{ "cells": [ { "cell_type": "code", "execution_count": null, "id": "b686543b", "metadata": {}, "outputs": [], "source": [ "import pandas as pd\n", "\n", "bluejay_filepath = \"/home/tony/Data/Preference/bluejay_t1/interesting_clips_bluejay_t1_20250811_public_only.pkl\"\n", "crow_filepath = \"/home/tony/Data/Preference/crow_t1/interesting_clips_crow_t1_20251020_public_only.pkl\"\n", "bluejay_df = pd.read_pickle(bluejay_filepath)\n", "print(len(bluejay_df))\n", "crow_df = pd.read_pickle(crow_filepath)\n", "print(len(crow_df))\n" ] }, { "cell_type": "code", "execution_count": null, "id": "1a5a4d71", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "4166823a", "metadata": {}, "outputs": [], "source": [ "print(len(discogs_title_terms_map))" ] }, { "cell_type": "code", "execution_count": null, "id": "7aedafe7", "metadata": {}, "outputs": [], "source": [ "# merge bluejay_df and crow_df but track the source, avoiding dataframe fragmentation\n", "bluejay_df_copy = bluejay_df.copy()\n", "crow_df_copy = crow_df.copy()\n", "bluejay_df_copy[\"model\"] = \"bluejay\"\n", "crow_df_copy[\"model\"] = \"crow\"\n", "df = pd.concat([bluejay_df_copy, crow_df_copy], ignore_index=True)\n", "print(len(df))\n" ] }, { "cell_type": "code", "execution_count": null, "id": "b1674e18", "metadata": {}, "outputs": [], "source": [ "# create a list of all prompts \n", "tags_df = df[\"metadata\"].apply(lambda x: x.get(\"tags\", \"\"))" ] }, { "cell_type": "code", "execution_count": null, "id": "556e7472", "metadata": {}, "outputs": [], "source": [ "# print the index of the df where tags are empty string\n", "empty_tags_indices = df[df[\"metadata\"].apply(lambda x: x.get(\"tags\", \"\") == \"\")].index\n", "print(empty_tags_indices)\n" ] }, { "cell_type": "code", "execution_count": null, "id": "fb06d972", "metadata": {}, "outputs": [], "source": [ "df.iloc[27]" ] }, { "cell_type": "code", "execution_count": null, "id": "55ec9466", "metadata": {}, "outputs": [], "source": [ "# for each row, split the string by comma and strip whitespace, but handle None safely\n", "tags = tags_df.apply(lambda x: [t.strip() for t in x.split(\",\")] if isinstance(x, str) else [])\n", "# flatten the list\n", "#\n", "# tags = tags.explode()\n", "# remove duplicates\n", "#tags = tags.unique()\n", "print(len(tags))" ] }, { "cell_type": "code", "execution_count": null, "id": "182a7ede", "metadata": {}, "outputs": [], "source": [ "import numpy as np\n", "rand_idx = np.random.randint(0, len(tags_df)) \n", "print(tags_list_sorted[rand_idx])" ] }, { "cell_type": "code", "execution_count": null, "id": "47d71ef0", "metadata": {}, "outputs": [], "source": [ "# sort tags_df by length of tags, handling None safely\n", "tags_list = tags_df.tolist()\n", "# filter out None to avoid TypeError\n", "tags_list_non_none = [x for x in tags_list if x is not None]\n", "# sort the non-None tags_list by length\n", "tags_list_sorted = sorted(tags_list_non_none, key=lambda x: len(x), reverse=False)\n" ] }, { "cell_type": "code", "execution_count": null, "id": "ddf4a27b", "metadata": {}, "outputs": [], "source": [ "# Create separate tags lists for model == \"crow\" and model == \"bluejay\"\n", "tags_df_crow = df[df[\"model\"] == \"crow\"][\"metadata\"].apply(lambda x: x.get(\"tags\", \"\"))\n", "tags_df_bluejay = df[df[\"model\"] == \"bluejay\"][\"metadata\"].apply(lambda x: x.get(\"tags\", \"\"))\n", "\n", "# Process crow tags\n", "tags_list_crow = tags_df_crow.tolist()\n", "tags_list_crow_non_none = [x for x in tags_list_crow if x is not None]\n", "tags_list_crow_sorted = sorted(tags_list_crow_non_none, key=lambda x: len(x), reverse=False)\n", "\n", "# Process bluejay tags\n", "tags_list_bluejay = tags_df_bluejay.tolist()\n", "tags_list_bluejay_non_none = [x for x in tags_list_bluejay if x is not None]\n", "tags_list_bluejay_sorted = sorted(tags_list_bluejay_non_none, key=lambda x: len(x), reverse=False)\n", "\n", "# Plot both as histograms of tag string lengths\n", "import matplotlib.pyplot as plt\n", "\n", "tags_list_lens_crow = [len(x) for x in tags_list_crow_sorted]\n", "tags_list_lens_bluejay = [len(x) for x in tags_list_bluejay_sorted]\n", "\n", "plt.hist(tags_list_lens_crow, bins=100, alpha=0.5, label=\"crow\", density=True)\n", "plt.hist(tags_list_lens_bluejay, bins=100, alpha=0.5, label=\"bluejay\", density=True)\n", "plt.legend()\n", "plt.title(\"Distribution of tag string lengths by model\")\n", "plt.xlabel(\"Tag string length\")\n", "plt.ylabel(\"Count\")\n", "plt.show()" ] }, { "cell_type": "code", "execution_count": null, "id": "a6fc5985", "metadata": {}, "outputs": [], "source": [ "tags_list_lens = [len(x) for x in tags_list_sorted]\n", "print(tags_list_lens)\n", "# plot the distribution of tags_list_lens\n", "plt.hist(tags_list_lens, bins=100)\n", "plt.show()\n", "\n" ] }, { "cell_type": "code", "execution_count": null, "id": "22ba2aa0", "metadata": {}, "outputs": [], "source": [ "import matplotlib.pyplot as plt\n", "# for each row, count the number of tags\n", "tags_count = tags_df.apply(lambda x: len(x) if x is not None else 0)\n", "# make a histogram of the number of tags per row\n", "#tags_count.value_counts().plot(kind=\"bar\")\n", "#plt.show()\n", "# sort from highest to lowest\n", "tags_count = tags_count.sort_values(ascending=False)\n", "# get the top 100 tags\n", "tags_count = tags_count.head(100)\n", "print(tags_count)\n" ] }, { "cell_type": "code", "execution_count": null, "id": "acfc446a", "metadata": {}, "outputs": [], "source": [ "# plot distrubution of reaction_play_count\n", "import matplotlib.pyplot as plt\n", "\n", "# filter to those with reaction_play_count > 10\n", "df = df[(df[\"reaction_play_count\"] > 5)]\n", "df = df[df[\"upvoted\"] == True]\n", "#df = df[df[\"is_pro_user\"] == True]\n", "df = df[df[\"deleted\"] == False]\n", "#df = df[df[\"source\"] == \"web\"]\n", "df = df[df[\"user_n_clips\"] >= 5]\n", "df = df[df[\"flag_count\"] == 0]\n", "df = df[df[\"is_public\"] == True]\n", "df = df[df[\"task\"].isin([\"cover\", \"\", \"artist_consistency\", \"artist_cover\", \"playlist_condition\"])]\n", "df = df[df[\"duration\"] > 60]\n", "print(len(df))\n", "\n", "# and ensure that we have a balanced number of clips per user, \n", "# for example, ensure we have no more than 3 clips per user\n", "df = df.groupby(\"user_id\").filter(lambda x: len(x) <= 100)\n", "print(len(df))\n" ] }, { "cell_type": "code", "execution_count": null, "id": "a1ac3d28", "metadata": {}, "outputs": [], "source": [ "df_subset = df.copy()" ] }, { "cell_type": "code", "execution_count": null, "id": "078df4c4", "metadata": {}, "outputs": [], "source": [ "# Sample a total of 200 clips: half English, half balanced among foreign languages\n", "\n", "total_clips = 50_000\n", "n_english = total_clips // 2\n", "n_foreign = total_clips - n_english\n", "\n", "# Get language for each row\n", "df[\"lang\"] = df[\"metadata\"].apply(lambda x: x.get(\"lang\", \"\"))\n", "\n", "# English clips\n", "df_en = df[df[\"lang\"] == \"English\"]\n", "if len(df_en) >= n_english:\n", " df_en_sample = df_en.sample(n_english, random_state=42)\n", "else:\n", " df_en_sample = df_en\n", "\n", "# Foreign languages\n", "foreign_langs = [l for l in df[\"lang\"].unique() if l != \"English\"]\n", "n_langs = len(foreign_langs)\n", "clips_per_lang = n_foreign // n_langs if n_langs > 0 else 0\n", "remainder = n_foreign % n_langs if n_langs > 0 else 0\n", "\n", "dfs = []\n", "for i, lang in enumerate(foreign_langs):\n", " n = clips_per_lang + (1 if i < remainder else 0)\n", " df_lang = df[df[\"lang\"] == lang]\n", " if len(df_lang) >= n:\n", " dfs.append(df_lang.sample(n, random_state=42))\n", " else:\n", " dfs.append(df_lang)\n", "\n", "df_foreign_sample = pd.concat(dfs) if dfs else pd.DataFrame(columns=df.columns)\n", "\n", "# Combine and shuffle\n", "df_subset = pd.concat([df_en_sample, df_foreign_sample]).sample(frac=1, random_state=42).reset_index(drop=True)" ] }, { "cell_type": "code", "execution_count": null, "id": "8c079e1e", "metadata": {}, "outputs": [], "source": [ "# plot the distribution of lang\n", "lang_counts = df_subset[\"metadata\"].apply(lambda x: x.get(\"lang\", \"\")).value_counts()\n", "plt.figure(figsize=(10, 5))\n", "lang_counts.plot(kind=\"bar\")\n", "plt.xlabel(\"Language\")\n", "plt.ylabel(\"Number of Clips\")\n", "plt.title(\"Distribution of Clips by Language\")\n", "plt.show()\n" ] }, { "cell_type": "code", "execution_count": null, "id": "76c7724f", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "f934f73f", "metadata": {}, "outputs": [], "source": [ "print(len(df_subset))\n", "# crop to exactly 30_000 rows\n", "df_subset = df_subset.sample(100_000, random_state=42)\n", "print(len(df_subset))\n" ] }, { "cell_type": "code", "execution_count": null, "id": "f8ff9282", "metadata": {}, "outputs": [], "source": [ "# now i need to create a jsonl file that has \n", "# text, tags, clip_ids, user_id, \n", "from suno_utils.utils.text import write_jsonl\n", "\n", "# iterate over the dataframe and create a jsonl file\n", "metas = []\n", "i = 0\n", "for index, row in df_subset.iterrows():\n", " meta = {\n", " \"text\": row[\"prompt_text\"],\n", " \"tags\": row[\"metadata\"].get(\"tags\", \"\"),\n", " \"id\": row[\"s3_id\"],\n", " \"user_id\": row[\"user_id\"],\n", " \"lang\" : row[\"metadata\"].get(\"lang\", \"\"),\n", " \"duration\": row[\"duration\"],\n", " }\n", " metas.append(meta)\n", "\n", "write_jsonl(metas, \"/home/christian/code/christian/metadata/sft/interesting_clips_bluejay_t1_20250811_public_only_100k.jsonl\")" ] }, { "cell_type": "code", "execution_count": null, "id": "bb15cc7e", "metadata": {}, "outputs": [], "source": [ "from suno_utils.utils.text import read_jsonl\n", "work_items = read_jsonl(\n", " \"/home/christian/code/christian/metadata/discogs_subset_sampled_metas.jsonl\"\n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "id": "10f0ec44", "metadata": {}, "outputs": [], "source": [ "print(len(work_items))" ] }, { "cell_type": "code", "execution_count": null, "id": "38707f6d", "metadata": {}, "outputs": [], "source": [ "# how many work items have \"text\" that is non-empty\n", "# how many work items have tags list that is non-empty\n", "text_count = 0\n", "tags_count = 0\n", "for item in work_items:\n", " if item[\"text\"] != \"\":\n", " text_count += 1\n", " \n", " if len(item[\"tags\"]) > 0:\n", " tags_count += 1\n", "print(text_count)\n", "print(tags_count)\n", " \n" ] }, { "cell_type": "code", "execution_count": null, "id": "6919d6a7", "metadata": {}, "outputs": [], "source": [ "# test some of the local data\n", "CODEC_FILEPATH = \"s3://suno-data/minz/models/dac_vae_tuned_25hz.pth\"\n", "\n", "from suno_utils.tasks.dac_vae_fixed_25hz import (\n", " preload_models as preload_codec_models,\n", " decode as codec_decode,\n", " encode as codec_encode,\n", " decode_stream_to_full_audio,\n", ")\n", "\n", "_ = preload_codec_models(CODEC_FILEPATH)\n" ] }, { "cell_type": "code", "execution_count": null, "id": "59d54fe6", "metadata": {}, "outputs": [], "source": [ "import os\n", "source_dir = \"/app2/suno/data/christian/outputs/v3-base-data-ctx-discogs-subset-t0\"\n", "\n", "# get all the dirs in the source_dir\n", "dirs = os.listdir(source_dir)\n", "\n", "# get the dirs that start with \"2025-09-29\"\n", "dirs = [d for d in dirs if os.path.isdir(os.path.join(source_dir, d))]\n", "print(dirs)\n" ] }, { "cell_type": "code", "execution_count": null, "id": "5942cdcf", "metadata": {}, "outputs": [], "source": [ "import numpy as np\n", "dir_idx = 1500\n", "dir_path = os.path.join(source_dir, dirs[dir_idx])\n", "model_name = \"16n_25hz_v45_infill_shared_flow_resume_1_75m\"\n", "\n", "# get all the files in the dir\n", "files = os.listdir(dir_path)\n", "# get the upsample vae and the origina_vae\n", "upsample_vae_filepath = os.path.join(dir_path, f\"{dirs[dir_idx]}_{model_name}_0_upsampled_vae.npz\")\n", "original_vae_filepath = os.path.join(dir_path, f\"{dirs[dir_idx]}_original_vae.npz\")\n", "\n", "upsample_vae = np.load(upsample_vae_filepath)[\"vae_latents\"]\n", "original_vae = np.load(original_vae_filepath)[\"vae_latents\"]\n", "\n", "print(upsample_vae.shape)\n", "print(original_vae.shape)\n", "\n", "# decode both\n", "print(\"upsample\")\n", "upsample_audio = codec_decode(upsample_vae[:750]).play()\n", "print(\"original\")\n", "original_audio = codec_decode(original_vae[:750]).play()\n", "\n", "\n", "\n" ] }, { "cell_type": "code", "execution_count": null, "id": "ce9472c5", "metadata": {}, "outputs": [], "source": [] } ], "metadata": { "kernelspec": { "display_name": "suno_diff", "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": "3.12.9" } }, "nbformat": 4, "nbformat_minor": 5 }