{ "cells": [ { "cell_type": "code", "execution_count": 1, "id": "de762114", "metadata": {}, "outputs": [], "source": [ "# import torch\n", "# print(torch.__version__)\n", "# print(torch.cuda.is_available())\n", "# print(torch.cuda.device_count())\n", "# # 1.10.0+cu113\n", "# # True\n", "# # 4" ] }, { "cell_type": "code", "execution_count": 1, "id": "1d45d58b", "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": 2, "id": "2fca657a", "metadata": {}, "outputs": [], "source": [ "import os\n", "os.environ[\"CUDA_VISIBLE_DEVICES\"] = \"3\"" ] }, { "cell_type": "code", "execution_count": null, "id": "43649943", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "b1ce93dc", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "d33dd9dd", "metadata": {}, "source": [ "## Make data" ] }, { "cell_type": "code", "execution_count": null, "id": "72b2e0c0", "metadata": {}, "outputs": [], "source": [ "# https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_classification/datasets.html" ] }, { "cell_type": "code", "execution_count": null, "id": "1a91d35e", "metadata": {}, "outputs": [], "source": [ "# Each line is a training example. audio_filepath contains path to the wav file, \n", "# duration is duration in seconds, offset is offset in seconds, and label is label (class):\n", "# {\n", "# \"audio_filepath\": \"/two/8aa35b0c_nohash_0.wav\", \n", "# \"duration\": 0.63, \"label\": \"speech\", \"offset\": 0.0,\n", "# }\n", "# {\n", "# \"audio_filepath\": \"/Emergency_vehicle/id_58368 simambulance.wav\", \n", "# \"duration\": 0.63, \"label\": \"background\", \"offset\": 4.0,\n", "# }\n", "\n", "# json manifest line example\n", "# {\"audio_filepath\": \"/path/to/audio_file\", \"offset\": 0, \"duration\": null, \"label\": \"infer\", \"text\": \"-\"}" ] }, { "cell_type": "code", "execution_count": 4, "id": "050f9a44", "metadata": {}, "outputs": [], "source": [ "import pandas as pd\n", "import json\n", "\n", "from suno_utils.utils.notebook import Audio\n", "from suno_utils.web.music import predict_music\n", "\n", "CC_DATA_DIR = \"/mnt/data-ssd-1/data/podcasts/cc_data/episode_audio/\"\n", "\n", "def _get_music_segments(music_p, duration_s, thres=0.45, max_merge_duration_s=1, min_music_duration_s=1):\n", " max_merge_duration_f = int(np.floor(len(music_p) / duration_s * max_merge_duration_s))\n", " min_music_duration_f = int(np.floor(len(music_p) / duration_s * min_music_duration_s))\n", " is_music_arr = music_p >= thres\n", " # merge separated music segments\n", " last_n = -1\n", " merged_arr = []\n", " for n, is_music in enumerate(is_music_arr):\n", " new_is_music = is_music\n", " if n - last_n <= max_merge_duration_f:\n", " new_is_music = True\n", " if is_music:\n", " last_n = n\n", " merged_arr.append(new_is_music)\n", " # only keep slices of min duration\n", " music_segments = []\n", " \n", " filtered_music_segments = []\n", " return filtered_music_segments\n", " \n", "cc_episodes_df = pd.read_csv(\"/mnt/data-ssd-1/data/podcasts/cc_data/episodes_with_audio.csv\")" ] }, { "cell_type": "code", "execution_count": 6, "id": "80c824ca", "metadata": {}, "outputs": [], "source": [ "data = []\n", "for _, row in cc_episodes_df.iterrows():\n", " filepath = CC_DATA_DIR + row[\"filename\"]\n", " audio = Audio.from_file(filepath)\n", " music_p = predict_music(audio, frame_duration_ms=50)\n", " music_segments = _get_music_segments(music_p, audio.duration_s)\n", " data.append((row[\"uuid\"], filepath, music_segments))\n", " break" ] }, { "cell_type": "code", "execution_count": 18, "id": "83d8e7b4", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "[]" ] }, "execution_count": 18, "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(music_p)" ] }, { "cell_type": "code", "execution_count": 17, "id": "d84ccc0f", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "[]" ] }, "execution_count": 17, "metadata": {}, "output_type": "execute_result" }, { "data": { "image/png": "iVBORw0KGgoAAAANSUhEUgAAAXQAAAD4CAYAAAD8Zh1EAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjUuMCwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy8/fFQqAAAACXBIWXMAAAsTAAALEwEAmpwYAAAgVUlEQVR4nO3de5wU5Z3v8c+PGW7KRS4DIrcBBV3QjcrEWxI1MSroLuxZk3N0XxqTNfFko0n2mE1C4h6PMWfPmmRPNnHXEzWbbC5n4zUmyzHkeI8xXhlQVEB0RNBBZQYEUUAHhmf/6BroGbqne7qrup+n6vt+veY11VXV1b+qeurXTz31dJU55xARkfANqncAIiISDyV0EZGUUEIXEUkJJXQRkZRQQhcRSYnGen3w+PHjXXNzc70+XkQkSMuXL9/snGsqNK1uCb25uZnW1tZ6fbyISJDMbEOxaWpyERFJCSV0EZGUUEIXEUkJJXQRkZRQQhcRSYmSCd3MfmxmHWb2XJHpZmbXmVmbmT1jZsfHH6aIiJRSTg39J8D8fqYvAGZFf5cCP6g+LBERGaiS/dCdc783s+Z+ZlkE/Mzl7sP7uJkdYmaTnHOvxxVkvj+8uJm/uX0lw4c08PLmHUwYOZQRwxo5eeY4xh08pOLlzmsey2mzC/bVL8sz7du4b/Um1m3ewV3P5Fb90FHD6HaOd3d3M2rYYHbt7ubNHV0AfPKUZv7tiQ3s7nb8xYnT+MUTr+xb1kFDGpg/91CmjBmOA/7pgTZapo/hlMPHce+aDl7qeIcz50ykfdsuTps1HoCnXt3GYy9tYeGxhzHlkOHs6Opm9WvbeX/zmH3LHTq4gaaRQzm86WDmTR/LI22beWLdln3T/9C2maljD2LS6OFs2LKDiaOG8cDzHezu3svsiSPpePs91ry+HYCW6bnlvrp1J5u2v8dZcyayZ6/jgec7OOXwcZw0cxx/ceI0xo8Y2ms7rXhlKx/7waPsdXBC81hmNh3MLcteBWD2xBEYxtpNb/NfT5vJ6te2c8SEEax4ZRsvd77DF86Yxf/8zRq+fPaRXPbhI3hwbQef+tdlHDSkgZ1d3fzTBcfx+ZufOmDfDG4wvnL2UTy4toNHX9rCWXMmAvC7tZ0sPPYwlqx8ja49e7n45Ol87sNHcOq3H+S9PXuZMHIoHW+/V3B/z2w6GAMuPXUmjYMGcfiEEQBs29nF6UdO4OePb+DNd7p4fN0WTpgxlss/cgSDG8pr4Vzz+nZ2dnUzb/r+fff7Fzr50u0rOfeYSYwatv+w3bKji5c63+GE5rEHLOe1t95lxNBGrl44t+hn7e7ey9fvfJb5Rx/Klne6OG/eFBoGGR1vv8v/f+4NLjxxOoMGGfeseoMRQxvZ0dXNmdH26/HWrt38akU7h44exvyjJxX9rP9269P86qmNfGjWeI6bekivaatf3859azoA+PPjJzPlkOGsbH+Lh17o5Lzjp9AwCG5rbccMZk8YydpNb+977ydPaWbUsEaue6CNUw4fx+CGQTz0Qidnz53I3as2AXDyzHE8llfWJ40exoKjJ/HVBUcytLGhaMzbdnbxv5au4bzjp3DizHFF5/OFlXM/9Cih3+WcO7rAtLuAa51zf4he3w981Tl3wK+GzOxScrV4pk2bNm/DhqL944t63zfu4a1du4vEOeDFAeAcHDFhBPddcVplCwA+/dNW7luzqeL3F2KWi63UPFB8vmLT1197Lude9zCrXttechmV+uaiuVx0cnOvcUd8fSl79lb/QeuvPZfmxb+pejl9HTFhBG0d7wz4fQ2Dchuxe6+j7e8WcMSVv+01/UtnzubzZ8wqa1k967X+2nMPGAe9y3n+Putb/numtf7tRw/4Yu3x4PMdfOony/a9/sbCuVx8SjNX3PY0d67YyJLLP8CcSaN6rU9+XACf/umyfcl45VVnMfqgwf2uV3+x5k+vxaMaPnva4SxecFTR6df+9nlueOgl4MD1rhczW+6cayk0raa/FHXO3QTcBNDS0lLR7jp1dhP/b+VrB4wfNayRZ64+u6K4Pn/zU6za+FZF7+2x1zmOmTyaZ6tcTr6X//5c7ljezt/cvhKA1deczZyr7j5gHuh9sJw1ZyL3rN7Ua/rzb2xn/vce7vXe7r2Os+ZM5KZPtBywjDh0F0jccSTzJK3rHHgyh97rWmgNt0RnZnHo2aewf58VSto90/b2kxl3dnX3er11Zy7O9Zt3ALkafKk99tq2d/cN79m7t8TcOfnrkB9r/vRyy+OcSaP48Sffz0l/f39Z8+fb/E7hM7Aem7a/2+9038TRy2UjMDXv9ZRonIiI1FAcCX0J8Imot8tJwFtJtZ+LiEhxJZtczOxm4HRgvJm1A/8DGAzgnLsBWAqcA7QBO4FPJRWsiIgUV04vlwtKTHfAZbFFJCIiFdEvRUVEUkIJPVJt34tyun9mTYhbJI6YvSsKvsUjiVFCByrsvn7gcuJaUJ5KvigKxWFF1jKJmEPmXTIuU5K7sdQ28WGTOS+iqD8ldJEayMoXpwW2omFFW1pwCb3YDtD3s0j1ip3JSRiCS+giIlKYErqISEoooUey2Esl6VXO4CYF/LtA51c08Qus2T5RSugxSftBIyL+U0Invm94XyoKhS5sFVtHXQRLh3r2Lsni2a2vlNBFRFJCCd1zqvukQ8hnQr5dEyhEJwk5SugiEpt6fG1Vk8zTdkE1uISeth0g4hMdX2ELLqEXVeUpVxbP2JI+lc7iNgX/Tv99i0eSk56EXme1OGhCbocVkeQpoRNju58n56uF77ZY/rwSHu1GASV0EZHUUEIXqQGdCUktKKGLSL98v6iqL8v90pPQ07pTPT+YRHygwyQnPQm9Sr7XQpKQ/N0WM7hRPVTLX3qGVltOW8+x4BK6r5tfqUvSwNfjS8oTXEIvqqqf/8ZTjH05GAp2Wyx2t0VfgpaqaD8KpCmhi4hknBK6iFRFl0r8oYQuUgNqEZFaCC6hqzIgkhwdX2ELLqEnxdeb+CcZl59rHD7fmiB8iycJ6iKbo4ROPKfDzjlvehoU7ltbOLi09cPNKl/2oy9xlMuXYzYuSugiIilRVkI3s/lmttbM2sxscYHp08zsQTN7ysyeMbNz4g81+qykFiwiOr4CVzKhm1kDcD2wAJgDXGBmc/rM9rfAbc6544Dzgf8Td6AiItK/cmroJwBtzrl1zrku4BZgUZ95HDAqGh4NvBZfiOXRJRGR+si/cO9r54KsKCehTwZezXvdHo3LdzVwoZm1A0uBzxdakJldamatZtba2dlZQbjJyeJF8qR7BmRxm0LhpFbPi2/97YZiCbhnbFb3Yajiuih6AfAT59wU4Bzg52Z2wLKdczc551qccy1NTU0xfXS65fcaSNsVeZG4xHU/ptCVk9A3AlPzXk+JxuW7BLgNwDn3GDAMGB9HgDURU1nwpkgN4OZc/gQtValwPxbrZtgzVnkyLOUk9GXALDObYWZDyF30XNJnnleAMwDM7I/IJXS/2lQCpTZJkdL0w6KckgndObcHuBy4G1hDrjfLKjO7xswWRrN9CfiMma0EbgY+6bSFRTJHPyyqr8ZyZnLOLSV3sTN/3FV5w6uBD8QbmoiIDERwvxTVxQ+R5OjwCltwCT0pWWwgSnqVs9r+71tZUvfU7FBCF6kBnVkmJ8lNG9peS01Cr6YWEseFHOf8OWjLv9dieAVWCvOk6AWnVNoI7eQjNQldRCTrUpPQfakdx03tk1JvIZTBEGKshdQkdBHxQGD1qrTVA5XQRURSQgk9yxI+Tc3qabBvq532/ZD29RuI4BJ6ys6QRLxSybWoNOfT0PJNcAk9CXG0ozmcNzu/0EFZ7EBN68Vk3yS9lbUXkxHal5USuohISiihi4ikRGoSuu7WK5JNajXcLzUJPa30NSX1ltWbrIVICT2SxRp+8ndbzKYslqUe4dWWgwu4X0roIiJFhJbuldCJZ6fl7rYYw4JioLstZk89u5+m+YwktDULL6ErA0mIVG6lBsJL6CIhCq2qJ0EKL6HrwBARKSi8hC4iXtHtI/yhhB7xteKf5PUmPTw4Gb6tdrX7IYT9GEKMtZCahF7N/oyrghHHs0njUGh9iq2jKlfp4Mtu9CWOcpUq/6GtT2oSer2phiCSPqEd1kroIlKVevdD9+XM2AdK6CK1oJyTGN1rZr/wEroODBGRgsJL6AnJYht48jfnyuBGxb+y1N9+KDatZ2xc61LvZpmsSE1CV8VdRLKurIRuZvPNbK2ZtZnZ4iLz/GczW21mq8zsF/GGmazYLqp48q1S+OZcRZ4pmmwoUiOVdj8tVS6y3q01tNVvLDWDmTUA1wNnAu3AMjNb4pxbnTfPLOBrwAecc1vNbEJSAfsqqeaFrDZbiD9CKIHJHX9hKaeGfgLQ5pxb55zrAm4BFvWZ5zPA9c65rQDOuY54wxSREIR2G4Cwoi2tnIQ+GXg173V7NC7fbGC2mT1iZo+b2fxCCzKzS82s1cxaOzs7K4u4iNC+SUXSIv/Y08XP+orromgjMAs4HbgA+KGZHdJ3JufcTc65FudcS1NTU0wfnW75bZyBVX4kT8g/fgkh8pC3b5zKSegbgal5r6dE4/K1A0ucc7udcy8DL5BL8MHIYlt10pWpzFbWPFvv/vZD7botxrMc6V85CX0ZMMvMZpjZEOB8YEmfeX5NrnaOmY0n1wSzLr4w99M3sUhydBYYtpIJ3Tm3B7gcuBtYA9zmnFtlZteY2cJotruBLWa2GngQ+LJzbktSQcctvrst+qHQhanid1v0JWqpRqUVnTR0W0yykhfA6vdSstsigHNuKbC0z7ir8oYdcEX0l0k6pRRJn9AO69T8UlREkhFaUssyJXTPZfFirQTGgyKq4yRHCV1EYhPaNZnAwi1JCT2SxTZw1WqSUWi71jNx+LCXfYghC1KT0LOYkEUkWaFV4FOT0KsRW7dFT/b+QMLwJGSpki9lrx50prlfcAnd14KrIiVpoB/u9RbacR1cQhcRkcKU0EVEUkIJ3XO62Cv1VuqWuPUuoobpOIkooUeyWB6Sv9tiFreqf1/CPuwHH2IoJG3XDJTQRWogXWkjO0Lbb0roQFy7zZtv+wJhFO0d5EnIIlI9JfS4+HlGKTIgWevTXWp9Q9saSugiIimhhB4Qb5p0JLVUxsKmhC4ikhJK6BFPe1UlSg+JToZvq11NPM6Vfn85XRJ92yZppYSO3zfnquRAKHTaXOw+1TrFTgdf73FUK0l9YYS2WZXQRWog6wnXV2mr0CihxyRr3b1EsiC0o1oJXYpSrVJCoHK6nxK6iEhKpCahV9/kEdrJVfKq7aWS1S3q242o+gun2HHj+vxPMgaflx2a4BK6zq5EkqPmi7AFl9CTEFcZ9uVgKBRH0Xtz9ROzL+sjpVXaW6PY+waytDRXkEM7BJTQfafzyVQIuXtcCEXQt2auelFCj4nKk0h4Sp2FhnZYK6GLiKSEErqISEoooUey2GSSdLtjFrcp+HiaXjyikt0WY1oZ/ZK6NspK6GY238zWmlmbmS3uZ77zzMyZWUt8IUq9hHsZTySbSiZ0M2sArgcWAHOAC8xsToH5RgJfBJ6IO8ikxXa3RU9SYKEoiq1jfxEXu0Oj+KfSXRVHt8V6S7KYhrQdoLwa+glAm3NunXOuC7gFWFRgvm8C3wLejTG+AxTbeVk9vReJUyUJLP/YU/fB+ionoU8GXs173R6N28fMjgemOud+09+CzOxSM2s1s9bOzs4BB+szFWPpj052whTacV31RVEzGwR8F/hSqXmdczc551qccy1NTU3VfnTmKClIXYSW1TKsnIS+EZia93pKNK7HSOBo4Hdmth44CViiC6Mi4ru01ZHKSejLgFlmNsPMhgDnA0t6Jjrn3nLOjXfONTvnmoHHgYXOudZEIk5IFishSa9zVruq+daMXE08cbWJ+7ZN0qpkQnfO7QEuB+4G1gC3OedWmdk1ZrYw6QBFROoltBp8YzkzOeeWAkv7jLuqyLynVx9WbcXV3dCXNu7Cd1ss0j2tv7stxhSPJE/7SiBFvxT1JZmKiNRLahJ6vaWx/2361kiSUO9rJYk+DSm5RSciuITuy68xRQYilFKrM92wBZfQi0lhBVnEC/WugUv5UpPQq+Vrk0mSUZVa5Wora55u0sT5lgCricavNSnMLIw4a0EJXUQyq9QN6EJrgVJCJ33thoWuMxS/22LxlU/bdqmnpGuQcd0ZM6tnVWkRXEL37XRWRMQXwSV0X+lrRrIqzT3PQjuuldBFpCo6a/aHErpIDaS3Dis+CS6hJ3V6l8U6RtI1qyxuU8C7FR/Ihc6+11Zje0i0Z9skrYJL6EmI6ysiiWdwVnIgFL45V/nz7n+P6pWhSHJPhZCMk4oxtCMgNQk9gDInIpKo1CT0LAittiAitaWEHpMQTksHTN8gEoBEm5sSXHYSlNBFRFJCCT2Syhp2CYmvcxY3Kv7V6vrrzVRsmisxvde8Zayw+qrXhhK6SC3oxjhSA0roxNfd0JdDtuDqFLs5ly9BS1Uq3Y9FnzVbRSy1lmTdP6TtAAEmdCUgkeQUSvBqLAlHcAk9a3x98IaIX3ScgBJ6bFScRNIntONaCV1EJCWU0CNZbNpIvNdiwsv3lW9Fqb94SnVbLGcnltVt0bNtklZK6CIiKaGEHiN/euAUeKZo8fstDmApUqmkt2WlXW/T0G0xSaFtByV0KcqfLyiR4lRM9wsuoSvJiCRHx1fYgkvo3kroqo+uJUm9hdBhIKkQ/V/z3spK6GY238zWmlmbmS0uMP0KM1ttZs+Y2f1mNj3+UEVE4pW2M5KSCd3MGoDrgQXAHOACM5vTZ7angBbn3B8DdwDfjjvQpIX2TRyHpCteAVTsEuHbnQWr6bYY15r4tUXSq5wa+glAm3NunXOuC7gFWJQ/g3PuQefczujl48CUeMMUkRAocddXOQl9MvBq3uv2aFwxlwC/LTTBzC41s1Yza+3s7Cw/ynJUUZLiOu3y5eyt4EOiK7jboh4SnX7qtti/0LZDrBdFzexCoAX4TqHpzrmbnHMtzrmWpqamOD86tfILVFy3+ZXa065LlrZvTmMZ82wEpua9nhKN68XMPgpcCZzmnHsvnvBERKRc5dTQlwGzzGyGmQ0BzgeW5M9gZscBNwILnXMd8Ycp9eDbxT2RQlRK9yuZ0J1ze4DLgbuBNcBtzrlVZnaNmS2MZvsOMAK43cyeNrMlRRYXAz/PrVSoJK2yXLZDW/dymlxwzi0FlvYZd1Xe8Edjjqv2PN1zyYbV/9KrvSia1Rq+b901q9kPhdbFOTfg6zlJ/zjJt21eLwH+UjT+PRdXbw5fLloO4JGi/a65J6sj0q9qimmpYz+0QyDAhC4iPgnh1gBZoYQuIpISSugiNaAfaUktpCahZ/UCnIhIj9Qk9HpTM6JI+oR2WCuhR0LbcXHQ3RaT4dtqV7MfCp35VrI837ZJWgWY0ONvi8zGzbmK3ISp35tziWT3ixnCOwYCTOjZkuWDSaRcOkxylNBFpCohJ9O0/XguNQld3cJEJOtSk9Alfr7cyiANtCkTpI27jxJ6pNqfL4fYDz7piMPbIvEI6afwJZ8pWujmXJV8TjibpJfQwlZCFxFJidQk9GpqyHGdsPly5lfoekLxuy0WD9qT1ZEE6Zmi/QttO6QmoWdBrQtXaKebklGhtuckILiE7kstuFZUVKWWCl0ID+H6kHJ6TnAJXUT8EnIyTVv9UAldRCQllNAj1VYyQqyl6DmPySi03vWsCfa3H0p2Wyy4vOzcnSu0sJXQRURSQgmdOC+0+tEiV/hui+XPu29aPOGIx9RtsX+hbQcl9ICEdvonIrWlhC7FhVY9kWzKWl/mfqQmoWf1ApxI4nRsBSO4hJ617+KQbvQk4avk+PLhh0c+xOCD4BJ6UqrNmyHm3eTvthjgRkmh/vZD6bstFnimaMwx1FOp1ho/oy5OCV2kBtTMK7WghE58D3Lw5aAtFEYl3dM8WR2pA+37nNC2gxK6iEhKKKGLiKREWQndzOab2VozazOzxQWmDzWzW6PpT5hZc+yRiohIv0omdDNrAK4HFgBzgAvMbE6f2S4BtjrnjgD+EfhW3IGKiEj/GsuY5wSgzTm3DsDMbgEWAavz5lkEXB0N3wH8s5mZS6AT9ZDGwt9Bwwc3VLXcXbu7OfO7D1X8/g1v7mTymOFlzz+4wdjdXXrzDG5IplXszO8+xPZduwtOG3vwEN7c0cWwwQ1sf3dPxZ9x67JXeWBNR8Xv7081+yppF/3oiQPG/cM9L/DvT782oOUUWsfxI4b2et2zr/pz2S9WMKyx8PHxdp/9e81dq7n5yVd4seMdAL5+57MM7XNsLfj+w70uFm7a/t6+4T+7/hEaCvQO6N6bXAfAoY2DGFRhj4QfPvwyv1vbWXR6z3aAeMvcF86YxZ++77DYltejnIQ+GXg173U7cGKxeZxze8zsLWAcsDl/JjO7FLgUYNq0aRUFfMWZs3l3dzevvLmTR9q2cM4xhzL3sNGcPXdiRcsDOPeYSWzctquqH/HMmjiC846fwidOns5FP3oSgAtPmsZ7u/fSU5Zf27aLx9ZtAeDhr3yEL9z8FCte2crSL36Is/7x972Wt/QLHwLg/BOm8n8f38BfnX44DYOML599JCs2bOXCk6fT+fb+A+nGi+Zxx/J2Pnva4YDjgbUdTBt70L7po4Y38pkPzeCXKzYyYeRQZjYdzOyJI1l47P5C9c1Fczl26hi27HiPe1dv4nMfPoKfPboeB8w9bBSvbNnJ/773BQAWLziKvc7xUscOfrminY/Pm8K4EUO54aGX+Pi8KUwaPYy2zv0Hw/44BrN8w9aKt3P+9t66czeb39m/DaaMGU771l1VLff0I5v6PcCLmT7uILr3Orbu6GLOYaNYv2Vnr+lnHDWBoYPL+3Ju37qLru69zJo4Yt+4nsTy0798f695f/iJeTzatqXgcr75Z0fz33/9HMdMHt3v5y199o19wx/9o4kMaTQmjxnO79Z2cuy0QwB4efMOhjQOomvPXmbnxQUwfEgDz7S/BcBRh44s+jnrNu8A4B8+/r4Dpv3dfzqaK3/1XC7uRXMBuPbPj2Hxnc8ypHEQ3/8vx/LFW5+ma8/eA9573fnHMXHUUGZNGMElH5xBt3P8/LENfHX+UVy15DlefXMXnz3tcG546KUD3nv6kU0cNKR4ZXDG+IO5Z/UmhjQM6rU/qjV6+ODYlpXPSiUxM/sYMN859+no9UXAic65y/PmeS6apz16/VI0z+ZCywRoaWlxra2tMayCiEh2mNly51xLoWnlVBk2AlPzXk+JxhWcx8wagdFA4WqDiIgkopyEvgyYZWYzzGwIcD6wpM88S4CLo+GPAQ8k0X4uIiLFlWxDj9rELwfuBhqAHzvnVpnZNUCrc24J8CPg52bWBrxJLumLiEgNlXNRFOfcUmBpn3FX5Q2/C3w83tBERGQg9EtREZGUUEIXEUkJJXQRkZRQQhcRSYmSPyxK7IPNOoENFb59PH1+heq5kOINKVYIK96QYoWw4g0pVqgu3unOuaZCE+qW0KthZq3Ffinlo5DiDSlWCCvekGKFsOINKVZILl41uYiIpIQSuohISoSa0G+qdwADFFK8IcUKYcUbUqwQVrwhxQoJxRtkG7qIiBwo1Bq6iIj0oYQuIpISwSX0Ug+sTvBzf2xmHdHDPHrGjTWze83sxej/mGi8mdl1UYzPmNnxee+5OJr/RTO7OG/8PDN7NnrPdWYVPlMrt6ypZvagma02s1Vm9kXP4x1mZk+a2coo3m9E42dEDx1vix5CPiQaX/Sh5Gb2tWj8WjM7O298rOXGzBrM7CkzuyuAWNdH++ppM2uNxnlZFqLlHWJmd5jZ82a2xsxO9jFeMzsy2qY9f9vN7K/rGqtzLpg/crfvfQmYCQwBVgJzavTZpwLHA8/ljfs2sDgaXgx8Kxo+B/gtYMBJwBPR+LHAuuj/mGh4TDTtyWhei967oIpYJwHHR8MjgRfIPeDb13gNGBENDwaeiJZ9G3B+NP4G4K+i4c8BN0TD5wO3RsNzojIxFJgRlZWGJMoNcAXwC+Cu6LXPsa4HxvcZ52VZiJb3U+DT0fAQ4BCf442W2QC8AUyvZ6yJJ8I4/4CTgbvzXn8N+FoNP7+Z3gl9LTApGp4ErI2GbwQu6DsfcAFwY974G6Nxk4Dn88b3mi+GuP8dODOEeIGDgBXknlu7GWjsu+/J3Zv/5Gi4MZrP+paHnvniLjfkntp1P/AR4K7os72MNVrGeg5M6F6WBXJPO3uZqMOG7/HmLecs4JF6xxpak0uhB1ZPrlMsABOdc69Hw28APU+qLhZnf+PbC4yvWnSKfxy5Wq+38UZNGE8DHcC95Gqp25xzPY+lz/+MXg8lB3oeSj7Q9ajU94CvAD1PLB7ncawADrjHzJZb7kHt4G9ZmAF0Av8aNWn9i5kd7HG8Pc4Hbo6G6xZraAndWy73FepVH1AzGwH8Evhr59z2/Gm+xeuc63bOHUuu9nsCcFR9IyrMzP4E6HDOLa93LAPwQefc8cAC4DIzOzV/omdloZFc0+YPnHPHATvINVvs41m8RNdLFgK3951W61hDS+jlPLC6ljaZ2SSA6H9HNL5YnP2Nn1JgfMXMbDC5ZP5vzrk7fY+3h3NuG/AguaaHQyz30PG+n1HsoeQDXY9KfABYaGbrgVvINbt839NYAXDObYz+dwC/IveF6WtZaAfanXNPRK/vIJfgfY0Xcl+UK5xzm6LX9Yu12rajWv6R+/ZeR+60rOeC0dwafn4zvdvQv0Pvix/fjobPpffFjyej8WPJtQ+Oif5eBsZG0/pe/DinijgN+BnwvT7jfY23CTgkGh4OPAz8CbkaT/6Fxs9Fw5fR+0LjbdHwXHpfaFxH7mJVIuUGOJ39F0W9jBU4GBiZN/woMN/XshAt72HgyGj46ihWn+O9BfiUD8dZTRJhnH/krhS/QK6N9coafu7NwOvAbnK1iEvItYXeD7wI3Je3Ewy4PorxWaAlbzl/CbRFf/mFoAV4LnrPP9PnotAAY/0gudO8Z4Cno79zPI73j4GnonifA66Kxs+MCnQbuYQ5NBo/LHrdFk2fmbesK6OY1pLXIyCJckPvhO5lrFFcK6O/VT3L87UsRMs7FmiNysOvySU5L+Ml9yW5BRidN65useqn/yIiKRFaG7qIiBShhC4ikhJK6CIiKaGELiKSEkroIiIpoYQuIpISSugiIinxH6xNGFA4mvr3AAAAAElFTkSuQmCC\n", "text/plain": [ "
" ] }, "metadata": { "needs_background": "light" }, "output_type": "display_data" } ], "source": [ "plt.plot(merged_arr)" ] }, { "cell_type": "code", "execution_count": 15, "id": "143dcdca", "metadata": {}, "outputs": [], "source": [ "# only keep slices of min duration\n", "\n", "music_arrays = []\n", "\n", "for n, is_music in enumerate(merged_arr):\n", " new_is_music = is_music\n", " if n - last_n <= min_music_duration_f:\n", " new_is_music = True\n", " if is_music:\n", " last_n = n\n", " merged_arr.append(new_is_music)" ] }, { "cell_type": "code", "execution_count": 123, "id": "69c3d2c8", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "podcast_id 1739\n", "audio_url https://eirecana.jellycast.com/files/audio/Epi...\n", "uuid b9da5193-ab1f-4c03-94f1-38959badcb07\n", "filesize_bytes 139857145\n", "title Episode XVI: Simply Good Music\n", "summary The latest episode of the Eirecana podcast fea...\n", "tags alt-country;irish;blues;folk;americana\n", "published_date 2017-08-07\n", "filename b9da5193-ab1f-4c03-94f1-38959badcb07.mp3\n", "Name: 0, dtype: object" ] }, "execution_count": 123, "metadata": {}, "output_type": "execute_result" } ], "source": [ "row" ] }, { "cell_type": "code", "execution_count": null, "id": "9d42a9d3", "metadata": {}, "outputs": [], "source": [ "save in nemo format\n", "# {\"audio_filepath\": \"/path/to/audio_file\", \"offset\": 0, \"duration\": 3.2, \"label\": \"music\", \"text\": \"-\"}" ] }, { "cell_type": "code", "execution_count": null, "id": "87998470", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "ed337ea9", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "1882683e", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "6ee5c327", "metadata": {}, "source": [ "## Training" ] }, { "cell_type": "code", "execution_count": null, "id": "f61f19dc", "metadata": {}, "outputs": [], "source": [ "docker run --gpus all \\\n", " -it \\\n", " -v /home/georg/code/NeMo:/NeMo \\\n", " -v /home/georg:/home/georg \\\n", " -v /mnt/data-ssd-1:/mnt/data-ssd-1 \\\n", " -v /mnt/data-ssd-2:/mnt/data-ssd-2 \\\n", " -p 8339:8339 \\\n", " --shm-size=8g \\\n", " --ulimit memlock=-1 \\\n", " --ulimit stack=67108864 \\\n", " nvcr.io/nvidia/pytorch:22.04-py3" ] }, { "cell_type": "code", "execution_count": null, "id": "b34c831b", "metadata": {}, "outputs": [], "source": [ "CUDA_VISIBLE_DEVICES=0,1,2,3 python examples/asr/speech_classification/speech_to_label.py \\\n", " --config-path=/NeMo/examples/asr/conf/marblenet\n", " --config-name=marblenet_3x2x64.yaml \\\n", " model.train_ds.manifest_filepath=/mnt/data-ssd-1/data/ljspeech/LJSpeech-1.1/ljspeech_train.json \\\n", " model.validation_ds.manifest_filepath=/mnt/data-ssd-1/data/ljspeech/LJSpeech-1.1/ljspeech_val.json \\\n", " model.train_ds.dataloader_params.batch_size=24 \\\n", " model.validation_ds.dataloader_params.batch_size=8 \\\n", " model.train_ds.dataloader_params.num_workers=12 \\\n", " model.validation_ds.dataloader_params.num_workers=8 \\\n", " trainer.max_epochs=200 \\\n", " trainer.check_val_every_n_epoch=1 \\\n", " model.optim.lr=1e-1 \\\n", " exp_manager.create_wandb_logger=True \\\n", " exp_manager.wandb_logger_kwargs.name=\"test_0\" \\\n", " exp_manager.wandb_logger_kwargs.project=\"vad_style_classification\"\n", " \n", " \n", "# sup_data_path=/mnt/data-ssd-1/tmp/nemo_sup_data \\\n", "# exp_manager.exp_dir=/mnt/data-ssd-1/checkpoints \\\n", "# +trainer.precision=16 \\\n", "# +trainer.amp_level=O1 # needed if using PyTorch < 1.6\n", "# trainer.devices=2 \\\n", "# trainer.accelerator=\"gpu\" \\\n", "# strategy=\"ddp\" \\\n", "# +model.train_ds.num_worker= \\\n", "# matchboxnet\n", "# model.train_ds.is_tarred=False \\\n", "# model.train_ds.tarred_audio_filepaths= \\" ] }, { "cell_type": "code", "execution_count": null, "id": "36057c73", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "ca250885", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "c08bfafc", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "6dc9e23a", "metadata": {}, "source": [ "## Use models" ] }, { "cell_type": "code", "execution_count": null, "id": "bb7b10bf", "metadata": {}, "outputs": [], "source": [ "# https://catalog.ngc.nvidia.com/orgs/nvidia/collections/nemo_asr/entities\n", "\n", "# TODO: use combo of the below to make it usable for inference \n", "# (and remove fraim append, instead return probs by frame)\n", "# (maybe also look at .transcribe)\n", "# https://github.com/NVIDIA/NeMo/blob/main/tutorials/asr/Online_Offline_Microphone_VAD_Demo.ipynb\n", "# https://github.com/NVIDIA/NeMo/blob/main/examples/asr/speech_classification/vad_infer.py" ] }, { "cell_type": "code", "execution_count": 60, "id": "c0520797", "metadata": {}, "outputs": [], "source": [ "# import nemo\n", "# import nemo.collections.asr as nemo_asr\n", "# from suno_utils.utils.display import suppress_logging\n", "\n", "# nemo_asr.models.EncDecClassificationModel.list_available_models()\n", "\n", "# # /home/georg/.cache/torch/NeMo/NeMo_1.10.0rc0/vad_telephony_marblenet/44c033a250118838ab1637e5a7ba06c2/vad_telephony_marblenet.nemo\n", "# with suppress_logging():\n", "# vad_model = nemo_asr.models.EncDecClassificationModel.from_pretrained(\"vad_telephony_marblenet\")\n", " \n", "# filepath = \"/mnt/data-ssd-1/data/southern-comfort/output/segments/01a36591-fd14-4d1e-9f51-099e938f436d.wav\"\n", "# vad_model.transcribe([filepath], logprobs=True)" ] }, { "cell_type": "code", "execution_count": 47, "id": "c46f1b0b", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "{\"audio_filepath\": \"/home/georg/notebooks/tasks/speech_classification/sample_data/data/chris-sample01_02.wav\", \"duration\": 123.7388125, \"offset\": 0,\"text\": \"this is the first sample using a headset this is the second sample using a headset\"}\r\n", "{\"audio_filepath\": \"/home/georg/notebooks/tasks/speech_classification/sample_data/data/chris-sample03.wav\", \"duration\": 63.464, \"offset\": 0, \"text\": \"this is the third sample using a laptop microphone\"}" ] } ], "source": [ "!cat /home/georg/notebooks/tasks/speech_classification/sample_data/chris_demo.json" ] }, { "cell_type": "code", "execution_count": 63, "id": "af923511", "metadata": {}, "outputs": [ { "data": { "text/html": [ "\n", " \n", " " ], "text/plain": [ "" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "Audio.from_file(\"/home/georg/notebooks/tasks/speech_classification/sample_data/data/chris-sample01_02.wav\").play()" ] }, { "cell_type": "code", "execution_count": null, "id": "a152f95e", "metadata": {}, "outputs": [], "source": [ "!rm -rf /home/georg/notebooks/tasks/speech_classification/tmp/*" ] }, { "cell_type": "code", "execution_count": null, "id": "0fdb1bf3", "metadata": {}, "outputs": [], "source": [ "CUDA_VISIBLE_DEVICES=\"3\" python /home/georg/code/NeMo/examples/asr/speech_classification/vad_infer.py \\\n", " --config-path=\"/home/georg/code/NeMo/examples/asr/conf/vad\" \\\n", " --config-name=\"vad_inference_postprocessing.yaml\" \\\n", " dataset=\"/home/georg/notebooks/tasks/speech_classification/sample_data/chris_demo.json\" \\\n", " vad.model_path=\"vad_telephony_marblenet\" \\\n", " frame_out_dir=\"/home/georg/notebooks/tasks/speech_classification/tmp/chris_demo_out\" \\\n", " vad.parameters.window_length_in_sec=0.63 \\\n", " gen_seg_table=False \\\n", " write_to_manifest=False \\\n", " out_manifest_filepath=\"/home/georg/notebooks/tasks/speech_classification/tmp/vad_out.json\"\n", "\n", "# vad.parameters.window_length_in_sec=0.63/31/15 \\\n", "# gen_seg_table=False \\\n", "# write_to_manifest=False \\\n", "# vad.parameters.postprocessing.onset=0.25 \\\n", "# vad.parameters.postprocessing.offset=0.25 \\\n", "# vad.parameters.postprocessing.min_duration_on=0.25 \\\n", "# vad.parameters.postprocessing.min_duration_off=0.25 \\" ] }, { "cell_type": "code", "execution_count": 119, "id": "bdd0fa86", "metadata": {}, "outputs": [ { "data": { "image/png": "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\n", "text/plain": [ "
" ] }, "metadata": { "needs_background": "light" }, "output_type": "display_data" } ], "source": [ "# careful, this will append... wtf??\n", "import pandas as pd\n", "with open(\"/home/georg/notebooks/tasks/speech_classification/tmp/chris_demo_out/chris-sample01_02.frame\") as f:\n", " a = 1 - np.array([float(n) for n in f.read().strip().split(\"\\n\")])\n", "duration_s = 123.7388125\n", "idx = np.linspace(0, duration_s, len(a))\n", "s = pd.Series(a, index=idx).rolling(100, center=True).mean()\n", "s.plot();" ] }, { "cell_type": "code", "execution_count": 112, "id": "806e3bfb", "metadata": {}, "outputs": [], "source": [ "# with open(\"/home/georg/notebooks/tasks/speech_classification/tmp/vad_out.json\") as f:\n", "# d = [json.loads(l) for l in f.read().strip().split(\"\\n\")]\n", "# d" ] }, { "cell_type": "code", "execution_count": null, "id": "fbf09576", "metadata": {}, "outputs": [], "source": [ "# sample_data/chris_demo.json\n", "\n", "# [{'audio_filepath': 'data/chris-sample01_02.wav',\n", "# 'duration': 123.7388125,\n", "# 'offset': 0,\n", "# 'text': 'this is the first sample using a headset this is the second sample using a headset'},\n", "# {'audio_filepath': 'data/chris-sample03.wav',\n", "# 'duration': 63.464,\n", "# 'offset': 0,\n", "# 'text': 'this is the third sample using a laptop microphone'}]" ] }, { "cell_type": "code", "execution_count": null, "id": "925bdbec", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "f2cf856e", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "82878d28", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "dc1060ee", "metadata": {}, "source": [ "## Playground" ] }, { "cell_type": "markdown", "id": "2a2b121d", "metadata": {}, "source": [ "## A) LJSpeech Fastpitch" ] }, { "cell_type": "markdown", "id": "efce9a11", "metadata": {}, "source": [ "### 1) From scratch" ] }, { "cell_type": "code", "execution_count": null, "id": "998146a5", "metadata": {}, "outputs": [], "source": [ "CUDA_VISIBLE_DEVICES=0,1,2,3 python examples/tts/fastpitch.py \\\n", "CUDA_VISIBLE_DEVICES=0 python examples/tts/fastpitch.py \\\n", " --config-path=/NeMo/examples/tts/conf/ \\\n", " --config-name=fastpitch_align_v1.05 \\\n", " train_dataset=/mnt/data-ssd-1/data/ljspeech/LJSpeech-1.1/ljspeech_train.json \\\n", " validation_datasets=/mnt/data-ssd-1/data/ljspeech/LJSpeech-1.1/ljspeech_val.json \\\n", " trainer.max_epochs=1000 \\\n", " trainer.check_val_every_n_epoch=1 \\\n", " model.optim.lr=1e-1 \\\n", " model.train_ds.dataloader_params.batch_size=24 \\\n", " model.validation_ds.dataloader_params.batch_size=8 \\\n", " model.train_ds.dataloader_params.num_workers=12 \\\n", " model.validation_ds.dataloader_params.num_workers=8 \\\n", " sup_data_path=/mnt/data-ssd-1/tmp/nemo_sup_data \\\n", " exp_manager.exp_dir=/mnt/data-ssd-1/checkpoints \\\n", " +exp_manager.create_wandb_logger=True \\\n", " +exp_manager.wandb_logger_kwargs.project=\"tts_demo\" \\\n", " +exp_manager.wandb_logger_kwargs.name=\"base\"" ] }, { "cell_type": "code", "execution_count": null, "id": "65a83bd7", "metadata": {}, "outputs": [], "source": [ "# large batch: first epoch ~15 mins\n", "# 4 gpu: stalls after a while with 3/4 gpus busy\n", "# 50s per epoch at 32 batch & 4gpu\n", "# 56s per epoch at 24 batch & 4gpu\n", "# 2m20s per epoch at 24 batch & 1gpu\n", "\n", "# ~20h train time for LJSpeech" ] }, { "cell_type": "markdown", "id": "31faf17b", "metadata": {}, "source": [ "### 2) Finetune" ] }, { "cell_type": "code", "execution_count": null, "id": "aff146cf", "metadata": {}, "outputs": [], "source": [ "# https://github.com/NVIDIA/NeMo/blob/main/tutorials/tts/FastPitch_Finetuning.ipynb" ] }, { "cell_type": "code", "execution_count": null, "id": "569a44e2", "metadata": {}, "outputs": [], "source": [ "# CUDA_VISIBLE_DEVICES=7 python /NeMo/examples/tts/fastpitch_finetune_custom.py \\\n", "#TODO\n", "# CUDA_VISIBLE_DEVICES=4,5,6,7 python /NeMo/examples/tts/fastpitch_finetune.py \\\n", "# --config-path=/data-ssd-2/georg/models/tts_en_fastpitch/ \\\n", "# --config-name=custom_finetune \\\n", "# train_dataset=/data-ssd-2/georg/data/spgi_mod/spgispeech_train.json \\\n", "# validation_datasets=/data-ssd-2/georg/data/spgi_mod/spgispeech_val.json \\\n", "# model.pitch_avg=131.0 \\\n", "# model.pitch_std=30.0 \\\n", "# model.pitch_fmin=30 \\\n", "# model.pitch_fmax=512 \\\n", "# model.n_speakers=1 \\\n", "# ~model.optim.sched \\\n", "# model.optim.name=adam \\\n", "# trainer.max_epochs=1000 \\\n", "# trainer.check_val_every_n_epoch=10 \\\n", "# model.optim.lr=2e-4 \\\n", "# model.train_ds.dataloader_params.batch_size=24 \\\n", "# model.validation_ds.dataloader_params.batch_size=24 \\\n", "# exp_manager.exp_dir=/data-ssd-2/georg/tmp/stuff_ft \\\n", "# prior_folder=/data-ssd-2/georg/tmp/prior \\\n", "# +exp_manager.create_wandb_logger=True \\\n", "# +exp_manager.wandb_logger_kwargs.project=\"tts_demo\" \\\n", "# +exp_manager.wandb_logger_kwargs.name=\"ft_base\"" ] }, { "cell_type": "code", "execution_count": null, "id": "b64b4ab2", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "b76fe1a8", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "edd98df9", "metadata": {}, "source": [ "# Test checkpoint" ] }, { "cell_type": "code", "execution_count": 10, "id": "c4c5bb08", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "[NeMo W 2021-12-13 19:11:04 optimizers:50] Apex was not found. Using the lamb or fused_adam optimizer will error out.\n", "################################################################################\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)" ] }, { "name": "stdout", "output_type": "stream", "text": [ "[NeMo I 2021-12-13 19:11:04 cloud:56] Found existing object /home/georg/.cache/torch/NeMo/NeMo_1.6.0rc0/tts_en_fastpitch_align/b50e16c5d695b00855ae53d6ba4e4f7f/tts_en_fastpitch_align.nemo.\n", "[NeMo I 2021-12-13 19:11:04 cloud:62] Re-using file from: /home/georg/.cache/torch/NeMo/NeMo_1.6.0rc0/tts_en_fastpitch_align/b50e16c5d695b00855ae53d6ba4e4f7f/tts_en_fastpitch_align.nemo\n", "[NeMo I 2021-12-13 19:11:04 common:729] Instantiating model from pre-trained checkpoint\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "\n", "################################################################################\n", "\n", "[NeMo W 2021-12-13 19:11:06 modelPT:130] If you intend to do training or fine-tuning, please call the ModelPT.setup_training_data() method and provide a valid configuration file to setup the train data loader.\n", " Train config : \n", " dataset:\n", " _target_: nemo.collections.asr.data.audio_to_text.AudioToCharWithPriorAndPitchDataset\n", " manifest_filepath: /raid/LJSpeech/nvidia_ljspeech_train.json\n", " max_duration: null\n", " min_duration: 0.1\n", " int_values: false\n", " normalize: true\n", " sample_rate: 22050\n", " trim: false\n", " sup_data_path: /raid/LJSpeech/prior\n", " n_window_stride: 256\n", " n_window_size: 1024\n", " pitch_fmin: 80\n", " pitch_fmax: 640\n", " pitch_avg: 211.27540199742586\n", " pitch_std: 52.1851002822779\n", " vocab:\n", " notation: phonemes\n", " punct: true\n", " spaces: true\n", " stresses: true\n", " add_blank_at: None\n", " pad_with_space: true\n", " chars: true\n", " improved_version_g2p: true\n", " dataloader_params:\n", " drop_last: false\n", " shuffle: true\n", " batch_size: 32\n", " num_workers: 12\n", " \n", "[NeMo W 2021-12-13 19:11:06 modelPT:137] If you intend to do validation, please call the ModelPT.setup_validation_data() or ModelPT.setup_multiple_validation_data() method and provide a valid configuration file to setup the validation data loader(s). \n", " Validation config : \n", " dataset:\n", " _target_: nemo.collections.asr.data.audio_to_text.AudioToCharWithPriorAndPitchDataset\n", " manifest_filepath: /raid/LJSpeech/nvidia_ljspeech_val.json\n", " max_duration: null\n", " min_duration: null\n", " int_values: false\n", " normalize: true\n", " sample_rate: 22050\n", " trim: false\n", " sup_data_path: /raid/LJSpeech/prior\n", " n_window_stride: 256\n", " n_window_size: 1024\n", " pitch_fmin: 80\n", " pitch_fmax: 640\n", " pitch_avg: 211.27540199742586\n", " pitch_std: 52.1851002822779\n", " vocab:\n", " notation: phonemes\n", " punct: true\n", " spaces: true\n", " stresses: true\n", " add_blank_at: None\n", " pad_with_space: true\n", " chars: true\n", " improved_version_g2p: true\n", " dataloader_params:\n", " drop_last: false\n", " shuffle: false\n", " batch_size: 32\n", " num_workers: 8\n", " \n", "[NeMo W 2021-12-13 19:11:06 fastpitch:108] AudioToCharWithPriorAndPitchDataset will be deprecated in 1.8 version. Please change your model to use Torch TTS Collection instead (e.g. see nemo.collections.tts.torch.data.TTSDataset).\n", "[NeMo E 2021-12-13 19:11:06 vocabs:323] Torch distributed needs to be initialized before you initialized . This class is prone to data access race conditions. Now downloading corpora from global rank 0. If other ranks pass this before rank 0, errors might result.\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "[NeMo I 2021-12-13 19:11:08 features:264] PADDING: 1\n", "[NeMo I 2021-12-13 19:11:08 features:281] STFT using torch\n", "[NeMo I 2021-12-13 19:11:09 save_restore_connector:149] Model FastPitchModel was successfully restored from /home/georg/.cache/torch/NeMo/NeMo_1.6.0rc0/tts_en_fastpitch_align/b50e16c5d695b00855ae53d6ba4e4f7f/tts_en_fastpitch_align.nemo.\n", "[NeMo I 2021-12-13 19:11:09 cloud:56] Found existing object /home/georg/.cache/torch/NeMo/NeMo_1.6.0rc0/tts_hifigan/e6da322f0f7e7dcf3f1900a9229a7e69/tts_hifigan.nemo.\n", "[NeMo I 2021-12-13 19:11:09 cloud:62] Re-using file from: /home/georg/.cache/torch/NeMo/NeMo_1.6.0rc0/tts_hifigan/e6da322f0f7e7dcf3f1900a9229a7e69/tts_hifigan.nemo\n", "[NeMo I 2021-12-13 19:11:09 common:729] Instantiating model from pre-trained checkpoint\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "[NeMo W 2021-12-13 19:11:11 modelPT:130] If you intend to do training or fine-tuning, please call the ModelPT.setup_training_data() method and provide a valid configuration file to setup the train data loader.\n", " Train config : \n", " dataset:\n", " _target_: nemo.collections.tts.data.datalayers.MelAudioDataset\n", " manifest_filepath: /home/fkreuk/data/train_finetune.txt\n", " min_duration: 0.75\n", " n_segments: 8192\n", " dataloader_params:\n", " drop_last: false\n", " shuffle: true\n", " batch_size: 64\n", " num_workers: 4\n", " \n", "[NeMo W 2021-12-13 19:11:11 modelPT:137] If you intend to do validation, please call the ModelPT.setup_validation_data() or ModelPT.setup_multiple_validation_data() method and provide a valid configuration file to setup the validation data loader(s). \n", " Validation config : \n", " dataset:\n", " _target_: nemo.collections.tts.data.datalayers.MelAudioDataset\n", " manifest_filepath: /home/fkreuk/data/val_finetune.txt\n", " min_duration: 3\n", " n_segments: 66150\n", " dataloader_params:\n", " drop_last: false\n", " shuffle: false\n", " batch_size: 5\n", " num_workers: 4\n", " \n", "[NeMo W 2021-12-13 19:11:11 features:241] Using torch_stft is deprecated and will be removed in 1.1.0. Please set stft_conv and stft_exact_pad to False for FilterbankFeatures and AudioToMelSpectrogramPreprocessor. Please set exact_pad to True as needed.\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "[NeMo I 2021-12-13 19:11:11 features:264] PADDING: 0\n", "[NeMo I 2021-12-13 19:11:11 features:281] STFT using torch\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "[NeMo W 2021-12-13 19:11:11 features:241] Using torch_stft is deprecated and will be removed in 1.1.0. Please set stft_conv and stft_exact_pad to False for FilterbankFeatures and AudioToMelSpectrogramPreprocessor. Please set exact_pad to True as needed.\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "[NeMo I 2021-12-13 19:11:11 features:264] PADDING: 0\n", "[NeMo I 2021-12-13 19:11:11 features:281] STFT using torch\n", "[NeMo I 2021-12-13 19:11:16 save_restore_connector:149] Model HifiGanModel was successfully restored from /home/georg/.cache/torch/NeMo/NeMo_1.6.0rc0/tts_hifigan/e6da322f0f7e7dcf3f1900a9229a7e69/tts_hifigan.nemo.\n" ] } ], "source": [ "from nemo.collections.tts.models import FastPitchModel\n", "from nemo.collections.tts.models import HifiGanModel\n", "\n", "spec_model = FastPitchModel.from_pretrained(model_name=\"tts_en_fastpitch\")\n", "vocoder_model = HifiGanModel.from_pretrained(model_name=\"tts_hifigan\")" ] }, { "cell_type": "code", "execution_count": 11, "id": "2082bf5d", "metadata": {}, "outputs": [], "source": [ "import os\n", "os.chdir('/home/georg/code/NeMo')" ] }, { "cell_type": "code", "execution_count": 215, "id": "ba98e29f", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "'FastPitch--v_loss=0.7980-epoch=25.ckpt'\r\n", "'FastPitch--v_loss=0.8053-epoch=20.ckpt'\r\n", "'FastPitch--v_loss=0.8057-epoch=15.ckpt'\r\n", "'FastPitch--v_loss=0.8791-epoch=212-last.ckpt'\r\n" ] } ], "source": [ "!ls /mnt/data-ssd-1/checkpoints/FastPitch/2021-12-12_04-42-46/checkpoints" ] }, { "cell_type": "code", "execution_count": 195, "id": "6126457f", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "[NeMo W 2021-12-15 08:53:13 modelPT:130] If you intend to do training or fine-tuning, please call the ModelPT.setup_training_data() method and provide a valid configuration file to setup the train data loader.\n", " Train config : \n", " dataset:\n", " _target_: nemo.collections.tts.torch.data.TTSDataset\n", " manifest_filepath: /mnt/data-ssd-1/data/nvidia_hifi_tts/custom_fix_6097_manifest_clean_train.json\n", " sample_rate: 22050\n", " sup_data_path: /mnt/data-ssd-1/tmp/nemo_sup_data\n", " sup_data_types:\n", " - duration_prior\n", " - pitch\n", " n_fft: 1024\n", " win_length: 1024\n", " hop_length: 256\n", " window: hann\n", " n_mels: 80\n", " lowfreq: 0\n", " highfreq: 8000\n", " max_duration: null\n", " min_duration: 0.1\n", " ignore_file: null\n", " trim: false\n", " pitch_fmin: 60\n", " pitch_fmax: 640\n", " pitch_norm: true\n", " pitch_avg: 117.1\n", " pitch_std: 26.8\n", " text_normalizer:\n", " _target_: nemo_text_processing.text_normalization.normalize.Normalizer\n", " lang: en\n", " input_case: cased\n", " whitelist: nemo_text_processing/text_normalization/en/data/whitelist_lj_speech.tsv\n", " text_normalizer_call_args:\n", " verbose: false\n", " punct_pre_process: true\n", " punct_post_process: true\n", " text_tokenizer:\n", " _target_: nemo.collections.tts.torch.tts_tokenizers.EnglishPhonemesTokenizer\n", " punct: true\n", " stresses: true\n", " chars: true\n", " space: ' '\n", " silence: null\n", " apostrophe: true\n", " sep: '|'\n", " add_blank_at: null\n", " pad_with_space: true\n", " g2p:\n", " _target_: nemo.collections.tts.torch.g2ps.EnglishG2p\n", " phoneme_dict: scripts/tts_dataset_files/cmudict-0.7b\n", " heteronyms: scripts/tts_dataset_files/heteronyms-030921\n", " dataloader_params:\n", " drop_last: false\n", " shuffle: true\n", " batch_size: 24\n", " num_workers: 12\n", " \n", "[NeMo W 2021-12-15 08:53:13 modelPT:137] If you intend to do validation, please call the ModelPT.setup_validation_data() or ModelPT.setup_multiple_validation_data() method and provide a valid configuration file to setup the validation data loader(s). \n", " Validation config : \n", " dataset:\n", " _target_: nemo.collections.tts.torch.data.TTSDataset\n", " manifest_filepath: /mnt/data-ssd-1/data/nvidia_hifi_tts/custom_6097_manifest_clean_dev.json\n", " sample_rate: 22050\n", " sup_data_path: /mnt/data-ssd-1/tmp/nemo_sup_data\n", " sup_data_types:\n", " - duration_prior\n", " - pitch\n", " n_fft: 1024\n", " win_length: 1024\n", " hop_length: 256\n", " window: hann\n", " n_mels: 80\n", " lowfreq: 0\n", " highfreq: 8000\n", " max_duration: null\n", " min_duration: null\n", " ignore_file: null\n", " trim: false\n", " pitch_fmin: 60\n", " pitch_fmax: 640\n", " pitch_norm: true\n", " pitch_avg: 117.1\n", " pitch_std: 26.8\n", " text_normalizer:\n", " _target_: nemo_text_processing.text_normalization.normalize.Normalizer\n", " lang: en\n", " input_case: cased\n", " whitelist: nemo_text_processing/text_normalization/en/data/whitelist_lj_speech.tsv\n", " text_normalizer_call_args:\n", " verbose: false\n", " punct_pre_process: true\n", " punct_post_process: true\n", " text_tokenizer:\n", " _target_: nemo.collections.tts.torch.tts_tokenizers.EnglishPhonemesTokenizer\n", " punct: true\n", " stresses: true\n", " chars: true\n", " space: ' '\n", " silence: null\n", " apostrophe: true\n", " sep: '|'\n", " add_blank_at: null\n", " pad_with_space: true\n", " g2p:\n", " _target_: nemo.collections.tts.torch.g2ps.EnglishG2p\n", " phoneme_dict: scripts/tts_dataset_files/cmudict-0.7b\n", " heteronyms: scripts/tts_dataset_files/heteronyms-030921\n", " dataloader_params:\n", " drop_last: false\n", " shuffle: false\n", " batch_size: 8\n", " num_workers: 8\n", " \n", "[NeMo W 2021-12-15 08:53:13 g2ps:78] apply_to_oov_word=None, it means that some of words will remain unchanged if they are not handled by one of rule in self.parse_one_word(). It is useful when you use tokenizer with set of phonemes and chars together, otherwise it can be not.\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "[NeMo I 2021-12-15 08:53:13 features:264] PADDING: 1\n", "[NeMo I 2021-12-15 08:53:13 features:281] STFT using torch\n" ] } ], "source": [ "ckpt_dir = \"/mnt/data-ssd-1/checkpoints/FastPitch/2021-12-12_04-42-46/checkpoints/\"\n", "best_ckpt = ckpt_dir + \"FastPitch--v_loss=0.7980-epoch=25.ckpt\"\n", "# ckpt_dir = \"/mnt/data-ssd-1/checkpoints/FastPitch/2021-12-13_16-34-10/checkpoints/\"\n", "# best_ckpt = ckpt_dir + \"FastPitch--v_loss=0.8528-epoch=75-last.ckpt\"\n", "spec_model_custom = FastPitchModel.load_from_checkpoint(best_ckpt)\n", "spec_model_custom.eval();" ] }, { "cell_type": "code", "execution_count": 214, "id": "5ca34e2f", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "'HifiGan--val_loss=0.1321-epoch=39.ckpt'\r\n", "'HifiGan--val_loss=0.1325-epoch=45.ckpt'\r\n", "'HifiGan--val_loss=0.1328-epoch=35.ckpt'\r\n", "'HifiGan--val_loss=0.1356-epoch=46-last.ckpt'\r\n" ] } ], "source": [ "!ls /mnt/data-ssd-1/checkpoints/HifiGan/2021-12-15_15-11-51/checkpoints/" ] }, { "cell_type": "code", "execution_count": 205, "id": "92ef8c25", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "[NeMo W 2021-12-15 14:01:56 modelPT:130] If you intend to do training or fine-tuning, please call the ModelPT.setup_training_data() method and provide a valid configuration file to setup the train data loader.\n", " Train config : \n", " dataset:\n", " _target_: nemo.collections.tts.data.datalayers.AudioDataset\n", " manifest_filepath: /mnt/data-ssd-1/data/nvidia_hifi_tts/custom_fix_6097_manifest_clean_train.json\n", " max_duration: null\n", " min_duration: 0.75\n", " n_segments: 8192\n", " trim: false\n", " dataloader_params:\n", " drop_last: false\n", " shuffle: true\n", " batch_size: 24\n", " num_workers: 12\n", " \n", "[NeMo W 2021-12-15 14:01:56 modelPT:137] If you intend to do validation, please call the ModelPT.setup_validation_data() or ModelPT.setup_multiple_validation_data() method and provide a valid configuration file to setup the validation data loader(s). \n", " Validation config : \n", " dataset:\n", " _target_: nemo.collections.tts.data.datalayers.AudioDataset\n", " manifest_filepath: /mnt/data-ssd-1/data/nvidia_hifi_tts/custom_6097_manifest_clean_dev.json\n", " max_duration: null\n", " min_duration: null\n", " n_segments: -1\n", " trim: false\n", " dataloader_params:\n", " drop_last: false\n", " shuffle: false\n", " batch_size: 8\n", " num_workers: 8\n", " \n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "[NeMo I 2021-12-15 14:01:56 features:264] PADDING: 0\n", "[NeMo I 2021-12-15 14:01:56 features:281] STFT using torch\n", "[NeMo I 2021-12-15 14:01:56 features:283] STFT using exact pad\n", "[NeMo I 2021-12-15 14:01:56 features:264] PADDING: 0\n", "[NeMo I 2021-12-15 14:01:56 features:281] STFT using torch\n", "[NeMo I 2021-12-15 14:01:56 features:283] STFT using exact pad\n" ] } ], "source": [ "# ckpt_dir = \"/mnt/data-ssd-1/checkpoints/HifiGan/2021-12-14_23-47-31/checkpoints/\"\n", "# best_ckpt = ckpt_dir + \"HifiGan--val_loss=0.1442-epoch=60.ckpt\"\n", "\n", "ckpt_dir = \"/mnt/data-ssd-1/checkpoints/HifiGan/2021-12-15_15-11-51/checkpoints/\"\n", "best_ckpt = ckpt_dir + \"HifiGan--val_loss=0.1333-epoch=21.ckpt\"\n", "\n", "vocoder_model_custom = HifiGanModel.load_from_checkpoint(best_ckpt)\n", "vocoder_model_custom.eval();" ] }, { "cell_type": "code", "execution_count": 206, "id": "bba3494f", "metadata": {}, "outputs": [], "source": [ "# durs=None, pitch=None\n", "# spec, _, durs_pred, _, pitch_pred, *_ = new_model(\n", "# text=tokens, durs=durs, pitch=pitch, speaker=None, pace=pace,\n", "# )\n", "\n", "def predict(sentence, pace=1.0, custom_spec=False, custom_vocoder=False):\n", " with torch.no_grad():\n", " if custom_spec:\n", " tokens = spec_model_custom.parse(sentence)\n", "# spec, *_ = spec_model_custom(\n", "# text=tokens, durs=None, pitch=None, speaker=None, pace=pace,\n", "# )\n", " spec = spec_model_custom.generate_spectrogram(\n", " tokens=tokens, pace=pace,\n", " )\n", " else:\n", " tokens = spec_model.parse(sentence)\n", " spec, *_ = spec_model(\n", " text=tokens, durs=None, pitch=None, speaker=None, pace=pace,\n", " )\n", " if custom_vocoder:\n", " pred_audio = vocoder_model_custom.convert_spectrogram_to_audio(spec=spec)\n", " audio = pred_audio.detach().to('cpu').numpy()\n", "# audio_denoised = vocoder_model_custom._bias_denoise(pred_audio, audio_mel[0]).squeeze(1)\n", " else:\n", " pred_audio = vocoder_model.convert_spectrogram_to_audio(spec=spec)\n", " audio = pred_audio.detach().to('cpu').numpy()\n", "# audio_denoised = vocoder_model._bias_denoise(pred_audio, audio_mel[0]).squeeze(1)\n", " \n", " spec = _array_to_numpy(spec)\n", " return spec, audio" ] }, { "cell_type": "code", "execution_count": 211, "id": "b910076c", "metadata": {}, "outputs": [], "source": [ "sentence = \"hey keenan and mikey, i thought a lot last night and i'm pretty sure that gayorg is taller than both of you.\"\n", "\n", "# sentence = \"to mitigate the volatile factors that impact our industry and are beyond our control. We are proud of the exceptional execution by our managers in the field. Our effective working capital management and consistent gross profit margins\"\n", "# sentence = \"The top half of the sales team, that is the top 40 sales professionals, generated revenue\"\n", "# sentence = \"which has a bit of a different margin profile. So while we continue to be pretty strong in terms of growth, we think our margins will be closer to the upper end of our stated range than what they've been over the last couple of quarters.\"\n", "# sentence = \"in terms of additional costs that are going into the business. The only observation that I have about M\"\n", "# sentence = \"I mentioned on our last call, consistent with prior years, we expect our earnings per share to be stronger in the second half of the year due to the normal seasonality of our businesses,\"\n", "\n", "spec, audio = predict(sentence, custom_spec=True, custom_vocoder=True, pace=1.2)\n", "# plt.plot(audio.squeeze())" ] }, { "cell_type": "code", "execution_count": 212, "id": "47771f9c", "metadata": {}, "outputs": [ { "data": { "text/html": [ "\n", " \n", " " ], "text/plain": [ "" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "play_array(audio)" ] }, { "cell_type": "code", "execution_count": null, "id": "9af5a2ea", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "8b955bd9", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "98888f1d", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "62db0cd7", "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 }