{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.10.12","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"gpu","dataSources":[{"sourceId":52324,"databundleVersionId":6229904,"sourceType":"competition"},{"sourceId":4143520,"sourceType":"datasetVersion","datasetId":2447262},{"sourceId":6212334,"sourceType":"datasetVersion","datasetId":3567413},{"sourceId":6637439,"sourceType":"datasetVersion","datasetId":3831703,"isSourceIdPinned":false}],"dockerImageVersionId":30528,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"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":"2024-09-03T22:35:33.730945Z","iopub.execute_input":"2024-09-03T22:35:33.731247Z","iopub.status.idle":"2024-09-03T22:36:44.484873Z","shell.execute_reply.started":"2024-09-03T22:35:33.731208Z","shell.execute_reply":"2024-09-03T22:36:44.483868Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"rm -r python-packages2 jiwer normalizer pyctcdecode pypikenlm","metadata":{"execution":{"iopub.status.busy":"2024-09-03T22:36:44.487092Z","iopub.execute_input":"2024-09-03T22:36:44.487871Z","iopub.status.idle":"2024-09-03T22:36:45.484366Z","shell.execute_reply.started":"2024-09-03T22:36:44.487813Z","shell.execute_reply":"2024-09-03T22:36:45.483280Z"},"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":"2024-09-03T22:36:45.486035Z","iopub.execute_input":"2024-09-03T22:36:45.486416Z","iopub.status.idle":"2024-09-03T22:36:57.827168Z","shell.execute_reply.started":"2024-09-03T22:36:45.486382Z","shell.execute_reply":"2024-09-03T22:36:57.826228Z"},"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 / \"wave-2-vec-2-bengali-ai\"\nLM_PATH = INPUT / \"arijitx-full-model/wav2vec2-xls-r-300m-bengali/language_model\"","metadata":{"execution":{"iopub.status.busy":"2024-09-03T22:36:57.829758Z","iopub.execute_input":"2024-09-03T22:36:57.830071Z","iopub.status.idle":"2024-09-03T22:36:57.836076Z","shell.execute_reply.started":"2024-09-03T22:36:57.830045Z","shell.execute_reply":"2024-09-03T22:36:57.835145Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"model = Wav2Vec2ForCTC.from_pretrained(MODEL_PATH)\nprocessor = Wav2Vec2Processor.from_pretrained(MODEL_PATH)","metadata":{"execution":{"iopub.status.busy":"2024-09-03T22:36:57.837167Z","iopub.execute_input":"2024-09-03T22:36:57.837440Z","iopub.status.idle":"2024-09-03T22:37:11.772645Z","shell.execute_reply.started":"2024-09-03T22:36:57.837417Z","shell.execute_reply":"2024-09-03T22:37:11.771677Z"},"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    str(LM_PATH / \"5gram.bin\"),\n)","metadata":{"execution":{"iopub.status.busy":"2024-09-03T22:37:11.774289Z","iopub.execute_input":"2024-09-03T22:37:11.774890Z","iopub.status.idle":"2024-09-03T22:37:51.403095Z","shell.execute_reply.started":"2024-09-03T22:37:11.774855Z","shell.execute_reply":"2024-09-03T22:37:51.402091Z"},"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":"2024-09-03T22:37:51.404318Z","iopub.execute_input":"2024-09-03T22:37:51.404682Z","iopub.status.idle":"2024-09-03T22:37:51.412970Z","shell.execute_reply.started":"2024-09-03T22:37:51.404650Z","shell.execute_reply":"2024-09-03T22:37:51.412111Z"},"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":"2024-09-03T22:37:51.414089Z","iopub.execute_input":"2024-09-03T22:37:51.414417Z","iopub.status.idle":"2024-09-03T22:37:51.422401Z","shell.execute_reply.started":"2024-09-03T22:37:51.414387Z","shell.execute_reply":"2024-09-03T22:37:51.421507Z"},"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":"2024-09-03T22:37:51.423518Z","iopub.execute_input":"2024-09-03T22:37:51.423788Z","iopub.status.idle":"2024-09-03T22:37:51.447046Z","shell.execute_reply.started":"2024-09-03T22:37:51.423764Z","shell.execute_reply":"2024-09-03T22:37:51.446161Z"},"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":"2024-09-03T22:37:51.449854Z","iopub.execute_input":"2024-09-03T22:37:51.450167Z","iopub.status.idle":"2024-09-03T22:37:51.456000Z","shell.execute_reply.started":"2024-09-03T22:37:51.450142Z","shell.execute_reply":"2024-09-03T22:37:51.455128Z"},"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=16, shuffle=False,\n    num_workers=2, collate_fn=collate_func, drop_last=False,\n    pin_memory=True,\n)","metadata":{"execution":{"iopub.status.busy":"2024-09-03T22:37:51.456941Z","iopub.execute_input":"2024-09-03T22:37:51.457227Z","iopub.status.idle":"2024-09-03T22:37:51.464810Z","shell.execute_reply.started":"2024-09-03T22:37:51.457186Z","shell.execute_reply":"2024-09-03T22:37:51.463887Z"},"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":"2024-09-03T22:37:51.466020Z","iopub.execute_input":"2024-09-03T22:37:51.466784Z","iopub.status.idle":"2024-09-03T22:37:51.498731Z","shell.execute_reply.started":"2024-09-03T22:37:51.466753Z","shell.execute_reply":"2024-09-03T22:37:51.497910Z"},"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":"2024-09-03T22:37:51.499852Z","iopub.execute_input":"2024-09-03T22:37:51.500145Z","iopub.status.idle":"2024-09-03T22:37:52.059397Z","shell.execute_reply.started":"2024-09-03T22:37:51.500105Z","shell.execute_reply":"2024-09-03T22:37:52.058540Z"},"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":"2024-09-03T22:37:52.060940Z","iopub.execute_input":"2024-09-03T22:37:52.061308Z","iopub.status.idle":"2024-09-03T22:38:02.556382Z","shell.execute_reply.started":"2024-09-03T22:37:52.061275Z","shell.execute_reply":"2024-09-03T22:38:02.555293Z"},"trusted":true},"execution_count":null,"outputs":[]},{"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":"2024-09-03T22:38:02.558207Z","iopub.execute_input":"2024-09-03T22:38:02.558500Z","iopub.status.idle":"2024-09-03T22:38:02.564876Z","shell.execute_reply.started":"2024-09-03T22:38:02.558470Z","shell.execute_reply":"2024-09-03T22:38:02.563991Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"pp_pred_sentence_list = [postprocess(s) for s in tqdm(pred_sentence_list)]","metadata":{"execution":{"iopub.status.busy":"2024-09-03T22:38:02.566032Z","iopub.execute_input":"2024-09-03T22:38:02.566311Z","iopub.status.idle":"2024-09-03T22:38:02.600959Z","shell.execute_reply.started":"2024-09-03T22:38:02.566288Z","shell.execute_reply":"2024-09-03T22:38:02.600115Z"},"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\ndisplay(test.head())","metadata":{"execution":{"iopub.status.busy":"2024-09-03T22:38:02.602010Z","iopub.execute_input":"2024-09-03T22:38:02.602251Z","iopub.status.idle":"2024-09-03T22:38:02.618317Z","shell.execute_reply.started":"2024-09-03T22:38:02.602229Z","shell.execute_reply":"2024-09-03T22:38:02.617482Z"},"trusted":true},"execution_count":null,"outputs":[]}]}