{"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":"

\nInf Kernel

","metadata":{}},{"cell_type":"markdown","source":"* based [Simayi Kernel(Inf&Train)](https://www.kaggle.com/code/bibanh/lb-0-72-resnet34-melspectrogram-stage-2-inference)\n* Changed timm model","metadata":{}},{"cell_type":"markdown","source":"\n

\nLibs Timm

\n","metadata":{}},{"cell_type":"code","source":"import sys\nsys.path.append('/kaggle/input/bird-lib-20230309095737/')\nimport timm","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:19.526193Z","iopub.execute_input":"2023-03-12T10:35:19.526841Z","iopub.status.idle":"2023-03-12T10:35:23.686946Z","shell.execute_reply.started":"2023-03-12T10:35:19.526803Z","shell.execute_reply":"2023-03-12T10:35:23.685464Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\n

\nLibs

\n","metadata":{}},{"cell_type":"code","source":"import torch\nimport torch.nn as nn\nfrom torch.utils.data import DataLoader, Dataset, random_split\nimport torch.nn.functional as F\nimport torchaudio\nfrom torchaudio import transforms\nfrom IPython.display import Audio\nimport torchvision\nfrom sklearn.preprocessing import OneHotEncoder, LabelEncoder\nfrom sklearn.model_selection import StratifiedKFold\nfrom sklearn.metrics import roc_auc_score\nimport numpy as np\nimport pandas as pd\nimport os\nimport glob\nimport math, random","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","execution":{"iopub.status.busy":"2023-03-12T10:35:23.689279Z","iopub.execute_input":"2023-03-12T10:35:23.689933Z","iopub.status.idle":"2023-03-12T10:35:25.039898Z","shell.execute_reply.started":"2023-03-12T10:35:23.689888Z","shell.execute_reply":"2023-03-12T10:35:25.038762Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def seed_everything(seed):\n random.seed(seed)\n os.environ['PYTHONHASHSEED'] = str(seed)\n np.random.seed(seed)\n torch.manual_seed(seed)\n torch.cuda.manual_seed(seed)\n torch.backends.cudnn.deterministic = True\n\nseed = 42\nseed_everything(seed)","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:25.041502Z","iopub.execute_input":"2023-03-12T10:35:25.042551Z","iopub.status.idle":"2023-03-12T10:35:25.053005Z","shell.execute_reply.started":"2023-03-12T10:35:25.042510Z","shell.execute_reply":"2023-03-12T10:35:25.051823Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\n

\nCFG

\n","metadata":{}},{"cell_type":"code","source":"class CFG:\n isOneHot = False\n rate = 32000\n num_classes = 264","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:25.055910Z","iopub.execute_input":"2023-03-12T10:35:25.056350Z","iopub.status.idle":"2023-03-12T10:35:25.071259Z","shell.execute_reply.started":"2023-03-12T10:35:25.056315Z","shell.execute_reply":"2023-03-12T10:35:25.069981Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

\nPreprocess

","metadata":{}},{"cell_type":"code","source":"class AudioUtil():\n @staticmethod\n def open(audio_file):\n sig, sr = torchaudio.load(audio_file)\n return (sig, sr)\n\n @staticmethod\n def rechannel(aud, new_channel):\n sig, sr = aud\n\n if (sig.shape[0] == new_channel):\n # Nothing to do\n return aud\n\n if (new_channel == 1):\n # Convert from stereo to mono by selecting only the first channel\n resig = sig[:1, :]\n else:\n # Convert from mono to stereo by duplicating the first channel\n resig = torch.cat([sig, sig, sig])\n\n return ((resig, sr))\n\n @staticmethod\n def resample(aud, newsr):\n sig, sr = aud\n\n if (sr == newsr):\n # Nothing to do\n return aud\n\n num_channels = sig.shape[0]\n # Resample first channel\n resig = torchaudio.transforms.Resample(sr, newsr)(sig[:1,:])\n if (num_channels > 1):\n # Resample the second channel and merge both channels\n retwo = torchaudio.transforms.Resample(sr, newsr)(sig[1:,:])\n resig = torch.cat([resig, retwo])\n\n return ((resig, newsr))\n\n @staticmethod\n def pad_trunc(aud, max_ms):\n sig, sr = aud\n num_rows, sig_len = sig.shape\n max_len = sr//1000 * max_ms\n\n if (sig_len > max_len):\n # Truncate the signal to the given length\n sig = sig[:,:max_len]\n\n elif (sig_len < max_len):\n # Length of padding to add at the beginning and end of the signal\n pad_begin_len = random.randint(0, max_len - sig_len)\n pad_end_len = max_len - sig_len - pad_begin_len\n\n # Pad with 0s\n pad_begin = torch.zeros((num_rows, pad_begin_len))\n pad_end = torch.zeros((num_rows, pad_end_len))\n\n sig = torch.cat((pad_begin, sig, pad_end), 1)\n \n return (sig, sr)\n\n @staticmethod\n def time_shift(aud, shift_limit):\n sig,sr = aud\n _, sig_len = sig.shape\n shift_amt = int(random.random() * shift_limit * sig_len)\n return (sig.roll(shift_amt), sr)\n\n @staticmethod\n def spectro_gram(aud, n_mels=64, n_fft=1024, hop_len=None):\n sig,sr = aud\n top_db = 80\n\n # spec has shape [channel, n_mels, time], where channel is mono, stereo etc\n spec = torchaudio.transforms.MelSpectrogram(sr, n_fft=n_fft, hop_length=hop_len, n_mels=n_mels)(sig)\n\n # Convert to decibels\n spec = torchaudio.transforms.AmplitudeToDB(top_db=top_db)(spec)\n return (spec)\n\n @staticmethod\n def spectro_augment(spec, max_mask_pct=0.1, n_freq_masks=1, n_time_masks=1):\n _, n_mels, n_steps = spec.shape\n mask_value = spec.mean()\n aug_spec = spec\n\n freq_mask_param = max_mask_pct * n_mels\n for _ in range(n_freq_masks):\n aug_spec = torchaudio.transforms.FrequencyMasking(freq_mask_param)(aug_spec, mask_value)\n\n time_mask_param = max_mask_pct * n_steps\n for _ in range(n_time_masks):\n aug_spec = torchaudio.transforms.TimeMasking(time_mask_param)(aug_spec, mask_value)\n\n return aug_spec","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:25.073088Z","iopub.execute_input":"2023-03-12T10:35:25.073494Z","iopub.status.idle":"2023-03-12T10:35:25.093320Z","shell.execute_reply.started":"2023-03-12T10:35:25.073434Z","shell.execute_reply":"2023-03-12T10:35:25.092094Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def preprocessing(aud):\n duration = 8000\n sr = 32000\n channel = 3\n shift_pct = 0.4\n reaud = AudioUtil.resample(aud, sr)\n rechan = AudioUtil.rechannel(reaud, channel)\n dur_aud = AudioUtil.pad_trunc(rechan, duration)\n shift_aud = AudioUtil.time_shift(dur_aud, shift_pct)\n sgram = AudioUtil.spectro_gram(shift_aud, n_mels=64, n_fft=1024, hop_len=None)\n aug_sgram = AudioUtil.spectro_augment(sgram, max_mask_pct=0.1, n_freq_masks=2, n_time_masks=2)\n aug_sgram_m, aug_sgram_s = aug_sgram.mean(), aug_sgram.std()\n aug_sgram = (aug_sgram - aug_sgram_m) / aug_sgram_s\n return aug_sgram","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:25.095276Z","iopub.execute_input":"2023-03-12T10:35:25.095698Z","iopub.status.idle":"2023-03-12T10:35:25.109890Z","shell.execute_reply.started":"2023-03-12T10:35:25.095647Z","shell.execute_reply":"2023-03-12T10:35:25.108707Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test_samples = list(glob.glob(\"/kaggle/input/birdclef-2023/test_soundscapes/*.ogg\"))\ntest_samples","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:25.111771Z","iopub.execute_input":"2023-03-12T10:35:25.112098Z","iopub.status.idle":"2023-03-12T10:35:25.130557Z","shell.execute_reply.started":"2023-03-12T10:35:25.112069Z","shell.execute_reply":"2023-03-12T10:35:25.129221Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"preprocessing(torchaudio.load(test_samples[0])).shape","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:25.132478Z","iopub.execute_input":"2023-03-12T10:35:25.132947Z","iopub.status.idle":"2023-03-12T10:35:26.425073Z","shell.execute_reply.started":"2023-03-12T10:35:25.132900Z","shell.execute_reply":"2023-03-12T10:35:26.423967Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"meta_df = pd.read_csv('/kaggle/input/birdclef-2023/train_metadata.csv')\nprint('data shape:',meta_df.shape)\nmeta_df.head(2)","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:26.426444Z","iopub.execute_input":"2023-03-12T10:35:26.426815Z","iopub.status.idle":"2023-03-12T10:35:26.568582Z","shell.execute_reply.started":"2023-03-12T10:35:26.426779Z","shell.execute_reply":"2023-03-12T10:35:26.567498Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"le = LabelEncoder().fit(meta_df['primary_label'])\ncompetition_classes = le.classes_","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:26.572516Z","iopub.execute_input":"2023-03-12T10:35:26.572868Z","iopub.status.idle":"2023-03-12T10:35:26.582710Z","shell.execute_reply.started":"2023-03-12T10:35:26.572837Z","shell.execute_reply":"2023-03-12T10:35:26.581328Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"forced_defaults = 0\ncompetition_class_map = []\nfor c in competition_classes:\n try:\n i = classes.index(c)\n competition_class_map.append(i)\n except:\n competition_class_map.append(0)\n forced_defaults += 1","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:26.583996Z","iopub.execute_input":"2023-03-12T10:35:26.584839Z","iopub.status.idle":"2023-03-12T10:35:26.593342Z","shell.execute_reply.started":"2023-03-12T10:35:26.584802Z","shell.execute_reply":"2023-03-12T10:35:26.592274Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"sample_sub = pd.read_csv(\"/kaggle/input/birdclef-2023/sample_submission.csv\")\nsample_sub[competition_classes] = sample_sub[competition_classes].astype(np.float32)\nsample_sub.head()","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:26.595001Z","iopub.execute_input":"2023-03-12T10:35:26.595323Z","iopub.status.idle":"2023-03-12T10:35:26.706120Z","shell.execute_reply.started":"2023-03-12T10:35:26.595293Z","shell.execute_reply":"2023-03-12T10:35:26.704981Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"for idx, col in enumerate(sample_sub.drop(columns = ['row_id'], axis = 1).columns):\n if col != competition_classes[idx]:\n print('Not fit class!')\nprint('If dont have any log, all fit')","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:26.707721Z","iopub.execute_input":"2023-03-12T10:35:26.708050Z","iopub.status.idle":"2023-03-12T10:35:26.718200Z","shell.execute_reply.started":"2023-03-12T10:35:26.708019Z","shell.execute_reply":"2023-03-12T10:35:26.716909Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

\nInference

","metadata":{}},{"cell_type":"markdown","source":"\n

\nModel Load & Selection

\n","metadata":{}},{"cell_type":"code","source":"import glob\nmodel_file_list = glob.glob(\"/kaggle/input/bird-train-a-121-20230309095332/*\")\nmodel_file_list","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:26.719912Z","iopub.execute_input":"2023-03-12T10:35:26.720236Z","iopub.status.idle":"2023-03-12T10:35:26.737780Z","shell.execute_reply.started":"2023-03-12T10:35:26.720206Z","shell.execute_reply":"2023-03-12T10:35:26.736838Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"model_file_list = [\n '/kaggle/input/bird-train-a-121-20230309095332/model_fold3_epoch9.pth',\n '/kaggle/input/bird-train-a-121-20230309095332/model_fold2_epoch9.pth',\n# '/kaggle/input/bird-train-a-121-20230309095332/model_fold4_epoch9.pth',\n# '/kaggle/input/bird-train-a-121-20230309095332/model_fold0_epoch9.pth',\n# '/kaggle/input/bird-train-a-121-20230309095332/model_fold1_epoch9.pth',\n]\nmodel_file_list","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:26.739282Z","iopub.execute_input":"2023-03-12T10:35:26.739618Z","iopub.status.idle":"2023-03-12T10:35:26.747236Z","shell.execute_reply.started":"2023-03-12T10:35:26.739567Z","shell.execute_reply":"2023-03-12T10:35:26.745882Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"model_list = []\nfor _model_path in model_file_list:\n model = timm.create_model(\"tf_efficientnetv2_s\", \n pretrained=False, \n num_classes=264)\n model.load_state_dict(torch.load(_model_path, map_location=torch.device('cpu')))\n model_list.append(model)","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:26.748500Z","iopub.execute_input":"2023-03-12T10:35:26.748826Z","iopub.status.idle":"2023-03-12T10:35:32.821112Z","shell.execute_reply.started":"2023-03-12T10:35:26.748795Z","shell.execute_reply":"2023-03-12T10:35:32.820097Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\n

\nPred

\n","metadata":{}},{"cell_type":"code","source":"def predict_for_sample(filename, sample_submission):\n file_id = filename.split(\".ogg\")[0].split(\"/\")[-1]\n lstFile = list(sample_submission[sample_submission.row_id.str.contains( file_id + \"_\")].row_id.unique())\n data, rat = torchaudio.load(filename)\n for file in lstFile:\n end = int(file.split('_')[-1])\n start = end - 5\n chunk = (data[:, start*rat:end*rat - 1], rat)\n samples = preprocessing(chunk).unsqueeze(0)\n \n \"\"\" Iter Inf Fold \"\"\"\n proba_list = []\n for _model in model_list:\n device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n myModel = _model.to(device)\n myModel = myModel.eval()\n probabilities = myModel(samples)\n probabilities = torch.nn.functional.softmax(probabilities).detach().numpy()\n proba_list.append(probabilities)\n \n \"\"\" Merge \"\"\"\n probabilities = [sum(column)/len(model_file_list) for column in zip(*proba_list)]\n sample_submission.loc[sample_submission.row_id == file, competition_classes] = probabilities","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:32.822550Z","iopub.execute_input":"2023-03-12T10:35:32.823335Z","iopub.status.idle":"2023-03-12T10:35:32.832761Z","shell.execute_reply.started":"2023-03-12T10:35:32.823300Z","shell.execute_reply":"2023-03-12T10:35:32.831893Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\n

\nCreateSub

\n","metadata":{}},{"cell_type":"code","source":"for sample_filename in test_samples:\n predict_for_sample(sample_filename, sample_sub)\nsample_sub","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:32.833974Z","iopub.execute_input":"2023-03-12T10:35:32.834542Z","iopub.status.idle":"2023-03-12T10:35:35.009315Z","shell.execute_reply.started":"2023-03-12T10:35:32.834509Z","shell.execute_reply":"2023-03-12T10:35:35.008078Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"sample_sub.to_csv(\"submission.csv\", index=False)\n","metadata":{"execution":{"iopub.status.busy":"2023-03-12T10:35:35.010757Z","iopub.execute_input":"2023-03-12T10:35:35.011110Z","iopub.status.idle":"2023-03-12T10:35:35.023221Z","shell.execute_reply.started":"2023-03-12T10:35:35.011078Z","shell.execute_reply":"2023-03-12T10:35:35.022102Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]}]}