streaming_gender_model.ipynb

{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.7.12","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"import os\nimport random\nimport warnings\nfrom shutil import copyfile\nfrom collections import Counter\nimport pickle\nimport multiprocessing\n\nimport tqdm\nimport numpy as np\nimport pandas as pd\nimport numpy as np\nimport librosa\nfrom IPython.display import display, Audio, FileLink\nimport torchaudio\nimport torch\nimport torch.nn as nn\nfrom torch.optim import Adam\nimport torch.nn.functional as F\nimport matplotlib.pyplot as plt","metadata":{"execution":{"iopub.status.busy":"2022-02-19T07:40:42.248271Z","iopub.execute_input":"2022-02-19T07:40:42.248921Z","iopub.status.idle":"2022-02-19T07:40:45.580915Z","shell.execute_reply.started":"2022-02-19T07:40:42.248823Z","shell.execute_reply":"2022-02-19T07:40:45.580075Z"},"trusted":true},"execution_count":1,"outputs":[]},{"cell_type":"code","source":"SEED = 1234\n\nrandom.seed(SEED)\nnp.random.seed(SEED)\ntorch.manual_seed(SEED)\ntorch.cuda.manual_seed(SEED)\ntorch.backends.cudnn.deterministic = True","metadata":{"execution":{"iopub.status.busy":"2022-02-19T07:40:45.582919Z","iopub.execute_input":"2022-02-19T07:40:45.583234Z","iopub.status.idle":"2022-02-19T07:40:45.591529Z","shell.execute_reply.started":"2022-02-19T07:40:45.583195Z","shell.execute_reply":"2022-02-19T07:40:45.590669Z"},"trusted":true},"execution_count":2,"outputs":[]},{"cell_type":"markdown","source":"# Constants","metadata":{}},{"cell_type":"code","source":"BATCH_SIZE = 128\nSAMPLE_RATE = 16000\nMAX_LENGTH_SEC = 7.\nMAX_LENGTH_SAMPLES = int(MAX_LENGTH_SEC * SAMPLE_RATE)\n\n\nspectrogram_params = {\n    'sample_rate': SAMPLE_RATE,\n    'n_fft': 1024,\n    'win_length': 1024,\n    'hop_length': 512,\n    'f_min': 0,\n    'f_max': 3800,\n    'window_fn': torch.hamming_window,\n    'power': 2.0,\n    'norm': 'slaney',\n    'n_mels': 64,\n    'mel_scale': 'htk',\n    'center': False,\n}\n\n\nmodel_params = {\n    'fc_size': 128,\n    'hidden_size': 128,\n    'num_layers': 3,\n    'dropout': 0.25,\n}\n\n\ndevice = \"cuda\" if torch.cuda.is_available() else \"cpu\"\ndevice","metadata":{"execution":{"iopub.status.busy":"2022-02-19T07:41:08.178965Z","iopub.execute_input":"2022-02-19T07:41:08.179415Z","iopub.status.idle":"2022-02-19T07:41:08.235909Z","shell.execute_reply.started":"2022-02-19T07:41:08.179379Z","shell.execute_reply":"2022-02-19T07:41:08.235091Z"},"trusted":true},"execution_count":3,"outputs":[{"execution_count":3,"output_type":"execute_result","data":{"text/plain":"'cuda'"},"metadata":{}}]},{"cell_type":"markdown","source":"# Utils","metadata":{}},{"cell_type":"code","source":"def convert_to_wav(args):\n    path_to_audio, path_to_save, new_sr = args\n    assert os.path.exists(path_to_audio)\n    assert path_to_save.endswith(\".wav\")\n    cmd = f\"ffmpeg -i {path_to_audio} -ar {new_sr} {path_to_save} -loglevel error\"\n    assert os.system(cmd) == 0, print(path_to_audio, path_to_save)","metadata":{"execution":{"iopub.status.busy":"2022-02-19T07:41:10.671137Z","iopub.execute_input":"2022-02-19T07:41:10.671868Z","iopub.status.idle":"2022-02-19T07:41:10.677462Z","shell.execute_reply.started":"2022-02-19T07:41:10.671818Z","shell.execute_reply":"2022-02-19T07:41:10.676633Z"},"trusted":true},"execution_count":4,"outputs":[]},{"cell_type":"code","source":"def get_balanced_dataset(df, feature):\n    assert feature in df\n    c = df[feature].value_counts()\n    keys = c.keys()\n    assert len(keys) == 2\n    m = min(c)\n    df_key_0 = df[df[feature] == keys[0]].sample(m)\n    df_key_1 = df[df[feature] == keys[1]].sample(m)\n    return pd.concat([df_key_0, df_key_1])","metadata":{"execution":{"iopub.status.busy":"2022-02-19T07:41:10.908411Z","iopub.execute_input":"2022-02-19T07:41:10.909257Z","iopub.status.idle":"2022-02-19T07:41:10.915129Z","shell.execute_reply.started":"2022-02-19T07:41:10.909214Z","shell.execute_reply":"2022-02-19T07:41:10.914168Z"},"trusted":true},"execution_count":5,"outputs":[]},{"cell_type":"code","source":"def validate_manifest(manifest: pd.DataFrame):\n    assert 'gender' in manifest\n    assert sum(manifest['gender'].apply(lambda x: x in ['male', 'female'])) == manifest.shape[0]\n    assert 'path_to_audio' in manifest\n    for p in tqdm.tqdm(manifest['path_to_audio'].values):\n        assert os.path.exists(p), print(f\"path doesn't exist: {p}\")","metadata":{"execution":{"iopub.status.busy":"2022-02-19T07:41:11.590816Z","iopub.execute_input":"2022-02-19T07:41:11.591358Z","iopub.status.idle":"2022-02-19T07:41:11.597405Z","shell.execute_reply.started":"2022-02-19T07:41:11.591320Z","shell.execute_reply":"2022-02-19T07:41:11.596711Z"},"trusted":true},"execution_count":6,"outputs":[]},{"cell_type":"code","source":"def get_amplitudes(path_to_audio, new_sample_rate):\n    amplitudes, sample_rate = torchaudio.load(path_to_audio)\n    amplitudes = torchaudio.transforms.Resample(sample_rate, new_sample_rate)(amplitudes)\n    return amplitudes.squeeze(0)","metadata":{"execution":{"iopub.status.busy":"2022-02-19T07:41:11.974802Z","iopub.execute_input":"2022-02-19T07:41:11.975332Z","iopub.status.idle":"2022-02-19T07:41:11.979964Z","shell.execute_reply.started":"2022-02-19T07:41:11.975294Z","shell.execute_reply":"2022-02-19T07:41:11.979213Z"},"trusted":true},"execution_count":7,"outputs":[]},{"cell_type":"markdown","source":"# CommonVoice","metadata":{}},{"cell_type":"code","source":"def common_voice_preprocessing(path_to_csv):\n    df = pd.read_csv(path_to_csv, sep=',')\n    ending = 'train' if path_to_csv.endswith('train.csv') else 'test'\n    dir_to_audio = \"/kaggle/input/common-voice/cv-valid-{}/\".format(ending)\n    df['path_to_audio'] = df['filename'].apply(lambda x: os.path.join(dir_to_audio, x))\n    df = df[~df['path_to_audio'].isna()]\n    df = df[df['gender'].isin(['male', 'female'])]\n    return df","metadata":{"execution":{"iopub.status.busy":"2022-02-19T07:41:12.977637Z","iopub.execute_input":"2022-02-19T07:41:12.978326Z","iopub.status.idle":"2022-02-19T07:41:12.986671Z","shell.execute_reply.started":"2022-02-19T07:41:12.978289Z","shell.execute_reply":"2022-02-19T07:41:12.985796Z"},"trusted":true},"execution_count":8,"outputs":[]},{"cell_type":"code","source":"class CommoVoiceDataset(torch.utils.data.Dataset):\n    def __init__(self, df, sample_rate=SAMPLE_RATE):\n        self.audio_paths = df['path_to_audio'].values\n        self.genders = df['gender'].apply(lambda x: 1 if x == 'male' else 0).values\n        self.sample_rate = sample_rate\n        \n    def __getitem__(self, idx):\n        amplitudes = get_amplitudes(self.audio_paths[idx], self.sample_rate)\n        return amplitudes, self.genders[idx], self.audio_paths[idx]\n\n    def __len__(self):\n        return len(self.genders)","metadata":{"execution":{"iopub.status.busy":"2022-02-19T07:41:13.573703Z","iopub.execute_input":"2022-02-19T07:41:13.574262Z","iopub.status.idle":"2022-02-19T07:41:13.581929Z","shell.execute_reply.started":"2022-02-19T07:41:13.574225Z","shell.execute_reply":"2022-02-19T07:41:13.581086Z"},"trusted":true},"execution_count":9,"outputs":[]},{"cell_type":"code","source":"df_train = common_voice_preprocessing('/kaggle/input/common-voice/cv-valid-train.csv')\ndf_test = common_voice_preprocessing('/kaggle/input/common-voice/cv-valid-test.csv')","metadata":{"execution":{"iopub.status.busy":"2022-02-19T07:41:14.515942Z","iopub.execute_input":"2022-02-19T07:41:14.516324Z","iopub.status.idle":"2022-02-19T07:41:15.496683Z","shell.execute_reply.started":"2022-02-19T07:41:14.516291Z","shell.execute_reply":"2022-02-19T07:41:15.495933Z"},"trusted":true},"execution_count":10,"outputs":[]},{"cell_type":"code","source":"df_train_balanced = get_balanced_dataset(df_train, 'gender')\ndf_test_balanced = get_balanced_dataset(df_test, 'gender')","metadata":{"execution":{"iopub.status.busy":"2022-02-19T07:41:15.498168Z","iopub.execute_input":"2022-02-19T07:41:15.499362Z","iopub.status.idle":"2022-02-19T07:41:15.562116Z","shell.execute_reply.started":"2022-02-19T07:41:15.499323Z","shell.execute_reply":"2022-02-19T07:41:15.561402Z"},"trusted":true},"execution_count":11,"outputs":[]},{"cell_type":"code","source":"df_train_balanced = df_train_balanced[df_train_balanced['up_votes'] >= 1]\ndf_test_balanced = df_test_balanced[df_test_balanced['up_votes'] >= 1]\n\ndf_train_balanced.shape, df_test_balanced.shape","metadata":{"execution":{"iopub.status.busy":"2022-02-19T07:41:15.563225Z","iopub.execute_input":"2022-02-19T07:41:15.563555Z","iopub.status.idle":"2022-02-19T07:41:15.583595Z","shell.execute_reply.started":"2022-02-19T07:41:15.563517Z","shell.execute_reply":"2022-02-19T07:41:15.582938Z"},"trusted":true},"execution_count":12,"outputs":[{"execution_count":12,"output_type":"execute_result","data":{"text/plain":"((36498, 9), (784, 9))"},"metadata":{}}]},{"cell_type":"code","source":"for mode in ['train', 'test']:\n    dir_to_save = f'/kaggle/output/common-voice/cv-valid-{mode}/cv-valid-{mode}/'\n    os.system(f\"rm -rf {dir_to_save}\")\n    os.makedirs(dir_to_save, exist_ok=True)\n    df = df_train_balanced if mode == 'train' else df_test_balanced\n    paths_to_audio = df['path_to_audio'].values\n    portion = 100\n    for j in tqdm.trange(0, len(paths_to_audio), portion):\n        query = [(path_to_audio, os.path.join(dir_to_save, path_to_audio.split(f'cv-valid-{mode}/')[-1] + \"__.wav\"), 16000) \n                 for path_to_audio in paths_to_audio[j: j+portion]]\n        with multiprocessing.Pool(4) as p:\n            for _ in p.imap_unordered(convert_to_wav, query):\n                pass\nprint(\"common voice dataset is ready\")","metadata":{"execution":{"iopub.status.busy":"2022-02-19T07:41:16.806876Z","iopub.execute_input":"2022-02-19T07:41:16.807369Z","iopub.status.idle":"2022-02-19T08:12:08.123851Z","shell.execute_reply.started":"2022-02-19T07:41:16.807325Z","shell.execute_reply":"2022-02-19T08:12:08.120688Z"},"trusted":true},"execution_count":13,"outputs":[{"name":"stderr","text":"100%|██████████| 365/365 [30:12<00:00,  4.97s/it]\n100%|██████████| 8/8 [00:38<00:00,  4.85s/it]","output_type":"stream"},{"name":"stdout","text":"common voice dataset is ready\n","output_type":"stream"},{"name":"stderr","text":"\n","output_type":"stream"}]},{"cell_type":"code","source":"df_train_balanced['path_to_audio'] = df_train_balanced['path_to_audio'].apply(lambda x: x.replace('input', 'output') + \"__.wav\")\ndf_test_balanced['path_to_audio'] = df_test_balanced['path_to_audio'].apply(lambda x: x.replace('input', 'output') + \"__.wav\")","metadata":{"execution":{"iopub.status.busy":"2022-02-19T08:12:08.125888Z","iopub.execute_input":"2022-02-19T08:12:08.126172Z","iopub.status.idle":"2022-02-19T08:12:08.199196Z","shell.execute_reply.started":"2022-02-19T08:12:08.126133Z","shell.execute_reply":"2022-02-19T08:12:08.198384Z"},"trusted":true},"execution_count":14,"outputs":[]},{"cell_type":"code","source":"train_dataset = CommoVoiceDataset(df_train_balanced)\ntest_dataset = CommoVoiceDataset(df_test_balanced)","metadata":{"execution":{"iopub.status.busy":"2022-02-19T08:12:08.203758Z","iopub.execute_input":"2022-02-19T08:12:08.205851Z","iopub.status.idle":"2022-02-19T08:12:08.245567Z","shell.execute_reply.started":"2022-02-19T08:12:08.205809Z","shell.execute_reply":"2022-02-19T08:12:08.244909Z"},"trusted":true},"execution_count":15,"outputs":[]},{"cell_type":"code","source":"def collate_fn(batch_data):\n    _, genders, audio_paths = zip(*batch_data)\n    max_length = min(MAX_LENGTH_SAMPLES, max((amplitude.shape[-1] for amplitude, _, _ in batch_data)))\n    amplitudes = torch.zeros((len(batch_data), max_length))\n    for i in range(len(batch_data)):\n        cur_length = min(batch_data[i][0].shape[-1], max_length)\n        amplitudes[i][:cur_length] = batch_data[i][0][:cur_length]\n    return {'amplitudes': amplitudes, \n            'genders': torch.tensor(genders).float(),\n            'audio_paths': audio_paths}\n\n\ntrain_loader = torch.utils.data.DataLoader(train_dataset, \n                                           batch_size=BATCH_SIZE, \n                                           shuffle=True, \n                                           num_workers=2,\n                                           collate_fn=collate_fn)\n\ntest_loader = torch.utils.data.DataLoader(test_dataset, \n                                          batch_size=BATCH_SIZE, \n                                          shuffle=False, \n                                          num_workers=2,\n                                          collate_fn=collate_fn)","metadata":{"execution":{"iopub.status.busy":"2022-02-19T08:12:08.250362Z","iopub.execute_input":"2022-02-19T08:12:08.252369Z","iopub.status.idle":"2022-02-19T08:12:08.265949Z","shell.execute_reply.started":"2022-02-19T08:12:08.252325Z","shell.execute_reply":"2022-02-19T08:12:08.265073Z"},"trusted":true},"execution_count":16,"outputs":[]},{"cell_type":"code","source":"# it's usefull to iterate over all dataset to be sure that data is ok\nfor batch_data in tqdm.tqdm(train_loader):\n    amplitudes = batch_data['amplitudes']\n    genders = batch_data['genders']\n    audio_paths = batch_data['audio_paths']\n    \nprint(amplitudes.shape)","metadata":{"execution":{"iopub.status.busy":"2022-02-19T08:12:08.270224Z","iopub.execute_input":"2022-02-19T08:12:08.272664Z","iopub.status.idle":"2022-02-19T08:12:39.772616Z","shell.execute_reply.started":"2022-02-19T08:12:08.272624Z","shell.execute_reply":"2022-02-19T08:12:39.771548Z"},"trusted":true},"execution_count":17,"outputs":[{"name":"stderr","text":"100%|██████████| 286/286 [00:31<00:00,  9.08it/s]","output_type":"stream"},{"name":"stdout","text":"torch.Size([18, 112000])\n","output_type":"stream"},{"name":"stderr","text":"\n","output_type":"stream"}]},{"cell_type":"markdown","source":"# Model","metadata":{}},{"cell_type":"markdown","source":"MelSpectrogram in Pytorch works not good, because of padding streaming results will be differerent from non-streaming fashion. But for the model it's not a problem.","metadata":{}},{"cell_type":"code","source":"class FeatureExtractor(nn.Module):\n    def __init__(self, spectrogram_params, device):\n        super(FeatureExtractor, self).__init__()\n        self.device = device\n        self.spectrogram_params = spectrogram_params\n        self.overlapping = spectrogram_params['win_length'] - spectrogram_params['hop_length']\n        self.spectrogram = torchaudio.transforms.MelSpectrogram(**spectrogram_params)        \n        self.spectrogram_db = torchaudio.transforms.AmplitudeToDB()\n        \n    def get_init_state(self, batch_size):\n        return torch.zeros((batch_size, self.overlapping)).to(self.device)\n\n    def forward(self, x, state):\n        \"\"\"\n        :param: x - audio signal, shape: [batch_size, signal_length]\n        :param: state - audio signal, shape: [batch_size, overlapping] or None\n        \n        returns: \n            spectrogram, shape: [batch_size, T, F]\n            new state, shape: [batch_size, overlapping]\n        \"\"\"\n        batch_size = x.shape[0]\n        if state is None:\n            state = self.get_init_state(batch_size)\n        x = torch.cat((state, x), dim=1)\n        new_state = x[:, -self.overlapping:]\n        x = self.spectrogram(x)\n        x = torch.permute(x, (0, 2, 1))\n        x = self.spectrogram_db(x)\n        return x, new_state","metadata":{"execution":{"iopub.status.busy":"2022-02-19T08:12:39.774479Z","iopub.execute_input":"2022-02-19T08:12:39.774750Z","iopub.status.idle":"2022-02-19T08:12:39.786541Z","shell.execute_reply.started":"2022-02-19T08:12:39.774710Z","shell.execute_reply":"2022-02-19T08:12:39.785921Z"},"trusted":true},"execution_count":18,"outputs":[]},{"cell_type":"code","source":"class Model(nn.Module):\n    def __init__(self, model_params, spectrogram_params, device):\n        super(Model, self).__init__()\n        self.model_params = model_params\n        self.spectrogram_params = spectrogram_params\n        self.device = device\n        self.tanh = torch.nn.Tanh()\n        self.relu = torch.nn.ReLU()\n        self.extractor_layer = FeatureExtractor(self.spectrogram_params, device=self.device)\n        self.lstm = torch.nn.LSTM(input_size=self.spectrogram_params['n_mels'],\n                                  hidden_size=self.model_params['hidden_size'],\n                                  batch_first=True,\n                                  num_layers=self.model_params['num_layers'],\n                                  dropout=self.model_params['dropout'])\n        self.fc_1 = nn.Linear(self.model_params['hidden_size'], self.model_params['fc_size'])\n        self.fc_2 = nn.Linear(self.model_params['fc_size'], self.model_params['fc_size'])\n        self.classifier = nn.Linear(self.model_params['fc_size'], 1)\n        \n    def get_init_lstm_states(self, batch_size):\n        shape = (self.model_params['num_layers'], batch_size, self.model_params['hidden_size'])\n        h_0 = torch.zeros(shape).to(self.device)\n        c_0 = torch.zeros(shape).to(self.device)\n        return (h_0, c_0)\n    \n    def forward(self, x, states):\n        \"\"\"\n        param: x: audio signal, shape: [batch_size]\n        param: states - dict, consists of lstm and signal states:\n            lstm states - (h_0, c_0) from docs, shape: [num_layers, batch_size, hidden_size]\n            signal states - overlapping, shape: [batch_size, win_length - hop_length]\n            \n        returns:\n            male logits, shape: [batch_size, 1]\n            new states, shape is identical states shape\n        \"\"\"\n        new_states = {}\n        batch_size = x.shape[0]\n        if 'samples' not in states:\n            states['samples'] = None\n        x, new_states['samples'] = self.extractor_layer(x, states['samples'])\n        if 'lstm' not in states:\n            states['lstm'] = self.get_init_lstm_states(batch_size)\n        x, new_states['lstm'] = self.lstm(x, states['lstm'])\n        x = self.relu(x[:, -1, :])\n        x = self.tanh(self.fc_1(x))\n        x = self.tanh(self.fc_2(x))\n        return self.classifier(x), new_states","metadata":{"execution":{"iopub.status.busy":"2022-02-19T08:12:39.789548Z","iopub.execute_input":"2022-02-19T08:12:39.791181Z","iopub.status.idle":"2022-02-19T08:12:39.827192Z","shell.execute_reply.started":"2022-02-19T08:12:39.791138Z","shell.execute_reply":"2022-02-19T08:12:39.826302Z"},"trusted":true},"execution_count":19,"outputs":[]},{"cell_type":"code","source":"model = Model(model_params=model_params,\n              spectrogram_params=spectrogram_params,\n              device=device).to(device)\nbatch_data = next(iter(train_loader))\namplitudes = batch_data['amplitudes']\n\ny, new_states = model(amplitudes.to(device), {})\nprint(y.shape)","metadata":{"execution":{"iopub.status.busy":"2022-02-19T09:17:01.431220Z","iopub.execute_input":"2022-02-19T09:17:01.431866Z","iopub.status.idle":"2022-02-19T09:17:02.218120Z","shell.execute_reply.started":"2022-02-19T09:17:01.431826Z","shell.execute_reply":"2022-02-19T09:17:02.217223Z"},"trusted":true},"execution_count":87,"outputs":[{"name":"stdout","text":"torch.Size([128, 1])\n","output_type":"stream"}]},{"cell_type":"markdown","source":"# Training","metadata":{}},{"cell_type":"code","source":"class Trainer:\n    def __init__(self, model, optimizer, loss_fn, train_loader, val_loader, device):\n        self.model = model\n        self.optimizer = optimizer\n        self.loss_fn = loss_fn\n        self.train_loader = train_loader\n        self.val_loader = val_loader\n        self.device = device\n        self.best_val_accuracy = 0\n        \n    def _batch_step(self, batch_data):\n        amplitudes = batch_data['amplitudes'].to(self.device)\n        genders = batch_data['genders'].to(self.device)\n        logits, new_states = self.model(amplitudes, states={})\n        logits = logits.view(-1)\n        loss_value = self.loss_fn(logits, genders)\n        probas = torch.sigmoid(logits)\n        return {\n            'batch_accuracy_sum': (probas.round().int() == genders).sum(),\n            'batch_loss': loss_value,\n            'batch_size': len(genders),\n        }\n    \n    def _print_logs(self, loss_value, accuracy_value, prefix=\"\"):\n        print(f\"{prefix}_loss:{loss_value:.5f}, \"\\\n              f\"{prefix}_accuracy:{accuracy_value:.5f}\")\n        \n    def _save_model(self):\n        path_to_save = f\"best_model_acc_{self.best_val_accuracy:.2f}.pt\"\n        print(f\"\\t model will be saved in {path_to_save}\")\n        torch.save(self.model.state_dict(), path_to_save)\n    \n    def train_epoch(self, verbose=0):\n        self.model.train()\n        loss_sum = 0\n        accuracy_sum = 0\n        num_elements = 0\n        for batch_data in self.train_loader:\n            optimizer.zero_grad()\n            batch_result = self._batch_step(batch_data)\n            batch_result['batch_loss'].backward()\n            optimizer.step()  \n            loss_sum += batch_result['batch_loss'].item()\n            accuracy_sum += batch_result['batch_accuracy_sum']\n            num_elements += batch_result['batch_size']\n            if verbose >= 1:\n                loss_value = batch_result['batch_loss'].item() / batch_result['batch_size']\n                accuracy_value = batch_result['batch_accuracy_sum'] / batch_result['batch_size']\n                self._print_logs(loss_value, accuracy_value, prefix=\"\\ttrain_batch\")\n        loss_value = loss_sum / num_elements\n        accuracy_value = accuracy_sum / num_elements\n        self._print_logs(loss_value, accuracy_value, prefix=\"train_epoch\")\n        \n        \n    def evaluate_epoch(self, verbose=0):\n        self.model.eval()\n        with torch.no_grad():\n            loss_sum = 0\n            accuracy_sum = 0\n            num_elements = 0\n            for batch_data in self.val_loader:\n                batch_result = self._batch_step(batch_data)\n                loss_sum += batch_result['batch_loss'].item()\n                accuracy_sum += batch_result['batch_accuracy_sum']\n                num_elements += batch_result['batch_size']\n                if verbose >= 1:\n                    loss_value = batch_result['batch_loss'].item() / batch_result['batch_size']\n                    accuracy_value = batch_result['batch_accuracy_sum'] / batch_result['batch_size']\n                    self._print_logs(loss_value, accuracy_value, prefix=\"\\tval_batch\")\n            loss_value = loss_sum / num_elements\n            accuracy_value = accuracy_sum / num_elements\n            self._print_logs(loss_value, accuracy_value, prefix=\"epoch_val\")\n            if self.best_val_accuracy < accuracy_value:\n                self.best_val_accuracy = accuracy_value\n                self._save_model()","metadata":{"execution":{"iopub.status.busy":"2022-02-19T09:17:05.360736Z","iopub.execute_input":"2022-02-19T09:17:05.361054Z","iopub.status.idle":"2022-02-19T09:17:05.381843Z","shell.execute_reply.started":"2022-02-19T09:17:05.361000Z","shell.execute_reply":"2022-02-19T09:17:05.380973Z"},"trusted":true},"execution_count":88,"outputs":[]},{"cell_type":"code","source":"optimizer = Adam(\n    [p for p in model.parameters() if p.requires_grad], lr=1e-3\n)\n\n\ntrainer = Trainer(model=model, \n                  optimizer=optimizer,\n                  loss_fn=torch.nn.BCEWithLogitsLoss(),\n                  train_loader=train_loader, \n                  val_loader=test_loader,\n                  device=device)\n\nfor epoch in range(10):\n    print(f\"epoch: {epoch}\")\n    trainer.train_epoch(verbose=0)\n    trainer.evaluate_epoch(verbose=0)\n    print(\"=\" * 25)\n    \nprint(\"training is finished!\")","metadata":{"execution":{"iopub.status.busy":"2022-02-19T09:17:10.616719Z","iopub.execute_input":"2022-02-19T09:17:10.616986Z","iopub.status.idle":"2022-02-19T09:24:23.158732Z","shell.execute_reply.started":"2022-02-19T09:17:10.616950Z","shell.execute_reply":"2022-02-19T09:24:23.157741Z"},"trusted":true},"execution_count":89,"outputs":[{"name":"stdout","text":"epoch: 0\ntrain_epoch_loss:0.00234, train_epoch_accuracy:0.88369\nepoch_val_loss:0.00171, epoch_val_accuracy:0.93240\n\t model will be saved in best_model_acc_0.93.pt\n=========================\nepoch: 1\ntrain_epoch_loss:0.00142, train_epoch_accuracy:0.93559\nepoch_val_loss:0.00229, epoch_val_accuracy:0.91454\n=========================\nepoch: 2\ntrain_epoch_loss:0.00129, train_epoch_accuracy:0.94372\nepoch_val_loss:0.00149, epoch_val_accuracy:0.94515\n\t model will be saved in best_model_acc_0.95.pt\n=========================\nepoch: 3\ntrain_epoch_loss:0.00139, train_epoch_accuracy:0.93748\nepoch_val_loss:0.00144, epoch_val_accuracy:0.94770\n\t model will be saved in best_model_acc_0.95.pt\n=========================\nepoch: 4\ntrain_epoch_loss:0.00135, train_epoch_accuracy:0.94052\nepoch_val_loss:0.00139, epoch_val_accuracy:0.94260\n=========================\nepoch: 5\ntrain_epoch_loss:0.00123, train_epoch_accuracy:0.94696\nepoch_val_loss:0.00110, epoch_val_accuracy:0.95281\n\t model will be saved in best_model_acc_0.95.pt\n=========================\nepoch: 6\ntrain_epoch_loss:0.00116, train_epoch_accuracy:0.94964\nepoch_val_loss:0.00104, epoch_val_accuracy:0.96429\n\t model will be saved in best_model_acc_0.96.pt\n=========================\nepoch: 7\ntrain_epoch_loss:0.00114, train_epoch_accuracy:0.95197\nepoch_val_loss:0.00135, epoch_val_accuracy:0.94770\n=========================\nepoch: 8\ntrain_epoch_loss:0.00107, train_epoch_accuracy:0.95605\nepoch_val_loss:0.00092, epoch_val_accuracy:0.96301\n=========================\nepoch: 9\ntrain_epoch_loss:0.00099, train_epoch_accuracy:0.96030\nepoch_val_loss:0.00158, epoch_val_accuracy:0.94643\n=========================\ntraining is finished!\n","output_type":"stream"}]},{"cell_type":"markdown","source":"Load model","metadata":{}},{"cell_type":"code","source":"model = Model(model_params=model_params,\n              spectrogram_params=spectrogram_params,\n              device=device).to(device)\n\nmodel.load_state_dict(torch.load(\"best_model_acc_0.96.pt\"))","metadata":{"execution":{"iopub.status.busy":"2022-02-19T09:24:33.232141Z","iopub.execute_input":"2022-02-19T09:24:33.232784Z","iopub.status.idle":"2022-02-19T09:24:33.256310Z","shell.execute_reply.started":"2022-02-19T09:24:33.232744Z","shell.execute_reply":"2022-02-19T09:24:33.255636Z"},"trusted":true},"execution_count":90,"outputs":[{"execution_count":90,"output_type":"execute_result","data":{"text/plain":"<All keys matched successfully>"},"metadata":{}}]},{"cell_type":"code","source":"FileLink('best_model_acc_0.96.pt')","metadata":{"execution":{"iopub.status.busy":"2022-02-19T09:24:48.380541Z","iopub.execute_input":"2022-02-19T09:24:48.380829Z","iopub.status.idle":"2022-02-19T09:24:48.386648Z","shell.execute_reply.started":"2022-02-19T09:24:48.380798Z","shell.execute_reply":"2022-02-19T09:24:48.385890Z"},"trusted":true},"execution_count":93,"outputs":[{"execution_count":93,"output_type":"execute_result","data":{"text/plain":"/kaggle/working/best_model_acc_0.96.pt","text/html":"<a href='best_model_acc_0.96.pt' target='_blank'>best_model_acc_0.96.pt</a><br>"},"metadata":{}}]},{"cell_type":"markdown","source":"# Streaming simulation","metadata":{}},{"cell_type":"markdown","source":"StreamingPredictor sends audio chunks (instead of full audio) and returns male proba for every chunk. It's\nusefull when we can't send full audio and want get results immediately (real-time).   \n**Hypothesis**: starting from some moment a model will be sure in predictions so we can don't wait all chunks and decide what is class *earlier*. Earlier we will understand in the next intuitive sense:  \n- if a model has the last *confidence_duration* ms big (small) proba values then we already know what is male and don't wait new chunks\n- big (small) values we will control using *proba_offset*: if proba > 0.5 + proba_offset then gender is male, if proba < 0.5 - proba_offset then gender is female\n- if we didn't stop earlier than the last proba we use the last prediction for gender (as in training)","metadata":{}},{"cell_type":"markdown","source":"**Hypothesis**: starting from some moment a model will be sure in predictions so we can don't wait all chunks and decide what is class *earlier*.","metadata":{}},{"cell_type":"code","source":"class StreamingPredictor:\n    def __init__(self, model):\n        self.model = model\n        self.device = model.device\n        \n    def streaming_request(self, path_to_audio, chunk_size_ms):\n        chunk_size = int(chunk_size_ms * 1e-3 * SAMPLE_RATE)\n        assert chunk_size >= self.model.spectrogram_params['win_length']\n        amplitudes = get_amplitudes(path_to_audio, new_sample_rate=SAMPLE_RATE)\n        amplitudes = torch.unsqueeze(amplitudes, 0)\n        new_states = {}\n        for i in range(0, amplitudes.shape[1], chunk_size):\n            chunk = amplitudes[:, i:i+chunk_size]\n            if chunk.shape[1] < spectrogram_params['win_length']:  # the last chunk \n                pad_size = spectrogram_params['win_length'] - chunk.shape[1]\n                chunk = torch.cat((chunk, torch.zeros(1, pad_size)), dim=1)\n            logits, new_states = model(chunk.to(self.device), states=new_states)\n            logits = logits.view(-1)\n            proba = torch.sigmoid(logits[-1])\n            yield float(proba.detach().cpu().numpy())","metadata":{"execution":{"iopub.status.busy":"2022-02-19T09:24:58.196601Z","iopub.execute_input":"2022-02-19T09:24:58.197154Z","iopub.status.idle":"2022-02-19T09:24:58.207067Z","shell.execute_reply.started":"2022-02-19T09:24:58.197117Z","shell.execute_reply":"2022-02-19T09:24:58.206244Z"},"trusted":true},"execution_count":94,"outputs":[]},{"cell_type":"markdown","source":"Firstly, it's interesting to know **what is the small chunk** we can send in order don't lose accuracy significantly","metadata":{}},{"cell_type":"code","source":"def test_streaming_requests(model, audio_paths, genders, chunk_size_ms=300):\n    streaming_predictor = StreamingPredictor(model=model)\n    accuracy_sum = 0\n    for audio_path, gender in tqdm.tqdm(list(zip(audio_paths, genders))):\n        probas = []\n        for proba in streaming_predictor.streaming_request(audio_path, chunk_size_ms=chunk_size_ms):\n            probas.append(proba)\n        gender_pred = int(probas[-1] > 0.5)\n        gender_gt = 1 if gender == 'male' else 0\n        accuracy_sum += int(gender_pred == gender_gt)\n    print(f\"accuracy: {accuracy_sum / len(genders)}\")","metadata":{"execution":{"iopub.status.busy":"2022-02-19T09:25:06.188561Z","iopub.execute_input":"2022-02-19T09:25:06.188821Z","iopub.status.idle":"2022-02-19T09:25:06.194865Z","shell.execute_reply.started":"2022-02-19T09:25:06.188792Z","shell.execute_reply":"2022-02-19T09:25:06.194180Z"},"trusted":true},"execution_count":96,"outputs":[]},{"cell_type":"code","source":"min_available_chunk_size_ms = (model.spectrogram_params['win_length'] / SAMPLE_RATE) * 1e3\n\nfor chunk_size_ms in [80, 100, 200, 300, 500, MAX_LENGTH_SEC * 1e3]:\n    if chunk_size_ms < min_available_chunk_size_ms:\n        print(f\"chunk size {chunk_size_ms} ms is very small and will be ignored\")\n        continue\n    print(f\"chunk size is {chunk_size_ms} ms\")\n    test_streaming_requests(model=model,\n                            audio_paths=df_test_balanced['path_to_audio'].values, \n                            genders=df_test_balanced['gender'].values,\n                            chunk_size_ms=chunk_size_ms)\n    print(\"=\" * 25)","metadata":{"execution":{"iopub.status.busy":"2022-02-19T09:25:09.785998Z","iopub.execute_input":"2022-02-19T09:25:09.786627Z","iopub.status.idle":"2022-02-19T09:27:25.968430Z","shell.execute_reply.started":"2022-02-19T09:25:09.786587Z","shell.execute_reply":"2022-02-19T09:27:25.967636Z"},"trusted":true},"execution_count":97,"outputs":[{"name":"stdout","text":"chunk size is 80 ms\n","output_type":"stream"},{"name":"stderr","text":"100%|██████████| 784/784 [00:44<00:00, 17.73it/s]\n","output_type":"stream"},{"name":"stdout","text":"accuracy: 0.9630102040816326\n=========================\nchunk size is 100 ms\n","output_type":"stream"},{"name":"stderr","text":"100%|██████████| 784/784 [00:37<00:00, 20.90it/s]\n","output_type":"stream"},{"name":"stdout","text":"accuracy: 0.9655612244897959\n=========================\nchunk size is 200 ms\n","output_type":"stream"},{"name":"stderr","text":"100%|██████████| 784/784 [00:21<00:00, 36.39it/s]\n","output_type":"stream"},{"name":"stdout","text":"accuracy: 0.9655612244897959\n=========================\nchunk size is 300 ms\n","output_type":"stream"},{"name":"stderr","text":"100%|██████████| 784/784 [00:16<00:00, 48.93it/s]\n","output_type":"stream"},{"name":"stdout","text":"accuracy: 0.9668367346938775\n=========================\nchunk size is 500 ms\n","output_type":"stream"},{"name":"stderr","text":"100%|██████████| 784/784 [00:11<00:00, 69.96it/s]\n","output_type":"stream"},{"name":"stdout","text":"accuracy: 0.9642857142857143\n=========================\nchunk size is 7000.0 ms\n","output_type":"stream"},{"name":"stderr","text":"100%|██████████| 784/784 [00:05<00:00, 139.23it/s]","output_type":"stream"},{"name":"stdout","text":"accuracy: 0.9604591836734694\n=========================\n","output_type":"stream"},{"name":"stderr","text":"\n","output_type":"stream"}]},{"cell_type":"markdown","source":"As we see, chunk size can be any size >= 80 ms (ideally, all predicitons don't depends on chunk size, but results depends on chunk size because of padding in spectrogram and db-scale). But we choice 300 ms for the next experiments","metadata":{}},{"cell_type":"markdown","source":"Secondly, let's find some interesting values for confidence_duration and proba_offset to accelarate classification.","metadata":{}},{"cell_type":"code","source":"def get_predictions_for_params(model, audio_paths, genders, chunk_size_ms): \n    streaming_predictor = StreamingPredictor(model=model)\n    results = {\n        'chunk_size_ms': chunk_size_ms,\n        'data': [],\n    }\n    for audio_path, gender in tqdm.tqdm(list(zip(audio_paths, genders))):\n        probas = []\n        for proba in streaming_predictor.streaming_request(audio_path, chunk_size_ms=chunk_size_ms):\n            probas.append(proba)\n        gender_gt = 1 if gender == 'male' else 0\n        results['data'].append({\n            'probas': probas,\n            'gender_gt': gender_gt,\n        })\n    return results","metadata":{"execution":{"iopub.status.busy":"2022-02-19T09:27:30.611568Z","iopub.execute_input":"2022-02-19T09:27:30.612262Z","iopub.status.idle":"2022-02-19T09:27:30.620705Z","shell.execute_reply.started":"2022-02-19T09:27:30.612224Z","shell.execute_reply":"2022-02-19T09:27:30.619998Z"},"trusted":true},"execution_count":98,"outputs":[]},{"cell_type":"code","source":"results = get_predictions_for_params(model=model, \n                                     audio_paths=df_test_balanced['path_to_audio'].values, \n                                     genders=df_test_balanced['gender'].values,\n                                     chunk_size_ms=300)","metadata":{"execution":{"iopub.status.busy":"2022-02-19T09:27:30.900303Z","iopub.execute_input":"2022-02-19T09:27:30.900788Z","iopub.status.idle":"2022-02-19T09:27:46.564668Z","shell.execute_reply.started":"2022-02-19T09:27:30.900751Z","shell.execute_reply":"2022-02-19T09:27:46.563801Z"},"trusted":true},"execution_count":99,"outputs":[{"name":"stderr","text":"100%|██████████| 784/784 [00:15<00:00, 50.08it/s]\n","output_type":"stream"}]},{"cell_type":"code","source":"with open('results.pkl', 'wb') as handle:\n    pickle.dump(results, handle, protocol=pickle.HIGHEST_PROTOCOL)","metadata":{"execution":{"iopub.status.busy":"2022-02-19T09:27:46.566535Z","iopub.execute_input":"2022-02-19T09:27:46.566822Z","iopub.status.idle":"2022-02-19T09:27:46.574737Z","shell.execute_reply.started":"2022-02-19T09:27:46.566783Z","shell.execute_reply":"2022-02-19T09:27:46.574009Z"},"trusted":true},"execution_count":100,"outputs":[]},{"cell_type":"code","source":"for idx in [1, 100, -100, -1]:\n    plt.title(f\"probas for some audio (gender_gt={results['data'][idx]['gender_gt']})\")\n    plt.grid()\n    plt.ylabel(\"male proba\")\n    plt.xlabel(\"chunk_id\")\n    plt.plot(results['data'][idx]['probas'])\n    plt.show();","metadata":{"execution":{"iopub.status.busy":"2022-02-19T09:27:46.576259Z","iopub.execute_input":"2022-02-19T09:27:46.576802Z","iopub.status.idle":"2022-02-19T09:27:47.305730Z","shell.execute_reply.started":"2022-02-19T09:27:46.576765Z","shell.execute_reply":"2022-02-19T09:27:47.305027Z"},"trusted":true},"execution_count":101,"outputs":[{"output_type":"display_data","data":{"text/plain":"<Figure size 432x288 with 1 Axes>","image/png":"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\n"},"metadata":{"needs_background":"light"}},{"output_type":"display_data","data":{"text/plain":"<Figure size 432x288 with 1 Axes>","image/png":"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\n"},"metadata":{"needs_background":"light"}},{"output_type":"display_data","data":{"text/plain":"<Figure size 432x288 with 1 Axes>","image/png":"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\n"},"metadata":{"needs_background":"light"}},{"output_type":"display_data","data":{"text/plain":"<Figure size 432x288 with 1 Axes>","image/png":"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\n"},"metadata":{"needs_background":"light"}}]},{"cell_type":"code","source":"with open('results.pkl', 'rb') as handle:\n    results = pickle.load(handle)\n\nchunk_size_ms = results['chunk_size_ms']\n\n\nfor confidence_duration in [chunk_size_ms, 2*chunk_size_ms, 3*chunk_size_ms]:\n    confidence_chunks = int(np.ceil(confidence_duration / chunk_size_ms))\n    for proba_offset in [0.2, 0.3, 0.4]:\n        accuracy_sum = 0\n        stats = {\n            'num_chunks': [],\n            'last_used_chunk': [],\n        }\n        for elem in tqdm.tqdm(results['data']):\n            probas = elem['probas']\n            stats['num_chunks'].append(len(probas))\n            earlier_stopping = False\n            for i in range(len(probas)):\n                last_probas = probas[i - confidence_chunks : i]\n                if len(last_probas) >= confidence_chunks:\n                    if np.min(last_probas) >= 0.5 + proba_offset:  # male\n                        stats['last_used_chunk'].append(i)\n                        accuracy_sum += int(elem['gender_gt'] == 1)\n                        earlier_stopping = True\n                        break\n                    if np.max(last_probas) <= 0.5 - proba_offset:  # female\n                        stats['last_used_chunk'].append(i)\n                        accuracy_sum += int(elem['gender_gt'] == 0)\n                        earlier_stopping = True\n                        break\n            if not earlier_stopping:\n                stats['last_used_chunk'].append(len(probas) - 1)\n                if probas[-1] >= 0.5:\n                    accuracy_sum += int(elem['gender_gt'] == 1)\n                else:\n                    accuracy_sum += int(elem['gender_gt'] == 0)\n        fraction = (np.array(stats['last_used_chunk']) + 1) / np.array(stats['num_chunks'])\n        print(f\"confidence_duration: {confidence_duration} ms\")\n        print(f\"proba_offset: {proba_offset}\")\n        print(f\"fraction of used chunks: {np.mean(fraction) * 100: .2f} %\")\n        print(f\"accuracy: {accuracy_sum / len(results['data']) : .3f}\")\n        print(\"=\" * 25)","metadata":{"execution":{"iopub.status.busy":"2022-02-19T09:27:47.308653Z","iopub.execute_input":"2022-02-19T09:27:47.309066Z","iopub.status.idle":"2022-02-19T09:27:47.918467Z","shell.execute_reply.started":"2022-02-19T09:27:47.309014Z","shell.execute_reply":"2022-02-19T09:27:47.917700Z"},"trusted":true},"execution_count":102,"outputs":[{"name":"stderr","text":"100%|██████████| 784/784 [00:00<00:00, 14897.61it/s]\n","output_type":"stream"},{"name":"stdout","text":"confidence_duration: 300 ms\nproba_offset: 0.2\nfraction of used chunks:  39.54 %\naccuracy:  0.921\n=========================\n","output_type":"stream"},{"name":"stderr","text":"100%|██████████| 784/784 [00:00<00:00, 13527.45it/s]\n","output_type":"stream"},{"name":"stdout","text":"confidence_duration: 300 ms\nproba_offset: 0.3\nfraction of used chunks:  43.61 %\naccuracy:  0.945\n=========================\n","output_type":"stream"},{"name":"stderr","text":"100%|██████████| 784/784 [00:00<00:00, 11963.93it/s]\n","output_type":"stream"},{"name":"stdout","text":"confidence_duration: 300 ms\nproba_offset: 0.4\nfraction of used chunks:  48.12 %\naccuracy:  0.955\n=========================\n","output_type":"stream"},{"name":"stderr","text":"100%|██████████| 784/784 [00:00<00:00, 14436.83it/s]\n","output_type":"stream"},{"name":"stdout","text":"confidence_duration: 600 ms\nproba_offset: 0.2\nfraction of used chunks:  48.56 %\naccuracy:  0.940\n=========================\n","output_type":"stream"},{"name":"stderr","text":"100%|██████████| 784/784 [00:00<00:00, 13146.13it/s]\n","output_type":"stream"},{"name":"stdout","text":"confidence_duration: 600 ms\nproba_offset: 0.3\nfraction of used chunks:  52.04 %\naccuracy:  0.949\n=========================\n","output_type":"stream"},{"name":"stderr","text":"100%|██████████| 784/784 [00:00<00:00, 11886.48it/s]\n","output_type":"stream"},{"name":"stdout","text":"confidence_duration: 600 ms\nproba_offset: 0.4\nfraction of used chunks:  56.49 %\naccuracy:  0.960\n=========================\n","output_type":"stream"},{"name":"stderr","text":"100%|██████████| 784/784 [00:00<00:00, 14491.99it/s]\n","output_type":"stream"},{"name":"stdout","text":"confidence_duration: 900 ms\nproba_offset: 0.2\nfraction of used chunks:  56.61 %\naccuracy:  0.946\n=========================\n","output_type":"stream"},{"name":"stderr","text":"100%|██████████| 784/784 [00:00<00:00, 12909.05it/s]\n","output_type":"stream"},{"name":"stdout","text":"confidence_duration: 900 ms\nproba_offset: 0.3\nfraction of used chunks:  60.09 %\naccuracy:  0.958\n=========================\n","output_type":"stream"},{"name":"stderr","text":"100%|██████████| 784/784 [00:00<00:00, 10876.32it/s]","output_type":"stream"},{"name":"stdout","text":"confidence_duration: 900 ms\nproba_offset: 0.4\nfraction of used chunks:  64.00 %\naccuracy:  0.963\n=========================\n","output_type":"stream"},{"name":"stderr","text":"\n","output_type":"stream"}]},{"cell_type":"markdown","source":"See, that accuracy is better when we use high proba_offset (it's expected) and use sufficient confidence_duration (it's expected).   \nFrom experiments, we can accelerate predictions on ~36% using confidence_duration=900ms and proba_offset=0.4","metadata":{}},{"cell_type":"markdown","source":"**Thx for reading!**","metadata":{}}]}