{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.10.13","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"none","dataSources":[{"sourceId":70203,"databundleVersionId":8068726,"sourceType":"competition"},{"sourceId":1071249,"sourceType":"datasetVersion","datasetId":595272}],"dockerImageVersionId":30698,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"\n
![](\"https://storage.googleapis.com/kaggle-media/competitions/BirdCLEF/sholicola-ian_lockwood_inset.jpeg\")
\n
đŠBirdCLEF 2024 - Bird call/song recognitionđŠ
\n","metadata":{}},{"cell_type":"markdown","source":"# **Importing all necessary libraries**","metadata":{}},{"cell_type":"code","source":"import numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\nimport regex as re\nimport os\n\nimport librosa # analyzing and processing audio and music signals. It provides tools for feature extraction, time-frequency analysis, and audio playback. \nimport IPython.display as ipd \nimport librosa.display as lid\n\nimport matplotlib.pyplot as plt\nimport matplotlib as mpl\nfrom plotly.subplots import make_subplots\nimport plotly.express as px\n\nfrom PIL import Image # Support for opening, manipulating, and saving many different image file formats\nimport os\nfrom sklearn.utils import shuffle\nimport plotly.graph_objects as go\n\ncmap = mpl.colormaps.get_cmap('coolwarm')","metadata":{"execution":{"iopub.status.busy":"2024-05-15T05:31:07.218488Z","iopub.execute_input":"2024-05-15T05:31:07.219356Z","iopub.status.idle":"2024-05-15T05:31:07.760534Z","shell.execute_reply.started":"2024-05-15T05:31:07.219323Z","shell.execute_reply":"2024-05-15T05:31:07.75971Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# This Python 3 environment comes with many helpful analytics libraries installed\n# It is defined by the kaggle/python Docker image: https://github.com/kaggle/docker-python\n\n# Input data files are available in the read-only \"../input/\" directory\n# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory\n\n# for dirname, _, filenames in os.walk('/kaggle/input'):\n# for filename in filenames:\n# print(os.path.join(dirname, filename))\n\n# You can write up to 20GB to the current directory (/kaggle/working/) that gets preserved as output when you create a version using \"Save & Run All\" \n# You can also write temporary files to /kaggle/temp/, but they won't be saved outside of the current session","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","_kg_hide-output":true,"_kg_hide-input":true,"execution":{"iopub.status.busy":"2024-05-15T04:13:30.02538Z","iopub.execute_input":"2024-05-15T04:13:30.026039Z","iopub.status.idle":"2024-05-15T04:13:30.032307Z","shell.execute_reply.started":"2024-05-15T04:13:30.025984Z","shell.execute_reply":"2024-05-15T04:13:30.031061Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# **EDA**\n","metadata":{}},{"cell_type":"markdown","source":"# The .csv file đ\n\n> đ**Note**:\n* `train.csv` contains information about the audio files available in `train_audio`. It contains 24459 datapoints in 12 unique columns.\n\n\nNote: The TRAIN data has 1 labeled bird species per recording (primary_label). However, in nature usually you can hear tens (even hundreds) of birds in one go, so in TEST set we need to predict 0, 1 or multiple species for one recording. Because of this, in TRAIN we have species column - or primary_label - (main bird) or secondarylabel \n
","metadata":{}},{"cell_type":"code","source":"data = pd.read_csv(\"/kaggle/input/birdclef-2024/train_metadata.csv\")\ndata.head()","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:30.033801Z","iopub.execute_input":"2024-05-15T04:13:30.034445Z","iopub.status.idle":"2024-05-15T04:13:30.262494Z","shell.execute_reply.started":"2024-05-15T04:13:30.034328Z","shell.execute_reply":"2024-05-15T04:13:30.261352Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## The 12 unique columns are:\n* primary_label : species of the bird ,\n* secondary_labels : background/other species heard in the respective audio file,\n* type : song or call,\n* latitude ,longitude : of the audio recording ,\n* scientific_name ,common_name ,\n* author : person who recorded,\n* license ,\n* rating ,\n* url : audio from the website xeno_canto ,\n* filename: audio file in train_audio dir*","metadata":{}},{"cell_type":"code","source":"data.shape","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:30.263904Z","iopub.execute_input":"2024-05-15T04:13:30.264388Z","iopub.status.idle":"2024-05-15T04:13:30.271515Z","shell.execute_reply.started":"2024-05-15T04:13:30.264288Z","shell.execute_reply":"2024-05-15T04:13:30.270281Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# **Primary label exploration**","metadata":{}},{"cell_type":"code","source":"data['primary_label'].value_counts()\n\n# There are 182 unique species across the dataset","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:30.274473Z","iopub.execute_input":"2024-05-15T04:13:30.274885Z","iopub.status.idle":"2024-05-15T04:13:30.295758Z","shell.execute_reply.started":"2024-05-15T04:13:30.274855Z","shell.execute_reply":"2024-05-15T04:13:30.294445Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"labels = data['primary_label'].nunique()\nprint(\"Number of species: \", labels)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:30.297381Z","iopub.execute_input":"2024-05-15T04:13:30.297822Z","iopub.status.idle":"2024-05-15T04:13:30.308508Z","shell.execute_reply.started":"2024-05-15T04:13:30.297781Z","shell.execute_reply":"2024-05-15T04:13:30.307057Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Plotting the number of samples for each species \nspecies = data['primary_label'].value_counts()\nfig = go.Figure(data=[go.Bar(y=species.values, x=species.index)],\n layout=go.Layout(margin=go.layout.Margin(l=0, r=0, b=10, t=50)))\n# number of recordings for each species in the training metadata\nfig.update_layout(title='Number of traning samples per species')\nfig.show()","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:30.310168Z","iopub.execute_input":"2024-05-15T04:13:30.31099Z","iopub.status.idle":"2024-05-15T04:13:30.808977Z","shell.execute_reply.started":"2024-05-15T04:13:30.310951Z","shell.execute_reply":"2024-05-15T04:13:30.808156Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Average species in the metadata\nspecies = data['primary_label'].value_counts()\nsumall = 0\nfor i in species.values:\n sumall += i\nsumall/182","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:30.810164Z","iopub.execute_input":"2024-05-15T04:13:30.810867Z","iopub.status.idle":"2024-05-15T04:13:30.822801Z","shell.execute_reply.started":"2024-05-15T04:13:30.810836Z","shell.execute_reply":"2024-05-15T04:13:30.821883Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# **Secondary label exploration**","metadata":{}},{"cell_type":"code","source":"# These secondary label attributes pertain to the background bird calls within the audio data. While not as prominent as the primary call, these secondary calls originate from distinct species.\ndata['secondary_labels'].value_counts()","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:30.823954Z","iopub.execute_input":"2024-05-15T04:13:30.824474Z","iopub.status.idle":"2024-05-15T04:13:30.845679Z","shell.execute_reply.started":"2024-05-15T04:13:30.824445Z","shell.execute_reply":"2024-05-15T04:13:30.844267Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Counting the number of calls in secondary label of each sample and adding a new column to keep track of it in the data frame\ndata['secondary_labels'] = data['secondary_labels'].astype(str)\ndata['secondary_labels'] = data['secondary_labels'].apply(lambda x: re.findall(r\"'(\\w+)'\", x))\ndata['len_sec_labels'] = data['secondary_labels'].map(len)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:30.847365Z","iopub.execute_input":"2024-05-15T04:13:30.848052Z","iopub.status.idle":"2024-05-15T04:13:31.217849Z","shell.execute_reply.started":"2024-05-15T04:13:30.848009Z","shell.execute_reply":"2024-05-15T04:13:31.216699Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"data[data.len_sec_labels>3].sample(3)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:31.219739Z","iopub.execute_input":"2024-05-15T04:13:31.22018Z","iopub.status.idle":"2024-05-15T04:13:31.250512Z","shell.execute_reply.started":"2024-05-15T04:13:31.22014Z","shell.execute_reply":"2024-05-15T04:13:31.24942Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#### **This is a configuration class CFG that holds various parameters and settings for a machine learning model, designed specifically for BirdCLEF 24 dataset.**","metadata":{}},{"cell_type":"code","source":"class CFG:\n seed = 42\n # Input image size and batch size\n img_size = [128, 384]\n# batch_size = 64\n \n # Audio duration, sample rate, and length\n duration = 15 # second\n sample_rate = 32000\n audio_len = duration*sample_rate\n \n # STFT parameters\n nfft = 2028\n window = 2048\n hop_length = audio_len // (img_size[1] - 1)\n fmin = 20\n fmax = 16000\n \n # Data augmentation parameters\n augment=True\n \n # Class Labels for BirdCLEF 24\n class_names = sorted(os.listdir('/kaggle/input/birdclef-2024/train_audio/')) # Lists all the files and directories in the train_audio directory of the BirdCLEF 24 dataset and it is sorted alphabetically\n num_classes = len(class_names) # Determines the number of classes by counting the elements in the class_names list\n class_labels = list(range(num_classes)) # Generating a list of numerical labels ranging from 0 to num_classes - 1.\n label2name = dict(zip(class_labels, class_names)) # mapping numerical labels to class names\n name2label = {v:k for k,v in label2name.items()} # mapping class names to numerical labels.\n ","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:31.252434Z","iopub.execute_input":"2024-05-15T04:13:31.252924Z","iopub.status.idle":"2024-05-15T04:13:31.274708Z","shell.execute_reply.started":"2024-05-15T04:13:31.252883Z","shell.execute_reply":"2024-05-15T04:13:31.273534Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# **Adding the corresponding audio file's path for every sample and creating a new column in the df**","metadata":{}},{"cell_type":"code","source":"BASE_PATH = '/kaggle/input/birdclef-2024'\ndata = pd.read_csv(f'{BASE_PATH}/train_metadata.csv')\ndata['filepath'] = BASE_PATH + '/train_audio/' + data.filename \ndata['target'] = data.primary_label.map(CFG.name2label)\ndata['filename'] = data.filepath.map(lambda x: x.split('/')[-1])\ndata['xc_id'] = data.filepath.map(lambda x: x.split('/')[-1].split('.')[0])\n\n# Display rwos\nprint(data.filepath[0])\nprint(data.target[24458]) # class index starts from 0 -181 (total of 182 classes)\nprint(data.filename[0])\nprint(data.xc_id[0])\n\ndata.head(2)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:31.276429Z","iopub.execute_input":"2024-05-15T04:13:31.277433Z","iopub.status.idle":"2024-05-15T04:13:31.496216Z","shell.execute_reply.started":"2024-05-15T04:13:31.277392Z","shell.execute_reply":"2024-05-15T04:13:31.49476Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def load_audio(filepath):\n audio, sr = librosa.load(filepath)\n return audio, sr\n\ndef get_spectrogram(audio):\n spec = librosa.feature.melspectrogram(y=audio, \n sr=CFG.sample_rate, \n n_mels=256,\n n_fft=2048,\n hop_length=512,\n fmax=CFG.fmax,\n fmin=CFG.fmin,\n )\n spec = librosa.power_to_db(spec, ref=1.0)\n min_ = spec.min()\n max_ = spec.max()\n if max_ != min_:\n spec = (spec - min_)/(max_ - min_)\n return spec\n\ndef display_audio(row):\n # Caption for viz\n caption = f'Id: {row.filename} | Name: {row.common_name} | Sci.Name: {row.scientific_name} | Rating: {row.rating}'\n # Read audio file\n audio, sr = load_audio(row.filepath)\n # Keep fixed length audio\n audio = audio[:CFG.audio_len]\n # Spectrogram from audio\n spec = get_spectrogram(audio)\n # Display audio\n print(\"# Audio:\")\n display(ipd.Audio(audio, rate=CFG.sample_rate))\n print('# Visualization:')\n fig, ax = plt.subplots(2, 1, figsize=(12, 2*3), sharex=True, tight_layout=True)\n fig.suptitle(caption)\n # Waveplot\n lid.waveshow(audio,\n sr=CFG.sample_rate,\n ax=ax[0],\n color= cmap(0.1))\n # Specplot\n lid.specshow(spec, \n sr = CFG.sample_rate, \n hop_length=512,\n n_fft=2048,\n fmin=CFG.fmin,\n fmax=CFG.fmax,\n x_axis = 'time', \n y_axis = 'mel',\n cmap = 'coolwarm',\n ax=ax[1])\n ax[0].set_xlabel('');\n fig.show()","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:31.501778Z","iopub.execute_input":"2024-05-15T04:13:31.502167Z","iopub.status.idle":"2024-05-15T04:13:31.51665Z","shell.execute_reply.started":"2024-05-15T04:13:31.502136Z","shell.execute_reply":"2024-05-15T04:13:31.515403Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"row = data.iloc[35]\n\n# Display audio\ndisplay_audio(row)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:31.518123Z","iopub.execute_input":"2024-05-15T04:13:31.518671Z","iopub.status.idle":"2024-05-15T04:13:47.591604Z","shell.execute_reply.started":"2024-05-15T04:13:31.518444Z","shell.execute_reply":"2024-05-15T04:13:47.590194Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"row = data.iloc[120]\n\n# Display audio\ndisplay_audio(row)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:47.593038Z","iopub.execute_input":"2024-05-15T04:13:47.593686Z","iopub.status.idle":"2024-05-15T04:13:49.419428Z","shell.execute_reply.started":"2024-05-15T04:13:47.593647Z","shell.execute_reply":"2024-05-15T04:13:49.418232Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"row = data.iloc[183]\n\n# Display audio\ndisplay_audio(row)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:49.420647Z","iopub.execute_input":"2024-05-15T04:13:49.420948Z","iopub.status.idle":"2024-05-15T04:13:51.542866Z","shell.execute_reply.started":"2024-05-15T04:13:49.420922Z","shell.execute_reply":"2024-05-15T04:13:51.542037Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"path = data.filepath[10] \npath","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:51.544284Z","iopub.execute_input":"2024-05-15T04:13:51.545095Z","iopub.status.idle":"2024-05-15T04:13:51.552089Z","shell.execute_reply.started":"2024-05-15T04:13:51.545062Z","shell.execute_reply":"2024-05-15T04:13:51.550744Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"signal, sr = librosa.load(path)\nprint('Signal:', signal)\nprint('Shape of signal:', np.shape(signal))\nprint('Sampling Rate (KHz):', sr)\nprint('Duration of audio:', np.shape(signal)[0]/sr)\nplt.plot(signal)\nplt.show()\nipd.Audio(path)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:51.553396Z","iopub.execute_input":"2024-05-15T04:13:51.553793Z","iopub.status.idle":"2024-05-15T04:13:51.913809Z","shell.execute_reply.started":"2024-05-15T04:13:51.553754Z","shell.execute_reply":"2024-05-15T04:13:51.91296Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def plot_dist(feature):\n feature_distribution = data[feature].value_counts().sort_index(ascending=False)\n colors = ['#2E8B57', '#3CB371', '#58D68D', '#74F9A4', '#9EF8C0',\n '#C7F8DB', '#E6F5D0', '#F9E79F', '#F7C677', '#F7A54A', '#F58231']\n rating_colors = {rating: color for rating, color in zip(feature_distribution.index, colors)}\n # Bar plot\n bar_data = go.Bar(x=feature_distribution.values, y=feature_distribution.index,\n orientation='h', marker=dict(color=[rating_colors[rating] for rating in feature_distribution.index]), showlegend=False, name='')\n # Donut chart\n labels = feature_distribution.index.tolist()\n values = feature_distribution.values.tolist()\n total_count = sum(values)\n percentages = [(value / total_count) * 100 for value in values]\n donut_data = go.Pie(labels=labels, values=values, marker=dict(colors=[rating_colors[rating] for rating in feature_distribution.index], line=dict(color='#ffffff', width=0.5)), hole=0.55,\n showlegend=False, name='', textinfo='label+percent', hoverinfo='label+percent', textposition='inside')\n fig = make_subplots(rows=1, cols=2, specs=[[{\"type\": \"bar\"}, {\"type\": \"pie\"}]])\n fig.add_trace(bar_data, row=1, col=1)\n fig.add_trace(donut_data, row=1, col=2)\n fig.update_layout(title=f'{feature} Distribution', plot_bgcolor='#1C1D20', paper_bgcolor='#1C1D20',\n title_font=dict(size=20, family=\"Lato, sans-serif\"), font=dict(color='#E1B12D'))\n fig.show()\nplot_dist(\"rating\")","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:15:28.328231Z","iopub.execute_input":"2024-05-15T04:15:28.328711Z","iopub.status.idle":"2024-05-15T04:15:28.651623Z","shell.execute_reply.started":"2024-05-15T04:15:28.328677Z","shell.execute_reply":"2024-05-15T04:15:28.650455Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"#Lets plot for western ghat\n# boundaries for the Western Ghats - as per Google search\nlower_latitude = 8\nupper_latitude = 22\nlower_longitude = 73\nupper_longitude = 79\nfiltered_data = data[(data['latitude'] >= lower_latitude) & \n (data['latitude'] <= upper_latitude) &\n (data['longitude'] >= lower_longitude) &\n (data['longitude'] <= upper_longitude)]\nfig = px.scatter_mapbox(filtered_data, lat=\"latitude\", lon=\"longitude\", color=\"common_name\",zoom=5, center={\"lat\": (lower_latitude + upper_latitude) / 2, \"lon\": (lower_longitude + upper_longitude) / 2})\nfig.update_layout( title=\"Distribution of Bird Species in the Western Ghats\", plot_bgcolor=None, paper_bgcolor=None,title_font=dict(size=20, family=\"Lato, sans-serif\", color='#E1B12D'), font=dict(color='#E1B12D'),mapbox_style=\"carto-positron\", margin=dict(t=50, r=50, b=50, l=50), hovermode='closest',\nshowlegend=True, legend=dict(title='Bird Species'), legend_title_font=dict(color='#E1B12D'))\nfig.show()","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:31:44.79859Z","iopub.execute_input":"2024-05-15T04:31:44.798992Z","iopub.status.idle":"2024-05-15T04:31:45.400064Z","shell.execute_reply.started":"2024-05-15T04:31:44.798961Z","shell.execute_reply":"2024-05-15T04:31:45.398842Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import librosa\nimport pandas as pd\nfrom IPython.display import Audio \ncolor_pal = plt.rcParams[\"axes.prop_cycle\"].by_key()[\"color\"]\n\naudio_1, sr_1= librosa.load(\"/kaggle/input/birdclef-2024/train_audio/asbfly/XC134896.ogg\", sr=32000)\nprint(audio_1.shape, sr_1)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T05:34:03.760593Z","iopub.execute_input":"2024-05-15T05:34:03.761007Z","iopub.status.idle":"2024-05-15T05:34:03.799632Z","shell.execute_reply.started":"2024-05-15T05:34:03.760978Z","shell.execute_reply":"2024-05-15T05:34:03.798504Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"pd.Series(audio_1).plot(figsize=(10, 5),\n lw=1,\n title='Raw Audio',\n color=color_pal[0])\nplt.show()","metadata":{"execution":{"iopub.status.busy":"2024-05-15T05:31:16.440012Z","iopub.execute_input":"2024-05-15T05:31:16.440378Z","iopub.status.idle":"2024-05-15T05:31:16.850142Z","shell.execute_reply.started":"2024-05-15T05:31:16.44035Z","shell.execute_reply":"2024-05-15T05:31:16.849131Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"Audio(data=audio_1, rate=sr_1)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T05:34:06.941435Z","iopub.execute_input":"2024-05-15T05:34:06.941823Z","iopub.status.idle":"2024-05-15T05:34:07.004602Z","shell.execute_reply.started":"2024-05-15T05:34:06.941791Z","shell.execute_reply":"2024-05-15T05:34:07.003256Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# UP UNTIL HERE","metadata":{}},{"cell_type":"code","source":"from sklearn.model_selection import train_test_split\nimport pandas as pd\n\ndef birds_stratified_split(df, target_col, test_size=0.2):\n class_counts = df[target_col].value_counts()\n low_count_classes = class_counts[class_counts < 2].index.tolist() ### Birds with single counts\n\n df['train'] = df[target_col].isin(low_count_classes)\n\n train_df, val_df = train_test_split(df[~df['train']], test_size=test_size, stratify=df[~df['train']][target_col], random_state=42)\n\n train_df = pd.concat([train_df, df[df['train']]], axis=0).reset_index(drop=True)\n\n # Remove the 'valid' column\n train_df.drop('train', axis=1, inplace=True)\n val_df.drop('train', axis=1, inplace=True)\n\n return train_df, val_df","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.080587Z","iopub.status.idle":"2024-05-15T04:13:52.08098Z","shell.execute_reply.started":"2024-05-15T04:13:52.080789Z","shell.execute_reply":"2024-05-15T04:13:52.080805Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_data, valid_data = birds_stratified_split(data, 'primary_label', 0.2)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.084089Z","iopub.status.idle":"2024-05-15T04:13:52.08505Z","shell.execute_reply.started":"2024-05-15T04:13:52.084743Z","shell.execute_reply":"2024-05-15T04:13:52.084769Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_data.head()","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.086665Z","iopub.status.idle":"2024-05-15T04:13:52.087431Z","shell.execute_reply.started":"2024-05-15T04:13:52.087104Z","shell.execute_reply":"2024-05-15T04:13:52.087129Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"data.primary_label.value_counts().tail(60)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.089062Z","iopub.status.idle":"2024-05-15T04:13:52.090243Z","shell.execute_reply.started":"2024-05-15T04:13:52.089951Z","shell.execute_reply":"2024-05-15T04:13:52.089975Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_data.primary_label.value_counts()","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.091775Z","iopub.status.idle":"2024-05-15T04:13:52.09238Z","shell.execute_reply.started":"2024-05-15T04:13:52.092089Z","shell.execute_reply":"2024-05-15T04:13:52.092112Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"valid_data.primary_label.value_counts()","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.094193Z","iopub.status.idle":"2024-05-15T04:13:52.094874Z","shell.execute_reply.started":"2024-05-15T04:13:52.094545Z","shell.execute_reply":"2024-05-15T04:13:52.094567Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Code adapted from: \n# https://www.kaggle.com/frlemarchand/bird-song-classification-using-an-efficientnet\n# Make sure to check out the entire notebook.\n\n# Load metadata file\ntrain = pd.read_csv('/kaggle/input/birdclef-2024/train_metadata.csv',)\n\n# Limit the number of training samples and classes\n# First, only use high quality samples\ntrain = train.query('rating>=3')\n\n# Second, assume that birds with the most training samples are also the most common\n# A species needs at least 200 recordings with a rating above 4 to be considered common\nbirds_count = {}\nfor bird_species, count in zip(train.primary_label.unique(), \n train.groupby('primary_label')['primary_label'].count().values):\n birds_count[bird_species] = count\nmost_represented_birds = [key for key,value in birds_count.items() if value >= 100] \n\nTRAIN = train.query('primary_label in @most_represented_birds')\nLABELS = sorted(TRAIN.primary_label.unique())\n\n# Let's see how many species and samples we have left\nprint('NUMBER OF SPECIES IN TRAIN DATA:', len(LABELS))\nprint('NUMBER OF SAMPLES IN TRAIN DATA:', len(TRAIN))\nprint('LABELS:', most_represented_birds)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.096182Z","iopub.status.idle":"2024-05-15T04:13:52.096596Z","shell.execute_reply.started":"2024-05-15T04:13:52.096402Z","shell.execute_reply":"2024-05-15T04:13:52.096419Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"TRAIN.head()","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.097617Z","iopub.status.idle":"2024-05-15T04:13:52.098003Z","shell.execute_reply.started":"2024-05-15T04:13:52.097807Z","shell.execute_reply":"2024-05-15T04:13:52.097823Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"TRAIN['primary_label'].value_counts()","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.099573Z","iopub.status.idle":"2024-05-15T04:13:52.099955Z","shell.execute_reply.started":"2024-05-15T04:13:52.099766Z","shell.execute_reply":"2024-05-15T04:13:52.099783Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"data_samp = TRAIN.sample(10)\ndata_samp.head()","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.101974Z","iopub.status.idle":"2024-05-15T04:13:52.102412Z","shell.execute_reply.started":"2024-05-15T04:13:52.102181Z","shell.execute_reply":"2024-05-15T04:13:52.102198Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import os\nfrom os import path\n\nimport numpy as np\nfrom numpy.linalg import norm\n\nimport pandas as pd\n\n# Signal processing\nfrom scipy.io import wavfile\nfrom scipy.interpolate import interp1d\nfrom scipy import signal\nfrom scipy.signal import periodogram\nfrom scipy.signal import find_peaks\n\n# Audio processing\nfrom pydub import AudioSegment\n\n# MFCC\nimport librosa, librosa.display\n\n# Wavelet \nimport pywt\nfrom pywt import wavedec\n\n# Visualization\nimport matplotlib.pyplot as plt\nfrom plotly.subplots import make_subplots\nimport plotly.graph_objects as go\nimport IPython\n\n# to display full dataframe information\npd.set_option('display.max_colwidth', None)\n\nfrom PIL import Image\n\nfrom mlxtend.preprocessing import minmax_scaling\n\nimport pickle\n\nfrom sklearn.decomposition import PCA\nfrom sklearn.cross_decomposition import CCA","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.104147Z","iopub.status.idle":"2024-05-15T04:13:52.105998Z","shell.execute_reply.started":"2024-05-15T04:13:52.105798Z","shell.execute_reply":"2024-05-15T04:13:52.105817Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def catstrvec_2_catnumvec(strcats, df_col):\n act_val = []\n for i in df_col.to_numpy():\n xy, x_ind, y_ind = np.intersect1d(strcats, i, return_indices=True)\n act_val.append(x_ind[0])\n \n return act_val\n\ndef convert_mp3_to_sig(audio_filepath):\n # files \n src = audio_filepath\n dst = \"test.wav\"\n\n # convert wav to mp3 \n sound = AudioSegment.from_mp3(src)\n sound.export(dst, format=\"wav\")\n \n return dst\n\n# Load a wav file\ndef get_wav_info(wav_file):\n rate, data = wavfile.read(wav_file)\n return rate, data\n\n# Calculate and plot spectrogram for a wav audio file\n# Binning the time-series and calculating the periodogram per time bin\n\ndef graph_spectrogram(wav_file):\n rate, data = get_wav_info(wav_file)\n # print('size of extracted data from sound file: ', data.shape)\n nfft = 100 # Length of each window segment - frequency data will be binned by nfft/2 \n fs = 8000 # Sampling frequencies\n noverlap = 0 #120 # Overlap between windows\n nchannels = data.ndim\n # print('nchannels: ', nchannels)\n \n # https://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.spectrogram.html\n \n # OR\n \n # pxx = periodogram per frequency (nfft/2 = num_of_freqs, bins=num_of_bins)\n # freqs = freqencies that correspond to each magnitude\n # bins = number of bins to group time-series\n # im = the image of the axis\n if nchannels == 1:\n pxx, freqs, bins, im = plt.specgram(data, nfft, fs, noverlap=noverlap)\n elif nchannels == 2:\n pxx, freqs, bins, im = plt.specgram(data[:,0], nfft, fs, noverlap=noverlap)\n \n # print('number of bins: ', bins.shape)\n # print('shape of pxx: ', pxx.shape)\n # print('shape of freqs: ', freqs.shape)\n \n plotORnot = 1\n if plotORnot == 1:\n plt.ylabel(\"Frequency\")\n plt.xlabel(\"Number of time windows\")\n plt.title(\"Spectrogram\")\n plt.show()\n\n return data, pxx, freqs\n\ndef make_a_properlist(vec):\n \n out = []\n for i in range(len(vec)):\n out = out + [np.ravel(vec[i])]\n \n if is_empty(out) == False:\n vecout = np.concatenate(out).ravel().tolist()\n else:\n vecout = list(np.ravel(out))\n \n return vecout\n\ndef is_empty(vec):\n vec = np.array(vec)\n if vec.shape[0] == 0:\n out = True\n else:\n out = False\n \n return out\n\ndef linear_intercurrentpt_makeshortSIGlong_interp1d(shortSIG, longSIG):\n\n x = np.linspace(shortSIG[0], len(shortSIG), num=len(shortSIG), endpoint=True)\n y = shortSIG\n # print('x : ', x)\n\n\n # -------------\n kind = 'linear'\n # kind : Specifies the kind of interpolation as a string or as an integer specifying the order of the spline interpolator to use. The string has to be one of âlinearâ, ânearestâ, ânearest-upâ, âzeroâ, âslinearâ, âquadraticâ, âcubicâ, âpreviousâ, or ânextâ. âzeroâ, âslinearâ, âquadraticâ and âcubicâ refer to a spline interpolation of zeroth, first, second or third order; âpreviousâ and ânextâ simply return the previous or next value of the point; ânearest-upâ and ânearestâ differ when interpolating half-integers (e.g. 0.5, 1.5) in that ânearest-upâ rounds up and ânearestâ rounds down. Default is âlinearâ.\n\n if kind == 'linear':\n f = interp1d(x, y)\n elif kind == 'cubic':\n f = interp1d(x, y, kind='cubic')\n # -------------\n\n xnew = np.linspace(shortSIG[0], len(shortSIG), num=len(longSIG), endpoint=True)\n # print('xnew : ', xnew)\n\n siglong = f(xnew)\n\n return siglong\n\ndef scale_feature_data(feat, plotORnot, type_scale):\n \n columns = ['0']\n dat = pd.DataFrame(data=feat, columns=columns)\n \n # which type of scaling\n if type_scale == 'minmax':\n # Values from 0 to 1\n scaled_data0 = minmax_scaling(dat, columns=columns)\n scaled_data = list(scaled_data0.to_numpy().ravel())\n \n elif type_scale == 'normalization':\n # normalization : same as mlxtend - Values from 0 to 1\n scaled_data = []\n for q in range(len(feat)):\n scaled_data.append( (feat[q] - np.min(feat))/(np.max(feat) - np.min(feat)) )\n \n elif type_scale == 'pos_normalization':\n # positive normalization : same as mlxtend - Values from 0 to 1\n shift_up = [i - np.min(feat) for i in feat]\n scaled_data = [q/np.max(shift_up) for q in shift_up]\n \n elif type_scale == 'standardization':\n # standardization : values are not restricted to a range, but scaled appropreately\n scaled_data = [(q - np.mean(feat))/np.std(feat) for q in feat]\n\n return scaled_data\n\n# level : the number of levels to decompose the time signal, le nombre des marquers par signale\ndef tsig_2_discrete_wavelet_transform(sig, waveletname, level, plotORnot):\n\n # On peut calculater dans deux façons: 0) dwt en boucle and then idwt, 1) wavedec et waverec\n # Mais le deux ne donnent pas le meme reponses, wavedec et waverec semble plus raisonable.\n coeff = wavedec(sig, waveletname, level)\n\n if plotORnot == 1:\n fig, axx = plt.subplots(nrows=len(coeff), ncols=1, figsize=(5,5))\n axx[0].set_title(\"coef\") # Pas certain si c'est coef0 ou coef1\n for k in range(len(coeff)):\n axx[k].plot(coeff[k], 'r') # output of the low pass filter (averaging filter) of the DWT\n plt.tight_layout()\n plt.show()\n\n return coeff\n\ndef tsig_2_spectrogram(sig, fs, nfft, noverlap, img_dim, plotORnot):\n\n # -----------------------------------\n fig,ax = plt.subplots(1)\n fig.subplots_adjust(left=0,right=1,bottom=0,top=1)\n ax.axis('off')\n # spectrum2D array : Columns are the periodograms of successive segments.\n # freqs1-D array : The frequencies corresponding to the rows in spectrum.\n # t1-D array : The times corresponding to midpoints of segments (i.e., the columns in spectrum).\n # imAxesImage : The image created by imshow containing the spectrogram.\n pxx, freqs, bins, img = ax.specgram(sig, nfft, fs, noverlap=noverlap)\n ax.axis('off')\n\n # -----------------------------------\n\n my_image = 'temp.png'\n fig.savefig(my_image)\n fname = os.path.abspath(os.getcwd()) + \"/\" + my_image\n \n # Convert image to an array:\n # Read image \n img = Image.open(fname) # PIL: img is not in array form, it is a PIL.PngImagePlugin.PngImageFile \n\n # -----------------------------------\n \n # Resize sectrogram image\n image = imgORmat_resize_imgORmat_CNN(img_dim, data_in=img, inpt='img3D', outpt='mat2D', norm='non', thresh='non')\n \n # -----------------------------------\n \n # Flatten image into a vector\n img_flatten = np.reshape(np.ravel(image), (img_dim*img_dim, ), order='F')\n # print('img_flatten.shape: ', img_flatten.shape)\n\n # -----------------------------------\n\n return img_flatten\nprint(\"skdlchn\")\n\ndef tsig_2_continuous_wavelet_transform(sig, fs, scales, waveletname, img_dim, plotORnot):\n \n # e.g.:\n # scales = np.arange(1, 128)\n # waveletname = 'mexh'\n # sig = data2[0:100]\n # print('len(sig): ', len(sig))\n \n dt = 1/fs\n # print('dt : ', dt)\n\n coefficients, frequencies = pywt.cwt(sig, scales, waveletname, dt)\n coefficients = np.array(coefficients)\n ylen, xlen = coefficients.shape\n\n # Time by frequency plot of cwt : then flatten and use as a feature\n stop_val = len(sig)/fs\n x = np.arange(0, stop_val, dt) # time\n y = frequencies # frequency \n X, Y = np.meshgrid(x, y)\n Z = coefficients\n\n # Each cwt versus frequencies\n # for i in range(len(coefficients)):\n # plt.plot(frequencies, coefficients[i])\n\n # \n fig = plt.figure()\n ax = plt.axes() # creates a 3D axis by using the keyword projection='3d'\n ax.contourf(X, Y, Z, xlen, cmap=plt.cm.seismic) # contour fill\n ax.axis('off')\n\n # -----------------------------------\n\n my_image = 'temp.png'\n fig.savefig(my_image)\n fname = os.path.abspath(os.getcwd()) + \"/\" + my_image\n\n # Convert image to an array:\n # Read image \n img = Image.open(fname) \n \n # -----------------------------------\n \n # Resize sectrogram image\n image = imgORmat_resize_imgORmat_CNN(img_dim, data_in=img, inpt='img3D', outpt='mat2D', norm='non', thresh='non')\n \n # -----------------------------------\n \n # Flatten image into a vector\n img_flatten = np.reshape(np.ravel(image), (img_dim*img_dim, ), order='F')\n # print('img_flatten.shape: ', img_flatten.shape)\n \n return img_flatten\n\ndef resize_img(img, img_dim):\n if type(img) == 'numpy.ndarray':\n # img is an array, retuns an image object\n rgb_image = Image.fromarray(img , 'RGB')\n else:\n # img is an image object, returns an image object\n try:\n rgb_image = img.convert('RGB')\n except AttributeError:\n rgb_image = Image.fromarray(img , 'RGB')\n\n # Resize image into a 64, 64, 3\n new_h, new_w = int(img_dim), int(img_dim)\n img3 = rgb_image.resize((new_w, new_h), Image.ANTIALIAS)\n w_resized, h_resized = img3.size[0], img3.size[1]\n return img3\n\ndef convert_img_a_mat(img, outpt):\n mat = np.array(img) # Convert image to an array\n if outpt == 'mat2D':\n # Transformer l'image de 3D Ă 2D\n # Convert image back to a 2D array\n matout = np.mean(mat, axis=2)\n elif outpt == 'img3D': # techniquement c'est un image parce qu'il y a trois RGB channels \n matout = mat\n return matout\n\ndef norm_mat(mat2Dor3D, norm):\n if norm == 'zero2one':\n # Normalizer l'image entre 0 et 1\n norout = mat2Dor3D/255\n elif norm == 'negone2posone':\n # Normalize the images to [-1, 1]\n norout = (mat2Dor3D - 127.5) / 127.5\n elif norm == 'non':\n norout = mat2Dor3D\n return norout\n\ndef threshold_mat(mat2D, thresh):\n # Threshold image\n val = 255/2\n if thresh == 'zero_moins_que_val':\n row, col = mat2D.shape\n mat_thresh = mat2D\n min_val = np.min(mat_thresh)\n for i in range(row):\n for j in range(col):\n if mat_thresh[i,j] < val:\n mat_thresh[i,j] = min_val\n elif thresh == 'non':\n mat_thresh = mat2D\n return mat_thresh\n\ndef imgORmat_resize_imgORmat_CNN(img_dim, data_in, inpt='img3D', outpt='mat2D', norm='non', thresh='non'):\n if inpt == 'img3D' and outpt=='mat2D':\n img = resize_img(data_in, img_dim)\n img3D = convert_img_a_mat(img, outpt)\n out = norm_mat(img3D, norm)\n elif inpt == 'mat2D' and outpt=='mat2D':\n data_in = np.array(data_in)\n img = Image.fromarray(data_in , 'L')\n img = resize_img(img, img_dim)\n mat2D = convert_img_a_mat(img, outpt)\n out = norm_mat(mat2D, norm)\n elif inpt == 'mat2D' and outpt=='img3D':\n data_in = np.array(data_in)\n img = Image.fromarray(data_in , 'L')\n img = resize_img(img, img_dim)\n img3D = convert_img_a_mat(img, outpt)\n out = norm_mat(img3D, norm)\n elif inpt == 'img3D' and outpt=='img3D':\n img = resize_img(data_in, img_dim)\n img3D = convert_img_a_mat(img, outpt)\n out = norm_mat(img3D, norm)\n\n return out\n\ndef save_dat_pickle(outSIG, file_name=\"outSIG.pkl\"):\n # Save data matrices to file\n open_file = open(file_name, \"wb\")\n pickle.dump(outSIG, open_file)\n open_file.close()\n\ndef load_dat_pickle(file_name=\"outSIG.pkl\"):\n open_file = open(file_name, \"rb\")\n dataout = pickle.load(open_file)\n open_file.close()\n return dataout\n\ndef binarize_Y1Dvec_2_Ybin(Y):\n \n # Transform a 1D Y vector (n_samples by 1) to a Y_bin (n_samples by n_classes) vector\n\n # Ensure vector is of integers\n Y = [int(i) for i in Y]\n\n # Number of samples\n m_examples = len(Y)\n\n # Number of classes\n temp = np.unique(Y)\n unique_classes = [int(i) for i in temp]\n # print('unique_classes : ', unique_classes)\n\n whichone = 2\n # Binarize the output\n if whichone == 0:\n from sklearn.preprocessing import label_binarize\n Y_bin = label_binarize(Y, classes=unique_classes) # seems to work now\n\n elif whichone == 1:\n from sklearn import preprocessing\n lb = preprocessing.LabelBinarizer()\n Y_bin = lb.fit_transform(Y) # seems to work now\n \n elif whichone == 2:\n # By hand\n Y_bin = np.zeros((m_examples, len(unique_classes)))\n for i in range(0, m_examples):\n if Y[i] == unique_classes[0]:\n Y_bin[i,0] = 1\n elif Y[i] == unique_classes[1]:\n Y_bin[i,1] = 1\n elif Y[i] == unique_classes[2]:\n Y_bin[i,2] = 1\n elif Y[i] == unique_classes[3]:\n Y_bin[i,3] = 1\n elif Y[i] == unique_classes[4]:\n Y_bin[i,4] = 1\n elif Y[i] == unique_classes[5]:\n Y_bin[i,5] = 1\n elif Y[i] == unique_classes[6]:\n Y_bin[i,6] = 1\n elif Y[i] == unique_classes[7]:\n Y_bin[i,7] = 1\n elif Y[i] == unique_classes[8]:\n Y_bin[i,8] = 1\n elif Y[i] == unique_classes[9]:\n Y_bin[i,9] = 1\n elif Y[i] == unique_classes[10]:\n Y_bin[i,10] = 1\n \n print('shape of Y_bin : ', Y_bin.shape)\n\n return Y_bin, unique_classes\n\ndef binarize_audio_signal(wind, data2):\n # wind = 1000\n a = int(np.floor(len(data2)/wind))\n vals = np.arange(0, a*wind, wind)\n sig = [np.max(data2[i:(i+wind)])*np.ones((wind)) for i in vals] \n sig = make_a_properlist(sig)\n \n return sig\n\ndef temporal_repeating_signatures(data2, timesteps, plotORnot):\n \n # ------------------------------------\n\n # choose wind such that sig is timesteps long\n size_of_mat = timesteps/2\n wind = int(len(data2)/size_of_mat)\n sig = binarize_audio_signal(wind, data2)\n\n min_border = np.mean(sig) + 1*np.std(sig)\n\n # On a besoin d'avoir au moins de 2 peaks\n peaks = []\n while len(peaks) < 2:\n # peaks, properties = find_peaks(sig, height=(min_border, np.max(sig)), prominence=100)\n peaks, properties = find_peaks(sig, height=(min_border, np.max(sig)))\n min_border = min_border - 10\n vv = [sig[i] for i in peaks]\n\n if plotORnot == 1:\n plt.plot(sig)\n plt.plot(peaks, vv, 'r*')\n plt.ylabel(\"Amplitude\")\n plt.xlabel(\"Data points\")\n plt.title(\"Bird call : binarized data2\")\n plt.show()\n\n # ------------------------------------\n\n # Calculate the difference in peaks and find the difference that repeats the most\n # This repeating difference is the ideal window to cut the data for bird call temporal signatures\n peakdiff = [peaks[i+1]-peaks[i] for i in range(len(peaks)-1)]\n z = sorted(peakdiff)\n o = [len(list(str(z[i]))) for i in range(len(z))]\n\n from collections import Counter\n c = Counter(o)\n mc = c.most_common()[0][0]\n\n for i in range(len(z)):\n if len(list(str(z[i]))) == mc:\n wind = z[i]\n\n # ------------------------------------\n\n # Cut the data and compare each binarized piece, using a metric\n # To see which pieces are similar\n a = int(np.floor(len(sig)/wind))\n vals = np.arange(0, a*wind, wind)\n\n dic = {}\n dic2 = {}\n for i in vals:\n out = []\n out2 = []\n\n A = sig[i:(i+wind)]\n A = [int(r) for r in A]\n for j in vals:\n # import required libraries\n B = sig[j:(j+wind)]\n B = [int(r) for r in B]\n\n # cosine similarity\n #cosine = np.dot(A,B)/(norm(A)*norm(B))\n #out.append(cosine)\n\n # absolute error\n err = [np.abs(A[q]-B[q]) for q in range(len(A))]\n out.append(err)\n\n # height value\n height = np.max(A)\n out2.append(height)\n\n dic[i] = out\n dic2[i] = out2\n\n # ------------------------------------\n\n # tot = np.array(tot)\n # tot.shape\n # import seaborn as sns\n # sns.heatmap(data=tot, annot=True)\n\n # ------------------------------------\n\n # Take the mean across similarity measures for each temporal piece\n zz = [np.mean(dic[vals[i]]) for i in range(len(dic))]\n zz2 = [np.mean(dic2[vals[i]]) for i in range(len(dic2))]\n\n # ------------------------------------\n\n # Find the temporal piece that are most similar to all other pieces\n select_crit = 2\n\n if select_crit == 0:\n # No height restriction\n # ind_piece = np.argmax(zz) # cosine similarity\n ind_piece = np.argmin(zz) # absolute error\n elif select_crit == 1:\n # With height restriction\n # use height as threshold for zz\n marg = 1*np.std(sig)\n thresh = np.max(zz2) - marg\n print('thresh : ', thresh)\n # OU\n # thresh = np.min(zz2)\n store = [zz for i in range(len(zz)) if zz2[i] > thresh]\n\n # the select index using metric (cosine, absolute error)\n # ind_piece = np.argmax(store) # cosine similarity\n ind_piece = np.argmin(store) # absolute error\n elif select_crit == 2:\n # Rien a faire avec metrics, choisi par rapport la hauteur\n ind_piece = np.argmax(zz2)\n \n # Plot selected temporal piece, with data\n st = vals[ind_piece]\n endd = st+wind\n sig2 = sig[st:endd]\n\n if plotORnot == 1:\n plt.plot(sig)\n plt.plot(np.arange(st,endd), sig2)\n plt.ylabel(\"Amplitude\")\n plt.xlabel(\"Data points\")\n plt.title(\"Non-downsampled : Bird call and call signature\")\n plt.show()\n\n # ------------------------------------\n\n # Downsampling\n v = downsample_sig(sig2, timesteps)\n\n # ------------------------------------\n \n return v\n\ndef downsample_sig(sig, timesteps):\n\n # Downsampling\n a = int(np.floor(len(sig)/timesteps))\n vals = np.arange(0, a*timesteps, a)\n v = [sig[i:(i+a)][0] for i in vals] # duh, pick the first point from each window, that is why [0] is there\n \n return v\n\n","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.107709Z","iopub.status.idle":"2024-05-15T04:13:52.108558Z","shell.execute_reply.started":"2024-05-15T04:13:52.108346Z","shell.execute_reply":"2024-05-15T04:13:52.108369Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"strcats = data.primary_label\nprint('Activities include: ', strcats)\nnum_of_birds = len(strcats)\nprint('Number of birds: ', num_of_birds)\n\ny = catstrvec_2_catnumvec(strcats, data.primary_label)\ny = np.array(y)\nprint('shape of y: ', y.shape)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.110673Z","iopub.status.idle":"2024-05-15T04:13:52.111077Z","shell.execute_reply.started":"2024-05-15T04:13:52.110881Z","shell.execute_reply":"2024-05-15T04:13:52.110901Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"path = \"/kaggle/input/birdclef-2024/train_audio/\"\n# Get file names \nfiles = os.listdir(path)\nall_fn = []\nfor f in files:\n fn = f.partition('.')[0]\n \n if fn[0:2] != '__':\n all_fn.append(fn)\nprint('FILE NAMES: ', all_fn)\nprint('length of all_fn: ', len(all_fn))","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.112171Z","iopub.status.idle":"2024-05-15T04:13:52.112579Z","shell.execute_reply.started":"2024-05-15T04:13:52.112391Z","shell.execute_reply":"2024-05-15T04:13:52.112408Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def convert_mp3_to_sig(audio_filepath):\n # files \n src = audio_filepath\n dst = \"test.wav\"\n\n # convert wav to mp3 \n sound = AudioSegment.from_ogg(src)\n sound.export(dst, format=\"wav\")\n \n return dst\n# Load a wav file\ndef get_wav_info(wav_file):\n rate, data = wavfile.read(wav_file)\n return rate, data\n# Calculate and plot spectrogram for a wav audio file\n# Binning the time-series and calculating the periodogram per time bin\n\ndef graph_spectrogram(wav_file):\n rate, data = get_wav_info(wav_file)\n # print('size of extracted data from sound file: ', data.shape)\n nfft = 100 # Length of each window segment - frequency data will be binned by nfft/2 \n fs = 8000 # Sampling frequencies\n noverlap = 0 #120 # Overlap between windows\n nchannels = data.ndim\n # print('nchannels: ', nchannels)\n \n # https://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.spectrogram.html\n \n # OR\n \n # pxx = periodogram per frequency (nfft/2 = num_of_freqs, bins=num_of_bins)\n # freqs = freqencies that correspond to each magnitude\n # bins = number of bins to group time-series\n # im = the image of the axis\n if nchannels == 1:\n pxx, freqs, bins, im = plt.specgram(data, nfft, fs, noverlap=noverlap)\n elif nchannels == 2:\n pxx, freqs, bins, im = plt.specgram(data[:,0], nfft, fs, noverlap=noverlap)\n \n # print('number of bins: ', bins.shape)\n # print('shape of pxx: ', pxx.shape)\n # print('shape of freqs: ', freqs.shape)\n \n plotORnot = 0\n if plotORnot == 1:\n plt.ylabel(\"Frequency\")\n plt.xlabel(\"Number of time windows\")\n plt.title(\"Spectrogram\")\n plt.show()\n\n return data, pxx, freqs\ndef make_a_properlist(vec):\n \n out = []\n for i in range(len(vec)):\n out = out + [np.ravel(vec[i])]\n \n if is_empty(out) == False:\n vecout = np.concatenate(out).ravel().tolist()\n else:\n vecout = list(np.ravel(out))\n \n return vecout\n\ndef is_empty(vec):\n vec = np.array(vec)\n if vec.shape[0] == 0:\n out = True\n else:\n out = False\n \n return out\ndef linear_intercurrentpt_makeshortSIGlong_interp1d(shortSIG, longSIG):\n\n x = np.linspace(shortSIG[0], len(shortSIG), num=len(shortSIG), endpoint=True)\n y = shortSIG\n # print('x : ', x)\n\n\n # -------------\n kind = 'linear'\n # kind : Specifies the kind of interpolation as a string or as an integer specifying the order of the spline interpolator to use. The string has to be one of âlinearâ, ânearestâ, ânearest-upâ, âzeroâ, âslinearâ, âquadraticâ, âcubicâ, âpreviousâ, or ânextâ. âzeroâ, âslinearâ, âquadraticâ and âcubicâ refer to a spline interpolation of zeroth, first, second or third order; âpreviousâ and ânextâ simply return the previous or next value of the point; ânearest-upâ and ânearestâ differ when interpolating half-integers (e.g. 0.5, 1.5) in that ânearest-upâ rounds up and ânearestâ rounds down. Default is âlinearâ.\n\n if kind == 'linear':\n f = interp1d(x, y)\n elif kind == 'cubic':\n f = interp1d(x, y, kind='cubic')\n # -------------\n\n xnew = np.linspace(shortSIG[0], len(shortSIG), num=len(longSIG), endpoint=True)\n # print('xnew : ', xnew)\n\n siglong = f(xnew)\n\n return siglong\ndef scale_feature_data(feat, plotORnot, type_scale):\n \n columns = ['0']\n dat = pd.DataFrame(data=feat, columns=columns)\n \n # which type of scaling\n if type_scale == 'minmax':\n # Values from 0 to 1\n scaled_data0 = minmax_scaling(dat, columns=columns)\n scaled_data = list(scaled_data0.to_numpy().ravel())\n \n elif type_scale == 'normalization':\n # normalization : same as mlxtend - Values from 0 to 1\n scaled_data = []\n for q in range(len(feat)):\n scaled_data.append( (feat[q] - np.min(feat))/(np.max(feat) - np.min(feat)) )\n \n elif type_scale == 'pos_normalization':\n # positive normalization : same as mlxtend - Values from 0 to 1\n shift_up = [i - np.min(feat) for i in feat]\n scaled_data = [q/np.max(shift_up) for q in shift_up]\n \n elif type_scale == 'standardization':\n # standardization : values are not restricted to a range, but scaled appropreately\n scaled_data = [(q - np.mean(feat))/np.std(feat) for q in feat]\n\n return scaled_data\n# level : the number of levels to decompose the time signal, le nombre des marquers par signale\ndef tsig_2_discrete_wavelet_transform(sig, waveletname, level, plotORnot):\n\n # On peut calculater dans deux façons: 0) dwt en boucle and then idwt, 1) wavedec et waverec\n # Mais le deux ne donnent pas le meme reponses, wavedec et waverec semble plus raisonable.\n coeff = wavedec(sig, waveletname, level)\n\n if plotORnot == 1:\n fig, axx = plt.subplots(nrows=len(coeff), ncols=1, figsize=(5,5))\n axx[0].set_title(\"coef\") # Pas certain si c'est coef0 ou coef1\n for k in range(len(coeff)):\n axx[k].plot(coeff[k], 'r') # output of the low pass filter (averaging filter) of the DWT\n plt.tight_layout()\n plt.show()\n\n return coeff\ndef tsig_2_spectrogram(sig, fs, nfft, noverlap, img_dim, plotORnot):\n\n # -----------------------------------\n fig,ax = plt.subplots(1)\n fig.subplots_adjust(left=0,right=1,bottom=0,top=1)\n ax.axis('off')\n # spectrum2D array : Columns are the periodograms of successive segments.\n # freqs1-D array : The frequencies corresponding to the rows in spectrum.\n # t1-D array : The times corresponding to midpoints of segments (i.e., the columns in spectrum).\n # imAxesImage : The image created by imshow containing the spectrogram.\n pxx, freqs, bins, img = ax.specgram(sig, nfft, fs, noverlap=noverlap)\n ax.axis('off')\n\n # -----------------------------------\n\n my_image = 'temp.png'\n fig.savefig(my_image)\n fname = os.path.abspath(os.getcwd()) + \"/\" + my_image\n \n # Convert image to an array:\n # Read image \n img = Image.open(fname) # PIL: img is not in array form, it is a PIL.PngImagePlugin.PngImageFile \n\n # -----------------------------------\n \n # Resize sectrogram image\n image = imgORmat_resize_imgORmat_CNN(img_dim, data_in=img, inpt='img3D', outpt='mat2D', norm='non', thresh='non')\n \n # -----------------------------------\n \n # Flatten image into a vector\n img_flatten = np.reshape(np.ravel(image), (img_dim*img_dim, ), order='F')\n # print('img_flatten.shape: ', img_flatten.shape)\n\n # -----------------------------------\n\n return img_flatten\ndef tsig_2_continuous_wavelet_transform(sig, fs, scales, waveletname, img_dim, plotORnot):\n \n # e.g.:\n # scales = np.arange(1, 128)\n # waveletname = 'mexh'\n # sig = data2[0:100]\n # print('len(sig): ', len(sig))\n \n dt = 1/fs\n # print('dt : ', dt)\n\n coefficients, frequencies = pywt.cwt(sig, scales, waveletname, dt)\n coefficients = np.array(coefficients)\n ylen, xlen = coefficients.shape\n\n # Time by frequency plot of cwt : then flatten and use as a feature\n stop_val = len(sig)/fs\n x = np.arange(0, stop_val, dt) # time\n y = frequencies # frequency \n X, Y = np.meshgrid(x, y)\n Z = coefficients\n\n # Each cwt versus frequencies\n # for i in range(len(coefficients)):\n # plt.plot(frequencies, coefficients[i])\n\n # \n fig = plt.figure()\n ax = plt.axes() # creates a 3D axis by using the keyword projection='3d'\n ax.contourf(X, Y, Z, xlen, cmap=plt.cm.seismic) # contour fill\n ax.axis('off')\n\n # -----------------------------------\n\n my_image = 'temp.png'\n fig.savefig(my_image)\n fname = os.path.abspath(os.getcwd()) + \"/\" + my_image\n\n # Convert image to an array:\n # Read image \n img = Image.open(fname) \n \n # -----------------------------------\n \n # Resize sectrogram image\n image = imgORmat_resize_imgORmat_CNN(img_dim, data_in=img, inpt='img3D', outpt='mat2D', norm='non', thresh='non')\n \n # -----------------------------------\n \n # Flatten image into a vector\n img_flatten = np.reshape(np.ravel(image), (img_dim*img_dim, ), order='F')\n # print('img_flatten.shape: ', img_flatten.shape)\n \n return img_flatten\ndef resize_img(img, img_dim):\n if type(img) == 'numpy.ndarray':\n # img is an array, retuns an image object\n rgb_image = Image.fromarray(img , 'RGB')\n else:\n # img is an image object, returns an image object\n try:\n rgb_image = img.convert('RGB')\n except AttributeError:\n rgb_image = Image.fromarray(img , 'RGB')\n\n # Resize image into a 64, 64, 3\n new_h, new_w = int(img_dim), int(img_dim)\n img3 = rgb_image.resize((new_w, new_h), Image.ANTIALIAS)\n w_resized, h_resized = img3.size[0], img3.size[1]\n return img3\n\ndef convert_img_a_mat(img, outpt):\n mat = np.array(img) # Convert image to an array\n if outpt == 'mat2D':\n # Transformer l'image de 3D Ă 2D\n # Convert image back to a 2D array\n matout = np.mean(mat, axis=2)\n elif outpt == 'img3D': # techniquement c'est un image parce qu'il y a trois RGB channels \n matout = mat\n return matout\n\ndef norm_mat(mat2Dor3D, norm):\n if norm == 'zero2one':\n # Normalizer l'image entre 0 et 1\n norout = mat2Dor3D/255\n elif norm == 'negone2posone':\n # Normalize the images to [-1, 1]\n norout = (mat2Dor3D - 127.5) / 127.5\n elif norm == 'non':\n norout = mat2Dor3D\n return norout\n\ndef threshold_mat(mat2D, thresh):\n # Threshold image\n val = 255/2\n if thresh == 'zero_moins_que_val':\n row, col = mat2D.shape\n mat_thresh = mat2D\n min_val = np.min(mat_thresh)\n for i in range(row):\n for j in range(col):\n if mat_thresh[i,j] < val:\n mat_thresh[i,j] = min_val\n elif thresh == 'non':\n mat_thresh = mat2D\n return mat_thresh\n\ndef imgORmat_resize_imgORmat_CNN(img_dim, data_in, inpt='img3D', outpt='mat2D', norm='non', thresh='non'):\n if inpt == 'img3D' and outpt=='mat2D':\n img = resize_img(data_in, img_dim)\n img3D = convert_img_a_mat(img, outpt)\n out = norm_mat(img3D, norm)\n elif inpt == 'mat2D' and outpt=='mat2D':\n data_in = np.array(data_in)\n img = Image.fromarray(data_in , 'L')\n img = resize_img(img, img_dim)\n mat2D = convert_img_a_mat(img, outpt)\n out = norm_mat(mat2D, norm)\n elif inpt == 'mat2D' and outpt=='img3D':\n data_in = np.array(data_in)\n img = Image.fromarray(data_in , 'L')\n img = resize_img(img, img_dim)\n img3D = convert_img_a_mat(img, outpt)\n out = norm_mat(img3D, norm)\n elif inpt == 'img3D' and outpt=='img3D':\n img = resize_img(data_in, img_dim)\n img3D = convert_img_a_mat(img, outpt)\n out = norm_mat(img3D, norm)\n\n return out\ndef save_dat_pickle(outSIG, file_name=\"outSIG.pkl\"):\n # Save data matrices to file\n open_file = open(file_name, \"wb\")\n pickle.dump(outSIG, open_file)\n open_file.close()\ndef load_dat_pickle(file_name=\"outSIG.pkl\"):\n open_file = open(file_name, \"rb\")\n dataout = pickle.load(open_file)\n open_file.close()\n return dataout\ndef binarize_Y1Dvec_2_Ybin(Y):\n \n # Transform a 1D Y vector (n_samples by 1) to a Y_bin (n_samples by n_classes) vector\n\n # Ensure vector is of integers\n Y = [int(i) for i in Y]\n\n # Number of samples\n m_examples = len(Y)\n\n # Number of classes\n temp = np.unique(Y)\n unique_classes = [int(i) for i in temp]\n # print('unique_classes : ', unique_classes)\n\n whichone = 2\n # Binarize the output\n if whichone == 0:\n from sklearn.preprocessing import label_binarize\n Y_bin = label_binarize(Y, classes=unique_classes) # seems to work now\n\n elif whichone == 1:\n from sklearn import preprocessing\n lb = preprocessing.LabelBinarizer()\n Y_bin = lb.fit_transform(Y) # seems to work now\n \n elif whichone == 2:\n # By hand\n Y_bin = np.zeros((m_examples, len(unique_classes)))\n for i in range(0, m_examples):\n if Y[i] == unique_classes[0]:\n Y_bin[i,0] = 1\n elif Y[i] == unique_classes[1]:\n Y_bin[i,1] = 1\n elif Y[i] == unique_classes[2]:\n Y_bin[i,2] = 1\n elif Y[i] == unique_classes[3]:\n Y_bin[i,3] = 1\n elif Y[i] == unique_classes[4]:\n Y_bin[i,4] = 1\n elif Y[i] == unique_classes[5]:\n Y_bin[i,5] = 1\n elif Y[i] == unique_classes[6]:\n Y_bin[i,6] = 1\n elif Y[i] == unique_classes[7]:\n Y_bin[i,7] = 1\n elif Y[i] == unique_classes[8]:\n Y_bin[i,8] = 1\n elif Y[i] == unique_classes[9]:\n Y_bin[i,9] = 1\n elif Y[i] == unique_classes[10]:\n Y_bin[i,10] = 1\n \n print('shape of Y_bin : ', Y_bin.shape)\n\n return Y_bin, unique_classes\ndef binarize_audio_signal(wind, data2):\n # wind = 1000\n a = int(np.floor(len(data2)/wind))\n vals = np.arange(0, a*wind, wind)\n sig = [np.max(data2[i:(i+wind)])*np.ones((wind)) for i in vals] \n sig = make_a_properlist(sig)\n \n return sig\ndef temporal_repeating_signatures(data2, timesteps, plotORnot):\n \n # ------------------------------------\n\n # choose wind such that sig is timesteps long\n size_of_mat = timesteps/2\n wind = int(len(data2)/size_of_mat)\n sig = binarize_audio_signal(wind, data2)\n\n min_border = np.mean(sig) + 1*np.std(sig)\n\n # On a besoin d'avoir au moins de 2 peaks\n peaks = []\n while len(peaks) < 2:\n # peaks, properties = find_peaks(sig, height=(min_border, np.max(sig)), prominence=100)\n peaks, properties = find_peaks(sig, height=(min_border, np.max(sig)))\n min_border = min_border - 10\n vv = [sig[i] for i in peaks]\n\n if plotORnot == 1:\n plt.plot(sig)\n plt.plot(peaks, vv, 'r*')\n plt.ylabel(\"Amplitude\")\n plt.xlabel(\"Data points\")\n plt.title(\"Bird call : binarized data2\")\n plt.show()\n\n # ------------------------------------\n\n # Calculate the difference in peaks and find the difference that repeats the most\n # This repeating difference is the ideal window to cut the data for bird call temporal signatures\n peakdiff = [peaks[i+1]-peaks[i] for i in range(len(peaks)-1)]\n z = sorted(peakdiff)\n o = [len(list(str(z[i]))) for i in range(len(z))]\n\n from collections import Counter\n c = Counter(o)\n mc = c.most_common()[0][0]\n\n for i in range(len(z)):\n if len(list(str(z[i]))) == mc:\n wind = z[i]\n\n # ------------------------------------\n\n # Cut the data and compare each binarized piece, using a metric\n # To see which pieces are similar\n a = int(np.floor(len(sig)/wind))\n vals = np.arange(0, a*wind, wind)\n\n dic = {}\n dic2 = {}\n for i in vals:\n out = []\n out2 = []\n\n A = sig[i:(i+wind)]\n A = [int(r) for r in A]\n for j in vals:\n # import required libraries\n B = sig[j:(j+wind)]\n B = [int(r) for r in B]\n\n # cosine similarity\n #cosine = np.dot(A,B)/(norm(A)*norm(B))\n #out.append(cosine)\n\n # absolute error\n err = [np.abs(A[q]-B[q]) for q in range(len(A))]\n out.append(err)\n\n # height value\n height = np.max(A)\n out2.append(height)\n\n dic[i] = out\n dic2[i] = out2\n\n # ------------------------------------\n\n # tot = np.array(tot)\n # tot.shape\n # import seaborn as sns\n # sns.heatmap(data=tot, annot=True)\n\n # ------------------------------------\n\n # Take the mean across similarity measures for each temporal piece\n zz = [np.mean(dic[vals[i]]) for i in range(len(dic))]\n zz2 = [np.mean(dic2[vals[i]]) for i in range(len(dic2))]\n\n # ------------------------------------\n\n # Find the temporal piece that are most similar to all other pieces\n select_crit = 2\n\n if select_crit == 0:\n # No height restriction\n # ind_piece = np.argmax(zz) # cosine similarity\n ind_piece = np.argmin(zz) # absolute error\n elif select_crit == 1:\n # With height restriction\n # use height as threshold for zz\n marg = 1*np.std(sig)\n thresh = np.max(zz2) - marg\n print('thresh : ', thresh)\n # OU\n # thresh = np.min(zz2)\n store = [zz for i in range(len(zz)) if zz2[i] > thresh]\n\n # the select index using metric (cosine, absolute error)\n # ind_piece = np.argmax(store) # cosine similarity\n ind_piece = np.argmin(store) # absolute error\n elif select_crit == 2:\n # Rien a faire avec metrics, choisi par rapport la hauteur\n ind_piece = np.argmax(zz2)\n \n # Plot selected temporal piece, with data\n st = vals[ind_piece]\n endd = st+wind\n sig2 = sig[st:endd]\n\n if plotORnot == 1:\n plt.plot(sig)\n plt.plot(np.arange(st,endd), sig2)\n plt.ylabel(\"Amplitude\")\n plt.xlabel(\"Data points\")\n plt.title(\"Non-downsampled : Bird call and call signature\")\n plt.show()\n\n # ------------------------------------\n\n # Downsampling\n v = downsample_sig(sig2, timesteps)\n\n # ------------------------------------\n \n return v\ndef downsample_sig(sig, timesteps):\n\n # Downsampling\n a = int(np.floor(len(sig)/timesteps))\n vals = np.arange(0, a*timesteps, a)\n v = [sig[i:(i+a)][0] for i in vals] # duh, pick the first point from each window, that is why [0] is there\n \n return v","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.113856Z","iopub.status.idle":"2024-05-15T04:13:52.114235Z","shell.execute_reply.started":"2024-05-15T04:13:52.114051Z","shell.execute_reply":"2024-05-15T04:13:52.114068Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# son_des_oiseaux2_4\ntry_some = 1 # 100, 300, 500\nsamps = np.random.permutation(try_some)\nsamps\n\nX = pd.DataFrame()\n\ntimesteps = 500 # 50, 100, 150 (best), 200\nlongsig = np.arange(timesteps)\n\nplotORnot = 0\n\nfor bird in samps:\n\n audio_filepath = path + all_fn[bird] + '.ogg'\n print(audio_filepath)\n # ------------------------------------\n\n # Do steps 1 and 2 to get the rate\n # 1) Converter mp3 Ă signale\n dst = convert_mp3_to_sig(audio_filepath)\n\n # 2) Obtenir le frĂ©quence d'Ă©chantillonnage\n rate, data = get_wav_info(dst)\n\n del dst, data\n\n # ------------------------------------\n\n # 3) Downsample such that the audio is not distorted\n sound = AudioSegment.from_file(audio_filepath, format='ogg', frame_rate=rate)\n fs = 10000\n sound = sound.set_frame_rate(fs)\n data2 = sound.get_array_of_samples()\n\n if plotORnot == 1:\n # Confirmer signale\n t = np.linspace(0, len(data2) / fs, len(data2))\n plt.plot(data2) # freqs is 101 long\n plt.ylabel(\"Amplitude\")\n plt.xlabel(\"Data points\")\n plt.title(\"Bird call : data2\")\n plt.show()\n\n del sound\n\n # ------------------------------------\n\n # Time-series downsampled\n useit = 1\n if useit == 1:\n timsig_down = downsample_sig(data2, timesteps)\n\n if plotORnot == 1:\n plt.plot(timsig_down)\n plt.ylabel(\"Amplitude\")\n plt.xlabel(\"Data points\")\n plt.title(\"Downsampled : data2\")\n plt.show()\n\n # ------------------------------------\n\n # Choose wind such that sig is timesteps long\n useit = 1\n if useit == 1:\n plotORnot = 1\n timeseries_repeat = temporal_repeating_signatures(data2, timesteps, plotORnot)\n\n # If the mp3 has all other sounds and one or two bird calls, it is likely to find a signature\n # for the other sounds because it partitions the data and looks for similar chunks of data.\n # If you have only bird calls, it will pick up a bird call signature.\n\n # ------------------------------------\n\n useit = 1\n if useit == 1:\n # Get the first 3principle componets : Principle Component Analysis (PCA)\n X_dat = [data2, data2]\n X_dat = np.reshape(X_dat, (len(data2), 2))\n feat_PCA = PCA(n_components=2).fit_transform(X_dat) # shape (len(data2), n_components)\n\n cols = feat_PCA.shape[1]\n pca_out = []\n for i in range(cols):\n sig = feat_PCA[:,i]\n pca_out.append(downsample_sig(sig, timesteps))\n\n if plotORnot == 1:\n plt.plot(data2)\n plt.ylabel(\"Magnitude\")\n plt.xlabel(\"Data points\")\n plt.title(\"data2\")\n plt.show()\n\n plt.plot(feat_PCA[:])\n plt.ylabel(\"Magnitude\")\n plt.xlabel(\"Data points\")\n plt.title(\"Non-downsampled PCA\")\n plt.show()\n\n plt.plot(pca_out[0])\n plt.plot(pca_out[1])\n plt.ylabel(\"Magnitude\")\n plt.xlabel(\"Data points\")\n plt.title(\"Downsampled PCA\")\n plt.show()\n\n # ------------------------------------\n\n useit = 0\n if useit == 1:\n # ....................\n # Binned periodogram\n # ....................\n # Plot directly the periodogram of the signal (Magnitude vs Frequency) : Unmanageable length = 149401\n detrend='linear'\n freqencies, spectrum = periodogram(data2, fs=fs, detrend=detrend, window=\"boxcar\", scaling='spectrum',)\n\n if plotORnot == 1:\n plt.plot(freqencies, spectrum)\n plt.ylabel(\"Magnitude\")\n plt.xlabel(\"Frequency\")\n plt.title(\"Periodogram\")\n plt.show()\n\n grp = 550\n spec_bin = [sum(spectrum[i:i+grp]) for i in range(0,len(spectrum), grp)]\n freq_bin = [freqencies[i] for i in range(0,len(freqencies), grp)]\n\n while len(spec_bin) > timesteps:\n grp = grp+5\n spec_bin = [sum(spectrum[i:i+grp]) for i in range(0,len(spectrum), grp)]\n freq_bin = [freqencies[i] for i in range(0,len(freqencies), grp)]\n\n if plotORnot == 1:\n plt.plot(freq_bin, spec_bin)\n plt.ylabel(\"Magnitude\")\n plt.xlabel(\"Frequency\")\n plt.title(\"Periodogram\")\n plt.show()\n\n del spectrum\n\n # ------------------------------------\n\n useit = 0\n if useit == 1:\n # ....................\n # MEL frequency\n # ....................\n data2 = np.array(data2)\n data2 = np.float32(data2)\n mfcc = librosa.feature.mfcc(data2)\n\n mfcc_mean = mfcc.mean(axis=1) # Manageable length = 20\n\n # ------------------------------------\n\n useit = 0\n if useit == 1:\n # ....................\n # Discrete time wavelets\n # ....................\n level = 5\n coeff = tsig_2_discrete_wavelet_transform(data2, 'sym5', level, plotORnot) # Manageable length = 16\n\n # Not sure why it gives more than level+1 coefficents\n # Only take level+1\n coeff = coeff[0:level+1]\n\n # Ensure that all wavelets are no longer than timesteps\n coeff_cut = [coeff[i] if len(coeff[i]) <= timesteps else coeff[i][0:timesteps] for i in range(len(coeff))]\n\n # ------------------------------------\n\n useit = 0\n if useit == 1:\n # Selectionez le taille d'image pour le STFFT and CTFFT par rapport le taille des autres marquers \n img_dim = int(np.sqrt(timesteps))\n\n # flattened short-time fourier transform : spectrogram\n nfft = 100 # Length of each window segment - frequency data will be binned by nfft/2 \n fs = 8000 # Sampling frequencies\n noverlap = 0 #120 # Overlap between windows - this give data consistency\n stfft = tsig_2_spectrogram(data2, fs, nfft, noverlap, img_dim, plotORnot) # Manageable length = img_dim*img_dim\n\n # ------------------------------------\n\n useit = 0\n if useit == 1:\n # Take a fixed small amount of the signal, for slower processing features\n part = 1000\n sig3 = data2[0:part]\n scales = np.arange(1, 128)\n waveletname = 'mexh'\n\n # Manageable length = img_dim*img_dim\n ctfft = tsig_2_continuous_wavelet_transform(sig3, fs, scales, waveletname, img_dim, plotORnot)\n\n # ------------------------------------\n\n # Not all the columns are the same length but we just need to store them somewhere together\n\n # Time features :\n col0 = pd.Series(timsig_down)\n col1 = pd.Series(timeseries_repeat)\n cols2 = pd.DataFrame(pca_out).T\n\n # Frequency features :\n #col3 = pd.Series(spec_bin)\n # col4 = pd.Series(mfcc_mean)\n # cols5 = pd.DataFrame(coeff_cut).T\n #col6 = pd.Series(stfft) \n #col7 = pd.Series(ctfft)\n\n # df = pd.concat([col0, col1, col2, cols3, col4, col5], axis=1)\n # df = pd.concat([col1, col2, cols3], axis=1)\n # df = pd.concat([col0, cols01, col2, cols3], axis=1)\n df = pd.concat([col0, col1, cols2], axis=1)\n\n # ------------------------------------\n\n # Interpolate toutes les vectors\n for i in range(len(df.columns)):\n shortsig = df.iloc[:,i].dropna().to_numpy()\n intrp = linear_intercurrentpt_makeshortSIGlong_interp1d(shortsig, longsig)\n # ------------------------------------\n\n # Scale toutes les vectors\n type_scale = 'standardization' # \n df.iloc[:,i] = scale_feature_data(intrp, plotORnot, type_scale)\n\n X = pd.concat([X, df], axis=0)\n\n\nX = X.to_numpy()\n\nsave_dat_pickle(X, file_name=\"X.pkl\")\nsave_dat_pickle(samps, file_name=\"samps.pkl\")","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.116154Z","iopub.status.idle":"2024-05-15T04:13:52.116557Z","shell.execute_reply.started":"2024-05-15T04:13:52.116367Z","shell.execute_reply":"2024-05-15T04:13:52.116384Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"","metadata":{}},{"cell_type":"markdown","source":"","metadata":{}},{"cell_type":"markdown","source":"","metadata":{}},{"cell_type":"code","source":"# Create Temp Folder\nfrom pathlib import Path\nTMP_DIR = Path('/kaggle/working/temp')\nTMP_DIR.mkdir(exist_ok=True)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.118112Z","iopub.status.idle":"2024-05-15T04:13:52.12001Z","shell.execute_reply.started":"2024-05-15T04:13:52.119774Z","shell.execute_reply":"2024-05-15T04:13:52.119793Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# load audio and split into chunks and gen mel spect \ninp_path = '/kaggle/input/birdclef-2024/train_audio/'\nout_path = '/kaggle/working/temp/'\n\ndef load_audio(filename):\n # 1) Load file\n signal, sr = librosa.load(inp_path+filename)\n soundscape = filename.split(\"/\")[-1].split(\".\")[0]\n paths = []\n \n # 2) Splitting into chunks of 5s\n splits = []\n for i in range(0, len(signal), 5*sr):\n split = signal[i:i + 5*sr]\n if len(split) < 5*sr:\n break \n splits.append(split)\n \n # 3) Generating a mel spectrogram for each 5s chunk\n for i, chunk in enumerate(splits):\n SPEC_SHAPE = (48, 150) \n n_fft=2048\n hop_length = int(5 * sr / (SPEC_SHAPE[1] - 1))\n mel_spec = librosa.feature.melspectrogram(y=chunk, sr=sr, n_fft=n_fft, hop_length=hop_length, \n n_mels=SPEC_SHAPE[0], fmin=500, fmax=12500)\n\n mel_spec = librosa.power_to_db(mel_spec, ref=np.max) \n mel_spec -= mel_spec.min()\n mel_spec /= mel_spec.max()\n plt.figure(figsize=(1, 1))\n librosa.display.specshow(mel_spec, sr=sr, hop_length=hop_length, x_axis='time', y_axis='mel')\n plt.axis('off')\n \n # 4) Saving audios in the form of filename_{endtime}.jpg\n path = f\"{out_path}{soundscape}_{(i+1)*5}\"\n plt.savefig(path, bbox_inches='tight', pad_inches=0)\n paths.append(path+\".png\")\n plt.close()\n return paths","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.121152Z","iopub.status.idle":"2024-05-15T04:13:52.122035Z","shell.execute_reply.started":"2024-05-15T04:13:52.121821Z","shell.execute_reply":"2024-05-15T04:13:52.121839Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"samples = []\nfor i, row in data_samp.iterrows():\n samples += load_audio(row.filename)\nsamples = shuffle(samples)\nsamples","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.123751Z","iopub.status.idle":"2024-05-15T04:13:52.124314Z","shell.execute_reply.started":"2024-05-15T04:13:52.123999Z","shell.execute_reply":"2024-05-15T04:13:52.124021Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"plt.figure(figsize=(15, 7))\nfor i in range(12):\n spec = Image.open(samples[i])\n plt.subplot(3, 4, i + 1)\n plt.title(samples[i].split(os.sep)[-1])\n plt.imshow(spec, origin='lower')\n plt.tight_layout()","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.125645Z","iopub.status.idle":"2024-05-15T04:13:52.126186Z","shell.execute_reply.started":"2024-05-15T04:13:52.125897Z","shell.execute_reply":"2024-05-15T04:13:52.125921Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### Exploring the audio data","metadata":{}},{"cell_type":"code","source":"from IPython.display import Audio \npath = '/kaggle/input/birdclef-2024/train_audio/' + data.iloc[5, -1]\nsignal, sr = librosa.load(path)\npath","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.127646Z","iopub.status.idle":"2024-05-15T04:13:52.128167Z","shell.execute_reply.started":"2024-05-15T04:13:52.12789Z","shell.execute_reply":"2024-05-15T04:13:52.127911Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"y, sr = librosa.load(path)\n\nprint('y:', y, '\\n')\nprint('y shape:', np.shape(y), '\\n')\nprint('Sample Rate (KHz):', sr, '\\n')\n\n# Verify length of the audio\nprint('Check Len of Audio:', np.shape(y)[0]/sr)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.129233Z","iopub.status.idle":"2024-05-15T04:13:52.129755Z","shell.execute_reply.started":"2024-05-15T04:13:52.129495Z","shell.execute_reply":"2024-05-15T04:13:52.129517Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### Loading Audio Files","metadata":{}},{"cell_type":"code","source":"\npath = '/kaggle/input/birdclef-2024/train_audio/' + data.iloc[200, -1]","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.131236Z","iopub.status.idle":"2024-05-15T04:13:52.131764Z","shell.execute_reply.started":"2024-05-15T04:13:52.131502Z","shell.execute_reply":"2024-05-15T04:13:52.131524Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"data.iloc[170, -1], path","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.133099Z","iopub.status.idle":"2024-05-15T04:13:52.133633Z","shell.execute_reply.started":"2024-05-15T04:13:52.133369Z","shell.execute_reply":"2024-05-15T04:13:52.133391Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"signal, sr = librosa.load(path)\nprint('Signal:', signal)\nprint('Shape of signal:', np.shape(signal))\nprint('Sampling Rate (KHz):', sr)\nprint('Duration of audio:', np.shape(signal)[0]/sr)\nplt.plot(signal)\nplt.show()\nipd.Audio(path)","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.135021Z","iopub.status.idle":"2024-05-15T04:13:52.135837Z","shell.execute_reply.started":"2024-05-15T04:13:52.135281Z","shell.execute_reply":"2024-05-15T04:13:52.135319Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"### Mel Spectrogram\ndata_samp = data.sample(10)\ndata_samp.head()\n# Create Temp Folder\nfrom pathlib import Path## Splitting data \nTMP_DIR = Path('/kaggle/working/temp')\nTMP_DIR.mkdir(exist_ok=True)\n# load audio and split into chunks and gen mel spect \ninp_path = '/kaggle/input/birdclef-2024/train_audio/'\nout_path = '/kaggle/working/temp/'\n\ndef load_audio(filename):\n # 1) Load file\n signal, sr = librosa.load(inp_path+filename)\n soundscape = filename.split(\"/\")[-1].split(\".\")[0]\n paths = []\n \n # 2) Splitting into chunks of 5s\n splits = []\n for i in range(0, len(signal), 5*sr):\n split = signal[i:i + 5*sr]\n if len(split) < 5*sr:\n break \n splits.append(split)\n \n # 3) Generating a mel spectrogram for each 5s chunk\n for i, chunk in enumerate(splits):\n SPEC_SHAPE = (48, 150) \n n_fft=2048\n hop_length = int(5 * sr / (SPEC_SHAPE[1] - 1))\n mel_spec = librosa.feature.melspectrogram(y=chunk, sr=sr, n_fft=n_fft, hop_length=hop_length, \n n_mels=SPEC_SHAPE[0], fmin=500, fmax=12500)\n\n mel_spec = librosa.power_to_db(mel_spec, ref=np.max) \n mel_spec -= mel_spec.min()\n mel_spec /= mel_spec.max()\n plt.figure(figsize=(1, 1))\n librosa.display.specshow(mel_spec, sr=sr, hop_length=hop_length, x_axis='time', y_axis='mel')\n plt.axis('off')\n \n # 4) Saving audios in the form of filename_{endtime}.jpg\n path = f\"{out_path}{soundscape}_{(i+1)*5}\"\n plt.savefig(path, bbox_inches='tight', pad_inches=0)\n paths.append(path+\".png\")\n plt.close()\n return paths\n \nsamples = []\nfor i, row in data_samp.iterrows():\n samples += load_audio(row.filename)\nsamples = shuffle(samples)\nprint(samples)\nplt.figure(figsize=(15, 7))\nfor i in range(12):\n spec = Image.open(samples[i])\n plt.subplot(3, 4, i + 1)\n plt.title(samples[i].split(os.sep)[-1])\n plt.imshow(spec, origin='lower')\n plt.tight_layout()\n### Splitting and storing spectrograms??\n### How to store along w labels??","metadata":{"execution":{"iopub.status.busy":"2024-05-15T04:13:52.138144Z","iopub.status.idle":"2024-05-15T04:13:52.138684Z","shell.execute_reply.started":"2024-05-15T04:13:52.138411Z","shell.execute_reply":"2024-05-15T04:13:52.138435Z"},"trusted":true},"execution_count":null,"outputs":[]}]}