{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"I improved the score of [this public notebook](https://www.kaggle.com/code/nischaydnk/bengali-finetuning-baseline-wav2vec2-inference) by simply training a 5-gram LM of my own. The training notebook can be found [here](https://www.kaggle.com/code/umongsain/build-an-n-gram-with-kenlm-macro).","metadata":{}},{"cell_type":"markdown","source":"## Import","metadata":{}},{"cell_type":"code","source":"!cp -r ../input/python-packages2 ./\n\n!tar xvfz ./python-packages2/jiwer.tgz\n!pip install ./jiwer/jiwer-2.3.0-py3-none-any.whl -f ./ --no-index\n!tar xvfz ./python-packages2/normalizer.tgz\n!pip install ./normalizer/bnunicodenormalizer-0.0.24.tar.gz -f ./ --no-index\n!tar xvfz ./python-packages2/pyctcdecode.tgz\n!pip install ./pyctcdecode/attrs-22.1.0-py2.py3-none-any.whl -f ./ --no-index --no-deps\n!pip install ./pyctcdecode/exceptiongroup-1.0.0rc9-py3-none-any.whl -f ./ --no-index --no-deps\n!pip install ./pyctcdecode/hypothesis-6.54.4-py3-none-any.whl -f ./ --no-index --no-deps\n!pip install ./pyctcdecode/numpy-1.21.6-cp37-cp37m-manylinux_2_12_x86_64.manylinux2010_x86_64.whl -f ./ --no-index --no-deps\n!pip install ./pyctcdecode/pygtrie-2.5.0.tar.gz -f ./ --no-index --no-deps\n!pip install ./pyctcdecode/sortedcontainers-2.4.0-py2.py3-none-any.whl -f ./ --no-index --no-deps\n!pip install ./pyctcdecode/pyctcdecode-0.4.0-py2.py3-none-any.whl -f ./ --no-index --no-deps\n\n!tar xvfz ./python-packages2/pypikenlm.tgz\n!pip install ./pypikenlm/pypi-kenlm-0.1.20220713.tar.gz -f ./ --no-index --no-deps","metadata":{"_kg_hide-output":true,"execution":{"iopub.status.busy":"2023-08-22T13:22:55.627396Z","iopub.execute_input":"2023-08-22T13:22:55.627767Z","iopub.status.idle":"2023-08-22T13:24:07.588817Z","shell.execute_reply.started":"2023-08-22T13:22:55.627736Z","shell.execute_reply":"2023-08-22T13:24:07.587659Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"rm -r python-packages2 jiwer normalizer pyctcdecode pypikenlm","metadata":{"execution":{"iopub.status.busy":"2023-08-22T13:24:07.591411Z","iopub.execute_input":"2023-08-22T13:24:07.592064Z","iopub.status.idle":"2023-08-22T13:24:08.535147Z","shell.execute_reply.started":"2023-08-22T13:24:07.592027Z","shell.execute_reply":"2023-08-22T13:24:08.533835Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import typing as tp\nfrom pathlib import Path\nfrom functools import partial\nfrom dataclasses import dataclass, field\n\nimport pandas as pd\nimport pyctcdecode\nimport numpy as np\nfrom tqdm.notebook import tqdm\n\nimport librosa\n\nimport pyctcdecode\nimport kenlm\nimport torch\nfrom transformers import Wav2Vec2Processor, Wav2Vec2ProcessorWithLM, Wav2Vec2ForCTC\nfrom bnunicodenormalizer import Normalizer\n\nimport cloudpickle as cpkl","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","execution":{"iopub.status.busy":"2023-08-22T13:24:08.537111Z","iopub.execute_input":"2023-08-22T13:24:08.537494Z","iopub.status.idle":"2023-08-22T13:24:20.628571Z","shell.execute_reply.started":"2023-08-22T13:24:08.537459Z","shell.execute_reply":"2023-08-22T13:24:20.627596Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"ROOT = Path.cwd().parent\nINPUT = ROOT / \"input\"\nDATA = INPUT / \"bengaliai-speech\"\nTRAIN = DATA / \"train_mp3s\"\nTEST = DATA / \"test_mp3s\"\n\nSAMPLING_RATE = 16_000\nMODEL_PATH = INPUT / \"bengali-wav2vec2-finetuned/\"\nLM_PATH = INPUT / \"bengali-sr-download-public-trained-models/wav2vec2-xls-r-300m-bengali/language_model/\"","metadata":{"execution":{"iopub.status.busy":"2023-08-22T13:25:20.951854Z","iopub.execute_input":"2023-08-22T13:25:20.952271Z","iopub.status.idle":"2023-08-22T13:25:20.958567Z","shell.execute_reply.started":"2023-08-22T13:25:20.952239Z","shell.execute_reply":"2023-08-22T13:25:20.957411Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### load model, processor, decoder","metadata":{}},{"cell_type":"code","source":"model = Wav2Vec2ForCTC.from_pretrained(MODEL_PATH)\nprocessor = Wav2Vec2Processor.from_pretrained(MODEL_PATH)","metadata":{"execution":{"iopub.status.busy":"2023-08-22T13:25:22.025245Z","iopub.execute_input":"2023-08-22T13:25:22.025625Z","iopub.status.idle":"2023-08-22T13:25:35.741936Z","shell.execute_reply.started":"2023-08-22T13:25:22.025594Z","shell.execute_reply":"2023-08-22T13:25:35.740952Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"vocab_dict = processor.tokenizer.get_vocab()\nsorted_vocab_dict = {k: v for k, v in sorted(vocab_dict.items(), key=lambda item: item[1])}\n\ndecoder = pyctcdecode.build_ctcdecoder(\n    list(sorted_vocab_dict.keys()),\n    \"/kaggle/input/build-an-n-gram-with-kenlm-macro/5gram_correct.arpa\",\n)","metadata":{"execution":{"iopub.status.busy":"2023-08-22T13:25:35.744044Z","iopub.execute_input":"2023-08-22T13:25:35.744452Z","iopub.status.idle":"2023-08-22T13:26:15.747055Z","shell.execute_reply.started":"2023-08-22T13:25:35.744417Z","shell.execute_reply":"2023-08-22T13:26:15.745711Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"processor_with_lm = Wav2Vec2ProcessorWithLM(\n    feature_extractor=processor.feature_extractor,\n    tokenizer=processor.tokenizer,\n    decoder=decoder\n)","metadata":{"execution":{"iopub.status.busy":"2023-08-22T13:26:15.748541Z","iopub.execute_input":"2023-08-22T13:26:15.748978Z","iopub.status.idle":"2023-08-22T13:26:15.761249Z","shell.execute_reply.started":"2023-08-22T13:26:15.748944Z","shell.execute_reply":"2023-08-22T13:26:15.760059Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## prepare dataloader","metadata":{}},{"cell_type":"code","source":"class BengaliSRTestDataset(torch.utils.data.Dataset):\n    \n    def __init__(\n        self,\n        audio_paths: list[str],\n        sampling_rate: int\n    ):\n        self.audio_paths = audio_paths\n        self.sampling_rate = sampling_rate\n        \n    def __len__(self,):\n        return len(self.audio_paths)\n    \n    def __getitem__(self, index: int):\n        audio_path = self.audio_paths[index]\n        sr = self.sampling_rate\n        w = librosa.load(audio_path, sr=sr, mono=False)[0]\n        \n        return w","metadata":{"execution":{"iopub.status.busy":"2023-08-22T13:26:15.763874Z","iopub.execute_input":"2023-08-22T13:26:15.764187Z","iopub.status.idle":"2023-08-22T13:26:15.77371Z","shell.execute_reply.started":"2023-08-22T13:26:15.764151Z","shell.execute_reply":"2023-08-22T13:26:15.772738Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test = pd.read_csv(DATA / \"sample_submission.csv\", dtype={\"id\": str})\nprint(test.head())","metadata":{"execution":{"iopub.status.busy":"2023-08-22T13:26:15.774969Z","iopub.execute_input":"2023-08-22T13:26:15.775386Z","iopub.status.idle":"2023-08-22T13:26:15.811242Z","shell.execute_reply.started":"2023-08-22T13:26:15.775334Z","shell.execute_reply":"2023-08-22T13:26:15.810209Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test_audio_paths = [str(TEST / f\"{aid}.mp3\") for aid in test[\"id\"].values]","metadata":{"execution":{"iopub.status.busy":"2023-08-22T13:26:15.81429Z","iopub.execute_input":"2023-08-22T13:26:15.81455Z","iopub.status.idle":"2023-08-22T13:26:15.821023Z","shell.execute_reply.started":"2023-08-22T13:26:15.814527Z","shell.execute_reply":"2023-08-22T13:26:15.820157Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test_dataset = BengaliSRTestDataset(\n    test_audio_paths, SAMPLING_RATE\n)\n\ncollate_func = partial(\n    processor_with_lm.feature_extractor,\n    return_tensors=\"pt\", sampling_rate=SAMPLING_RATE,\n    padding=True,\n)\n\ntest_loader = torch.utils.data.DataLoader(\n    test_dataset, batch_size=8, shuffle=False,\n    num_workers=2, collate_fn=collate_func, drop_last=False,\n    pin_memory=True,\n)","metadata":{"execution":{"iopub.status.busy":"2023-08-22T13:26:15.824461Z","iopub.execute_input":"2023-08-22T13:26:15.824861Z","iopub.status.idle":"2023-08-22T13:26:15.8313Z","shell.execute_reply.started":"2023-08-22T13:26:15.824837Z","shell.execute_reply":"2023-08-22T13:26:15.830362Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Inference","metadata":{}},{"cell_type":"code","source":"if not torch.cuda.is_available():\n    device = torch.device(\"cpu\")\nelse:\n    device = torch.device(\"cuda\")\nprint(device)","metadata":{"execution":{"iopub.status.busy":"2023-08-22T13:26:15.832507Z","iopub.execute_input":"2023-08-22T13:26:15.833324Z","iopub.status.idle":"2023-08-22T13:26:15.865124Z","shell.execute_reply.started":"2023-08-22T13:26:15.833292Z","shell.execute_reply":"2023-08-22T13:26:15.86417Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"model = model.to(device)\nmodel = model.eval()\nmodel = model.half()","metadata":{"execution":{"iopub.status.busy":"2023-08-22T13:26:15.869811Z","iopub.execute_input":"2023-08-22T13:26:15.870084Z","iopub.status.idle":"2023-08-22T13:26:20.65418Z","shell.execute_reply.started":"2023-08-22T13:26:15.87006Z","shell.execute_reply":"2023-08-22T13:26:20.653153Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"pred_sentence_list = []\n\nwith torch.no_grad():\n    for batch in tqdm(test_loader):\n        x = batch[\"input_values\"]\n        x = x.to(device, non_blocking=True)\n        with torch.cuda.amp.autocast(True):\n            y = model(x).logits\n        y = y.detach().cpu().numpy()\n        \n        for l in y:  \n            sentence = processor_with_lm.decode(l, beam_width=512).text\n            pred_sentence_list.append(sentence)","metadata":{"execution":{"iopub.status.busy":"2023-08-22T13:26:20.655515Z","iopub.execute_input":"2023-08-22T13:26:20.656285Z","iopub.status.idle":"2023-08-22T13:26:34.922258Z","shell.execute_reply.started":"2023-08-22T13:26:20.656249Z","shell.execute_reply":"2023-08-22T13:26:34.921068Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Make Submission","metadata":{}},{"cell_type":"code","source":"bnorm = Normalizer()\n\ndef postprocess(sentence):\n    period_set = set([\".\", \"?\", \"!\", \"।\"])\n    _words = [bnorm(word)['normalized']  for word in sentence.split()]\n    sentence = \" \".join([word for word in _words if word is not None])\n    try:\n        if sentence[-1] not in period_set:\n            sentence+=\"।\"\n    except:\n        # print(sentence)\n        sentence = \"।\"\n    return sentence","metadata":{"execution":{"iopub.status.busy":"2023-08-22T13:26:34.924155Z","iopub.execute_input":"2023-08-22T13:26:34.924539Z","iopub.status.idle":"2023-08-22T13:26:34.931809Z","shell.execute_reply.started":"2023-08-22T13:26:34.924501Z","shell.execute_reply":"2023-08-22T13:26:34.930726Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"pp_pred_sentence_list = [\n    postprocess(s) for s in tqdm(pred_sentence_list)]","metadata":{"execution":{"iopub.status.busy":"2023-08-22T13:26:34.933376Z","iopub.execute_input":"2023-08-22T13:26:34.933712Z","iopub.status.idle":"2023-08-22T13:26:34.97934Z","shell.execute_reply.started":"2023-08-22T13:26:34.93368Z","shell.execute_reply":"2023-08-22T13:26:34.978447Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test[\"sentence\"] = pp_pred_sentence_list\n\ntest.to_csv(\"submission.csv\", index=False)\n\nprint(test.head())","metadata":{"execution":{"iopub.status.busy":"2023-08-22T13:26:34.980826Z","iopub.execute_input":"2023-08-22T13:26:34.981466Z","iopub.status.idle":"2023-08-22T13:26:34.994084Z","shell.execute_reply.started":"2023-08-22T13:26:34.981433Z","shell.execute_reply":"2023-08-22T13:26:34.993015Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## EOF","metadata":{}},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]}]}