chunk)])\n", " # collect relevant parts of meta file to avoid copying all to subprocesses\n", " tmp_uid_chunks = Parallel(n_jobs=n_jobs, prefer=\"threads\")(\n", " delayed(_collect_uids)(\n", " [\n", " f\"s3://suno-data/datasets/bundles/{dset_version}/{dset_name}/{SEMANTIC_EMBED_DIR}/\" + \n", " f\"metas/part_{idx_idx}.jsonl\"\n", " for idx_idx in range(idx*n_sem, (idx+1)*n_sem)\n", " ],\n", " [\n", " f\"s3://suno-data/datasets/bundles/{dset_version}/{dset_name}/{CODEC_EMBED_DIR}/\" + \n", " f\"metas/part_{idx_idx}.jsonl\"\n", " for idx_idx in range(idx*n_coarse, (idx+1)*n_coarse)\n", " ],\n", " )\n", " for idx in idx_chunk\n", " ) \n", " uids_per_part = {idx: tmp_uid_chunks[n] for n, idx in enumerate(idx_chunk)}\n", " # collect data\n", " encoded_arrays_list = Parallel(n_jobs=n_jobs, prefer=\"processes\")(\n", " delayed(_process_archives)(\n", " dset_name,\n", " [\n", " f\"s3://suno-data/datasets/bundles/{dset_version}/{dset_name}/{SEMANTIC_EMBED_DIR}/\" + \n", " f\"part_{idx_idx}.npz\"\n", " for idx_idx in range(idx*n_sem, (idx+1)*n_sem)\n", " ],\n", " [\n", " f\"s3://suno-data/datasets/bundles/{dset_version}/{dset_name}/{CODEC_EMBED_DIR}/\" + \n", " f\"part_{idx_idx}.npz\"\n", " for idx_idx in range(idx*n_coarse, (idx+1)*n_coarse)\n", " ],\n", " {\n", " uid: meta_info_map[dset_name][uid] \n", " for uid in uids_per_part[idx] \n", " if uid in meta_info_map[dset_name]\n", " },\n", " )\n", " for idx in idx_chunk\n", " ) \n", " for encoded_arrays in encoded_arrays_list:\n", " to_write_len = np.sum([arr.size for arr, _ in encoded_arrays])\n", " if to_write_len == 0:\n", " continue\n", " out_mm = np.memmap(\n", " out_mm_filepath, dtype=np.uint16, mode=\"r+\", shape=(n_offs+to_write_len,)\n", " )\n", " for arr, arr_meta in encoded_arrays:\n", " out_mm[n_offs:n_offs+arr.size] = arr.reshape(-1,)\n", " n_offs += arr.size\n", " dataset_str = dset_name\n", " if \"dset_suffix\" in arr_meta:\n", " dataset_str += f\"_{arr_meta['dset_suffix']}\"\n", " add_meta = {\n", " \"dataset\": dataset_str,\n", " \"id\": arr_meta[\"id\"],\n", " \"start_s\": round(arr_meta[\"start_s\"], 2),\n", " \"end_s\": round(arr_meta[\"end_s\"], 2),\n", " \"original_duration_s\": arr_meta[\"original_duration_s\"],\n", " }\n", " if \"original_id\" in arr_meta:\n", " add_meta[\"original_id\"] = arr_meta[\"original_id\"]\n", " if \"tags\" in arr_meta:\n", " add_meta[\"tags\"] = arr_meta[\"tags\"]\n", " if \"tags_private\" in arr_meta:\n", " add_meta[\"tags_private\"] = arr_meta[\"tags_private\"]\n", " if \"text\" in arr_meta:\n", " add_meta[\"text\"] = arr_meta[\"text\"]\n", " if \"text_private\" in arr_meta:\n", " add_meta[\"text_private\"] = arr_meta[\"text_private\"]\n", " if \"text_lang\" in arr_meta:\n", " add_meta[\"text_lang\"] = arr_meta[\"text_lang\"]\n", " if \"text_aligned\" in arr_meta:\n", " add_meta[\"text_aligned\"] = arr_meta[\"text_aligned\"]\n", " if \"views\" in arr_meta:\n", " add_meta[\"views\"] = arr_meta[\"views\"]\n", " tot_duration_dict[dataset_str] += arr_meta[\"end_s\"] - arr_meta[\"start_s\"]\n", " with open(out_metas_filepath, \"a\") as f:\n", " f.write(json.dumps(add_meta) + \"\\n\")\n", " out_mm.flush()\n", " del out_mm\n", " del encoded_arrays_list\n", " gc.collect()\n", " for k, v in tot_duration_dict.items():\n", " print(f\"{round(v / 60 / 60):,} hours of {k}\")\n", " return n_offs\n", "\n", "def prep_data(\n", " datasets,\n", " is_val=False,\n", " njobs=5,\n", " chunksize=10,\n", "):\n", " n_offs = 0\n", " dset_type = \"val\" if is_val else \"tr\"\n", " out_mm_filepath = os.path.join(OUT_DATA_DIR, f\"data_{dset_type}.bin\")\n", " out_metas_filepath = os.path.join(OUT_DATA_DIR, f\"metas_{dset_type}.jsonl\")\n", " out_info_filepath = os.path.join(OUT_DATA_DIR, f\"info_{dset_type}.json\")\n", " out_mm = np.memmap(out_mm_filepath, dtype=np.uint16, mode=\"w+\", shape=(1,))\n", " with open(out_metas_filepath, \"w\") as f:\n", " f.write(\"\")\n", " for dataset in datasets:\n", " n_offs = _prep_data(\n", " dataset,\n", " njobs=njobs,\n", " chunksize=chunksize,\n", " is_val=is_val,\n", " n_offs=n_offs,\n", " )\n", " datasets_info = {}\n", " with open(out_metas_filepath) as f:\n", " n = 0\n", " for line in f:\n", " line = line.strip()\n", " if len(line) == 0:\n", " continue\n", " m = json.loads(line)\n", " if m[\"dataset\"] not in datasets_info:\n", " datasets_info[m[\"dataset\"]] = {\"idx_list\": []}\n", " datasets_info[m[\"dataset\"]][\"idx_list\"].append(n)\n", " n += 1\n", " write_json(datasets_info, out_info_filepath)" ] }, { "cell_type": "code", "execution_count": 26, "id": "2a01e35c", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████████| 1/1 [02:23<00:00, 143.32s/it]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "24 hours of youtube_music\n", "2 hours of youtube_music_lyrics_foreign\n", "3 hours of youtube_music_lyrics\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████████| 1/1 [02:22<00:00, 142.39s/it]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "5 hours of genius_hq_lyrics_foreign\n", "4 hours of genius_hq\n", "9 hours of genius_hq_lyrics\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████████| 1/1 [02:28<00:00, 148.48s/it]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "17 hours of imslp\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████████| 1/1 [02:23<00:00, 143.39s/it]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "30 hours of jamendo\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████████| 1/1 [02:22<00:00, 142.76s/it]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "14 hours of pond5_music\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████████| 1/1 [02:22<00:00, 142.60s/it]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "8 hours of deezer_lyrics_foreign\n", "12 hours of deezer\n", "7 hours of deezer_lyrics\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████████| 1/1 [02:32<00:00, 152.06s/it]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "30 hours of ytm_tagged\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████████| 1/1 [02:27<00:00, 147.79s/it]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "3 hours of musescore_lyrics\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████████| 1/1 [02:22<00:00, 142.87s/it]" ] }, { "name": "stdout", "output_type": "stream", "text": [ "30 hours of ytm_mb\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "\n" ] } ], "source": [ "# (start_idx, end_idx), n_archives_semantic, n_archives_coarse\n", "datasets = [\n", " (\"youtube_music\", \"v1\", (0, 1), 1, 1),\n", " (\"genius_hq\", \"v1\", (0, 1), 1, 1),\n", " (\"imslp\", \"v1\", (0, 1), 1, 1),\n", " (\"jamendo\", \"v1\", (0, 1), 1, 1),\n", " (\"pond5_music\", \"v2\", (0, 1), 1, 1),\n", " (\"deezer\", \"v2\", (0, 1), 1, 1),\n", " (\"ytm_tagged\", \"v2\", (0, 1), 1, 1),\n", " (\"musescore\", \"v2\", (0, 1), 1, 1),\n", " (\"ytm_mb\", \"v2\", (0, 1), 1, 1),\n", "]\n", "prep_data(\n", " datasets,\n", " is_val=True,\n", ")\n", "# 24 hours of youtube_music\n", "# 3 hours of youtube_music_lyrics\n", "# 2 hours of youtube_music_lyrics_foreign\n", "# 4 hours of genius_hq\n", "# 9 hours of genius_hq_lyrics\n", "# 5 hours of genius_hq_lyrics_foreign\n", "# 17 hours of imslp\n", "# 30 hours of jamendo\n", "# 14 hours of pond5_music\n", "# 12 hours of deezer\n", "# 7 hours of deezer_lyrics\n", "# 8 hours of deezer_lyrics_foreign\n", "# 30 hours of ytm_tagged\n", "# 3 hours of musescore_lyrics\n", "# 30 hours of ytm_mb" ] }, { "cell_type": "code", "execution_count": 9, "id": "956bd1f3", "metadata": {}, "outputs": [], "source": [ "# Stats:\n", "# tot expected: 4.5Tb\n", "# speed for ytm_mb: 53129/200*(3*60+40)/60/60 == 16 hours" ] }, { "cell_type": "code", "execution_count": null, "id": "a2dfa127", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████| 22/22 [1:00:15<00:00, 164.33s/it]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "10,363 hours of youtube_music_lyrics_foreign\n", "105,498 hours of youtube_music\n", "12,168 hours of youtube_music_lyrics\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████████████████████████████████████████████████████████████████████████| 22/22 [59:35<00:00, 162.51s/it]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "17,469 hours of genius_hq_lyrics_foreign\n", "17,784 hours of genius_hq\n", "41,544 hours of genius_hq_lyrics\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████████| 3/3 [08:15<00:00, 165.26s/it]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "12,175 hours of imslp\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████████| 1/1 [02:07<00:00, 127.69s/it]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "3,251 hours of jamendo\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████████████████████████████████████████████████████████████████████████| 21/21 [53:48<00:00, 153.75s/it]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "58,592 hours of pond5_music\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████████| 8/8 [20:14<00:00, 151.82s/it]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "6,770 hours of deezer_lyrics_foreign\n", "12,181 hours of deezer\n", "10,132 hours of deezer_lyrics\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████| 28/28 [1:22:32<00:00, 176.88s/it]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "141,476 hours of ytm_tagged\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|█████████████████████████████████████████████████████████████████████████████████| 1/1 [01:39<00:00, 99.35s/it]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "103 hours of musescore_lyrics\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ " 98%|███████████████████████████████████████████████████████████████████████▋ | 261/266 [12:41:38<14:44, 176.97s/it]" ] } ], "source": [ "# (start_idx, end_idx), n_archives_semantic, n_archives_coarse\n", "datasets = [\n", " (\"youtube_music\", \"v1\", (1, 4204), 1, 1),\n", " (\"genius_hq\", \"v1\", (1, 4302), 1, 1),\n", " (\"imslp\", \"v1\", (1, 558), 1, 1),\n", " (\"jamendo\", \"v1\", (1, 112), 1, 1),\n", " (\"pond5_music\", \"v2\", (1, 4138), 1, 1),\n", " (\"deezer\", \"v2\", (1, 1538), 1, 1),\n", " (\"ytm_tagged\", \"v2\", (1, 5545), 1, 1),\n", " (\"musescore\", \"v2\", (1, 33), 1, 1),\n", "# (\"ytm_mb\", \"v2\", (1, 53129), 1, 1),\n", " (\"ytm_mb\", \"v2\", (1, 50001), 1, 1),\n", "]\n", "prep_data(\n", " datasets,\n", " is_val=False,\n", " njobs=50,\n", " chunksize=200,\n", ")\n", "# youtube_music: ~1h runtime\n", "# 105,498 hours of youtube_music\n", "# 12,168 hours of youtube_music_lyrics\n", "# 10,363 hours of youtube_music_lyrics_foreign\n", "# genius_hq: ~1h runtime\n", "# 17,784 hours of genius_hq\n", "# 41,544 hours of genius_hq_lyrics\n", "# 17,469 hours of genius_hq_lyrics_foreign\n", "# imslp: ~8m runtime\n", "# 12,175 hours of imslp\n", "# jamendo: ~27m runtime\n", "# 3,251 hours of jamendo\n", "# pond5_music: ~51m runtime\n", "# 58,592 hours of pond5_music\n", "# deezer: ~19m runtime\n", "# 12,181 hours of deezer\n", "# 10,132 hours of deezer_lyrics\n", "# 6,770 hours of deezer_lyrics_foreign\n", "# ytm_tagged: ~1h runtime\n", "# 141,476 hours of ytm_tagged\n", "# musescore: ~5m runtime\n", "# 103 hours of ytm_tagged\n", "# ytm_mb: ~12h runtime\n", "# 1,642,164 hours of ytm_mb" ] }, { "cell_type": "code", "execution_count": 29, "id": "2ab57df8", "metadata": {}, "outputs": [], "source": [ "# # verify\n", "# mm = np.memmap(\"/mnt/localdisk/data/data_val.bin\", dtype=np.uint16, mode=\"r\")\n", "# metas = read_jsonl(\"/mnt/localdisk/test_data/metas_val.jsonl\")\n", "# mm = mm.reshape(-1, 3008, 13)\n", "# assert(len(mm) == len(metas))\n", "# assert(mm[:100,:,0].min() >= 0)\n", "# assert(mm[:100,:,0].max() <= 4000)\n", "# assert(mm[:100,:,1:].min() >= 0)\n", "# assert(mm[:100,:,1:].max() <= 4096)" ] }, { "cell_type": "code", "execution_count": null, "id": "f5bdf8ac", "metadata": {}, "outputs": [], "source": [ "# !du -hs /mnt/data/georg/data/chirp_v2/data_tr.bin\n", "# # 1003G" ] }, { "cell_type": "code", "execution_count": 33, "id": "cbcbca32", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "8187\n" ] } ], "source": [ "# # randomly listen to some stuff\n", "# from suno_utils.tasks.dac_2c_12cb import preload_models as preload_codec_models\n", "# from suno_utils.tasks.dac_2c_12cb import (\n", "# encode as codec_encode,\n", "# decode as codec_decode, \n", "# EMBEDDING_RATE as CODEC_EMBEDDING_RATE,\n", "# )\n", "# _ = preload_codec_models(\"/app/suno/models/dac_2c_25x12.pt\")\n", "# mm = np.memmap(\"/mnt/localdisk/test_data/data_val.bin\", dtype=np.uint16, mode=\"r\")\n", "# mm = mm.reshape(-1, 3008, 13)\n", "# test_metas = read_jsonl(\"/mnt/localdisk/data/metas_val.jsonl\")\n", "# test_info = read_json(\"/mnt/localdisk/data/info_val.json\")\n", "# assert(len(test_metas) == len(mm))\n", "# idx_list = list(range(len(test_metas)))\n", "# #random.shuffle(idx_list)\n", "# #idx_list = [idx for idx in idx_list if \"text\" in test_metas[idx]]\n", "# print(len(mm))" ] }, { "cell_type": "code", "execution_count": 40, "id": "7008ddc7", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "tags: ['Rap', 'West Coast', 'Trap']\n" ] } ], "source": [ "# # show text and audio\n", "# # idx = random.choice(idx_list)\n", "# # idx = random.choice(test_info[\"pond5_music\"][\"idx_list\"])\n", "# idx = random.choice(test_info[\"genius_hq_lyrics\"][\"idx_list\"])\n", "# # idx = random.choice(test_info[\"musescore_lyrics\"][\"idx_list\"])\n", "# # idx = random.choice(test_info[\"deezer_lyrics\"][\"idx_list\"])\n", "# # idx = random.choice(test_info[\"youtube_music_lyrics_foreign\"][\"idx_list\"])\n", "# # idx_key = random.choice(list(test_info.keys()))\n", "# # print(idx_key)\n", "# # idx = random.choice(test_info[idx_key][\"idx_list\"])\n", "# assert(\"original_duration_s\" in test_metas[idx])\n", "# print(\"tags:\", test_metas[idx].get(\"tags\"))\n", "# arr = mm[idx,1:].copy().astype(np.int16)[:,1:]\n", "# pad_idx_arr = np.where(arr == 4096)[0]\n", "# if len(pad_idx_arr) > 0:\n", "# arr = arr[:pad_idx_arr[0],:]\n", "# a = codec_decode(arr)\n", "# a.play(compress=False)\n", "# print(\"text:\", test_metas[idx].get(\"text\"))\n", "# plt.plot(a.array_float[0])" ] }, { "cell_type": "code", "execution_count": null, "id": "f818ab33", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "87b856f6", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "62fb16fc", "metadata": {}, "source": [ "### try same with train" ] }, { "cell_type": "code", "execution_count": 27, "id": "f987040e", "metadata": {}, "outputs": [], "source": [ "%matplotlib inline\n", "from matplotlib import pyplot as plt" ] }, { "cell_type": "code", "execution_count": 23, "id": "4471a7fb", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "80000000\n" ] } ], "source": [ "# randomly listen to some stuff\n", "from suno_utils.tasks.dac_2c_12cb import preload_models as preload_codec_models\n", "from suno_utils.tasks.dac_2c_12cb import (\n", " encode as codec_encode,\n", " decode as codec_decode, \n", " EMBEDDING_RATE as CODEC_EMBEDDING_RATE,\n", ")\n", "from suno_utils.utils.text import read_jsonl, read_json\n", "_ = preload_codec_models(\"/app/suno/models/dac_2c_25x12.pt\")\n", "mm = np.memmap(\"/mnt/localdisk/data/data_tr.bin\", dtype=np.uint16, mode=\"r\")\n", "mm = mm.reshape(-1, 3008, 13)\n", "test_metas = read_jsonl(\"/mnt/localdisk/data/metas_tr.jsonl\")\n", "test_info = read_json(\"/mnt/localdisk/data/info_tr.json\")\n", "assert(len(test_metas) == len(mm))\n", "idx_list = list(range(len(test_metas)))\n", "#random.shuffle(idx_list)\n", "#idx_list = [idx for idx in idx_list if \"text\" in test_metas[idx]]\n", "print(len(mm))" ] }, { "cell_type": "code", "execution_count": 29, "id": "ecb73f42", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "tags: None\n" ] }, { "data": { "text/html": [ "\n", "\n" ], "text/plain": [ "" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "text: None\n" ] }, { "data": { "text/plain": [ "[]" ] }, "execution_count": 29, "metadata": {}, "output_type": "execute_result" }, { "data": { "image/png": "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", "text/plain": [ "
" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "import random\n", "# show text and audio\n", "# idx = random.choice(idx_list)\n", "# idx = random.choice(test_info[\"pond5_music\"][\"idx_list\"])\n", "# idx = random.choice(test_info[\"genius_hq_lyrics\"][\"idx_list\"])\n", "idx = random.choice(test_info[\"ytm_mb\"][\"idx_list\"])\n", "# idx = random.choice(test_info[\"musescore_lyrics\"][\"idx_list\"])\n", "# idx = random.choice(test_info[\"deezer_lyrics\"][\"idx_list\"])\n", "# idx = random.choice(test_info[\"youtube_music_lyrics_foreign\"][\"idx_list\"])\n", "# idx_key = random.choice(list(test_info.keys()))\n", "# print(idx_key)\n", "# idx = random.choice(test_info[idx_key][\"idx_list\"])\n", "assert(\"original_duration_s\" in test_metas[idx])\n", "print(\"tags:\", test_metas[idx].get(\"tags\"))\n", "arr = mm[idx,1:].copy().astype(np.int16)[:,1:]\n", "pad_idx_arr = np.where(arr == 4096)[0]\n", "if len(pad_idx_arr) > 0:\n", " arr = arr[:pad_idx_arr[0],:]\n", "a = codec_decode(arr)\n", "a.play(compress=False)\n", "print(\"text:\", test_metas[idx].get(\"text\"))\n", "plt.plot(a.array_float[0])" ] }, { "cell_type": "code", "execution_count": null, "id": "ee88a2eb", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "1d977073", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "c2653b41", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "067a561b", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "73c92bda", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "08665ec0", "metadata": {}, "source": [ "## Playground" ] }, { "cell_type": "markdown", "id": "747f0363", "metadata": {}, "source": [ "#### get all tags" ] }, { "cell_type": "code", "execution_count": 23, "id": "73286512", "metadata": {}, "outputs": [], "source": [ "test_metas = read_jsonl(\"/mnt/data/georg/data/chirp_v2/metas_val.jsonl\")" ] }, { "cell_type": "code", "execution_count": 54, "id": "a3ed7922", "metadata": {}, "outputs": [], "source": [ "tags = []\n", "for e in train_metas:\n", " tags.extend(e.get(\"tags\", []))" ] }, { "cell_type": "code", "execution_count": 55, "id": "2eed69a8", "metadata": {}, "outputs": [], "source": [ "vc_tags = pd.Series(tags).value_counts()" ] }, { "cell_type": "code", "execution_count": 56, "id": "bf1c83e8", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "Pop 2495781\n", "Rap 2431682\n", "Rock 1193692\n", "commute 1089137\n", "energy boosters 940473\n", "Name: count, dtype: int64" ] }, "execution_count": 56, "metadata": {}, "output_type": "execute_result" } ], "source": [ "vc_tags.head(5)" ] }, { "cell_type": "code", "execution_count": 63, "id": "f76f1e83", "metadata": {}, "outputs": [], "source": [ "vc_tags = vc_tags.to_frame()" ] }, { "cell_type": "code", "execution_count": 72, "id": "6766efb1", "metadata": {}, "outputs": [ { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
count
lonely castle.0.0
Strings.0.0
asianbeatz-riff-1.0.0
dnb loops 001 only guitar wet 180 bpm.0.0
yuung.0.0
\n", "
" ], "text/plain": [ " count\n", "lonely castle. 0.0\n", "Strings. 0.0\n", "asianbeatz-riff-1. 0.0\n", "dnb loops 001 only guitar wet 180 bpm. 0.0\n", "yuung. 0.0" ] }, "execution_count": 72, "metadata": {}, "output_type": "execute_result" } ], "source": [ "vc_tags.tail()" ] }, { "cell_type": "code", "execution_count": 78, "id": "fece8109", "metadata": {}, "outputs": [], "source": [ "a = vc_tags[vc_tags[\"count\"]>=3][\"count\"].to_dict()" ] }, { "cell_type": "code", "execution_count": 4, "id": "52e1a982", "metadata": {}, "outputs": [], "source": [ "test_metas = read_jsonl(\"/mnt/data/georg/data/chirp_v2/metas_val.jsonl\")\n", "test_info = read_json(\"/mnt/data/georg/data/chirp_v2/info_val.json\")" ] }, { "cell_type": "code", "execution_count": 21, "id": "b6b9f19e", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "7854\n" ] }, { "data": { "text/plain": [ "\" So, so you think you can tell heaven from hell, blue skies from pain? Can you tell a green field from a cold steel rail? A smile from a viel? Do you think you can tell? Did they get you to trade your heroes for ghosts? Hot ashfor trees, hot air for a cool breeze? Cold comfort for change Did you exchange a walkon part in the war for a lead role in a cage? How I wish, how I wish you were here, we're just two lost souls swiming in a fishbowl, year after year. Running over the same old ground, what have we found? The same old fears, wish you were here.\"" ] }, "execution_count": 21, "metadata": {}, "output_type": "execute_result" } ], "source": [ "idx = random.choice(test_info[\"musescore_lyrics\"][\"idx_list\"])\n", "print(idx)\n", "m = test_metas[idx]\n", "re.sub(r\"\\[.*?\\]\", \"\", m[\"text\"])" ] }, { "cell_type": "code", "execution_count": null, "id": "5dffab2d", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "4531aac2", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "f0e1aafb", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "143ac0a7", "metadata": {}, "outputs": [], "source": [] } ], "metadata": { "kernelspec": { "display_name": "Python 3 (ipykernel)", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.10.13" } }, "nbformat": 4, "nbformat_minor": 5 }