{ "cells": [ { "cell_type": "code", "execution_count": 1, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:58:21.040680Z", "start_time": "2024-05-16T13:58:19.777010Z" }, "execution": { "iopub.execute_input": "2024-06-30T22:24:22.269200Z", "iopub.status.busy": "2024-06-30T22:24:22.269051Z", "iopub.status.idle": "2024-06-30T22:24:23.707123Z", "shell.execute_reply": "2024-06-30T22:24:23.706603Z", "shell.execute_reply.started": "2024-06-30T22:24:22.269180Z" } }, "outputs": [], "source": [ "%load_ext autoreload\n", "%autoreload 2\n", "\n", "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", "from preference_data_preparation_4min_13b_extend import *\n", "\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": "2024-05-16T13:58:21.082172Z", "start_time": "2024-05-16T13:58:21.041926Z" }, "execution": { "iopub.execute_input": "2024-06-30T22:24:23.707962Z", "iopub.status.busy": "2024-06-30T22:24:23.707772Z", "iopub.status.idle": "2024-06-30T22:24:23.769512Z", "shell.execute_reply": "2024-06-30T22:24:23.769057Z", "shell.execute_reply.started": "2024-06-30T22:24:23.707945Z" } }, "outputs": [], "source": [ "OUT_DATA_DIR = \"/app/suno/data/dpo/13b_extend_mix/\"\n", "os.makedirs(OUT_DATA_DIR, exist_ok=True)\n", "shutil.copyfile(\"/app/suno/data/dpo/7v_v20_full/tokenizer_60k.json\", os.path.join(OUT_DATA_DIR, \"tokenizer_60k.json\"))\n", "NPZ_DIR = \"/app/suno/data/dpo/13b_npz\"" ] }, { "cell_type": "code", "execution_count": 3, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:58:53.962528Z", "start_time": "2024-05-16T13:58:21.105919Z" }, "execution": { "iopub.execute_input": "2024-06-30T22:24:23.770322Z", "iopub.status.busy": "2024-06-30T22:24:23.770186Z", "iopub.status.idle": "2024-06-30T22:24:27.840448Z", "shell.execute_reply": "2024-06-30T22:24:27.839859Z", "shell.execute_reply.started": "2024-06-30T22:24:23.770307Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Preference data shape (110244, 50)\n" ] } ], "source": [ "df = pd.read_csv(\"/home/tony/Data/Preference/13b_v0/interesting_clips_20240623_extend.csv\") # , engine='python')\n", "print(\"Preference data shape\", df.shape)" ] }, { "cell_type": "code", "execution_count": 4, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:58:56.199480Z", "start_time": "2024-05-16T13:58:53.963687Z" }, "execution": { "iopub.execute_input": "2024-06-30T22:24:27.842153Z", "iopub.status.busy": "2024-06-30T22:24:27.841757Z", "iopub.status.idle": "2024-06-30T22:26:17.679003Z", "shell.execute_reply": "2024-06-30T22:26:17.678413Z", "shell.execute_reply.started": "2024-06-30T22:24:27.842134Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "4815452\n", "4815452\n", "pre-downloaded df (110244, 50)\n", "downloaded df (110244, 50)\n" ] } ], "source": [ "converted_paths = os.listdir(NPZ_DIR)\n", "print(len(converted_paths))\n", "\n", "converted_paths = set([f.replace(\".npz\", \"\") for f in converted_paths])\n", "print(len(converted_paths))\n", "\n", "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": 5, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:58:56.467253Z", "start_time": "2024-05-16T13:58:56.207647Z" }, "execution": { "iopub.execute_input": "2024-06-30T22:26:17.679901Z", "iopub.status.busy": "2024-06-30T22:26:17.679744Z", "iopub.status.idle": "2024-06-30T22:26:17.943577Z", "shell.execute_reply": "2024-06-30T22:26:17.943102Z", "shell.execute_reply.started": "2024-06-30T22:26:17.679884Z" } }, "outputs": [ { "data": { "text/plain": [ "is_13b\n", "True 110244\n", "Name: count, dtype: int64" ] }, "execution_count": 5, "metadata": {}, "output_type": "execute_result" } ], "source": [ "df[\"is_13b\"] = df[\"model_name\"].str.contains(\"v3p5\")\n", "df[\"is_13b\"].value_counts()" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "# LET's do the data prep" ] }, { "cell_type": "code", "execution_count": 6, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:58:56.592883Z", "start_time": "2024-05-16T13:58:56.470781Z" }, "execution": { "iopub.execute_input": "2024-06-30T22:26:17.944350Z", "iopub.status.busy": "2024-06-30T22:26:17.944199Z", "iopub.status.idle": "2024-06-30T22:26:17.987633Z", "shell.execute_reply": "2024-06-30T22:26:17.987139Z", "shell.execute_reply.started": "2024-06-30T22:26:17.944333Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "preference model_name \n", "False chirp-v3p5-engine-upload 55908\n", "True chirp-v3p5-engine-upload 54336\n", "Name: count, dtype: int64\n", "(110244, 51)\n", "(110244, 51)\n" ] } ], "source": [ "## for 13b this is easy for now\n", "print(df.groupby([\"preference\"])[\"model_name\"].value_counts())\n", "print(df.shape)\n", "df = df[df[\"model_name\"].isin([\"chirp-v3p5-engine-upload\"])]\n", "print(df.shape)" ] }, { "cell_type": "code", "execution_count": 7, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:58:56.909539Z", "start_time": "2024-05-16T13:58:56.595736Z" }, "execution": { "iopub.execute_input": "2024-06-30T22:26:17.989554Z", "iopub.status.busy": "2024-06-30T22:26:17.989420Z", "iopub.status.idle": "2024-06-30T22:26:18.083800Z", "shell.execute_reply": "2024-06-30T22:26:18.083269Z", "shell.execute_reply.started": "2024-06-30T22:26:17.989538Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "(110244, 51)\n", "(98602, 51)\n", "preference model_name \n", "False chirp-v3p5-engine-upload 49301\n", "True chirp-v3p5-engine-upload 49301\n", "Name: count, dtype: int64\n" ] } ], "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": 8, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:36.043975Z", "start_time": "2024-05-16T13:58:56.910958Z" }, "execution": { "iopub.execute_input": "2024-06-30T22:26:18.084589Z", "iopub.status.busy": "2024-06-30T22:26:18.084438Z", "iopub.status.idle": "2024-06-30T22:26:39.620647Z", "shell.execute_reply": "2024-06-30T22:26:39.620055Z", "shell.execute_reply.started": "2024-06-30T22:26:18.084572Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "unique_requests 49301\n" ] } ], "source": [ "# Let's use the old selection for now -- for quality assurance\n", "# 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": 9, "metadata": { "execution": { "iopub.execute_input": "2024-06-30T22:26:39.621477Z", "iopub.status.busy": "2024-06-30T22:26:39.621316Z", "iopub.status.idle": "2024-06-30T22:26:43.017590Z", "shell.execute_reply": "2024-06-30T22:26:43.017010Z", "shell.execute_reply.started": "2024-06-30T22:26:39.621460Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "pre-downloaded audio_prompt df (98602, 73)\n", "downloaded df (98602, 73)\n" ] } ], "source": [ "print(\"pre-downloaded audio_prompt df\", df.shape)\n", "df[df[\"audio_prompt_id\"].isin(converted_paths)].shape\n", "df = df[df[\"audio_prompt_id\"].isin(converted_paths)].copy()\n", "print(\"downloaded df\", df.shape)" ] }, { "cell_type": "code", "execution_count": 10, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:36.393047Z", "start_time": "2024-05-16T13:59:36.048831Z" }, "execution": { "iopub.execute_input": "2024-06-30T22:26:43.018413Z", "iopub.status.busy": "2024-06-30T22:26:43.018261Z", "iopub.status.idle": "2024-06-30T22:26:43.043526Z", "shell.execute_reply": "2024-06-30T22:26:43.043044Z", "shell.execute_reply.started": "2024-06-30T22:26:43.018396Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "unique_requests 49301\n" ] } ], "source": [ "# GPT requests are also fine for now\n", "print(\"unique_requests\", df[\"request_id\"].nunique())" ] }, { "cell_type": "code", "execution_count": 11, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:40.799375Z", "start_time": "2024-05-16T13:59:36.394236Z" }, "execution": { "iopub.execute_input": "2024-06-30T22:26:43.044272Z", "iopub.status.busy": "2024-06-30T22:26:43.044128Z", "iopub.status.idle": "2024-06-30T22:26:46.175299Z", "shell.execute_reply": "2024-06-30T22:26:46.174704Z", "shell.execute_reply.started": "2024-06-30T22:26:43.044256Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "32415\n", "good_continue_at\n", "True 98602\n", "Name: count, dtype: int64\n", "\n", " Check some basics... \n", " preference\n", "False 49301\n", "True 49301\n", "Name: count, dtype: int64 is_13b\n", "True 98602\n", "Name: count, dtype: int64 model_name\n", "chirp-v3p5-engine-upload 98602\n", "Name: count, dtype: int64 preference model_name \n", "False chirp-v3p5-engine-upload 49301\n", "True chirp-v3p5-engine-upload 49301\n", "Name: count, dtype: int64\n" ] } ], "source": [ "# get the original duration of the clips, if they are concacted\n", "df[\"original_duration_s\"] = df[\"total_start_s\"] + df[\"duration\"]\n", "# 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 most 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())\n", "\n", "\n", "print(\n", " \"\\n Check some basics... \\n\",\n", " df[\"preference\"].value_counts(),\n", " df[\"is_13b\"].value_counts(),\n", " df[\"model_name\"].value_counts(),\n", " df.groupby([\"preference\"])[\"model_name\"].value_counts()\n", ")\n", "\n", "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": 12, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:41.035167Z", "start_time": "2024-05-16T13:59:40.801098Z" }, "execution": { "iopub.execute_input": "2024-06-30T22:26:46.177386Z", "iopub.status.busy": "2024-06-30T22:26:46.176973Z", "iopub.status.idle": "2024-06-30T22:26:46.282760Z", "shell.execute_reply": "2024-06-30T22:26:46.282197Z", "shell.execute_reply.started": "2024-06-30T22:26:46.177368Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "negative 48529 positive 45924\n", "total pair requests 49301 selected pair requests 45704 frac 0.927\n" ] } ], "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\n", " # & (df[\"play_count\"] <= 3) # 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\"] <= 240) # can't be badly long\n", " & (df[\"has_continue_and_start_continue_at\"].isna()) # won't have any continues\n", " # & (df[\"dislike_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[\"reaction_play_count\"] >= 1)\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\"] <= 240) # can't be badly long\n", " & (df[\"dislike_count\"] == 0) # can't have dislikes\n", " & (df[\"flag_count\"] == 0) # can't have issues\n", " # & (\n", " # (\n", " # (df[\"part_of_concat\"] == True)\n", " # & (df[\"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[\"play_count\"] >= normal_pos_play_count)\n", " # )\n", " # )\n", " # & (df[\"user_n_clips\"] >= 20) # user needs to have genereated at least 20\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": 13, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:41.250737Z", "start_time": "2024-05-16T13:59:41.036434Z" }, "execution": { "iopub.execute_input": "2024-06-30T22:26:46.283554Z", "iopub.status.busy": "2024-06-30T22:26:46.283403Z", "iopub.status.idle": "2024-06-30T22:26:46.384935Z", "shell.execute_reply": "2024-06-30T22:26:46.384359Z", "shell.execute_reply.started": "2024-06-30T22:26:46.283537Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "requests 45704 clips 91408 total khrs 4.565; N gpus for 1000 iters 5.713; n unique users 11960\n" ] } ], "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 1000 iters {df_slice.shape[0] / 8 / 2 / 1000:.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": 14, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:41.277006Z", "start_time": "2024-05-16T13:59:41.252105Z" }, "execution": { "iopub.execute_input": "2024-06-30T22:26:46.385768Z", "iopub.status.busy": "2024-06-30T22:26:46.385612Z", "iopub.status.idle": "2024-06-30T22:26:46.407482Z", "shell.execute_reply": "2024-06-30T22:26:46.406990Z", "shell.execute_reply.started": "2024-06-30T22:26:46.385751Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "positive in playlist (3759, 78)\n" ] } ], "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": 15, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:41.323409Z", "start_time": "2024-05-16T13:59:41.278278Z" }, "execution": { "iopub.execute_input": "2024-06-30T22:26:46.408338Z", "iopub.status.busy": "2024-06-30T22:26:46.408196Z", "iopub.status.idle": "2024-06-30T22:26:46.433713Z", "shell.execute_reply": "2024-06-30T22:26:46.433306Z", "shell.execute_reply.started": "2024-06-30T22:26:46.408321Z" } }, "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) # & (df_slice[\"dislike_count\"] < 1)\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": 16, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:41.392244Z", "start_time": "2024-05-16T13:59:41.324472Z" }, "execution": { "iopub.execute_input": "2024-06-30T22:26:46.434399Z", "iopub.status.busy": "2024-06-30T22:26:46.434254Z", "iopub.status.idle": "2024-06-30T22:26:46.469137Z", "shell.execute_reply": "2024-06-30T22:26:46.468724Z", "shell.execute_reply.started": "2024-06-30T22:26:46.434383Z" } }, "outputs": [], "source": [ "# save positive ids\n", "# positive_preference_ids = df_slice[df_slice[\"preference\"] == False][\"s3_id\"].to_json(orient='values')\n", "# with open('/home/tony/Data/Preference/7b_v2/7v_v20_full_recut_id_negative.json', 'w') as file:\n", "# file.write(positive_preference_ids)" ] }, { "cell_type": "code", "execution_count": 17, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T14:00:20.866354Z", "start_time": "2024-05-16T14:00:12.443344Z" }, "execution": { "iopub.execute_input": "2024-06-30T22:26:46.469815Z", "iopub.status.busy": "2024-06-30T22:26:46.469682Z", "iopub.status.idle": "2024-06-30T22:26:46.502690Z", "shell.execute_reply": "2024-06-30T22:26:46.502277Z", "shell.execute_reply.started": "2024-06-30T22:26:46.469800Z" } }, "outputs": [], "source": [ "# df_slice.to_csv(\"/home/tony/Data/Preference/7b_v0/interesting_clips_v3_processed.csv\")" ] }, { "cell_type": "code", "execution_count": 18, "metadata": { "execution": { "iopub.execute_input": "2024-06-30T22:26:46.503356Z", "iopub.status.busy": "2024-06-30T22:26:46.503216Z", "iopub.status.idle": "2024-06-30T22:26:46.538713Z", "shell.execute_reply": "2024-06-30T22:26:46.538303Z", "shell.execute_reply.started": "2024-06-30T22:26:46.503341Z" } }, "outputs": [], "source": [ "# df_slice_2 = pd.read_csv(\"/home/tony/Data/Preference/13b_v0/interesting_clips_20240623_slice.csv\")\n", "# print(df_slice_2.shape)\n", "# df_slice_mix = pd.concat([df_slice, df_slice_2])\n", "# print(df_slice_mix.shape)\n", "# df_slice = df_slice_mix.copy()" ] }, { "cell_type": "code", "execution_count": 19, "metadata": { "execution": { "iopub.execute_input": "2024-06-30T22:26:46.539513Z", "iopub.status.busy": "2024-06-30T22:26:46.539259Z", "iopub.status.idle": "2024-06-30T22:26:57.545880Z", "shell.execute_reply": "2024-06-30T22:26:57.545286Z", "shell.execute_reply.started": "2024-06-30T22:26:46.539496Z" } }, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "/tmp/ipykernel_2956732/1464495589.py:2: DtypeWarning: Columns (72,73) have mixed types. Specify dtype option on import or set low_memory=False.\n", " df_slice = pd.read_csv(\"/home/tony/Data/Preference/13b_v0/interesting_clips_20240624_mix.csv\")\n" ] } ], "source": [ "# df_slice.to_csv(\"/home/tony/Data/Preference/13b_v0/interesting_clips_20240624_mix.csv\")\n", "df_slice = pd.read_csv(\"/home/tony/Data/Preference/13b_v0/interesting_clips_20240624_mix.csv\")" ] }, { "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": 4, "metadata": { "execution": { "iopub.execute_input": "2024-06-24T05:06:06.690061Z", "iopub.status.busy": "2024-06-24T05:06:06.689893Z", "iopub.status.idle": "2024-06-24T05:06:06.709106Z", "shell.execute_reply": "2024-06-24T05:06:06.708672Z", "shell.execute_reply.started": "2024-06-24T05:06:06.690044Z" } }, "outputs": [ { "data": { "text/plain": [ "0 NaN\n", "1 NaN\n", "2 17.84\n", "3 17.84\n", "4 9.80\n", " ... \n", "197115 NaN\n", "197116 8.00\n", "197117 8.00\n", "197118 NaN\n", "197119 NaN\n", "Name: continue_at, Length: 197120, dtype: float64" ] }, "execution_count": 4, "metadata": {}, "output_type": "execute_result" } ], "source": [ "df_slice[\"continue_at\"]" ] }, { "cell_type": "code", "execution_count": 5, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:41.932296Z", "start_time": "2024-05-16T13:59:41.932287Z" }, "execution": { "iopub.execute_input": "2024-06-24T05:06:06.709952Z", "iopub.status.busy": "2024-06-24T05:06:06.709803Z", "iopub.status.idle": "2024-06-24T05:06:06.795797Z", "shell.execute_reply": "2024-06-24T05:06:06.795349Z", "shell.execute_reply.started": "2024-06-24T05:06:06.709936Z" } }, "outputs": [ { "data": { "text/plain": [ "47221" ] }, "execution_count": 5, "metadata": {}, "output_type": "execute_result" } ], "source": [ "# don't have continue at\n", "df_slice[df_slice[\"continue_at\"].isna()][\"request_id\"].nunique()" ] }, { "cell_type": "code", "execution_count": 6, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:41.932966Z", "start_time": "2024-05-16T13:59:41.932957Z" }, "execution": { "iopub.execute_input": "2024-06-24T05:06:06.797050Z", "iopub.status.busy": "2024-06-24T05:06:06.796893Z", "iopub.status.idle": "2024-06-24T05:06:06.826505Z", "shell.execute_reply": "2024-06-24T05:06:06.826042Z", "shell.execute_reply.started": "2024-06-24T05:06:06.797034Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "98560\n" ] } ], "source": [ "final_filtered_requests = df_slice[\"request_id\"].unique()\n", "print(len(final_filtered_requests))" ] }, { "cell_type": "code", "execution_count": 7, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:41.933558Z", "start_time": "2024-05-16T13:59:41.933550Z" }, "execution": { "iopub.execute_input": "2024-06-24T05:06:06.827229Z", "iopub.status.busy": "2024-06-24T05:06:06.827088Z", "iopub.status.idle": "2024-06-24T05:06:06.854942Z", "shell.execute_reply": "2024-06-24T05:06:06.854530Z", "shell.execute_reply.started": "2024-06-24T05:06:06.827214Z" } }, "outputs": [], "source": [ "# df_slice.to_csv(\"/home/tony/Data/Preference/7b_v2/7b_before_recode_20240412\", index=False)" ] }, { "cell_type": "code", "execution_count": 8, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:41.934277Z", "start_time": "2024-05-16T13:59:41.934268Z" }, "execution": { "iopub.execute_input": "2024-06-24T05:06:06.855662Z", "iopub.status.busy": "2024-06-24T05:06:06.855523Z", "iopub.status.idle": "2024-06-24T05:06:07.392627Z", "shell.execute_reply": "2024-06-24T05:06:07.392065Z", "shell.execute_reply.started": "2024-06-24T05:06:06.855648Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "97574 986\n", "(195148, 83) (1972, 83)\n" ] } ], "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", "df_slice = df_slice.sort_values(by=[\"request_id\", \"preference\"])\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": "markdown", "metadata": {}, "source": [ "# Actually make" ] }, { "cell_type": "code", "execution_count": 9, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:41.934954Z", "start_time": "2024-05-16T13:59:41.934946Z" }, "execution": { "iopub.execute_input": "2024-06-24T05:06:07.393450Z", "iopub.status.busy": "2024-06-24T05:06:07.393298Z", "iopub.status.idle": "2024-06-24T05:06:07.408232Z", "shell.execute_reply": "2024-06-24T05:06:07.407788Z", "shell.execute_reply.started": "2024-06-24T05:06:07.393433Z" } }, "outputs": [], "source": [ "# val_df[[\"request_id\", \"metadata\", \"updated_at\", \"user_id\", \"preference\"]].head()" ] }, { "cell_type": "code", "execution_count": 10, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:41.935620Z", "start_time": "2024-05-16T13:59:41.935613Z" }, "execution": { "iopub.execute_input": "2024-06-24T05:06:07.408966Z", "iopub.status.busy": "2024-06-24T05:06:07.408822Z", "iopub.status.idle": "2024-06-24T05:06:13.247977Z", "shell.execute_reply": "2024-06-24T05:06:13.247339Z", "shell.execute_reply.started": "2024-06-24T05:06:07.408950Z" } }, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████████████████████████████| 195148/195148 [00:05<00:00, 33662.09it/s]" ] }, { "name": "stdout", "output_type": "stream", "text": [ "9,988 hours of 195148 clips, 17.423928571428572 nodes\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "\n" ] } ], "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 / 2 / 700} nodes\")" ] }, { "cell_type": "code", "execution_count": 11, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:41.936268Z", "start_time": "2024-05-16T13:59:41.936260Z" }, "execution": { "iopub.execute_input": "2024-06-24T05:06:13.248915Z", "iopub.status.busy": "2024-06-24T05:06:13.248755Z", "iopub.status.idle": "2024-06-24T05:06:41.315731Z", "shell.execute_reply": "2024-06-24T05:06:41.315174Z", "shell.execute_reply.started": "2024-06-24T05:06:13.248899Z" } }, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████| 1972/1972 [00:28<00:00, 70.38it/s]" ] }, { "name": "stdout", "output_type": "stream", "text": [ "Total 1972 clips\n", "53 hours of False\n", "54 hours of True\n", "Done\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "\n" ] } ], "source": [ "make_dataset(val_df, OUT_DATA_DIR, is_val=True, npz_dir=NPZ_DIR)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:41.936964Z", "start_time": "2024-05-16T13:59:41.936957Z" }, "execution": { "iopub.execute_input": "2024-06-24T05:06:41.317623Z", "iopub.status.busy": "2024-06-24T05:06:41.317285Z" } }, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ " 15%|███████████████▏ | 28482/195148 [06:28<39:00, 71.20it/s]" ] } ], "source": [ "make_dataset(train_df, OUT_DATA_DIR, is_val=False, npz_dir=NPZ_DIR)" ] }, { "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": "2024-05-16T13:59:41.937879Z", "start_time": "2024-05-16T13:59:41.937870Z" } }, "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, 6016, 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": "2024-05-16T13:59:41.938629Z", "start_time": "2024-05-16T13:59:41.938621Z" } }, "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\"] = \"0\"\n", "# _ = preload_codec_models(\"/app/suno/data/dpo/models/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": "2024-05-16T13:59:41.939205Z", "start_time": "2024-05-16T13:59:41.939198Z" } }, "outputs": [], "source": [ "# import random\n", "# 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(\"\\n negative example \\n\", test_metas[idx])\n", "# a.play(compress=False)\n", "# pos_a = decode(pos_arr)\n", "# print(\"\\n positive example \\n\", 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": "2024-05-16T13:59:41.939977Z", "start_time": "2024-05-16T13:59:41.939969Z" } }, "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": "2024-05-16T13:59:41.940610Z", "start_time": "2024-05-16T13:59:41.940603Z" } }, "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": "2024-05-16T13:59:41.941167Z", "start_time": "2024-05-16T13:59:41.941159Z" } }, "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": "2024-05-16T13:59:41.941801Z", "start_time": "2024-05-16T13:59:41.941793Z" } }, "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", "validation_on_metas(test_metas)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:41.942520Z", "start_time": "2024-05-16T13:59:41.942511Z" } }, "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": "2024-05-16T13:59:41.943072Z", "start_time": "2024-05-16T13:59:41.943065Z" } }, "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": "2024-05-16T13:59:41.944246Z", "start_time": "2024-05-16T13:59:41.944237Z" } }, "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": "2024-05-16T13:59:41.945249Z", "start_time": "2024-05-16T13:59:41.945241Z" } }, "outputs": [], "source": [ "print(\"1 epoch per batch 4, total\", total_iters / 8 / 8 / 2)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:41.945972Z", "start_time": "2024-05-16T13:59:41.945964Z" } }, "outputs": [], "source": [ "!cd /home/tony/Work/tony/slurm/extend/ && sbatch sbatch_ipo_13b_extend" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "# some gymathtics loading prev data" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "ExecuteTime": { "end_time": "2024-05-16T13:59:41.946562Z", "start_time": "2024-05-16T13:59:41.946555Z" } }, "outputs": [], "source": [ "# prev_v3_data = \"/app/suno/data/dpo/7v_v20_full/\"\n", "\n", "# test_val_metas = read_jsonl(os.path.join(prev_v3_data, f\"meta_val.jsonl\"))\n", "# test_tr_metas = read_jsonl(os.path.join(prev_v3_data, f\"meta_tr.jsonl\"))\n", "\n", "# all_ids = set()\n", "# for meta in test_val_metas:\n", "# all_ids.add(meta[\"id\"])\n", "# for meta in test_tr_metas:\n", "# all_ids.add(meta[\"id\"])\n", "# print(len(all_ids), len(test_val_metas) + len(test_tr_metas))\n", "\n", "# all_ids = list(all_ids)\n", "# with open(\"/home/tony/Data/Preference/7b_v2/7v_v20_full_recut_id.json\", \"w\") as fp:\n", "# json.dump(all_ids, fp)" ] }, { "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": "3.10.14" }, "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": 4 }