{ "cells": [ { "cell_type": "code", "execution_count": 1, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T04:31:33.325671Z", "start_time": "2024-04-25T04:31:31.954721Z" } }, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "/tmp/ipykernel_331409/4201459797.py:1: DeprecationWarning: \n", "Pyarrow will become a required dependency of pandas in the next major release of pandas (pandas 3.0),\n", "(to allow more performant data types, such as the Arrow string type, and better interoperability with other libraries)\n", "but was not found to be installed on your system.\n", "If this would cause problems for you,\n", "please provide us feedback at https://github.com/pandas-dev/pandas/issues/54466\n", " \n", " import pandas as pd\n" ] } ], "source": [ "import pandas as pd\n", "import numpy as np\n", "import os\n", "from tqdm import tqdm\n", "from sklearn.model_selection import train_test_split\n", "from suno_utils.utils.s3 import download_s3_files\n", "import sys\n", "from collections import defaultdict\n", "from suno_utils.utils.text import (\n", " write_jsonl,\n", " read_jsonl,\n", " write_json,\n", " read_json,\n", ")\n", "import shutil\n", "import ast\n", "from preference_helper import *\n", "\n", "sys.path.insert(0, \"/home/tony/Work/glockenspiel/sunoGPT/scripts/\")\n", "\n", "from data_preparation_7b import *\n", "import numpy as np\n", "\n", "pd.set_option('display.max_rows', 500)\n", "pd.set_option('display.max_columns', 500)\n", "pd.set_option('display.width', 1000)" ] }, { "cell_type": "code", "execution_count": 2, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T04:31:33.366658Z", "start_time": "2024-04-25T04:31:33.327115Z" } }, "outputs": [], "source": [ "OUT_DATA_DIR = \"/app/suno/data/dpo/7v_r1_v16/\"\n", "# OUT_DATA_DIR = \"/home/tony/Data/test/7v_v6_full/\"\n", "os.makedirs(OUT_DATA_DIR, exist_ok=True)\n", "shutil.copyfile(\"/app/suno/data/dpo/7v_v1_full/tokenizer_60k.json\", os.path.join(OUT_DATA_DIR, \"tokenizer_60k.json\"))\n", "NPZ_DIR = \"/app/suno/data/dpo/v3_npz\"" ] }, { "cell_type": "code", "execution_count": 3, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T04:31:33.390755Z", "start_time": "2024-04-25T04:31:33.367720Z" } }, "outputs": [], "source": [ "# df_origin = pd.read_csv(\"/home/tony/Data/Preference/7b_v1/interesting_clips.csv\")\n", "# df_origin[df_origin[\"id\"] == \"75f69d46-1d81-4327-9253-a8a4886777d1\"]\n", "# df_origin[df_origin[\"request_id\"] == \"9414f356-88d6-40ba-b95e-9f4c5004a4aa\"]\n", "# df_origin[df_origin[\"request_id\"] == \"b2a150c1-f058-4bff-a013-1961e574631d\"]\n", "# df_origin[df_origin[\"id\"] == \"574ba431-18a1-4e28-ac7c-d3e401dba6fe\"]" ] }, { "cell_type": "code", "execution_count": 4, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T04:31:38.107843Z", "start_time": "2024-04-25T04:31:33.392226Z" } }, "outputs": [ { "data": { "text/plain": [ "(122254, 43)" ] }, "execution_count": 4, "metadata": {}, "output_type": "execute_result" } ], "source": [ "df = pd.read_csv(\"/home/tony/Data/Preference/7b_v2/interesting_clips_20240424.csv\")\n", "df.shape\n", "# v0: (68746, 31)\n", "# v5: (938008, 37)\n", "# v6: (1097024, 38)\n", "# v21: (1822704, 41)\n", "# r1 \n", "# v2: 2620268 (pre fixes...lots of imperfections...)\n", "# v3: 1552512" ] }, { "cell_type": "code", "execution_count": 5, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T04:31:38.139332Z", "start_time": "2024-04-25T04:31:38.109015Z" } }, "outputs": [], "source": [ "# download all clips; 883k\n", "from suno_utils.utils.s3 import download_s3_files\n", "\n", "s3_ids = df[\"s3_id\"].values\n", "s3_paths = [f\"s3://suno-data-uploads/studio/uploads/{s3_id}.npz\" for s3_id in s3_ids]\n", "local_paths = [f\"{NPZ_DIR}/{s3_id}.npz\" for s3_id in s3_ids]" ] }, { "cell_type": "code", "execution_count": 6, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T04:31:47.677387Z", "start_time": "2024-04-25T04:31:38.141746Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "jobs to be done 107712 107712\n" ] } ], "source": [ "finished_paths = os.listdir(NPZ_DIR)\n", "finished_paths_set = set(finished_paths)\n", "unfinished_s3_paths = [\n", " path for path in s3_paths if os.path.basename(path) not in finished_paths_set\n", "]\n", "unfinished_paths = [\n", " path for path in local_paths if os.path.basename(path) not in finished_paths_set\n", "]\n", "print(\"jobs to be done\", len(unfinished_paths), len(unfinished_s3_paths))" ] }, { "cell_type": "code", "execution_count": 7, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:09:45.461115Z", "start_time": "2024-04-25T04:31:47.678585Z" } }, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████| 108/108 [1:37:55<00:00, 54.40s/it]\n" ] } ], "source": [ "_ = download_s3_files(unfinished_s3_paths, unfinished_paths, n_cores=64)" ] }, { "cell_type": "code", "execution_count": 8, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:15:51.776033Z", "start_time": "2024-04-25T06:09:45.462579Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "jobs to be done 14 14\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:03<00:00, 3.41s/it]\n" ] } ], "source": [ "# # deleted files are at: deleted; try to get them as well\n", "finished_paths = os.listdir(NPZ_DIR)\n", "finished_paths_set = set(finished_paths)\n", "unfinished_s3_paths = [\n", " path for path in s3_paths if os.path.basename(path) not in finished_paths_set\n", "]\n", "unfinished_paths = [\n", " path for path in local_paths if os.path.basename(path) not in finished_paths_set\n", "]\n", "print(\"jobs to be done\", len(unfinished_paths), len(unfinished_s3_paths))\n", "unfinished_deleted_s3_paths = [\n", " path.replace(\"/uploads/\", \"/deleted/\") for path in unfinished_s3_paths\n", "]\n", "_ = download_s3_files(unfinished_deleted_s3_paths, unfinished_paths, n_cores=32)" ] }, { "cell_type": "code", "execution_count": 9, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:15:51.780128Z", "start_time": "2024-04-25T06:15:51.777944Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Finish downloads\n" ] } ], "source": [ "print(\"Finish downloads\")" ] }, { "cell_type": "code", "execution_count": 10, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:37.445993Z", "start_time": "2024-04-25T06:15:51.782824Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "11256105\n" ] } ], "source": [ "converted_paths = os.listdir(NPZ_DIR)\n", "print(len(converted_paths))" ] }, { "cell_type": "code", "execution_count": 11, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:37.450034Z", "start_time": "2024-04-25T06:21:37.447265Z" } }, "outputs": [], "source": [ "# # don't run this unless you kill some job accidentally\n", "# for s3_processed_file in tqdm.tqdm(converted_paths):\n", "# processed_file = os.path.basename(s3_processed_file)\n", "# processed_file_path = os.path.join(\"/app/suno/data/dpo/v3_npz\", processed_file)\n", "# # could have been removed already\n", "# if os.path.exists(processed_file_path):\n", "# file_size = os.stat(processed_file_path).st_size\n", "# # print(processed_file_path, file_size)\n", "# # break\n", "# if file_size < 2000:\n", "# print(processed_file_path, file_size)\n", "# os.remove(processed_file_path)\n", "# converted_paths = os.listdir(NPZ_DIR)" ] }, { "cell_type": "code", "execution_count": 12, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:42.079918Z", "start_time": "2024-04-25T06:21:37.451122Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "11256105\n" ] } ], "source": [ "converted_paths = set([f.replace(\".npz\", \"\") for f in converted_paths])\n", "print(len(converted_paths))" ] }, { "cell_type": "code", "execution_count": 13, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:54.215286Z", "start_time": "2024-04-25T06:21:42.081900Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "pre-downloaded df (122254, 43)\n", "downloaded df (122254, 43)\n" ] } ], "source": [ "print(\"pre-downloaded df\", df.shape)\n", "df[df[\"s3_id\"].isin(converted_paths)].shape\n", "df = df[df[\"s3_id\"].isin(converted_paths)].copy()\n", "print(\"downloaded df\", df.shape)" ] }, { "cell_type": "code", "execution_count": 14, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:54.246764Z", "start_time": "2024-04-25T06:21:54.217242Z" } }, "outputs": [ { "data": { "text/plain": [ "is_7b\n", "True 122254\n", "Name: count, dtype: int64" ] }, "execution_count": 14, "metadata": {}, "output_type": "execute_result" } ], "source": [ "df[\"is_7b\"] = df[\"model_name\"].str.contains(\"v3\")\n", "df[\"is_7b\"].value_counts()" ] }, { "cell_type": "code", "execution_count": 15, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.120955Z", "start_time": "2024-04-25T06:21:54.248283Z" } }, "outputs": [ { "ename": "NameError", "evalue": "name 'BREAK' is not defined", "output_type": "error", "traceback": [ "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m", "\u001b[0;31mNameError\u001b[0m Traceback (most recent call last)", "Cell \u001b[0;32mIn[15], line 1\u001b[0m\n\u001b[0;32m----> 1\u001b[0m \u001b[43mBREAK\u001b[49m\n", "\u001b[0;31mNameError\u001b[0m: name 'BREAK' is not defined" ] } ], "source": [ "BREAK" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "# LET's do the data prep" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.123921Z", "start_time": "2024-04-25T06:21:56.123911Z" } }, "outputs": [], "source": [ "date_cut = '2024-03-22 04:30:00' # v3 launch test time" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.124816Z", "start_time": "2024-04-25T06:21:56.124808Z" } }, "outputs": [], "source": [ "# # let's also kick out the ... ipo and ipo-dpoed model for now?\n", "print(df.groupby([\"preference\"])[\"model_name\"].value_counts())\n", "print(df.shape)\n", "# df = df[df[\"model_name\"].isin([\"chirp-v3-engine-d\", \"chirp-v3-engine-v0\", \"chirp-v3-engine-i\"])]\n", "# df = df[df[\"model_name\"].isin([\"chirp-v3-engine-d\", \"chirp-v3-engine-v0\"])]\n", "# only IPO\n", "# df = df[(df[\"model_name\"].isin([\"chirp-v3-engine-i\"]) ) & (df[\"created_at\"] >= date_cut)]\n", "# for some reason...we can't train dpo on the ipo data...it just doesn't follow lyrics...X.x\n", "# df = df[\n", "# (\n", "# (df[\"model_name\"].isin([\"chirp-v3-engine-d\", \"chirp-v3-engine-v0\"]))\n", "# & (df[\"preference\"] == True)\n", "# )\n", "# | (df[\"preference\"] == False)\n", "# ]\n", "print(df.shape)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.125652Z", "start_time": "2024-04-25T06:21:56.125644Z" } }, "outputs": [], "source": [ "print(df.shape)\n", "df = df[\n", " df[\"request_id\"].isin(\n", " df[\"request_id\"].value_counts().index[df[\"request_id\"].value_counts() == 2]\n", " )\n", "]\n", "print(df.shape)\n", "print(df.groupby([\"preference\"])[\"model_name\"].value_counts())\n", "assert df.shape[0] == df[\"request_id\"].nunique() * 2" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.126651Z", "start_time": "2024-04-25T06:21:56.126643Z" } }, "outputs": [], "source": [ "# expand the metadata columns -- this takes forever...~ 6 mins\n", "test_slice = df[\"metadata\"].apply(lambda x: ast.literal_eval(x))\n", "test_slice_series = test_slice.apply(pd.Series)\n", "df = pd.concat([df, test_slice_series], axis=1, join=\"inner\")\n", "print(\"unique_requests\", df[\"request_id\"].nunique())" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.127391Z", "start_time": "2024-04-25T06:21:56.127384Z" } }, "outputs": [], "source": [ "# cut on v3 time\n", "df[\"has_bad_gpt_prompt\"] = (~df[\"gpt_description_prompt\"].isna()) & (df[\"created_at\"] <= \"2024-03-22 04:30:00\")\n", "df[\"has_bad_gpt_prompt\"].value_counts()" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.128297Z", "start_time": "2024-04-25T06:21:56.128289Z" } }, "outputs": [], "source": [ "# double check we removed the gpt prompted ones for now \n", "# TODO: add these data back...\n", "df = df[df[\"has_bad_gpt_prompt\"] == False]\n", "print(\"unique_requests\", df[\"request_id\"].nunique())" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.128992Z", "start_time": "2024-04-25T06:21:56.128984Z" } }, "outputs": [], "source": [ "# get the original duration of the clips, if they are concacted\n", "df[\"original_duration_s\"] = df[\"total_start_s\"] + df[\"duration\"]" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.129697Z", "start_time": "2024-04-25T06:21:56.129689Z" } }, "outputs": [], "source": [ "# classify the continue at behavoirs by the duration choice\n", "audio_prompt_id_to_continue_at = {}\n", "for _, row in df[~df[\"audio_prompt_id\"].isna()].iterrows():\n", " audio_prompt_id = row[\"audio_prompt_id\"]\n", " if audio_prompt_id not in audio_prompt_id_to_continue_at:\n", " audio_prompt_id_to_continue_at[audio_prompt_id] = row[\"continue_at\"]\n", " else:\n", " # pick the max\n", " audio_prompt_id = max(\n", " audio_prompt_id_to_continue_at[audio_prompt_id], row[\"continue_at\"]\n", " )\n", "print(len(audio_prompt_id_to_continue_at))\n", "df[\"has_continue_and_start_continue_at\"] = df[\"id\"].apply(\n", " lambda x: audio_prompt_id_to_continue_at.get(x)\n", ")\n", "# we want continue at to be at least half of the clip...\n", "df[\"good_continue_at\"] = (\n", " (df[\"has_continue_and_start_continue_at\"] / df[\"duration\"]) > 0.9\n", ") | df[\"has_continue_and_start_continue_at\"].isna()\n", "print(df[\"good_continue_at\"].value_counts())" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.130639Z", "start_time": "2024-04-25T06:21:56.130631Z" } }, "outputs": [], "source": [ "print(\n", " df[\"preference\"].value_counts(),\n", " df[\"is_7b\"].value_counts(),\n", " df[\"model_name\"].value_counts(),\n", " df.groupby([\"preference\"])[\"model_name\"].value_counts()\n", ")" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.131407Z", "start_time": "2024-04-25T06:21:56.131399Z" } }, "outputs": [], "source": [ "df = df.sort_values(by=[\"request_id\", \"preference\"])\n", "df[\"duration_rel_diff\"] = df['duration'].diff() \n", "df[\"play_rel_diff\"] = df['reaction_play_count'].diff() " ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.132339Z", "start_time": "2024-04-25T06:21:56.132332Z" } }, "outputs": [], "source": [ "df[\n", " [\n", " \"request_id\",\n", " \"preference\",\n", " \"duration\",\n", " \"duration_rel_diff\",\n", " \"reaction_play_count\",\n", " \"play_rel_diff\",\n", " ]\n", "].tail(n=10)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.133305Z", "start_time": "2024-04-25T06:21:56.133298Z" } }, "outputs": [], "source": [ "df[df[\"preference\"] == True].describe()" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.134003Z", "start_time": "2024-04-25T06:21:56.133995Z" } }, "outputs": [], "source": [ "df[df[\"preference\"] == False].describe()" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.135202Z", "start_time": "2024-04-25T06:21:56.135194Z" } }, "outputs": [], "source": [ "normal_pos_play_count = 3\n", "# this is lower, cause a concat is probably already ensuring that it is good\n", "concat_pos_play_count = 1\n", "# this is a filter on the concated clip\n", "concat_total_play_count = 3\n", "\n", "neg_filter_selection_mask = (\n", " (df[\"preference\"] == False) # get basics aligned\n", " & (df[\"reaction_play_count\"] >= 1) # has to be played once -- just villian?\n", " & (\n", " df[\"reaction_play_count\"] <= 3\n", " ) # if it is actually bad, shouldn't be listened often\n", " & (df[\"duration\"] >= 10) # can't be too short, otherwise it is obvious\n", " & (df[\"duration\"] <= 120) # can't be badly long\n", " & (df[\"has_continue_and_start_continue_at\"].isna()) # won't have any continues\n", " & ((df[\"dislike_count\"] >= 1) | (df[\"flag_count\"] >= 1)) # this is kinda strict\n", " # & (\n", " # (df_slice[\"is_in_playlist\"] == False)\n", " # & (df_slice[\"concat_in_playlist\"] == False)\n", " # ) # can't be part of a playlist -- otherwise there are some like signal in it?\n", ")\n", "pos_filter_selectin_mask = (\n", " (df[\"preference\"] == True) # get basics aligned\n", " & (\n", " df[\"good_continue_at\"] == True\n", " ) # if continue, needs to continue off a certain percentage\n", " & (df[\"continue_at\"].isna())\n", " & (\n", " (\n", " (df[\"part_of_concat\"] == True)\n", " & (df[\"reaction_play_count\"] >= concat_pos_play_count)\n", " & (df[\"concat_play_counts\"] >= concat_total_play_count)\n", " )\n", " | (\n", " (df[\"part_of_concat\"] == False)\n", " & (df[\"reaction_play_count\"] >= normal_pos_play_count)\n", " )\n", " )\n", " & (df[\"play_rel_diff\"] >= 0) # this is more like quality assurance\n", " & (df[\"duration\"] >= 10) # can't be too short, otherwise it is obvious\n", " & (df[\"duration\"] <= 120) # can't be badly long\n", " & (df[\"dislike_count\"] == 0) # can't have dislikes\n", " & (df[\"flag_count\"] == 0) # can't have issues\n", " & (df[\"user_n_clips\"] >= 10) # user needs to have genereated at least 40\n", " # & (df[\"duration_rel_diff\"] < 10) # positive isn't just longer\n", " # & (df[\"upvote_count\"] >= 1)\n", ")\n", "print(\n", " \"negative\",\n", " sum(neg_filter_selection_mask),\n", " \"positive\",\n", " sum(pos_filter_selectin_mask),\n", ")\n", "\n", "neg_filter_requests = df[neg_filter_selection_mask][\"request_id\"].unique()\n", "pos_filter_requests = df[pos_filter_selectin_mask][\"request_id\"].unique()\n", "# looking for very strong signal here:\n", "# listen to the positive/negative more than once\n", "# disliked one of the clips\n", "unique_requests = set(pos_filter_requests).intersection(neg_filter_requests)\n", "print(\n", " \"total pair requests\",\n", " df[\"request_id\"].nunique(),\n", " \"selected pair requests\",\n", " len(unique_requests),\n", " f\"frac {len(unique_requests) / df['request_id'].nunique():.3f}\",\n", ")" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.136048Z", "start_time": "2024-04-25T06:21:56.136041Z" } }, "outputs": [], "source": [ "df_slice = df[df[\"request_id\"].isin(set(unique_requests))].copy()\n", "print(\n", " \"requests\",\n", " df_slice[\"request_id\"].nunique(),\n", " \"clips\",\n", " df_slice.shape[0],\n", " f\"total khrs {sum(df_slice['duration'] / 3600 / 1000):.3f};\",\n", " f\"N gpus for 1250 iters {df_slice.shape[0] / 8 / 4 / 1250:.3f};\",\n", " f\"n unique users {df_slice['user_id'].nunique()}\",\n", ")\n", "# 76171 152342 total khrs 2.880 n gpus for 1250 iters 3.809" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.137135Z", "start_time": "2024-04-25T06:21:56.137127Z" } }, "outputs": [], "source": [ "last_continue_requests = df[\n", " ((df[\"has_continue_and_start_continue_at\"] / df[\"duration\"]) > 0.99)\n", "][\"request_id\"].unique()\n", "early_continue_requests = df[\n", " ((df[\"has_continue_and_start_continue_at\"] / df[\"duration\"]) < 0.99)\n", " & ((df[\"has_continue_and_start_continue_at\"] / df[\"duration\"]) > 0.01)\n", "][\"request_id\"].unique()\n", "not_continue_requests = df[df[\"has_continue_and_start_continue_at\"].isna()][\n", " \"request_id\"\n", "].unique()" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.137847Z", "start_time": "2024-04-25T06:21:56.137839Z" } }, "outputs": [], "source": [ "print(df_slice.groupby([\"preference\"])[\"model_name\"].value_counts())" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.138543Z", "start_time": "2024-04-25T06:21:56.138536Z" } }, "outputs": [], "source": [ "get_preferfence_counts(df_slice)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.139257Z", "start_time": "2024-04-25T06:21:56.139249Z" } }, "outputs": [], "source": [ "# get_preferfence_counts(df_slice[df_slice[\"request_id\"].isin(not_continue_requests)])" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.139996Z", "start_time": "2024-04-25T06:21:56.139988Z" } }, "outputs": [], "source": [ "# get_preferfence_counts(df_slice[df_slice[\"request_id\"].isin(last_continue_requests)])" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.140825Z", "start_time": "2024-04-25T06:21:56.140818Z" } }, "outputs": [], "source": [ "# get_preferfence_counts(df_slice[df_slice[\"request_id\"].isin(early_continue_requests)])" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.141705Z", "start_time": "2024-04-25T06:21:56.141698Z" } }, "outputs": [], "source": [ "df_slice[df_slice[\"preference\"] == True].describe()" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.142271Z", "start_time": "2024-04-25T06:21:56.142264Z" } }, "outputs": [], "source": [ "df_slice[df_slice[\"preference\"] == False].describe()" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.143647Z", "start_time": "2024-04-25T06:21:56.143638Z" } }, "outputs": [], "source": [ "test_mask = (df_slice[\"preference\"] == True) & (\n", " (df_slice[\"is_in_playlist\"] == True) | (df_slice[\"concat_in_playlist\"] == True)\n", ")\n", "print(\"positive in playlist\", df_slice[test_mask].shape)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.144363Z", "start_time": "2024-04-25T06:21:56.144356Z" } }, "outputs": [], "source": [ "interesting_clips_must_be_positive_mask = (\n", " (df_slice[\"upvoted\"] == True)\n", " | (df_slice[\"has_action\"] == True)\n", " | (df_slice[\"part_of_concat\"] == True)\n", ")\n", "interesting_clips_must_be_not_negative_mask = df_slice[\"downvoted\"] == False\n", "interesting_clips_mask = interesting_clips_must_be_positive_mask & interesting_clips_must_be_not_negative_mask\n", "assert interesting_clips_mask.eq(df_slice[\"preference\"]).all()" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.145149Z", "start_time": "2024-04-25T06:21:56.145142Z" } }, "outputs": [], "source": [ "BREAK" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Need to kick out the ones has gpt prompt -- these are pairs with different text inputs" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.145714Z", "start_time": "2024-04-25T06:21:56.145707Z" } }, "outputs": [], "source": [ "# don't have continue at\n", "df_slice[df_slice[\"continue_at\"].isna()][\"request_id\"].nunique()" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.146380Z", "start_time": "2024-04-25T06:21:56.146372Z" } }, "outputs": [], "source": [ "final_filtered_requests = df_slice[\"request_id\"].unique()\n", "print(len(final_filtered_requests))" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.147108Z", "start_time": "2024-04-25T06:21:56.147100Z" } }, "outputs": [], "source": [ "train_requests, val_requests = train_test_split(\n", " sorted(list(final_filtered_requests)), test_size=0.01, random_state=42\n", ")\n", "print(len(train_requests), len(val_requests))\n", "\n", "train_df = df_slice[df_slice[\"request_id\"].isin(set(train_requests))].copy()\n", "val_df = df_slice[df_slice[\"request_id\"].isin(set(val_requests))].copy()\n", "train_df = train_df.sort_values(by=[\"request_id\", \"preference\"])\n", "train_df = train_df.reset_index()\n", "val_df = val_df.sort_values(by=[\"request_id\", \"preference\"])\n", "val_df = val_df.reset_index()\n", "\n", "print(train_df.shape, val_df.shape)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.148104Z", "start_time": "2024-04-25T06:21:56.148096Z" } }, "outputs": [], "source": [ "def reshift(arr):\n", " sem_start_idx = 0\n", " sem_end_idx = len(arr) - 1\n", " semantic_arr = arr[:, :SEMANTIC_N_CODEBOOKS]\n", " coarse_arr = arr[:, SEMANTIC_N_CODEBOOKS:]\n", "\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", " assert not (sem_end_idx > len(semantic_arr) or coarse_end_idx > len(coarse_arr))\n", "\n", " # get array segments\n", " arr_s = semantic_arr[sem_start_idx:sem_end_idx, :].copy()\n", " arr_c = coarse_arr[coarse_start_idx:coarse_end_idx, :].copy()\n", " assert arr_s.max() <= SEMANTIC_PAD_TOKEN\n", " assert arr_c.max() <= COARSE_PAD_TOKEN\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", " return arr\n", "\n", "\n", "def make_dataset(input_df, is_val=False):\n", " dset_type = \"val\" if is_val else \"tr\"\n", " out_mmap_path = os.path.join(OUT_DATA_DIR, f\"data_{dset_type}.bin\")\n", " out_metas_path = os.path.join(OUT_DATA_DIR, f\"meta_{dset_type}.jsonl\")\n", " out_info_filepath = os.path.join(OUT_DATA_DIR, f\"info_{dset_type}.json\")\n", "\n", " # gather the data\n", " _ = np.memmap(out_mmap_path, dtype=np.uint16, mode=\"w+\", shape=(1,))\n", " n_offs = 0\n", " tot_duration_dict = defaultdict(int)\n", " datasets_info = defaultdict(dict)\n", " n = 0\n", " for i, row in tqdm.tqdm(input_df.iterrows(), total=len(input_df)):\n", " # we need to alternate between preference: neg, pos\n", " # print(i, row)\n", " assert row[\"preference\"] == (i % 2 == 1)\n", " # make mmap -- two different paths\n", " local_path = (\n", " f\"{NPZ_DIR}/{row['s3_id']}.npz\"\n", " if row[\"is_7b\"]\n", " else f\"/app/suno/data/dpo/7b_npz/{row['s3_id']}.npz\"\n", " )\n", " if not os.path.exists(local_path):\n", " # print(row, local_path)\n", " raise ValueError()\n", " # print(local_path)\n", " # print( np.load(local_path))\n", " try:\n", " arr = (\n", " np.load(local_path)[\"v3.0_raw\"]\n", " if row[\"is_7b\"]\n", " else np.load(local_path)[\"v2_raw\"]\n", " )\n", " except Exception as e:\n", " print(local_path)\n", " raise e\n", " assert arr.shape[0] <= 3000\n", " assert arr.shape[1] == 13\n", " arr_duration = arr.shape[0] / 25\n", " # print(arr.shape)\n", " arr = reshift(arr)\n", " # print(\"after shift and pad\", arr.shape)\n", " arr = arr.reshape(\n", " -1,\n", " )\n", " # print(arr.shape)\n", " out_mm = np.memmap(\n", " out_mmap_path,\n", " dtype=np.uint16,\n", " mode=\"r+\",\n", " shape=(n_offs + arr.size,),\n", " )\n", " out_mm[n_offs : n_offs + arr.size] = arr\n", " # print(f\"offset is: {n_offs}\")\n", " # break\n", " # write it once\n", " out_mm.flush()\n", " del out_mm\n", "\n", " add_metas = []\n", " add_meta = {\n", " \"dataset\": f\"perference_{int(row['preference'])}\",\n", " \"id\": row[\"s3_id\"], # this is the row s3_id\n", " \"start_s\": row[\"total_start_s\"] if row[\"total_start_s\"] >= 0 else None,\n", " \"end_s\": (\n", " row[\"total_clip_s\"] if row[\"total_clip_s\"] >= 0 else None\n", " ), # for full clips we do know it has an edding, other wise, we don't know\n", " \"original_duration_s\": (\n", " row[\"original_duration_s\"]\n", " if row[\"original_duration_s\"] >= 0\n", " else arr_duration\n", " ), # this nees to be... a bit more complicated, only works with concat!\n", " \"vocal_start_s\": None, # these are unfortunately missing for now\n", " \"vocal_end_s\": None, # these are unfortunately missing for now\n", " \"tags\": [\n", " row[\"tags\"] if row[\"tags\"] else \"\"\n", " ], # tags is a list, do you know :)\n", " \"text\": row[\"prompt\"] if row[\"prompt\"] else \"\",\n", " }\n", " add_metas.append(add_meta)\n", " tot_duration_dict[row[\"preference\"]] += arr_duration\n", " write_jsonl(\n", " add_metas,\n", " os.path.join(out_metas_path),\n", " do_append=bool(n_offs != 0),\n", " )\n", " if \"idx_list\" not in datasets_info[add_meta[\"dataset\"]]:\n", " datasets_info[add_meta[\"dataset\"]][\"idx_list\"] = [n]\n", " else:\n", " datasets_info[add_meta[\"dataset\"]][\"idx_list\"].append(n)\n", " n += 1\n", " n_offs += arr.size\n", "\n", " write_json(datasets_info, out_info_filepath)\n", " print(f\"Total {n} clips\")\n", " for k, v in tot_duration_dict.items():\n", " print(f\"{round(v / 60 / 60):,} hours of {k}\")\n", " print(f\"Done\")" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "# Actually make" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.148638Z", "start_time": "2024-04-25T06:21:56.148631Z" } }, "outputs": [], "source": [ "# val_df[[\"request_id\", \"metadata\", \"updated_at\", \"user_id\", \"preference\"]].head()" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.149612Z", "start_time": "2024-04-25T06:21:56.149605Z" } }, "outputs": [], "source": [ "total_duration = 0\n", "for i, row in tqdm.tqdm(train_df.iterrows(), total=len(train_df)):\n", " # we need to alternate between preference: neg, pos\n", " # print(i, row)\n", " try:\n", " assert row[\"preference\"] == (i % 2 == 1)\n", " except:\n", " print(i, row)\n", " total_duration += row[\"duration\"]\n", "print(f\"{round(total_duration / 60 / 60):,} hours of {train_df.shape[0]} clips, {train_df.shape[0] / 8 / 4 / 1000} nodes\")" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.150399Z", "start_time": "2024-04-25T06:21:56.150391Z" } }, "outputs": [], "source": [ "make_dataset(val_df, is_val=True)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.151261Z", "start_time": "2024-04-25T06:21:56.151254Z" } }, "outputs": [], "source": [ "make_dataset(train_df, is_val=False)" ] }, { "cell_type": "markdown", "metadata": { "ExecuteTime": { "end_time": "2024-01-29T19:46:47.549860Z", "start_time": "2024-01-29T19:46:47.548015Z" } }, "source": [ "# Validation" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.152000Z", "start_time": "2024-04-25T06:21:56.151992Z" } }, "outputs": [], "source": [ "# verify\n", "mm = np.memmap(os.path.join(OUT_DATA_DIR, f\"data_val.bin\"), dtype=np.uint16, mode=\"r\")\n", "test_metas = read_jsonl(os.path.join(OUT_DATA_DIR, f\"meta_val.jsonl\"))\n", "test_info = read_json(os.path.join(OUT_DATA_DIR, f\"info_val.json\"))\n", "mm = mm.reshape(-1, 3008, 13)\n", "assert len(mm) == len(test_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() <= 2048" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.152708Z", "start_time": "2024-04-25T06:21:56.152701Z" } }, "outputs": [], "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_stream_to_full_audio as codec_decode,\n", " EMBEDDING_RATE as CODEC_EMBEDDING_RATE,\n", " decode as decode\n", ")\n", "os.environ[\"CUDA_VISIBLE_DEVICES\"] = \"3\"\n", "_ = preload_codec_models(\"/app/suno/tony/v3/dac_2c_25x12.pt\", device=\"cuda\")\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": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.153616Z", "start_time": "2024-04-25T06:21:56.153607Z" } }, "outputs": [], "source": [ "idx = random.choice(test_info[\"perference_0\"][\"idx_list\"])\n", "assert \"original_duration_s\" in test_metas[idx]\n", "# positive index should be shifted by 1\n", "pos_idx = idx + 1\n", "print(\n", " \"tags:\",\n", " test_metas[idx].get(\"tags\") == test_metas[pos_idx].get(\"tags\"),\n", " test_metas[idx].get(\"tags\"),\n", ")\n", "arr = mm[idx, 1:].copy().astype(np.int16)[:, 1:]\n", "pos_arr = mm[pos_idx, 1:].copy().astype(np.int16)[:, 1:]\n", "pad_idx_arr = np.where(arr == COARSE_PAD_TOKEN)[0]\n", "if len(pad_idx_arr) > 0:\n", " arr = arr[: pad_idx_arr[0], :]\n", "pos_pad_idx_arr = np.where(pos_arr == COARSE_PAD_TOKEN)[0]\n", "if len(pos_pad_idx_arr) > 0:\n", " pos_arr = pos_arr[: pos_pad_idx_arr[0], :]\n", "a = decode(arr)\n", "print(\"negative example\", test_metas[idx])\n", "a.play(compress=False)\n", "pos_a = decode(pos_arr)\n", "print(\"positive example\", test_metas[pos_idx])\n", "pos_a.play(compress=False)\n", "print(\n", " \"text:\",\n", " test_metas[idx].get(\"text\") == test_metas[pos_idx].get(\"text\"),\n", " test_metas[idx].get(\"text\"),\n", ")" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.154538Z", "start_time": "2024-04-25T06:21:56.154530Z" } }, "outputs": [], "source": [ "# val_df[val_df[\"tags\"] == 'a vibrant blend of experimental jazz fusion, drum-and-bass and swagger fuzzed-out guitars']" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.155210Z", "start_time": "2024-04-25T06:21:56.155203Z" } }, "outputs": [], "source": [ "# from collections import Counter\n", "# c = Counter()\n", "# for _, row in df_slice.iterrows():\n", "# # print(row[\"metadata\"])\n", "# for k in ast.literal_eval(row[\"metadata\"]).keys():\n", "# c[k] += 1" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.155970Z", "start_time": "2024-04-25T06:21:56.155962Z" } }, "outputs": [], "source": [ "# original_npz_path = f\"/app/suno/data/dpo/7b_npz/{test_metas[idx]['id']}.npz\"\n", "# original_npz_path = \"/app/suno/data/dpo/7b_npz/729c3011-f672-4ccd-8d82-1cbf2b52ff69.npz\"\n", "# original_arr = np.load(original_npz_path)[\"v2_raw\"]" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.156596Z", "start_time": "2024-04-25T06:21:56.156589Z" } }, "outputs": [], "source": [ "def validation_on_metas(input_metas):\n", "\n", " total_bad = 0\n", " total_good = 0\n", " for idx in range(len(input_metas)):\n", " if idx % 2 == 0:\n", " pos_idx = idx + 1\n", " if input_metas[idx].get(\"tags\") != input_metas[pos_idx].get(\"tags\"):\n", " # print(test_metas[idx].get(\"text\") == test_metas[pos_idx].get(\"text\"), test_metas[idx].get(\"tags\"), test_metas[pos_idx].get(\"tags\"))\n", " total_bad += 1\n", " else:\n", " total_good += 1\n", " print(total_good, total_bad)\n", " return\n", "\n", "\n", "validation_on_metas(test_metas)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.157293Z", "start_time": "2024-04-25T06:21:56.157285Z" } }, "outputs": [], "source": [ "train_info = read_json(os.path.join(OUT_DATA_DIR, f\"info_tr.json\"))" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.157976Z", "start_time": "2024-04-25T06:21:56.157969Z" } }, "outputs": [], "source": [ "n_neg_tr = train_info[\"perference_0\"][\"idx_list\"]\n", "n_pos_tr = train_info[\"perference_1\"][\"idx_list\"]\n", "assert len(n_pos_tr) == len(n_neg_tr)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.158986Z", "start_time": "2024-04-25T06:21:56.158978Z" } }, "outputs": [], "source": [ "total_iters = (len(n_neg_tr) + len(n_pos_tr))\n", "print(\"total samples\", total_iters, train_df.shape)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.159486Z", "start_time": "2024-04-25T06:21:56.159478Z" } }, "outputs": [], "source": [ "print(\"1 epoch per batch 4, total\", total_iters / 8 / 4 / 4)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-04-25T06:21:56.160332Z", "start_time": "2024-04-25T06:21:56.160325Z" } }, "outputs": [], "source": [ "!cd /home/tony/Work/tony/slurm && sbatch sbatch_dpo" ] }, { "cell_type": "code", "execution_count": null, "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" }, "toc": { "base_numbering": 1, "nav_menu": {}, "number_sections": true, "sideBar": true, "skip_h1_title": false, "title_cell": "Table of Contents", "title_sidebar": "Contents", "toc_cell": false, "toc_position": {}, "toc_section_display": true, "toc_window_display": false } }, "nbformat": 4, "nbformat_minor": 2 }