{ "cells": [ { "cell_type": "code", "execution_count": 7, "id": "taken-budapest", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Populating the interactive namespace from numpy and matplotlib\n" ] } ], "source": [ "%pylab inline" ] }, { "cell_type": "code", "execution_count": 8, "id": "ongoing-cooper", "metadata": {}, "outputs": [], "source": [ "import os\n", "os.environ['CUDA_VISIBLE_DEVICES'] = ''" ] }, { "cell_type": "code", "execution_count": 9, "id": "creative-temple", "metadata": {}, "outputs": [], "source": [ "import re\n", "import pandas as pd\n", "import numpy as np\n", "import librosa\n", "from scipy.io import wavfile\n", "import librosa.display\n", "from pydub import AudioSegment\n", "from IPython.display import display_html\n", "import torch\n", "\n", "\n", "def _convert_array(arr, force_int=False, force_float=False):\n", " if isinstance(arr, torch.Tensor):\n", " arr = arr.cpu().detach().numpy()\n", " assert(len(arr.shape) in (1, 2))\n", " if len(arr.shape) == 2:\n", " assert(arr.shape[0] == 1)\n", " arr = arr[0]\n", " if force_int and \"float\" in str(arr.dtype):\n", " arr = np.floor(arr.copy() * 32768).astype(np.int16)\n", " if force_float and \"int\" in str(arr.dtype):\n", " arr = arr.copy().astype(np.float32) / 32768\n", " return arr\n", "\n", "\n", "def plot_ft(arr, window_size=1024, hop_length=512, n_mels=128):\n", " arr = _convert_array(arr, force_float=True)\n", " window = np.hanning(window_size)\n", " stft= librosa.core.spectrum.stft(arr, n_fft=window_size, hop_length=hop_length, window=window)\n", " out = 2 * np.abs(stft) / np.sum(window)\n", " data = librosa.amplitude_to_db(out, ref=np.max)\n", "\n", " plt.figure(figsize=(16, 10))\n", " librosa.display.specshow(data, x_axis='time', y_axis='mel', sr=16000)\n", " plt.set_cmap('afmhot')\n", "\n", "\n", "def play_array(arr):\n", " arr = _convert_array(arr, force_int=True)\n", " display_html(AudioSegment(\n", " arr.tobytes(), \n", " frame_rate=16000,\n", " sample_width=2, \n", " channels=1\n", " ))" ] }, { "cell_type": "code", "execution_count": null, "id": "heavy-fraction", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "treated-bunny", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "intelligent-christmas", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "egyptian-arcade", "metadata": {}, "source": [ "## Training" ] }, { "cell_type": "code", "execution_count": null, "id": "regular-thumb", "metadata": {}, "outputs": [], "source": [ "cd /data-ssd-3/georg/code/denoiser\n", "\n", "# finetune\n", "./train.py continue_pretrained=dns48 ddp=1\n", "\n", "# from scratch\n", "./train.py ddp=1 dummy=onus\n", "\n", "# CUDA_VISIBLE_DEVICES=4,5,6,7 \n", "# 500h hours of train\n", "# ~10m per epoch for small\n", "# ~2.2h per epoch for large\n", "\n", "# large: hidden=64, stride=2, resample=2, causal=False\n", "\n", "# epoch time and first val loss\n", "# (10 mins -- base, but higher batch) - 0.00983\n", "\n", "# (20 mins -- 1/3 gpu memory, nc) - 0.00883\n", "# (30 mins -- 1/2 gpu memory, h=64, nc) - 0.00863\n", "# (1.5h -- 4/5 gpu memory, rs=2, st2=2, nc) - 0.00832\n", "# (25 mins -- 1/4 gpu memory, h=64, rs=2, nc) - 0.00906\n", "# (2.2h -- 1 gpu memory, h=64, rs=2, st=2, nc) - 0.00822" ] }, { "cell_type": "code", "execution_count": null, "id": "shaped-dance", "metadata": {}, "outputs": [], "source": [ "# stft\n", "# 2.8h -- 1.2 gpu memory, h=64, rs=2, st=2, nc" ] }, { "cell_type": "code", "execution_count": null, "id": "monthly-boston", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "social-speaking", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "unsigned-lawyer", "metadata": {}, "source": [ "## Plot metrics" ] }, { "cell_type": "code", "execution_count": 10, "id": "higher-dover", "metadata": {}, "outputs": [], "source": [ "LOG_DIR = \"/data-ssd-3/georg/code/denoiser/outputs/\"\n", "\n", "def get_metric_df(exp_name):\n", " metrics_df = pd.read_json(LOG_DIR + exp_name + \"/history.json\")\n", " return metrics_df" ] }, { "cell_type": "code", "execution_count": 15, "id": "systematic-reliance", "metadata": {}, "outputs": [], "source": [ "# 25m per epoch on 4 gpu for default\n", "data_map = {\n", "# \"only_amr\": \"exp_dummy=base\", # data with just amr\n", " \"base\": \"exp_dummy=new_data\",\n", "# \"lg\": \"exp_demucs.depth=6\",\n", " \"only_stft\": \"exp_dummy=only_spec\",\n", "# \"only_stft_narrower\": \"exp_dummy=only_spec_narrower\",\n", "}\n", "\n", "df_container = {k: get_metric_df(v) for k, v in data_map.items()}" ] }, { "cell_type": "code", "execution_count": 16, "id": "dedicated-pacific", "metadata": {}, "outputs": [ { "data": { "image/png": "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\n", "text/plain": [ "
" ] }, "metadata": { "needs_background": "light" }, "output_type": "display_data" } ], "source": [ "plt.figure(figsize=(12, 4))\n", "\n", "plt.subplot(1, 2, 1)\n", "for k, df in df_container.items():\n", " df[\"train\"].plot(marker='.', markersize=12, label=k)\n", "plt.title(\"tr_loss\")\n", "plt.legend();\n", "\n", "plt.subplot(1, 2, 2)\n", "for k, df in df_container.items():\n", " df[\"valid\"].plot(marker='.', markersize=12, label=k)\n", "plt.title(\"val_loss\")\n", "plt.legend();" ] }, { "cell_type": "code", "execution_count": 17, "id": "public-environment", "metadata": {}, "outputs": [], "source": [ "data_map = {\n", "# \"only_amr\": \"exp_dummy=base\", # data with just amr\n", " \"base\": \"exp_dummy=new_data\",\n", "# \"lg\": \"exp_demucs.depth=6\",\n", " \"only_stft\": \"exp_dummy=only_spec\",\n", "# \"only_stft_narrower\": \"exp_dummy=only_spec_narrower\",\n", "}\n", "\n", "df_container = {k: get_metric_df(v) for k, v in data_map.items()}" ] }, { "cell_type": "code", "execution_count": 18, "id": "insured-alloy", "metadata": {}, "outputs": [ { "data": { "image/png": "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\n", "text/plain": [ "
" ] }, "metadata": { "needs_background": "light" }, "output_type": "display_data" } ], "source": [ "plt.figure(figsize=(12, 4))\n", "\n", "# pesq: 0=very bad, 5=very good\n", "plt.subplot(1, 2, 1)\n", "# ref_pesq = 2.06\n", "ref_pesq = 1.68 # with packet loss\n", "max_epoch = max([df.index.max() for _, df in df_container.items()])\n", "plt.plot([1, max_epoch], [ref_pesq, ref_pesq], '--k', alpha=0.6)\n", "for k, df in df_container.items():\n", " df[\"pesq\"].dropna().plot(marker='.', markersize=12, label=k)\n", "plt.title(\"pesq\")\n", "plt.legend();\n", "\n", "# stoi: 0=very bad, 1=very good\n", "plt.subplot(1, 2, 2)\n", "ref_stoi = 0.86\n", "# ref_stoi = 0.85 # with packet loss\n", "max_epoch = max([df.index.max() for _, df in df_container.items()])\n", "plt.plot([1, max_epoch], [ref_stoi, ref_stoi], '--k', alpha=0.6)\n", "for k, df in df_container.items():\n", " df[\"stoi\"].dropna().plot(marker='.', markersize=12, label=k)\n", "plt.title(\"stoi\")\n", "plt.legend();" ] }, { "cell_type": "code", "execution_count": null, "id": "planned-twins", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "accepting-content", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "junior-cloud", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "entertaining-spending", "metadata": {}, "source": [ "## test checkpoint" ] }, { "cell_type": "code", "execution_count": 1, "id": "complete-proof", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "################################################################################\n", "### WARNING, path does not exist: KALDI_ROOT=/mnt/matylda5/iveselyk/Tools/kaldi-trunk\n", "### (please add 'export KALDI_ROOT=' in your $HOME/.profile)\n", "### (or run as: KALDI_ROOT= python .py)\n", "################################################################################\n", "\n" ] } ], "source": [ "from denoiser import pretrained\n", "import numpy as np\n", "import torchaudio\n", "from denoiser.demucs import Demucs\n", "import torch\n", "from pydub import AudioSegment\n", "from scipy.io import wavfile\n", "from IPython.display import display_html" ] }, { "cell_type": "code", "execution_count": 2, "id": "bottom-words", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "34.2M model params\n" ] } ], "source": [ "# model = Demucs(hidden=48)\n", "model = Demucs(hidden=48, causal=False)#, depth=6)\n", "# model = Demucs(hidden=64, stride=2, resample=2, causal=False)\n", "\n", "n_params = np.sum([e.numel() for e in model.state_dict().values()]) / 1e6\n", "print(\"{}M model params\".format(round(n_params, 1)))" ] }, { "cell_type": "code", "execution_count": 3, "id": "interested-sponsorship", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "" ] }, "execution_count": 3, "metadata": {}, "output_type": "execute_result" } ], "source": [ "# cpp = \"/data-ssd-3/georg/code/denoiser/outputs/exp_dummy=base/best.th\"\n", "cpp = \"/data-ssd-3/georg/code/denoiser/outputs/exp_dummy=new_data/best.th\"\n", "# cpp = \"/data-ssd-3/georg/code/denoiser/outputs/exp_dummy=only_spec/best.th\"\n", "state_dict = torch.load(cpp, map_location='cpu')[\"state\"]\n", "model.load_state_dict(state_dict)" ] }, { "cell_type": "code", "execution_count": 635, "id": "incredible-coordination", "metadata": {}, "outputs": [], "source": [ "import pandas as pd\n", "\n", "train_df = pd.read_csv(data_dir + \"meta_tr.csv\")\n", "val_df = pd.read_csv(data_dir + \"meta_val.csv\")" ] }, { "cell_type": "code", "execution_count": 636, "id": "effective-kelly", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "17.7% packet loss\n" ] } ], "source": [ "# n = 1\n", "n = 17\n", "\n", "# row = train_df.iloc[n]\n", "row = val_df.iloc[n]\n", "\n", "clean_fp = row['clean_path']\n", "noisy_fp = row['noisy_path']\n", "print(\"{}% packet loss\".format(round(row['packet_loss'] * 100, 1))) \n", "\n", "arr_clean, _ = torchaudio.load(clean_fp)\n", "arr_noisy, _ = torchaudio.load(noisy_fp)\n", "out = model(arr_noisy)\n", "arr_enh = (out / max(out.abs().max().item(), 1)).cpu().detach()[0]" ] }, { "cell_type": "code", "execution_count": 645, "id": "preceding-soviet", "metadata": {}, "outputs": [], "source": [ "# plot_ft(arr_noisy)" ] }, { "cell_type": "code", "execution_count": 646, "id": "blank-maine", "metadata": {}, "outputs": [], "source": [ "# plot_ft(arr_enh)" ] }, { "cell_type": "code", "execution_count": 631, "id": "norman-pendant", "metadata": {}, "outputs": [ { "data": { "text/html": [ "\n", " \n", " " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "play_array(arr_noisy)" ] }, { "cell_type": "code", "execution_count": 632, "id": "forced-reception", "metadata": {}, "outputs": [ { "data": { "text/html": [ "\n", " \n", " " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "play_array(arr_enh)" ] }, { "cell_type": "code", "execution_count": 633, "id": "correct-nerve", "metadata": {}, "outputs": [ { "data": { "text/html": [ "\n", " \n", " " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "play_array(arr_clean)" ] }, { "cell_type": "code", "execution_count": 601, "id": "static-extraction", "metadata": {}, "outputs": [], "source": [ "# wavfile.write(\"tmp/noisy.wav\", 16000, np.floor(arr_noisy.numpy()[0] * 32768).astype(np.int16))\n", "# wavfile.write(\"tmp/enh.wav\", 16000, np.floor(arr_enh.numpy()[0] * 32768).astype(np.int16))\n", "# wavfile.write(\"tmp/clean.wav\", 16000, np.floor(arr_clean.numpy()[0] * 32768).astype(np.int16))" ] }, { "cell_type": "code", "execution_count": 187, "id": "marked-print", "metadata": {}, "outputs": [], "source": [ "# plot_ft(arr_enh)" ] }, { "cell_type": "code", "execution_count": null, "id": "special-exhibit", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": 605, "id": "suspended-damages", "metadata": {}, "outputs": [ { "data": { "text/html": [ "\n", " \n", " " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# best new_data model\n", "play_array(arr_enh)" ] }, { "cell_type": "code", "execution_count": null, "id": "awful-incidence", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "thorough-territory", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": 540, "id": "sufficient-dublin", "metadata": {}, "outputs": [], "source": [ "arr, _ = torchaudio.load(\"tmp/test.wav\")\n", "out = model(arr)\n", "arr_enh = (out / max(out.abs().max().item(), 1)).cpu().detach()[0]" ] }, { "cell_type": "code", "execution_count": 528, "id": "academic-beatles", "metadata": {}, "outputs": [ { "data": { "text/html": [ "\n", " \n", " " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "play_array(arr)" ] }, { "cell_type": "code", "execution_count": 529, "id": "democratic-cover", "metadata": {}, "outputs": [ { "data": { "text/html": [ "\n", " \n", " " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "play_array(arr_enh)" ] }, { "cell_type": "code", "execution_count": null, "id": "vulnerable-dragon", "metadata": {}, "outputs": [], "source": [ "# TODO: this has artifacts....WHY?\n", "# TODO: try upsampling that blindly adds higher freqs?\n", "# TODO: score things with scribe" ] }, { "cell_type": "code", "execution_count": null, "id": "valid-drink", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "transparent-phenomenon", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "exact-second", "metadata": {}, "source": [ "## TODO: Short Term" ] }, { "cell_type": "code", "execution_count": null, "id": "exciting-order", "metadata": {}, "outputs": [], "source": [ "# TODO:\n", "# shift invariant time loss\n", "# continue fully trained with lower LR\n", "# better architecture (smaller first conv, squeeze excite)\n", "# look at train pesq std and compare to val (and look at worst val preds and why)\n", "\n", "# train TTS on spgispeech\n", "# mos eval on amzn\n", "# train a deep demucs to fix up all sorts of things" ] }, { "cell_type": "code", "execution_count": null, "id": "parliamentary-andorra", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "unexpected-flesh", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "bacterial-confidentiality", "metadata": {}, "source": [ "## TODO: Long Term Improvements" ] }, { "cell_type": "code", "execution_count": null, "id": "dimensional-jenny", "metadata": {}, "outputs": [], "source": [ "# update hydra core (nemo needs newer)\n", "# convert to PTL\n", "# add snr/lsd metrics\n", "# https://github.com/facebookresearch/denoiser/issues/81\n", "# do concat instead of sum\n", "# add logging on model size, and profile speed during forward and backprop\n", "# multiple val sets\n", "# log different losses separately\n", "# evaluate mid-epoch\n", "# wandb\n", "# add phase loss and randomized stft loss\n", "# simplify folder naming scheme\n", "# prevent nans, and cuda fft errors that cause crash\n", "# change local path save naming convention\n", "# bandmask applies to both input/output - make specaugment\n", "# add more augmentation - we are overfitting\n", "# model add loss - prob of correct output (normalized pesq?)\n", "# add input lever for different types of improvement (echo removal, equalizing etc)" ] }, { "cell_type": "code", "execution_count": null, "id": "honey-waste", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "horizontal-queens", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "treated-commissioner", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "corresponding-vinyl", "metadata": {}, "source": [ "## Playground" ] }, { "cell_type": "code", "execution_count": 12, "id": "pending-cabinet", "metadata": {}, "outputs": [], "source": [ "import ipywidgets as widgets\n", "from scipy.io import wavfile\n", "import pyaudio\n", "import numpy as np" ] }, { "cell_type": "code", "execution_count": 33, "id": "existing-announcement", "metadata": {}, "outputs": [], "source": [ "_, in_array = wavfile.read(\"tmp/noisy.wav\")\n", "in_array = np.array([in_array])\n", "_, out_array = wavfile.read(\"tmp/clean.wav\")\n", "out_array = np.array([out_array])" ] }, { "cell_type": "code", "execution_count": 13, "id": "loved-tooth", "metadata": {}, "outputs": [ { "data": { "application/vnd.jupyter.widget-view+json": { "model_id": "426e2b9c5a754094a1d431b5a763f70a", "version_major": 2, "version_minor": 0 }, "text/plain": [ "FloatSlider(value=0.0, description='Boost:', max=1.0, readout=False, step=0.01)" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "boost_slider = widgets.FloatSlider(\n", " value=0,\n", " min=0,\n", " max=1,\n", " step=0.01,\n", " description='Boost:',\n", "# disabled=False,\n", "# continuous_update=False,\n", " orientation='horizontal',\n", " readout=False,\n", "# readout_format='d'\n", ")\n", "boost_slider" ] }, { "cell_type": "code", "execution_count": 21, "id": "extreme-england", "metadata": {}, "outputs": [], "source": [ "n = 0\n", "def callback(in_data, frame_count, time_info, status):\n", " global n\n", " # loop\n", " if n >= in_array.shape[1] // frame_count:\n", " n = 0\n", " frac = (1 - boost_slider.value)\n", " audio_array = np.floor(in_array * frac + out_array * (1 - frac)).astype(np.int16)\n", " data = audio_array[:, n * frame_count : (n + 1) * frame_count]\n", " n += 1\n", " return (bytes(data), pyaudio.paContinue)\n", "\n", "chunk_size = 1600\n", "audio = pyaudio.PyAudio()\n", "stream = audio.open(\n", " format=8,\n", " channels=1,\n", " rate=16000,\n", " output=True,\n", " stream_callback=callback,\n", ")" ] }, { "cell_type": "code", "execution_count": 22, "id": "iraqi-produce", "metadata": {}, "outputs": [], "source": [ "# stream.start_stream()\n", "# time.sleep(0.1)\n", "stream.stop_stream()\n", "stream.close()" ] }, { "cell_type": "code", "execution_count": null, "id": "multiple-offset", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "unlikely-contractor", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": 468, "id": "opposite-thought", "metadata": {}, "outputs": [], "source": [ "# cycle through amr\n", "# !sox tmp/conference_bad_2.wav -r 8k -C 0 -c 1 tmp/tmp.amr-nb\n", "# !sox tmp/tmp.amr-nb -r 16k -b 16 tmp/tmp.wav" ] }, { "cell_type": "code", "execution_count": 469, "id": "engaging-cannon", "metadata": {}, "outputs": [], "source": [ "# arr_noisy, _ = torchaudio.load(\"tmp/sks_hard.wav\")\n", "# arr_noisy, _ = torchaudio.load(\"tmp/k_call3.wav\")\n", "arr_noisy, _ = torchaudio.load(\"tmp/tmp.wav\")\n", "out = model(arr_noisy)\n", "arr_enh = (out / max(out.abs().max().item(), 1)).cpu().detach()[0]" ] }, { "cell_type": "code", "execution_count": 465, "id": "impressive-anderson", "metadata": {}, "outputs": [ { "data": { "text/html": [ "\n", " \n", " " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "play_array(arr_noisy)" ] }, { "cell_type": "code", "execution_count": 470, "id": "romance-notification", "metadata": {}, "outputs": [ { "data": { "text/html": [ "\n", " \n", " " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "play_array(arr_noisy)" ] }, { "cell_type": "code", "execution_count": 471, "id": "confused-state", "metadata": {}, "outputs": [ { "data": { "text/html": [ "\n", " \n", " " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "play_array(arr_enh)" ] }, { "cell_type": "code", "execution_count": null, "id": "intermediate-classroom", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "suffering-paint", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "charged-nickname", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": 41, "id": "accomplished-thought", "metadata": {}, "outputs": [], "source": [ "# Load model from cloud\n", "\n", "# \"--dns48\" - Use pre-trained real time H=48 model trained on DNS.\n", "# \"--dns64\" - Use pre-trained real time H=64 model trained on DNS.\n", "# \"--master64\" - Use pre-trained real time H=64 model trained on DNS and Valentini.\n", "# \"--valentini_nc\" - Use pre-trained H=64 model trained on Valentini, non causal.\n", "\n", "class Args():\n", " def __init__(self, model_path=None, dns48=None, dns64=None, master64=None, valentini_nc=None):\n", " self.model_path = model_path\n", " self.dns48 = dns48\n", " self.dns64 = dns64\n", " self.master64 = master64\n", " self.valentini_nc = valentini_nc\n", "\n", "args = Args(dns64=True)\n", "\n", "pretrained_model = pretrained.get_model(args)" ] }, { "cell_type": "code", "execution_count": 42, "id": "interracial-clause", "metadata": {}, "outputs": [], "source": [ "out = pretrained_model(arr_enh)\n", "arr_enh_2 = (out / max(out.abs().max().item(), 1)).cpu().detach()[0]" ] }, { "cell_type": "code", "execution_count": 43, "id": "horizontal-subcommittee", "metadata": {}, "outputs": [ { "data": { "text/html": [ "\n", " \n", " " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "play_array(arr_enh_2)" ] }, { "cell_type": "code", "execution_count": 399, "id": "demonstrated-nicaragua", "metadata": {}, "outputs": [], "source": [ "d = \"/data-ssd-3/georg/data/DNS-Challenge/datasets/\"\n", "\n", "# _, clean_arr = wavfile.read(d + \"clean/read_speech/book_05767_chp_0015_reader_06592_6.wav\")\n", "# _, noisy_arr = wavfile.read(d + \"training_data-09_02_2021/slices/0.wav\")\n", "\n", "_, clean_arr = wavfile.read(d + \"clean/read_speech/book_02668_chp_0001_reader_06016_18.wav\")\n", "# _, noisy_arr = wavfile.read(d + \"training_data-09_02_2021/slices/65346.wav\")\n", "_, noisy_arr = wavfile.read(d + \"training_data-09_02_2021/slices_shift/65346.wav\")" ] }, { "cell_type": "code", "execution_count": 400, "id": "bottom-plate", "metadata": {}, "outputs": [], "source": [ "# arr = np.floor(arr.copy() * 32768).astype(np.int16)\n", "clean_arr = clean_arr.copy().astype(np.float32) / 32768\n", "noisy_arr = noisy_arr.copy().astype(np.float32) / 32768" ] }, { "cell_type": "code", "execution_count": 423, "id": "exact-shakespeare", "metadata": {}, "outputs": [], "source": [ "vl = []\n", "sl = list(range(-5000, 5000))\n", "for n in sl:\n", "# v = np.abs(clean_arr[1000:-1000] - noisy_arr[1000+n+78:-1000+n+78]).mean()\n", " v = np.abs(clean_arr[10000:-10000] - noisy_arr[10000+n:-10000+n]).mean()\n", " vl.append(v)" ] }, { "cell_type": "code", "execution_count": 424, "id": "useful-thursday", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "[]" ] }, "execution_count": 424, "metadata": {}, "output_type": "execute_result" }, { "data": { "image/png": "iVBORw0KGgoAAAANSUhEUgAAAXoAAAD4CAYAAADiry33AAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjMuMywgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/Il7ecAAAACXBIWXMAAAsTAAALEwEAmpwYAAAstklEQVR4nO3deXhV1b3/8fc3MyEhTGEGGQVRBDUiVsEBB9RWtHXAXq1ttVqH3lp7a6Fap9rr0F5r/dXr0NZaZ616lVYUUZyqKARkliHMMwGSEDIP6/fH2UlODifJTsi0w+f1POdhn73XOVk77HzOOmuvvbY55xARkY4rpq0rICIiLUtBLyLSwSnoRUQ6OAW9iEgHp6AXEeng4tq6ApF69uzpBg8e3NbVEBEJlIULF+5xzqVH29bugn7w4MFkZma2dTVERALFzDbVtU1dNyIiHZyCXkSkg1PQi4h0cAp6EZEOTkEvItLBKehFRDo4Bb2ISAenoBfxKaeglLeX7mjraog0moJexKefvPQVN724iO25RW1dFZFGUdCL+LTNC/iisoo2rolI4yjoRXwy71/dlE2CRkEv4lOll/Ax1kBBkXZGQS8i0sEp6EV8irFQU149NxI0CnoRv7wuG6dOegkYBb2IT1Vd85XKeQkYBb2IT+Z13VSqRS8Bo6AX8am6RV/ZptUQaTQFvYhPMWrRS0Ap6EV8suqTsW1bD5HGUtCLNJJa9BI0voLezKaY2WozyzKz6VG2TzKzRWZWbmaXRGy72szWeo+rm6viIq1NJ2MlqBoMejOLBR4DzgNGA1eY2eiIYpuB7wMvRry2O3AXcBIwHrjLzLoderVFWl/V1AcaXilB46dFPx7Ics6td86VAi8DU8MLOOc2OueWApHjEc4F5jjn9jnncoA5wJRmqLdIqzNdMCUB5Sfo+wNbwp5v9db54eu1ZnadmWWaWWZ2drbPtxZpXUZV100bV0SkkdrFyVjn3FPOuQznXEZ6enpbV0ckKqvuulHSS7D4CfptwMCw5wO8dX4cymtF2pWaKRAU9BIsfoJ+ATDCzIaYWQIwDZjp8/1nA+eYWTfvJOw53jqRwKkadaOcl6BpMOidc+XAzYQC+mvgVefcCjO718wuBDCzE81sK3Ap8KSZrfBeuw/4DaEPiwXAvd46kcBR140EVZyfQs65WcCsiHV3hi0vINQtE+21TwNPH0IdRdqFqq6bCp2NlYBpFydjRYLAdOMRCSgFvYhPNTcHV9RLsCjoRXyq7qPXNMUSMAp6EZ+qLphSe16CRkEv4pdG3UhAKehFfFIfvQSVgl7EJ914RIJKQS/ikyY1k6BS0Iv4pCtjJagU9CKNpJiXoFHQi/ikG49IUCnoRXyq6aNX0EuwKOhFfNKoGwkqBb1II2nUjQSNgl7Ep6rZK9V1I0GjoBfxqerKWA27kaBR0Iv4pHH0ElQKehGfam4O3qbVEGk0Bb1IIzn13UjAKOhFfKo5GdvGFRFppA4f9MVlFSzclFP9PK+ojPzisjaskQSVpimWKkG7QXyHC/ov1+/l+3+bT3FZBQBPfbKe7zz+Oat27gfgosc+49QHP2zLKkpA6YKpw0thaTm3vrKY9dkHqKx0vL9yFyXlFWTtzmfM3bOZu2pXva+vahBs2FPAlEc+4dXMLWzaW8C67ANs2FPA1Mc+4y+fridrd+j9P1y1m8oW+gDpcEE//Y1lfLQ6m0c/WMvsFTt5eM4aAJZuzWO99wvOKyqr1Sp7e+kOnv9iEwBfbc4ht7AU5xzPfLaBLfsKySssY+GmfQAs3pLLxIfmsmFPQevvnLQxjaM/FKXllbyxaCtlFaGb7m7PLWLLvsKDyu0vLuPPn6xnZ14xd761POo38K05hazcvr9F65u5MYc3vtrGb/61kreX7eDaZzN54qP1fLU5l8LSCn74TCafrs2mtLyS215bwufr9rD3QAlzVu7iN/9ayQn3vc/iLbn831fbWLUzn9teW8ppv/uIyf/zMc98toElW3K57+2vufrp+by2aCs/eGYB3/rTv1tkX8zP11AzmwL8EYgF/uKceyBieyLwLHACsBe43Dm30cwSgCeBDKAS+Klz7qP6flZGRobLzMxswq5Adn4JJ/72fV9le6YkcsGYPqzYvp9Mr2sn846zyLgv9PqbzxjOnz7MomdKAn3Skli+bT/zZpzJ819s4rEP13HtqUO44fRhvL1sB1PH9SclMY4FG/dxVJ8udOkUV92f++7yHQBMOaZvo/dn094CPlmTzZUTjqh+Pwh1RyXGxWBmfLo2m27JCRzTP63R7y+N86NnM5mzche//uZorjl1SFtXp83kFJRy2ZPzuHfqMaQkxrFkay65haUcKKlgSM9kLj9xECXlFXy+bi+TRqTjnGPxllzW7DrAr/5vGV2T47nnwqN56N3VbMstYv1/n8/lT83j+EHdKK90/PXfGwAY2TuV1bvymTiiJw9dcix90zoB8Mj7a/jjB2txDjbcf36tv40q23OLKCqrYFh6Cv/I3MLby3bwt++fSH5JOf/74Tqy80sY2L0T/dI6sXFvAVm7DzCqTyrHHdGNZz7bSLfkeLp3TuTpz0J1+cawHny+bi/9u3biyN4pfLg6u0V+t69efzLjh3Rv0mvNbKFzLiPqtoaC3sxigTXA2cBWYAFwhXNuZViZG4FjnXM/NrNpwMXOucvN7CYgwzn3AzPrBbwDnOicq6zr5x1K0M94Yykvzd9Ct+R4cgpbph9+/ODuzN+476B1E4Z259G5WQCcMrwHN5w2nPTURM595BMANj5wAR+u2s1bi7fx0CVjWb49jw++3sWVE46oPoAh1OrJLSylV5ckLntiHvM37mPiiJ4MS0/h8hMHMrxXCiNuf4drTx3C7RccxZAZs+iXlsTnMyazeEsuqUlxDO3ZmYpKx7JteYzpn8ajc7PYvb+YUX1S+f4pBwdUcVkF8bExxMYYJeUVLN6cS9+0TvTqksjeglK6doonNsZIio8F4J1lO+jeOYGThvZokd9xe1UV9HdccBTXThza1tVpVf9csp28ojLGDezKKwu28Jz3DTiai8b1483F2wEY1SeVyUf14rEP19VZvipEG9IlKY79xeW11vXv2okJQ3swuEcynRJi+XTtHn529pHc9dZylmzNY+EdZ3GC13j76L9O5+aXFrF8W8t9EwjPnuevOYlfvLaEHXnFtcp0io+lyOta/uO0cfz05cUAjOiVwpxbT2vyzz7UoD8ZuNs5d673fAaAc+7+sDKzvTLzzCwO2AmkA38CvnDOPeeV+wCY4ZybX9fPa2rQL9y0j+88Pg8IfcqPvONdSr2viPdddAx3vLm8Vvlj+ndh+bb9nD4ynRtPH85lT4ZeG2PQKzWJnftr/+ccqvCDf8Z5o7j/nVUAXDlhEKlJ8Tz+0Tr+65wjydyUE+p6uuI4/vOlrw56nxG9Uli7+wAAYwd2ZcmWXADiYoxyr3/v/m+PYenWPF6av5lfThnFg++uqn79w5eNpU9aEnNW7uKMkb0Y3iuFn778FZv2FvLAd8bw7vKdvJq5FQh9YM1bt5fEuNCB+f6tpzG8VwqDp79NamIcy+45t1l/R+3d9c9lMnvFLm4//yh+NKnjBv2BknI27S3gX0t3cP2koby3Yhe3vb4UgKT4GIrL6myndQh1NRSHpndmfXZNl23V3/RDlxzL+WP68lnWHs4Z3Zt7/rmS7blFPPW9DNbuymd/cTlbcwp5d/lOHr3iOHbnl3DKA3P5xrAePH/NSTz5yXqSE2I5Y2QvBvVIbnK9DzXoLwGmOOeu9Z5fBZzknLs5rMxyr8xW7/k64CTg24S+CVwBDAS+Aq5xzr0e8TOuA64DGDRo0AmbNtXdWqhLXlEZpz4wl1+eN4orJxzB5U/O48sN+5jzs0kM75XCtx//nH5pnThxcDeuOnkwAB+t3s03hvWkU0Iss1fsJC7GmHxUbwAe+zCL7blFvLN8J6Xllfz9hyfyxw+y+GRNNkf17cLXO0KtgowjurG3oPSw6LOfcd4osvNL+Iv31fqK8QMpKKngqL5duH7SUGJiQl+hS8ormL9hH6cO7xn1a3VQ/fi5hby7YiczzhvF9acNa+vqNIvHP1rH1pxC3l62g0cuH8ekEenc+MIi3l2xEwi1mLflFjXpvf/fFcfxp7lZrN6Vz79/eQY/e2UxCzbWjID7zUXH8GuvAfbuLROZ8sinTfo5E0f05NO1ew5aH974SYiLobS87g+oOy44ircWb2f1znz+fHUGpwzrwYT759I5MZZuyQkUlJTzu0vH0i05ntN+9xHjB3fnrgtH8/B7a3hk2jhSk+IbVec5K3dxwhHd6N45oXE7W4+2DPpc4HfAGcAmIB54yjn3Zl0/71C6bsorKomLDZ1f3rCngHeW7+CG04Y1W9iUV1TywpebuXBsP97/ehe/eG0pPzlzOD8/ZyRFpRW8OH8zH6/JZtKIniTExfDmV9tYtDm3+vUnDu5W60D/8/cy+NGzoX09fWQ6H7VQv1+kKycMYltOUbP2M557dG9+/c3RDOiWzDOfbeDuf648pP7G9qgq6KefN4ofByzow/82fvbKYgZ268QX6/fV6oYc2rMzxWUVbM9r+rfZd2+ZyF1vreDLDfvY+MAFOOdq/f09+fE6lm7N48FLjqVzQiynPvghJw/rwe8vHUtRaQVzV+1m9a58vli/l417CigprySvqIzRfbvw5k2nMP2NpbyxaFutn7nh/vM58bcf0CUpjpeum8B7K3ayZtcBLjquP9NfX8o1pw7houP6c+bvP+Jb4/px/aRhHP+bOaH6XHUCJxzRjZ4piQftS25hKZ0SYkmMi621fntuEWmd4umcGNfk31NLaLOuGxfx5mb2OXBteP9+pEMJ+tZUWFrOEx+t40eThtb5ab5lXyHPzttIeaVj6rj+jB2QxoGScsbc/R790pL4bPqZXPLEPGLN+N8rj+fXby5nR14x91x4NKlJcUx/fVmtP8Q+XZL47cXHcN/bX1d/g3jyqhO4/rmFQChsTx7ag/+etYpl95zDut0F5BaW8vXOfH7zr5UM75XCD04ZzH+cdATOOe7550qOG9SV9dkF7M4v4aX5m+mcEEtBaUWTfifpqYm89uOT+dGzmazZdYAuSXF8+/gBdEtOYEtOITvyinjh2gkA7D1QQnJCHJ0SYht41/bjhucX8s7yndw2ZSQ3nj68rasTlXOOsgrHeyt38vWO/fzi3FG8+dU2bnllMb86fxRPfLyefQWlTXrv4b1SiIsxVu3Mr1533aShPPXJem45awRj+qcx+aje5BWWUVBaTr+unep5t5DwgQXhyisqKa2oZMOeAnILyzhleE8A8ovL2HuglE+z9vDwe6uZOCI91B2yv5hOCbH1tqxLyiuqQ/vxj9axa38xd194dFN+Fe3SoQZ9HKGTsZOBbYROxn7XObcirMxNwJiwk7Hfds5dZmbJ3s8oMLOzgV875ybV9/OCEvSHYntuEckJsXRNTqh1IjSaUx6YS05hKcvvPre6a2TtrnzumrmCx757PGmd4rng//2byzIG8APvRGtFpav1fs45Ssorq0+m1qUqJB54ZxXPzttI984J7M4vqd6ekhjHgZLyet6hYYvvPJuuyQkMnfE2nRPiuGfq0Vx8XH/yS8rpEvZHWlpeSXF5Ra11be3GFxYya9lOfnHuSG46o/0EfV5hGbGxRkpiHE9+vI5HP1hb/WHdNTme3EMYmPDbi4/h4ffWsLeglM+nn0m/rp3IKyojOSGWnMJS0lMSyS0so1szdkH4tTu/mC5J8Q0e14eL+oK+we8ezrlyM7sZmE1oeOXTzrkVZnYvkOmcmwn8FXjOzLKAfcA07+W9gNlmVknoQ+KqQ9+d4Atv6TR0kL5y/QQSYmOqQx5gRO9UXvzRhOrn7/x0Yq3XRH5omJmvPwYzIyHOuOOCo7htykicgyVbc5n21Bf8+puj+e74QRx157sNvk99pjzyKScN7U6lg/yScm59dQnZ+SXc/84q/jhtHCcN6UGftCT+NHctj87NYvV9Uw766txWqm4l2N6ujL3ofz8jJTGOf/7kVP7nvTXVgxCARoX8Q5ccS5ekON5buYsje6cyum8XJh2ZTlqneN78ajt9uiQBkNYp9OHbKzX0vC1CPvznS8N8jaNvTYdDiz5InHMs2pzD2AFdiYuNYfrrS/nnku28e8skVu3M50fPZpKaGMcfrxjH/qJynvh4Xa2v9o3VMyWBi8b1rz7hO/fnpzE0PaW5dueQ3PTiIt5euoNbzz6S/5w8ok3qkFdYRufE2Or+9spKx9BfzQJCrfeCknLKKhr/N93evqVI4x1Si14Ob2bGCUfUnFC976JjuOtbR9MpIbZ6RMMPTx3CmaNCo5UmDO3Bi/M38/wXm5rUF7znQGl1yAPs3F/M0PQUyioqiY9tHxdyt9WVsZWVjrH3vse3j+/Pw5eN47l5G9mSUzMipq7W+6g+qQd9+J4zujfjBnXl/ZW7ePm6k0mIax+/W2kZCnpplLjYGKp6Uob07Fw9fLVKn7Qkbj37SM4Ymc6clbuYu2o3vbskcd2kofzgbwsorajkF+eO5IOvd9UakVSX7PwS8orKGP/b9+nftRMnDe3O/d8+toX2rn41k5q13s90zlFYWkHnxDj2FITOl7yxaBsPXzaOX7+1ooFXh7z9nxO57bWlvL5oKxeN68egHp259ewjAdrtSWVpXgp6OSQjeqdGXX/coG4cN6gbt559JGZGbIxx38XH8Mairdxw2jDGDujKlX/9ssH3n7tqNwAl5ZWs31PA+rDrFXIKyrjpjOGMGdC60z+0Znfn/321jVtfXcK/f3kG+4tqToTfPbPukL9+0lBOHtaD7/9tARA6Z/Pbi4/htikj6d1F/dqHIwW9tKi4sO6WyzIGclnGQABOOKIbN5w+jMc/qvvSeIC3Fm/nLe+K4iovzd9SvbxhTwGzf1bvQK5mUxXvVRMMllVUYtTex0Px0vzNbNxbwIzzjgLgT3PX8tQn6wH4y6cbeObzjdVlw5cjHdk7lVO94YhVkuJjNTrlMKaglzbRKSGWX04ZRXxsDMu35TG6bxdW7thf3YL3q7i8aWP+m8RV/RNaGHH7O4wb2JU3bzqlWd5+xhvLQv+edxQ78or4/XtrqrfVF+wAb950Cr/510oWbsqhV5dE4mJjuHBsv4MCXw5PCnppU1V9xQAb9xRUB/3Zo3szZ2X9830DxJpVX33pnOO215Zy9ujenHN0n2ava1XAh08Zvtiba6g5FZVW1JpTxY9xA7tyy1kj+P3s1RzbvysAj15xXLPXTYJJp9ql3ejVJXQZ+klDuvPn72XUeRFZuPV7Cjj6rtmUlFewNaeIfyzcynXelcItpaVH3ewrLCU77EK1aH46eQRf/moyU47uw30XHQPAxBHpvHXzqaQlt5+LzKR9UIte2o3khDjevWUi/b0Lyn521gie+XwjH/z8dB58dxUvfrk56usKSyv4wd8WsHRrXovWrzrfW/hcbE5B/UHfp0sS3xrbj95dknjiqhNatjLSISjopV0Z1adL9fKNpw/nuknDSIiLYeyANF6sZ5COn/nMD1VV0Ld0i37mku31zrQ4b8aZHWpWUGl5Cnppt2JijASv++bSEwbSN60T33u6zlsZ1LJ7fzGPfLCW2ct3svDXZx9yXV5dsKV66t7mvK3ngo376J2aVGse8qqRNpG+c/wAThzcTSEvjaagl0CIiTEmHZnOD04ZzN8+29hg+fH//UH1cnNcVVt14w1o3gumLn1iHolxMSy7u+GbuNz5rdHV88yINIZOxkqg/HLKKF4Km9DtqglHcPFx/et9za79xTjnyC1s/JQMCzftY132gVrrmqvrpsQbGlpSXsnu/IPngB/cI5l3fjqR8YNDU1Ao5KWp1KKXQEmKj2V031A//vGDuvKbi47hkffX1Puauat2k5oUx89eWcK7t0ysdR6gIVW3pwznnKOyGfpv9h6o+eA59cEPD9o+vFcKR/XtwjM/PJGiJt4jQAQU9BJAacnxPPbd4zlxSDeABifkuvOtFXzz2L4ArN6Z36igj8YBFU1o1e8vLiOnoJTFW3LpmpzA1fWcb3jx2pMY3S9Uz+SEOJIT9KcqTaejRwLpAi+4AUbWMd9OuKpcLm/CFL6RKp2jogkt+pteWFR9b9Oj+9X/YfMNXdEqzUh99BJ4Z47qVavfPtyg7qHRLG8v2wGEpj0+1EnJKl3tfnq/7xd+A+v84uh36po4oie/OHfkIdVPJJKCXgLPzDh5WI/q55dlDKherurPr/K72au555+hWxbvzi9mzF2zWdLIaQyco1aLvik3+ti8rzDq+rsvPFo3AJFmp6CXDuehS8ZWL5835uA5b6omCJu3bi/5JeU8+Un0GTTr6p4JnYyteR5+6768ojK2RIT4/mJ/t/Nbcuc5DGsnd9OSjkVBLx1G37Skg4YgDuiWzJ3fHH1Q2T/MWcNPX14MUD3G/vkvNvHvsO6VsoroV6dWOlfrZGz4Vaw3vbCIiQ+FRtA451i+LY9j736Pl+dHn75hxnmjqpc1R420FJ2MlQ7j41+cUT3DZJXunRNISTz4MP/jB2urlxO8oL/jzeUAbHzgAqDuoI/sugkP+n9nhT4oissq+Mun66unGp65pPac+gCnj0zn+tOGcf87q5gwtPtB20Wai4JeOozwYZZ905LYkVdMr9REOkcJ+nDxcTFRQ72uvvfIk7HR5qXJKSxl+bb91c/DPxhiDB777vGMHdgVgMw7zqJLklrz0nIU9NIh/f2H45m/YR+dE+PonFhzZ6XzjunDlpzCWiFcWenIiXIj87pb9LWHV5ZWHHwxU25hGV3DumLyimr66R1w3pia4aE9UxL97ZRIE/nqozezKWa22syyzGx6lO2JZvaKt/1LMxvsrY83s7+b2TIz+9rMZjRz/UWiOrJ3KldOOAKguusmrVM8j195At2SE2qVLSitYF+U6RHqDHpqt9BLorTocwvLyC8pZ1h6Z8YO7MqqnfnV2/7jpEGN3h+RQ9Fgi97MYoHHgLOBrcACM5vpnFsZVuwaIMc5N9zMpgEPApcDlwKJzrkxZpYMrDSzl5xzG5t7R0Tqkup1i1RNW9A1MuhLytkXNh1BZaWjrLKSn7z0VdT3q3SuVtfNgeJyKisdj39cM3ont7CU/OJyUpLiSQrrUnrx2pOYMLQHIq3JT4t+PJDlnFvvnCsFXgamRpSZCvzdW34NmGyhuVQd0NnM4oBOQCmwH5FW1DMlFOyJ8aHDfV9B6KYed31rNBOGdie/uIy9YV03B0rLWbgxh68250Z9v8qIk7G78ktYsHEfv5u9unrdDS8s4pM12aQmxtU6PTywezIxPu6cJdKc/AR9f2BL2POt3rqoZZxz5UAe0INQ6BcAO4DNwO+dc/sif4CZXWdmmWaWmZ2d3eidEKlP984JXDlhEE99LwOAfmmhO1hNOaYPXZLiWbYtr1brvaCkvN45551ztbbv3l9MblH0sfKpSXHcctaIWnURaW0tPY5+PFAB9AOGAD83s6GRhZxzTznnMpxzGenp6S1cJTncmBn3XTSG4weFJkG745ujeeumU+ib1om0TvEUl9XuY//bZxvZlhv9ylUIDa8M77rZnV9S5x2hUhLj+Mawntx4+jBiY6zBEUAiLcHPUbcNGBj2fIC3LlqZrV43TRqwF/gu8K5zrgzYbWafARlA9FvoiLSCtE7x1UMb46PMfFnXHZ6qRE5qlldYVuc0wilJoT+x26aM0hw20mb8tOgXACPMbIiZJQDTgJkRZWYCV3vLlwBzXWimp83AmQBm1hmYAKxqjoqLNIfhTZhyIPKCqf3FZRSVRQ/61LAWvG4BKG2lwaD3+txvBmYDXwOvOudWmNm9ZnahV+yvQA8zywJuBaqGYD4GpJjZCkIfGH9zzi1FpJ246uQjuH7SQb2J9YocdbNpbyGFYS36hLDbFqbqQihpB3x1GDrnZgGzItbdGbZcTGgoZeTrDkRbL9JexMfGMOP8o7hi/CCKyyuY8sinDb4mctTNyh372bi3AIDOCbGhE7Ve7ld13Yi0JU1qJgIM7tmZgd2SfZauadHHeUMlC0srMIPPpp9Za8Kz5ITYqO8g0poU9CKezolxLLzjLE47sv6RX6EWfWj5PyeHDZ1MTqBrcgJ/uGxc9brEBm5zKNIadBSKhOmRktjgbQLDR91kDO5G/66hcflJ8aHW+wXH9uWmM4ZVv59IW1MHokiEhgbHhI+jjzWr7p4J76a58fThjB3QlYwjurVYPUX8UoteJEJcxBQFqREnVItKK6pb9LExRrI3hLJTWNB3TozjnKP7aEiltAsKepEIP5k8gvTUmi6XTvG1T6jO37iPD77eBUBMjJHiTYOcFK8Tr9I+KehFIhw/qBsLbj+r+nnVZGjh3lwcumNUrBmdE0It+qrJ00TaGwW9SAOS4g5uqVf1yMTG1PTRV52UFWlvFPQiDYjWoq/qeY8xo9Qba6mgl/ZKQS/SgMQoLfqcwtC0xLExRq633E9BL+2Ugl6kAfVd9BRjNbcq7N0lqbWqJNIoGkcv0oD42HqCPsa45awjSYqPZWSf1FaslYh/CnqRBtTXoo81Y3S/Ljx6xXGtWCORxlHXjUgD4mJrLnpKjbhDVKzu/yoBoKAXqcPI3qGumPCrW/t2rd0Prxt9SxAo6EXq8NoNJ/PJL84gpp5pDJTzEgTqoxepQ2pSPKlJ8fWGeazmspEAUItepAHhLXoXMYOxum4kCBT0Io3ggMszBlY/V4tegkBBL9KAyMnKfnvxMdXLatFLECjoRRpw3aRhtaYtjgu7gErDKyUIFPQiDUhPTeRX54+Kuk05L0GgoBfxwbz5Kl3E2dj6hl6KtBe+gt7MppjZajPLMrPpUbYnmtkr3vYvzWywt/4/zGxx2KPSzMY17y6ItLy68lxdNxIEDQa9mcUCjwHnAaOBK8xsdESxa4Ac59xw4A/AgwDOuRecc+Occ+OAq4ANzrnFzVd9kdYVMbryoPvLirRHflr044Es59x651wp8DIwNaLMVODv3vJrwGQ7+K7IV3ivFQmcum7yrZt/SxD4Cfr+wJaw51u9dVHLOOfKgTygR0SZy4GXov0AM7vOzDLNLDM7O9tPvUVaVXWcRzbpRQKgVU7GmtlJQKFzbnm07c65p5xzGc65jPT09NaokkijqOEuQeYn6LcBA8OeD/DWRS1jZnFAGrA3bPs06mjNiwRB9agb7/ko3WREAsTPpGYLgBFmNoRQoE8DvhtRZiZwNTAPuASY67xxaGYWA1wGTGyuSou0tsgW/Rs3foOi0oq2qYxIIzUY9M65cjO7GZgNxAJPO+dWmNm9QKZzbibwV+A5M8sC9hH6MKgyCdjinFvf/NUXaR1VOV81jj45IY7kBE3+KsHg60h1zs0CZkWsuzNsuRi4tI7XfgRMaHoVRdqe+uglyHRlrIgvtfvoRYJEQS/ig1r0EmQKehEfavro27QaIk2ioBfxQVfASpAp6EV8qG7Rq5deAkhBL+KDGvQSZAp6ER+qgl599BJECnoRH2puPNLGFRFpAgW9iB/qupEAU9CL+KCclyBT0Iv4oOGVEmQKehEfFPMSZAp6ER9qRt3obKwEj4JexAdTm14CTEEv4oO66CXIFPQiPtRMgSASPAp6ET/UopcAU9CL+KArYyXIFPQiPqiPXoJMQS8i0sEp6EV8UINegkxBL+JD1RQIuvGIBJGCXsQH9dFLkPkKejObYmarzSzLzKZH2Z5oZq942780s8Fh2441s3lmtsLMlplZUjPWX6RV6ObgEmQNBr2ZxQKPAecBo4ErzGx0RLFrgBzn3HDgD8CD3mvjgOeBHzvnjgZOB8qarfYirUQtegkyPy368UCWc269c64UeBmYGlFmKvB3b/k1YLKFOjXPAZY655YAOOf2OucqmqfqIq2pqo9eJHj8BH1/YEvY863euqhlnHPlQB7QAzgScGY228wWmdlt0X6AmV1nZplmlpmdnd3YfRBpcWrRS5C19MnYOOBU4D+8fy82s8mRhZxzTznnMpxzGenp6S1cJZHGUx+9BJmfoN8GDAx7PsBbF7WM1y+fBuwl1Pr/xDm3xzlXCMwCjj/USou0Nt1hSoLMT9AvAEaY2RAzSwCmATMjyswErvaWLwHmutAdGmYDY8ws2fsAOA1Y2TxVF2k9NTGvJr0ET1xDBZxz5WZ2M6HQjgWeds6tMLN7gUzn3Ezgr8BzZpYF7CP0YYBzLsfMHib0YeGAWc65t1toX0RajBr0EmQNBj2Ac24WoW6X8HV3hi0XA5fW8drnCQ2xFAkszV4pQaYrY0V8UItegkxBL9IIatBLECnoRXxQi16CTEEv4kNNH73a9BI8CnoRH9SilyBT0Iv4oKCXIFPQi/hgmtRMAkxBL+KDWvQSZAp6EZEOTkEv4oMa9BJkCnoRH6q6bjS6UoJIQS/ii9r0ElwKehEfalr0atJL8CjoRXxQe16CTEEv4kPVHabUnpcgUtCL+KAWvQSZgl7Eh+oLptSklwBS0Iv4YGrTS4Ap6EV8qB5107bVEGkSBb2ISAenoBfxQePoJcgU9CI+mKavlADzFfRmNsXMVptZlplNj7I90cxe8bZ/aWaDvfWDzazIzBZ7jyeauf4irUKDbiTI4hoqYGaxwGPA2cBWYIGZzXTOrQwrdg2Q45wbbmbTgAeBy71t65xz45q32iKtSw16CTI/LfrxQJZzbr1zrhR4GZgaUWYq8Hdv+TVgsum7rnQgNTcHb+OKiDSBn6DvD2wJe77VWxe1jHOuHMgDenjbhpjZV2b2sZlNjPYDzOw6M8s0s8zs7OxG7YBIa1CzRYKspU/G7gAGOeeOA24FXjSzLpGFnHNPOecynHMZ6enpLVwlkcar6aNXk16Cx0/QbwMGhj0f4K2LWsbM4oA0YK9zrsQ5txfAObcQWAcceaiVFml1atFLgPkJ+gXACDMbYmYJwDRgZkSZmcDV3vIlwFznnDOzdO9kLmY2FBgBrG+eqou0HvXRS5A1OOrGOVduZjcDs4FY4Gnn3AozuxfIdM7NBP4KPGdmWcA+Qh8GAJOAe82sDKgEfuyc29cSOyLSktRHL0HWYNADOOdmAbMi1t0ZtlwMXBrlda8Drx9iHUXaDQW+BJGujBXxofpkrLpuJIAU9CI+6LIQCTIFvYgPmgJBgkxBL+KDGvQSZAp6ER+q7zClJr0EkIJexA+16CXAFPQiPtTcSlBNegkeBb2ID2rQS5Ap6EV8qBpeqXH0EkQKehEf1KKXIFPQi/hgGnQjAaagF/HB1KaXAFPQi/hQ3aJXJ70EkIJeRKSDU9CL+FDVotfkZhJECnoRH9RHL0GmoBfxobpF37bVEGkSBb1IIyQnxLZ1FUQazdetBEUOd/GxMfzq/FGcOapXW1dFpNEU9CI+XTdpWFtXQaRJ1HUjItLBKehFRDo4Bb2ISAfnK+jNbIqZrTazLDObHmV7opm94m3/0swGR2wfZGYHzOy/mqneIiLiU4NBb2axwGPAecBo4AozGx1R7Bogxzk3HPgD8GDE9oeBdw69uiIi0lh+WvTjgSzn3HrnXCnwMjA1osxU4O/e8mvAZPOuFTezi4ANwIpmqbGIiDSKn6DvD2wJe77VWxe1jHOuHMgDephZCvBL4J76foCZXWdmmWaWmZ2d7bfuIiLiQ0ufjL0b+INz7kB9hZxzTznnMpxzGenp6S1cJRGRw4ufC6a2AQPDng/w1kUrs9XM4oA0YC9wEnCJmT0EdAUqzazYOfenun7YwoUL95jZJv+70G70BPa0dSVamfb58HC47XNQ9/eIujb4CfoFwAgzG0Io0KcB340oMxO4GpgHXALMdaE7NEysKmBmdwMH6gt5AOdcIJv0ZpbpnMto63q0Ju3z4eFw2+eOuL8NBr1zrtzMbgZmA7HA0865FWZ2L5DpnJsJ/BV4zsyygH2EPgxERKQd8DXXjXNuFjArYt2dYcvFwKUNvMfdTaifiIgcIl0Z23yeausKtAHt8+HhcNvnDre/ppsdi4h0bGrRi4h0cAp6EZEOTkHfCGb2czNzZtbTe25m9qg3mdtSMzs+rOzVZrbWe1wdtv4EM1vmvebRqqki2hsz+52ZrfL26//MrGvYthle/Veb2blh66NOfmdmQ7zJ7rK8ye8SWnl3DllDE/sFiZkNNLMPzWylma0ws59667ub2RzvmJ1jZt289Y0+ztsjM4s1s6/M7F/e86jHZX2TNNZ17Ld7zjk9fDwIXRA2G9gE9PTWnU9osjYDJgBfeuu7A+u9f7t5y928bfO9sua99ry23rc69vccIM5bfhB40FseDSwBEoEhwDpCw25jveWhQIJXZrT3mleBad7yE8ANbb1/jfxd1LlvQXwAfYHjveVUYI33//oQMN1bPz3s/7zRx3l7fAC3Ai8C//KeRz0ugRuBJ7zlacAr3nLUY7+t98vPQy16//4A3AaEn72eCjzrQr4AuppZX+BcYI5zbp9zLgeYA0zxtnVxzn3hQkfOs8BFrboXPjnn3nOheYsAviB0RTSE9vll51yJc24DkEVo4ruok99531jOJDTZHYQmv7uolXajufiZ2C8wnHM7nHOLvOV84GtC81WFT04Y/v/UqOO89fbEPzMbAFwA/MV7Xt9xWdckjXUd++2egt4HM5sKbHPOLYnYVNeEb/Wt3xplfXv3Q2qmmW7sPvcAcsM+NIKyz+H8TOwXSF63xHHAl0Bv59wOb9NOoLe33Nj/8/boEUINtUrveX3HZdRJGgnW/taim4N7zOx9oE+UTbcDvyLUldGh1LfPzrm3vDK3A+XAC61ZN2l53uyyrwO3OOf2h58ucs45M+sQY6/N7JvAbufcQjM7vY2r0yYU9B7n3FnR1pvZGEL9cUu8P4QBwCIzG0/dE75tA06PWP+Rt35AlPJtoq59rmJm3we+CUz2upqg/knuoq3fS+irfpzXOmrTfW4iPxP7BYqZxRMK+Recc294q3eZWV/n3A6va2a3t76xx3l7cwpwoZmdDyQBXYA/UvdxWdckjcE9Dtr6JEHQHsBGak7GXkDtk1TzvfXdCd1spZv32AB097ZFnow9v633qY79nAKsBNIj1h9N7RNS6wmdrIzzlodQc8LyaO81/6D2Sa8b23r/Gvm7qHPfgvjwjr1ngUci1v+O2idjH/KWG32ct9cHoQ+mqpOxUY9L4CZqn4x91VuOeuy39T752u+2rkDQHhFBb4Rus7gOWAZkhJX7IaGTNVnAD8LWZwDLvdf8Ce/q5Pb28Oq9BVjsPZ4I23a7V//VhI0aIjQ6Y4237faw9UO9D7gs748rsa33rwm/j6j7FsQHcCqhQQVLw/5/zyfUD/0BsBZ4n5rGSaOP8/b6iAj6qMcloVb/P7z184GhYa+Peuy394emQBAR6eA06kZEpINT0IuIdHAKehGRDk5BLyLSwSnoRUQ6OAW9iEgHp6AXEeng/j+zky/MmMnu+gAAAABJRU5ErkJggg==\n", "text/plain": [ "
" ] }, "metadata": { "needs_background": "light" }, "output_type": "display_data" } ], "source": [ "plt.plot(np.array(sl), vl)" ] }, { "cell_type": "code", "execution_count": 425, "id": "prescribed-church", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "[]" ] }, "execution_count": 425, "metadata": {}, "output_type": "execute_result" }, { "data": { "image/png": "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\n", "text/plain": [ "
" ] }, "metadata": { "needs_background": "light" }, "output_type": "display_data" } ], "source": [ "plt.plot(sl[5000-500:5000+500], vl[5000-500:5000+500])" ] }, { "cell_type": "code", "execution_count": null, "id": "twelve-master", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "democratic-wyoming", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "lesser-timber", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "remarkable-grass", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "metropolitan-motel", "metadata": {}, "source": [ "## Shift data" ] }, { "cell_type": "code", "execution_count": 387, "id": "matched-fantasy", "metadata": {}, "outputs": [], "source": [ "!mkdir /data-ssd-3/georg/data/DNS-Challenge/datasets/training_data-09_02_2021/slices_shift" ] }, { "cell_type": "code", "execution_count": 388, "id": "helpful-engine", "metadata": {}, "outputs": [], "source": [ "fns = os.listdir(\"/data-ssd-3/georg/data/DNS-Challenge/datasets/training_data-09_02_2021/slices\")" ] }, { "cell_type": "code", "execution_count": 397, "id": "sorted-cookbook", "metadata": {}, "outputs": [], "source": [ "import tqdm" ] }, { "cell_type": "code", "execution_count": 398, "id": "welsh-sheep", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 65347/65347 [01:25<00:00, 764.09it/s] \n" ] } ], "source": [ "from_dir = \"/data-ssd-3/georg/data/DNS-Challenge/datasets/training_data-09_02_2021/slices/\"\n", "to_dir = \"/data-ssd-3/georg/data/DNS-Challenge/datasets/training_data-09_02_2021/slices_shift/\"\n", "for fn in tqdm.tqdm(fns):\n", " if \"wav\" not in fn:\n", " continue\n", " rate, arr = wavfile.read(from_dir + fn)\n", " arr = np.hstack([arr[78:], arr[-78:][::-1]])\n", " wavfile.write(to_dir + fn, rate, arr)" ] }, { "cell_type": "code", "execution_count": null, "id": "auburn-thomas", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "retired-queensland", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "annual-chart", "metadata": {}, "source": [ "### Explore parameters" ] }, { "cell_type": "code", "execution_count": 548, "id": "dominant-night", "metadata": {}, "outputs": [], "source": [ "import torchaudio\n", "from denoiser.demucs import *" ] }, { "cell_type": "code", "execution_count": 549, "id": "incorporate-pleasure", "metadata": {}, "outputs": [], "source": [ "mix, _ = torchaudio.load(noisy_fp)\n", "# mix = torch.cat([mix.view(1,1,-1), mix.view(1,1,-1)])" ] }, { "cell_type": "code", "execution_count": 550, "id": "rolled-variance", "metadata": {}, "outputs": [], "source": [ "def log_n_params(m):\n", " print(round(m / 1e6, 1), \"M\")" ] }, { "cell_type": "code", "execution_count": 551, "id": "primary-pleasure", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Model params:\n", "34.2 M\n" ] } ], "source": [ "print(\"Model params:\")\n", "log_n_params(sum([e.numel() for e in model.state_dict().values()]))" ] }, { "cell_type": "code", "execution_count": 552, "id": "dental-thursday", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "0.5 M\n" ] } ], "source": [ "if mix.dim() == 2:\n", " mix = mix.unsqueeze(1)\n", "\n", "if model.normalize:\n", " mono = mix.mean(dim=1, keepdim=True)\n", " std = mono.std(dim=-1, keepdim=True)\n", " mix = mix / (model.floor + std)\n", "else:\n", " std = 1\n", "x = mix\n", "log_n_params(x.numel())" ] }, { "cell_type": "code", "execution_count": 553, "id": "serial-sterling", "metadata": {}, "outputs": [], "source": [ "# torch.nn.functional.interpolate(\n", "# input, size=None, scale_factor=None, mode='nearest', align_corners=None, recompute_scale_factor=None\n", "# )" ] }, { "cell_type": "code", "execution_count": 554, "id": "further-editing", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "2.0 M\n" ] } ], "source": [ "if model.resample == 2:\n", " x = upsample2(x)\n", "# x = nn.functional.interpolate(x, scale_factor=2, mode='nearest')\n", "elif model.resample == 4:\n", " x = upsample2(x)\n", " x = upsample2(x)\n", "# x = nn.functional.interpolate(x, scale_factor=2, mode='nearest')\n", "# x = nn.functional.interpolate(x, scale_factor=2, mode='nearest')\n", "log_n_params(x.numel())" ] }, { "cell_type": "code", "execution_count": 555, "id": "premium-general", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "Sequential(\n", " (0): Conv1d(1, 48, kernel_size=(8,), stride=(4,))\n", " (1): ReLU()\n", " (2): Conv1d(48, 96, kernel_size=(1,), stride=(1,))\n", " (3): GLU(dim=1)\n", ")" ] }, "execution_count": 555, "metadata": {}, "output_type": "execute_result" } ], "source": [ "model.encoder[0]" ] }, { "cell_type": "code", "execution_count": 556, "id": "social-champion", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "Sequential(\n", " (0): Conv1d(48, 96, kernel_size=(8,), stride=(4,))\n", " (1): ReLU()\n", " (2): Conv1d(96, 192, kernel_size=(1,), stride=(1,))\n", " (3): GLU(dim=1)\n", ")" ] }, "execution_count": 556, "metadata": {}, "output_type": "execute_result" } ], "source": [ "model.encoder[1]" ] }, { "cell_type": "code", "execution_count": 559, "id": "valid-playlist", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "torch.Size([1, 1, 1984000])" ] }, "execution_count": 559, "metadata": {}, "output_type": "execute_result" } ], "source": [ "x.shape" ] }, { "cell_type": "code", "execution_count": 564, "id": "greenhouse-object", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "23807952" ] }, "execution_count": 564, "metadata": {}, "output_type": "execute_result" } ], "source": [ "model.encoder[0][:1](x).numel()" ] }, { "cell_type": "code", "execution_count": 568, "id": "trained-freight", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "23807952" ] }, "execution_count": 568, "metadata": {}, "output_type": "execute_result" } ], "source": [ "nn.Conv1d(1, 48, kernel_size=(8,), stride=(4,))(x).numel()" ] }, { "cell_type": "code", "execution_count": 569, "id": "several-participation", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "7935984" ] }, "execution_count": 569, "metadata": {}, "output_type": "execute_result" } ], "source": [ "nn.Conv1d(1, 16, kernel_size=(8,), stride=(4,))(x).numel()" ] }, { "cell_type": "code", "execution_count": null, "id": "neural-company", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": 563, "id": "funny-rugby", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "torch.Size([1, 48, 495999])" ] }, "execution_count": 563, "metadata": {}, "output_type": "execute_result" } ], "source": [ "model.encoder[0][:4](x).shape" ] }, { "cell_type": "code", "execution_count": 558, "id": "ultimate-coverage", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "Sequential(\n", " (0): Conv1d(1, 48, kernel_size=(8,), stride=(4,))\n", ")" ] }, "execution_count": 558, "metadata": {}, "output_type": "execute_result" } ], "source": [ "model.encoder[0][:1]" ] }, { "cell_type": "code", "execution_count": 25, "id": "independent-parade", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "23.8 M\n", "11.9 M\n", "6.0 M\n", "3.0 M\n", "1.5 M\n" ] } ], "source": [ "skips = []\n", "for encode in model.encoder:\n", " x = encode(x)\n", " log_n_params(x.numel())\n", " skips.append(x)\n", "x = x.permute(2, 0, 1)" ] }, { "cell_type": "code", "execution_count": 26, "id": "renewable-nutrition", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "1.5 M\n" ] } ], "source": [ "x, _ = model.lstm(x)\n", "x = x.permute(1, 2, 0)\n", "log_n_params(x.numel())" ] }, { "cell_type": "code", "execution_count": 20, "id": "median-chosen", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "3.0 M\n", "6.0 M\n", "11.9 M\n", "23.8 M\n", "2.0 M\n" ] } ], "source": [ "for decode in model.decoder:\n", " skip = skips.pop(-1) # last element in list\n", "# x = th.cat([x, skip[..., :x.shape[-1]]], dim=1)\n", " x = x + skip[..., :x.shape[-1]]\n", " x = decode(x)\n", " log_n_params(x.numel())\n", "# 3.0 M\n", "# 6.0 M\n", "# 11.9 M\n", "# 23.8 M\n", "# 2.0 M" ] }, { "cell_type": "code", "execution_count": 21, "id": "religious-balloon", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "0.5 M\n" ] } ], "source": [ "if model.resample == 2:\n", " x = downsample2(x)\n", "# x = nn.functional.interpolate(x, scale_factor=0.5, mode='nearest')\n", "elif model.resample == 4:\n", " x = downsample2(x)\n", " x = downsample2(x)\n", "# x = nn.functional.interpolate(x, scale_factor=0.5, mode='nearest')\n", "# x = nn.functional.interpolate(x, scale_factor=0.5, mode='nearest')\n", "log_n_params(x.numel())" ] }, { "cell_type": "code", "execution_count": null, "id": "diagnostic-duration", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "direct-influence", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "studied-commerce", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "latest-anaheim", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "communist-dispatch", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "sufficient-valuable", "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": "3.8.10" } }, "nbformat": 4, "nbformat_minor": 5 }