{"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":"code","source":"# !pip install datasets evaluate jiwer transformers fastparquet seaborn pydub transformers torchmetrics torch torchaudio torchvision accelerate -q","metadata":{"scrolled":true,"execution":{"iopub.status.busy":"2023-09-30T18:27:16.326698Z","iopub.execute_input":"2023-09-30T18:27:16.327611Z","iopub.status.idle":"2023-09-30T18:27:16.334186Z","shell.execute_reply.started":"2023-09-30T18:27:16.327571Z","shell.execute_reply":"2023-09-30T18:27:16.333063Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!pip install jiwer bitsandbytes","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:16.336558Z","iopub.execute_input":"2023-09-30T18:27:16.337656Z","iopub.status.idle":"2023-09-30T18:27:30.684068Z","shell.execute_reply.started":"2023-09-30T18:27:16.337627Z","shell.execute_reply":"2023-09-30T18:27:30.682980Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import pandas as pd\nimport numpy as np\nimport librosa\nimport matplotlib.pyplot as plt\nimport seaborn as sb\nfrom pydub import AudioSegment\nimport os\nfrom datasets import load_dataset, DatasetDict, Dataset\nfrom tqdm import tqdm\nimport pyarrow as pa\nimport pyarrow.parquet as pq\nfrom datasets import Audio, load_from_disk\nfrom transformers import Wav2Vec2Tokenizer, Wav2Vec2Processor, Wav2Vec2ProcessorWithLM, Wav2Vec2FeatureExtractor, Wav2Vec2CTCTokenizer, Wav2Vec2ForCTC\nimport torch\nimport torchaudio\nimport torchaudio.transforms as tat\n# import evaluate\nfrom dataclasses import dataclass\nfrom typing import Any, Dict, List, Union, Optional, Collection\nfrom torchmetrics.text import WordErrorRate\nfrom transformers import Seq2SeqTrainingArguments, TrainingArguments, EarlyStoppingCallback\nfrom transformers import Seq2SeqTrainer, Trainer\nimport gc\nimport os, psutil\nfrom datasets import load_dataset, load_metric, Audio\nimport bitsandbytes as bnb","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","execution":{"iopub.status.busy":"2023-09-30T18:27:30.685805Z","iopub.execute_input":"2023-09-30T18:27:30.686184Z","iopub.status.idle":"2023-09-30T18:27:42.507621Z","shell.execute_reply.started":"2023-09-30T18:27:30.686153Z","shell.execute_reply":"2023-09-30T18:27:42.506692Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"TRAIN_PATH = \"/kaggle/input/bengaliai-speech/train.csv\"\nTRAIN_AUDIO_FILES = \"/kaggle/input/bengaliai-speech/train_mp3s\"\nTEST_AUDIO_FILES = \"/kaggle/input/bengaliai-speech/test_mp3s\"\nEXAMPLES_AUDIO_FILES = \"/kaggle/input/bengaliai-speech/examples\"\nSUBMISSION_PATH = \"/kaggle/input/bengaliai-speech/sample_submission.csv\"","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:42.509781Z","iopub.execute_input":"2023-09-30T18:27:42.510118Z","iopub.status.idle":"2023-09-30T18:27:42.515594Z","shell.execute_reply.started":"2023-09-30T18:27:42.510087Z","shell.execute_reply":"2023-09-30T18:27:42.514193Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"### hyper-parameters\nSR = 16000\ntorch.backends.cudnn.benchmark = True\noutput_dir = \"./bengali-ex007\"\nMODEL_PATH = \"/kaggle/input/bengali-ex006/checkpoint-95000/\"\n# LM_PATH = \"/kaggle/input/arijitx-full-model/wav2vec2-xls-r-300m-bengali/language_model\"","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:42.516996Z","iopub.execute_input":"2023-09-30T18:27:42.517616Z","iopub.status.idle":"2023-09-30T18:27:42.528211Z","shell.execute_reply.started":"2023-09-30T18:27:42.517585Z","shell.execute_reply":"2023-09-30T18:27:42.527348Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#### Train Dataframe","metadata":{}},{"cell_type":"code","source":"train_df = pd.read_csv(\"/kaggle/input/bengaliai-speech-normalized-data/normalized.csv\")\ntrain_df","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:42.529252Z","iopub.execute_input":"2023-09-30T18:27:42.530055Z","iopub.status.idle":"2023-09-30T18:27:49.801284Z","shell.execute_reply.started":"2023-09-30T18:27:42.530001Z","shell.execute_reply":"2023-09-30T18:27:49.800402Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_df['split'].value_counts()","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:49.802651Z","iopub.execute_input":"2023-09-30T18:27:49.803201Z","iopub.status.idle":"2023-09-30T18:27:49.870270Z","shell.execute_reply.started":"2023-09-30T18:27:49.803168Z","shell.execute_reply":"2023-09-30T18:27:49.869315Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_data = train_df[train_df['split']=='train'].reset_index(drop=True)\nvalid_data = train_df[train_df['split']=='valid'].reset_index(drop=True)","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:49.871622Z","iopub.execute_input":"2023-09-30T18:27:49.872239Z","iopub.status.idle":"2023-09-30T18:27:50.212862Z","shell.execute_reply.started":"2023-09-30T18:27:49.872208Z","shell.execute_reply":"2023-09-30T18:27:50.211939Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_data","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:50.214065Z","iopub.execute_input":"2023-09-30T18:27:50.214620Z","iopub.status.idle":"2023-09-30T18:27:50.227386Z","shell.execute_reply.started":"2023-09-30T18:27:50.214590Z","shell.execute_reply":"2023-09-30T18:27:50.226206Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"indexes = set(pd.read_csv(\"/kaggle/input/dataset-overlaps-with-commonvoice-11-bn/indexes.csv\")['id'])\nindexes","metadata":{"scrolled":true,"execution":{"iopub.status.busy":"2023-09-30T18:27:50.231224Z","iopub.execute_input":"2023-09-30T18:27:50.231455Z","iopub.status.idle":"2023-09-30T18:27:50.328255Z","shell.execute_reply.started":"2023-09-30T18:27:50.231435Z","shell.execute_reply":"2023-09-30T18:27:50.327393Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_data_v2 = train_data[~train_data.index.isin(indexes)].reset_index(drop=True)\ntrain_data_v2","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:50.329657Z","iopub.execute_input":"2023-09-30T18:27:50.330184Z","iopub.status.idle":"2023-09-30T18:27:50.535950Z","shell.execute_reply.started":"2023-09-30T18:27:50.330150Z","shell.execute_reply":"2023-09-30T18:27:50.535079Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"len(train_data), len(train_data_v2)","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:50.537172Z","iopub.execute_input":"2023-09-30T18:27:50.537962Z","iopub.status.idle":"2023-09-30T18:27:50.544859Z","shell.execute_reply.started":"2023-09-30T18:27:50.537930Z","shell.execute_reply":"2023-09-30T18:27:50.543806Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_data_v2.sample(frac=0.04, random_state=42)","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:50.546310Z","iopub.execute_input":"2023-09-30T18:27:50.547244Z","iopub.status.idle":"2023-09-30T18:27:50.592247Z","shell.execute_reply.started":"2023-09-30T18:27:50.547210Z","shell.execute_reply":"2023-09-30T18:27:50.591313Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_data_v2.sample(frac=0.04, random_state=42).sample(frac=0.08, random_state=42)","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:50.593542Z","iopub.execute_input":"2023-09-30T18:27:50.594056Z","iopub.status.idle":"2023-09-30T18:27:50.633742Z","shell.execute_reply.started":"2023-09-30T18:27:50.594012Z","shell.execute_reply":"2023-09-30T18:27:50.632826Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"valid_0 = valid_data.sample(frac=0.1, random_state=42)\ntrain_0 = valid_data[~valid_data.index.isin(valid_0.index)]\n\ndata_1 = train_data_v2.sample(frac=0.05, random_state=42)\nvalid_1 = data_1.sample(frac=0.08, random_state=42)\ntrain_1 = data_1[~data_1.index.isin(valid_1.index)]\n\ntrain = pd.concat([train_0, train_1], axis=0).sample(frac=1, random_state=42).reset_index(drop=True)\nvalid = pd.concat([valid_0, valid_1], axis=0).sample(frac=1, random_state=42).reset_index(drop=True)\n\ndel data_1, valid_0, valid_1, train_0, train_1\nall_ids = train_data_v2['id'].to_list()\ntrain_ids = train['id'].to_list()\nvalid_ids = valid['id'].to_list()\n\n# in kaggle notebook, validating is very time-consuming, so here I use a very small validation set, rather than 5667.\nvalid = valid.sample(n=500, random_state=42)\n\nprint(len(all_ids))\nprint(len(train_ids))\nprint(len(valid_ids))","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:50.634910Z","iopub.execute_input":"2023-09-30T18:27:50.635203Z","iopub.status.idle":"2023-09-30T18:27:50.725799Z","shell.execute_reply.started":"2023-09-30T18:27:50.635175Z","shell.execute_reply":"2023-09-30T18:27:50.724887Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class W2v2Dataset(torch.utils.data.Dataset):\n    def __init__(self, df):\n        self.df = df\n        self.pathes = df['id'].values\n        self.sentences = df['normalized'].values\n        # self.resampler = tat.Resample(32000, SR)\n\n    def __getitem__(self, idx):\n        apath = f'/kaggle/input/bengaliai-speech/train_mp3s/{self.pathes[idx]}.mp3'\n        # waveform, sample_rate = torchaudio.load(apath, format=\"mp3\")\n        waveform, sample_rate = librosa.load(apath, sr=16000)\n        # waveform = self.resampler(waveform)\n        batch = dict()\n        y = processor(waveform.reshape(-1), sampling_rate=SR).input_values[0] \n        batch[\"input_values\"] = y\n        with processor.as_target_processor():\n            batch[\"labels\"] = processor(self.sentences[idx]).input_ids       \n        \n        return batch\n\n    def __len__(self):\n        return len(self.df)","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:50.727234Z","iopub.execute_input":"2023-09-30T18:27:50.727772Z","iopub.status.idle":"2023-09-30T18:27:50.734835Z","shell.execute_reply.started":"2023-09-30T18:27:50.727740Z","shell.execute_reply":"2023-09-30T18:27:50.733944Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_dataset = W2v2Dataset(train)\nvalid_dataset = W2v2Dataset(valid)","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:50.736168Z","iopub.execute_input":"2023-09-30T18:27:50.736727Z","iopub.status.idle":"2023-09-30T18:27:50.750549Z","shell.execute_reply.started":"2023-09-30T18:27:50.736698Z","shell.execute_reply":"2023-09-30T18:27:50.749587Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"processor = Wav2Vec2Processor.from_pretrained(MODEL_PATH)\nvocab_dict = processor.tokenizer.get_vocab()\nsorted_vocab_dict = {k: v for k, v in sorted(vocab_dict.items(), key=lambda item: item[1])}\n\n# decoder = pyctcdecode.build_ctcdecoder(\n#     list(sorted_vocab_dict.keys()),\n#     str(LM_PATH+\"/5gram.bin\"),\n#     str(LM_PATH+\"/unigrams.txt\"),\n# )\n\n# processor_with_lm = Wav2Vec2ProcessorWithLM(\n#     feature_extractor=processor.feature_extractor,\n#     tokenizer=processor.tokenizer,\n#     decoder=decoder\n# )","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:50.751984Z","iopub.execute_input":"2023-09-30T18:27:50.752713Z","iopub.status.idle":"2023-09-30T18:27:50.785830Z","shell.execute_reply.started":"2023-09-30T18:27:50.752683Z","shell.execute_reply":"2023-09-30T18:27:50.785073Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"@dataclass\nclass DataCollatorCTCWithPadding:\n    \"\"\"\n    Data collator that will dynamically pad the inputs received.\n    Args:\n        processor (:class:`~transformers.Wav2Vec2Processor`)\n            The processor used for proccessing the data.\n        padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):\n            Select a strategy to pad the returned sequences (according to the model's padding side and padding index)\n            among:\n            * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single\n              sequence if provided).\n            * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the\n              maximum acceptable input length for the model if that argument is not provided.\n            * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of\n              different lengths).\n        max_length (:obj:`int`, `optional`):\n            Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).\n        max_length_labels (:obj:`int`, `optional`):\n            Maximum length of the ``labels`` returned list and optionally padding length (see above).\n        pad_to_multiple_of (:obj:`int`, `optional`):\n            If set will pad the sequence to a multiple of the provided value.\n            This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=\n            7.5 (Volta).\n    \"\"\"\n\n    processor: Wav2Vec2Processor\n    padding: Union[bool, str] = True\n    max_length: Optional[int] = None\n    max_length_labels: Optional[int] = None\n    pad_to_multiple_of: Optional[int] = None\n    pad_to_multiple_of_labels: Optional[int] = None\n\n    def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:\n        # split inputs and labels since they have to be of different lenghts and need\n        # different padding methods\n        input_features = [{\"input_values\": feature[\"input_values\"]} for feature in features]\n        label_features = [{\"input_ids\": feature[\"labels\"]} for feature in features]\n\n        batch = self.processor.pad(\n            input_features,\n            padding=self.padding,\n            max_length=self.max_length,\n            pad_to_multiple_of=self.pad_to_multiple_of,\n            return_tensors=\"pt\",\n        )\n        with self.processor.as_target_processor():\n            labels_batch = self.processor.pad(\n                label_features,\n                padding=self.padding,\n                max_length=self.max_length_labels,\n                pad_to_multiple_of=self.pad_to_multiple_of_labels,\n                return_tensors=\"pt\",\n            )\n\n        # replace padding with -100 to ignore loss correctly\n        labels = labels_batch[\"input_ids\"].masked_fill(labels_batch.attention_mask.ne(1), -100)\n\n        batch[\"labels\"] = labels\n\n        return batch","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:50.787098Z","iopub.execute_input":"2023-09-30T18:27:50.787387Z","iopub.status.idle":"2023-09-30T18:27:50.809547Z","shell.execute_reply.started":"2023-09-30T18:27:50.787360Z","shell.execute_reply":"2023-09-30T18:27:50.808614Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"data_collator = DataCollatorCTCWithPadding(processor=processor, padding=True)","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:50.810803Z","iopub.execute_input":"2023-09-30T18:27:50.811697Z","iopub.status.idle":"2023-09-30T18:27:50.824549Z","shell.execute_reply.started":"2023-09-30T18:27:50.811665Z","shell.execute_reply":"2023-09-30T18:27:50.823614Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"wer_metric = load_metric(\"wer\")\n\ndef compute_metrics(pred):\n    pred_logits = pred.predictions\n    pred_ids = np.argmax(pred_logits, axis=-1)\n\n    pred.label_ids[pred.label_ids == -100] = processor.tokenizer.pad_token_id\n\n    pred_str = processor.batch_decode(pred_ids)\n    # we do not want to group tokens when computing the metrics\n    label_str = processor.batch_decode(pred.label_ids, group_tokens=False)\n\n    wer = wer_metric.compute(predictions=pred_str, references=label_str)\n\n    return {\"wer\": wer}","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:50.825687Z","iopub.execute_input":"2023-09-30T18:27:50.826558Z","iopub.status.idle":"2023-09-30T18:27:51.584338Z","shell.execute_reply.started":"2023-09-30T18:27:50.826513Z","shell.execute_reply":"2023-09-30T18:27:51.583471Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:51.585621Z","iopub.execute_input":"2023-09-30T18:27:51.585913Z","iopub.status.idle":"2023-09-30T18:27:51.591211Z","shell.execute_reply.started":"2023-09-30T18:27:51.585884Z","shell.execute_reply":"2023-09-30T18:27:51.589480Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"model = Wav2Vec2ForCTC.from_pretrained(\n    MODEL_PATH,\n    attention_dropout=0.1,\n    hidden_dropout=0.1,\n    feat_proj_dropout=0.0,\n    mask_time_prob=0.05,\n    layerdrop=0.1,\n    # gradient_checkpointing=True, \n    ctc_loss_reduction=\"mean\", \n    pad_token_id=processor.tokenizer.pad_token_id,\n    vocab_size=len(processor.tokenizer),\n    ctc_zero_infinity=True,\n    diversity_loss_weight=100 \n).to(device)","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:27:51.592694Z","iopub.execute_input":"2023-09-30T18:27:51.593376Z","iopub.status.idle":"2023-09-30T18:28:07.358891Z","shell.execute_reply.started":"2023-09-30T18:27:51.593346Z","shell.execute_reply":"2023-09-30T18:28:07.357937Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"model.freeze_feature_extractor()","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:28:07.360389Z","iopub.execute_input":"2023-09-30T18:28:07.361076Z","iopub.status.idle":"2023-09-30T18:28:07.366701Z","shell.execute_reply.started":"2023-09-30T18:28:07.361041Z","shell.execute_reply":"2023-09-30T18:28:07.365842Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"training_args = TrainingArguments(\n    output_dir=output_dir,\n    overwrite_output_dir=True,\n    group_by_length=False,\n    lr_scheduler_type='cosine',\n    weight_decay=0.01,\n    per_device_train_batch_size=4,\n    per_device_eval_batch_size=16,\n    gradient_accumulation_steps=4,\n    gradient_checkpointing=True,\n    evaluation_strategy=\"steps\",\n    optim=\"adamw_bnb_8bit\",\n    # dataloader_pin_memory=True,\n    # dataloader_num_workers=4,\n    save_strategy=\"steps\",\n    num_train_epochs=3.5,\n    max_steps=-1, # you can change to \"num_train_epochs\"\n    fp16=True,\n    save_steps=500,\n    eval_steps=500,\n    logging_steps=500,\n    learning_rate=5e-5,\n    warmup_steps=600,\n    save_total_limit=1,\n    load_best_model_at_end=True,\n    metric_for_best_model=\"wer\",\n    greater_is_better=False,\n    prediction_loss_only=False,\n    auto_find_batch_size=True,\n    report_to=\"none\"\n)","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:30:23.807246Z","iopub.execute_input":"2023-09-30T18:30:23.807574Z","iopub.status.idle":"2023-09-30T18:30:23.813714Z","shell.execute_reply.started":"2023-09-30T18:30:23.807547Z","shell.execute_reply":"2023-09-30T18:30:23.812716Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"trainer = Trainer(\n    model=model,\n    data_collator=data_collator,\n    args=training_args,\n    compute_metrics=compute_metrics,\n    train_dataset=train_dataset,\n    eval_dataset=valid_dataset,\n    tokenizer=processor.feature_extractor,\n    callbacks=[EarlyStoppingCallback(early_stopping_patience=5)],\n)","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:30:24.486362Z","iopub.execute_input":"2023-09-30T18:30:24.486682Z","iopub.status.idle":"2023-09-30T18:30:24.503190Z","shell.execute_reply.started":"2023-09-30T18:30:24.486657Z","shell.execute_reply":"2023-09-30T18:30:24.502394Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"trainer.train()","metadata":{"execution":{"iopub.status.busy":"2023-09-30T18:30:24.869236Z","iopub.execute_input":"2023-09-30T18:30:24.869530Z","iopub.status.idle":"2023-09-30T18:30:36.916232Z","shell.execute_reply.started":"2023-09-30T18:30:24.869506Z","shell.execute_reply":"2023-09-30T18:30:36.915101Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"trainer.save_model(f'wav2vec2_bengali_ex007')","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!zip -r wav2vec2_bengali_ex007.zip /kaggle/working/wav2vec2_bengali_ex007","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]}]}