├── predict.sh ├── preprocess ├── resample_wavs_to_16k.sh ├── xmltodict.py └── loadData.py ├── doall.sh ├── train ├── io_utils_mod.py ├── generateImages.py ├── log.py ├── model-BirdClef.py ├── MapCallback.py ├── model-AlexNet.py └── trainModel.py ├── README.md ├── predict └── predict.py └── LICENSE /predict.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | cd predict 3 | python predict.py 4 | cd .. 5 | -------------------------------------------------------------------------------- /preprocess/resample_wavs_to_16k.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | ORIGINALDIR=$(dirname $0) 3 | DIR=$1 4 | cd $DIR 5 | mkdir ../wav_16khz 6 | FILES=* 7 | for f in $FILES 8 | do 9 | echo "Processing ./$f file to ../wav_16khz/$f" 10 | # take action on each file. $f store current file name 11 | sox ./$f -r 16000 ../wav_16khz/$f 12 | done 13 | cd $ORIGINALDIR 14 | -------------------------------------------------------------------------------- /doall.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | mkdir -p birdclef_data 3 | cd birdclef_data 4 | wget -nc http://otmedia.lirmm.fr/LifeCLEF/BirdCLEF2016/BirdCLEF2016TestSet.tar.gz 5 | wget -nc http://otmedia.lirmm.fr/LifeCLEF/BirdCLEF2016/BirdCLEF2016TrainingSet.tar.gz 6 | unp *.tar.gz 7 | cd .. 8 | preprocess/resample_wavs_to_16k.sh birdclef_data/TrainingSet/wav 9 | preprocess/resample_wavs_to_16k.sh birdclef_data/test/wav2015 10 | cd preprocess 11 | python loadData.py 12 | cd .. 13 | cd train 14 | mkdir modelWeights 15 | python trainModel.py 16 | cd .. 17 | -------------------------------------------------------------------------------- /train/io_utils_mod.py: -------------------------------------------------------------------------------- 1 | from __future__ import absolute_import 2 | import h5py 3 | import numpy as np 4 | from collections import defaultdict 5 | 6 | 7 | class HDF5Matrix(): 8 | refs = defaultdict(int) 9 | 10 | def __init__(self, datapath, dataset, start, end, normalizer=None): 11 | if datapath not in list(self.refs.keys()): 12 | f = h5py.File(datapath) 13 | self.refs[datapath] = f 14 | else: 15 | f = self.refs[datapath] 16 | self.start = start 17 | self.end = end 18 | self.data = f[dataset] 19 | self.normalizer = normalizer 20 | 21 | def __len__(self): 22 | return self.end - self.start 23 | 24 | def __getitem__(self, key): 25 | if isinstance(key, slice): 26 | if key.stop + self.start <= self.end: 27 | idx = slice(key.start+self.start, key.stop + self.start) 28 | else: 29 | raise IndexError 30 | elif isinstance(key, int): 31 | if key + self.start < self.end: 32 | idx = key+self.start 33 | else: 34 | raise IndexError 35 | elif isinstance(key, np.ndarray): 36 | if np.max(key) + self.start < self.end: 37 | idx = (self.start + key).tolist() 38 | else: 39 | raise IndexError 40 | elif isinstance(key, list): 41 | if max(key) + self.start < self.end: 42 | idx = [x + self.start for x in key] 43 | else: 44 | raise IndexError 45 | if self.normalizer is not None: 46 | return self.normalizer(self.data[idx]) 47 | else: 48 | return self.data[idx] 49 | 50 | @property 51 | def shape(self): 52 | ret = [] 53 | ret.append(self.end - self.start) 54 | ret = ret + [dim for dim in list(self.data.shape[1:])] 55 | return tuple(ret) 56 | 57 | 58 | def save_array(array, name): 59 | import tables 60 | f = tables.open_file(name, 'w') 61 | atom = tables.Atom.from_dtype(array.dtype) 62 | ds = f.createCArray(f.root, 'data', atom, array.shape) 63 | ds[:] = array 64 | f.close() 65 | 66 | 67 | def load_array(name): 68 | import tables 69 | f = tables.open_file(name) 70 | array = f.root.data 71 | a = np.empty(shape=array.shape, dtype=array.dtype) 72 | a[:] = array[:] 73 | f.close() 74 | return a -------------------------------------------------------------------------------- /train/generateImages.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | # 3 | # Birdsong classificatione in noisy environment with convolutional neural nets in Keras 4 | # Copyright (C) 2017 Báint Czeba, Bálint Pál Tóth (toth.b@tmit.bme.hu) 5 | # 6 | # This program is free software: you can redistribute it and/or modify 7 | # it under the terms of the GNU General Public License as published by 8 | # the Free Software Foundation, either version 3 of the License, or 9 | # (at your option) any later version. 10 | # 11 | # This program is distributed in the hope that it will be useful, 12 | # but WITHOUT ANY WARRANTY; without even the implied warranty of 13 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 14 | # GNU General Public License for more details. 15 | # 16 | # You should have received a copy of the GNU General Public License 17 | # along with this program. If not, see . (c) Balint Czeba, Balint Pal Toth 18 | # 19 | # Please cite the following paper if this code was useful for your research: 20 | # 21 | # Bálint Pál Tóth, Bálint Czeba, 22 | # "Convolutional Neural Networks for Large-Scale Bird Song Classification in Noisy Environment", 23 | # In: Working Notes of Conference and Labs of the Evaluation Forum, Évora, Portugália, 2016, p. 8 24 | 25 | # this script generates images from the preprocessed spectograms 26 | 27 | from scipy import io 28 | import pandas as pd 29 | import numpy as np 30 | import time 31 | import pickle 32 | import os 33 | import h5py 34 | 35 | hdf5path = '../birdclef_data/data_top999_nozero.hdf5' # it takes long if we save all the spectorgram, maybe it is better to call processNMostCommon from loadData.py with a small N (eg. =3) 36 | todirrootpath = '../images_bw_3/' 37 | 38 | f = h5py.File(hdf5path, 'r') 39 | X = f.get('X') 40 | y = f.get('y') 41 | mediaId = f.get('MediaId') 42 | classId = f.get('ClassId') 43 | print(mediaId[0]) 44 | 45 | y_class = np.empty((0,1)) 46 | for row in y: 47 | for i in range(len(row)): 48 | if row[i]==1: 49 | y_class=np.vstack((y_class,i)) 50 | 51 | import matplotlib 52 | matplotlib.use('Agg') 53 | import matplotlib.pyplot as plt 54 | plt.ioff() 55 | 56 | fig = plt.figure(frameon=False) 57 | ax = fig.add_axes([0, 0, 1, 1]) 58 | ax.axis('off') 59 | 60 | i=0 61 | lenX=X.shape[0] 62 | img_artist = ax.imshow(np.flipud(X[0][0]), cmap=plt.cm.binary) 63 | 64 | if not os.path.exists(todirrootpath): 65 | os.makedirs(todirrootpath) 66 | 67 | for i in range(X.shape[0]): 68 | directory=os.path.join(todirrootpath,'{}'.format(classId[i][0])) 69 | if not os.path.exists(directory): 70 | os.makedirs(directory) 71 | print('{}/{}'.format(i, lenX)) 72 | img_artist.set_data(np.flipud(X[i][0])) 73 | fileNumber=0 74 | while(os.path.isfile(os.path.join(directory, '{}_{}.png'.format(int(mediaId[i][0]), fileNumber)))): 75 | fileNumber=fileNumber+1; 76 | print(fileNumber) 77 | 78 | with open(os.path.join(directory, '{}_{}.png'.format(int(mediaId[i][0]), fileNumber)), 'w') as outfile: 79 | fig.canvas.print_png(outfile) 80 | 81 | f.close() 82 | -------------------------------------------------------------------------------- /train/log.py: -------------------------------------------------------------------------------- 1 | import pandas as pd 2 | import numpy as np 3 | 4 | def appendDfToPickle(df, filePath): 5 | import os 6 | import pandas as pd 7 | if not os.path.isfile(filePath): 8 | df.to_pickle(filePath) 9 | else: 10 | tempDF=pd.read_pickle(filePath) 11 | tempDF=tempDF.append(df, ignore_index = True) 12 | tempDF.to_pickle(filePath) 13 | 14 | def appendDfToExcel(df, excelFilePath): 15 | import os 16 | if not os.path.isfile(excelFilePath): 17 | df.to_excel(excelFilePath, index=False) 18 | else: 19 | tempDF=pd.read_excel(excelFilePath) 20 | tempDF=tempDF.append(df) 21 | tempDF.to_excel(excelFilePath, index=False) 22 | 23 | def appendDfToCSV(df, csvFilePath): 24 | import os 25 | if not os.path.isfile(csvFilePath): 26 | df.to_csv(csvFilePath, index=False, sep='\t') 27 | else: 28 | tempDF=pd.read_csv(csvFilePath, sep='\t') 29 | tempDF=tempDF.append(df) 30 | tempDF.to_csv(csvFilePath, index=False, sep='\t') 31 | 32 | def modelToDict(model): 33 | layerNumber=0; 34 | d=dict() 35 | for layer in model.layers: 36 | columnHeader="zz_layer({:02d})".format(layerNumber) #zz: to put it to the last columns 37 | layerNumber+=1; 38 | d[columnHeader]=str(layerToDict(layer)) 39 | 40 | #for k,v in model.get_config().iteritems(): ##not working since keras 1.0.0 41 | # if k in ['loss', 'optimizer']: 42 | # d[k]=str(v) 43 | d['optimizer']=str(model.optimizer.get_config()) 44 | d['loss']=str(model.loss) 45 | d['output_dim'] = model.layers[-1].get_config()['output_dim'] 46 | return d 47 | 48 | def layerToDict(layer): 49 | d=dict() 50 | for k,v in layer.get_config().iteritems(): 51 | if k in [\ 52 | #Basic 53 | 'input_shape', 'init', 'name', 'output_dim', 'activation', 'p', \ 54 | #Convolution1D 55 | 'nb_filter', 'pool_length', 'filter_length', 'border_mode', \ 56 | #Convolution2D 57 | 'nb_row','nb_col', 'subsample', 'pool_size']: 58 | d[k]=v 59 | d['output_shape']=layer.output_shape 60 | return d 61 | 62 | def resultToDict(result): 63 | d=dict() 64 | d['epochs']=len(result.epoch) 65 | minIsGood=['loss', 'val_loss'] 66 | maxIsGood=['acc', 'val_acc'] 67 | for k,v in result.history.iteritems(): 68 | if k in minIsGood: 69 | d['final_' + k]=v[-1] 70 | d['best_' + k]=np.min(v) 71 | d['best_' + k + '_epoch']=np.argmin(v) + 1 #!!!!! 72 | elif k in maxIsGood: 73 | d['final_' + k]=v[-1] 74 | d['best_' + k]=np.max(v) 75 | d['best_' + k + '_epoch']=np.argmax(v) + 1 #!!!!! 76 | return d 77 | 78 | def logToDataFrame(model=None, fitting_result=None, otherDict=None): 79 | d=dict() 80 | if model is not None: 81 | d.update(modelToDict(model)) 82 | if fitting_result is not None: 83 | d.update(resultToDict(fitting_result)) 84 | if otherDict is not None: 85 | d.update(otherDict) 86 | return pd.DataFrame(d, index=[0]) 87 | 88 | def logToXLS(filepath, model=None, fitting_result=None, otherDict=None): 89 | appendDfToExcel(logToDataFrame(model, fitting_result, otherDict), filepath) 90 | 91 | def logToCSV(filepath, model=None, fitting_result=None, otherDict=None): 92 | appendDfToCSV(logToDataFrame(model, fitting_result, otherDict), filepath) -------------------------------------------------------------------------------- /train/model-BirdClef.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | # 3 | # Birdsong classificatione in noisy environment with convolutional neural nets in Keras 4 | # Copyright (C) 2017 Báint Czeba, Bálint Pál Tóth (toth.b@tmit.bme.hu) 5 | # 6 | # This program is free software: you can redistribute it and/or modify 7 | # it under the terms of the GNU General Public License as published by 8 | # the Free Software Foundation, either version 3 of the License, or 9 | # (at your option) any later version. 10 | # 11 | # This program is distributed in the hope that it will be useful, 12 | # but WITHOUT ANY WARRANTY; without even the implied warranty of 13 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 14 | # GNU General Public License for more details. 15 | # 16 | # You should have received a copy of the GNU General Public License 17 | # along with this program. If not, see . (c) Balint Czeba, Balint Pal Toth 18 | # 19 | # Please cite the following paper if this code was useful for your research: 20 | # 21 | # Bálint Pál Tóth, Bálint Czeba, 22 | # "Convolutional Neural Networks for Large-Scale Bird Song Classification in Noisy Environment", 23 | # In: Working Notes of Conference and Labs of the Evaluation Forum, Évora, Portugália, 2016, p. 8 24 | 25 | from keras.models import Sequential 26 | from keras.layers.core import Dense, Activation, Dropout, Flatten 27 | from keras.optimizers import SGD, RMSprop 28 | from keras.layers.recurrent import LSTM 29 | from keras.callbacks import EarlyStopping, ModelCheckpoint 30 | from keras.layers.convolutional import Convolution2D, MaxPooling2D 31 | from keras.layers import BatchNormalization 32 | 33 | model = Sequential() 34 | 35 | # convolutional layers 36 | model.add(Convolution2D(input_shape=(1,200,310), 37 | nb_filter=128, 38 | nb_row=16, 39 | nb_col=16, 40 | border_mode='valid', 41 | init='lecun_uniform', 42 | activation='relu', 43 | subsample=(8, 8) 44 | )) 45 | model.add(MaxPooling2D(pool_size=(2,2))) 46 | 47 | model.add(Convolution2D(nb_filter=256, 48 | nb_row=5, 49 | nb_col=3, 50 | border_mode='valid', 51 | init='lecun_uniform', 52 | activation='relu' 53 | )) 54 | model.add(BatchNormalization()) 55 | model.add(MaxPooling2D(pool_size=(2,2))) 56 | 57 | model.add(Convolution2D(nb_filter=384, 58 | nb_row=3, 59 | nb_col=3, 60 | border_mode='same', 61 | init='lecun_uniform', 62 | activation='relu' 63 | )) 64 | 65 | model.add(Convolution2D(nb_filter=384, 66 | nb_row=3, 67 | nb_col=3, 68 | border_mode='same', 69 | init='lecun_uniform', 70 | activation='relu' 71 | )) 72 | 73 | model.add(BatchNormalization()) 74 | model.add(MaxPooling2D(pool_size=(2,2))) 75 | 76 | # dense layers 77 | model.add(Flatten()) 78 | model.add(Dense(2048, activation='relu')) 79 | model.add(Dropout(0.5)) 80 | model.add(Dense(2048, activation='relu')) 81 | model.add(Dropout(0.5)) 82 | model.add(Dense(output_dim=output_dim, activation='softmax')) 83 | -------------------------------------------------------------------------------- /train/MapCallback.py: -------------------------------------------------------------------------------- 1 | from sklearn.metrics import average_precision_score 2 | class Callback(object): 3 | '''Abstract base class used to build new callbacks. 4 | # Properties 5 | params: dict. Training parameters 6 | (eg. verbosity, batch size, number of epochs...). 7 | model: instance of `keras.models.Model`. 8 | Reference of the model being trained. 9 | The `logs` dictionary that callback methods 10 | take as argument will contain keys for quantities relevant to 11 | the current batch or epoch. 12 | Currently, the `.fit()` method of the `Sequential` model class 13 | will include the following quantities in the `logs` that 14 | it passes to its callbacks: 15 | on_epoch_end: logs include `acc` and `loss`, and 16 | optionally include `val_loss` 17 | (if validation is enabled in `fit`), and `val_acc` 18 | (if validation and accuracy monitoring are enabled). 19 | on_batch_begin: logs include `size`, 20 | the number of samples in the current batch. 21 | on_batch_end: logs include `loss`, and optionally `acc` 22 | (if accuracy monitoring is enabled). 23 | ''' 24 | def __init__(self): 25 | pass 26 | 27 | def set_params(self, params): 28 | self.params = params 29 | 30 | def set_model(self, model): 31 | self.model = model 32 | 33 | def on_epoch_begin(self, epoch, logs={}): 34 | pass 35 | 36 | def on_epoch_end(self, epoch, logs={}): 37 | pass 38 | 39 | def on_batch_begin(self, batch, logs={}): 40 | pass 41 | 42 | def on_batch_end(self, batch, logs={}): 43 | pass 44 | 45 | def on_train_begin(self, logs={}): 46 | pass 47 | 48 | def on_train_end(self, logs={}): 49 | pass 50 | 51 | class MapCallback(Callback): 52 | 53 | '''Abstract base class used to build new callbacks. 54 | # Properties 55 | params: dict. Training parameters 56 | (eg. verbosity, batch size, number of epochs...). 57 | model: instance of `keras.models.Model`. 58 | Reference of the model being trained. 59 | The `logs` dictionary that callback methods 60 | take as argument will contain keys for quantities relevant to 61 | the current batch or epoch. 62 | Currently, the `.fit()` method of the `Sequential` model class 63 | will include the following quantities in the `logs` that 64 | it passes to its callbacks: 65 | on_epoch_end: logs include `acc` and `loss`, and 66 | optionally include `val_loss` 67 | (if validation is enabled in `fit`), and `val_acc` 68 | (if validation and accuracy monitoring are enabled). 69 | on_batch_begin: logs include `size`, 70 | the number of samples in the current batch. 71 | on_batch_end: logs include `loss`, and optionally `acc` 72 | (if accuracy monitoring is enabled). 73 | ''' 74 | def __init__(self): 75 | super(Callback, self).__init__() 76 | 77 | def set_params(self, params): 78 | self.params = params 79 | 80 | def set_model(self, model): 81 | self.model = model 82 | 83 | def on_epoch_begin(self, epoch, logs={}): 84 | pass 85 | 86 | def on_epoch_end(self, epoch, logs={}): 87 | X_validation = self.model.validation_data[0] 88 | y_validation = self.model.validation_data[1] 89 | y_result=self.model.predict(X_validation) 90 | 91 | map = average_precision_score(y_validation.data[y_validation.start: y_validation.end], y_result, average='micro') 92 | 93 | logs['val_map']=map 94 | 95 | print("val_MAP: {}\n".format(map)) 96 | 97 | def on_batch_begin(self, batch, logs={}): 98 | pass 99 | 100 | def on_batch_end(self, batch, logs={}): 101 | pass 102 | 103 | def on_train_begin(self, logs={}): 104 | pass 105 | 106 | def on_train_end(self, logs={}): 107 | pass 108 | -------------------------------------------------------------------------------- /train/model-AlexNet.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | # 3 | # Birdsong classificatione in noisy environment with convolutional neural nets in Keras 4 | # Copyright (C) 2017 Báint Czeba, Bálint Pál Tóth (toth.b@tmit.bme.hu) 5 | # 6 | # This program is free software: you can redistribute it and/or modify 7 | # it under the terms of the GNU General Public License as published by 8 | # the Free Software Foundation, either version 3 of the License, or 9 | # (at your option) any later version. 10 | # 11 | # This program is distributed in the hope that it will be useful, 12 | # but WITHOUT ANY WARRANTY; without even the implied warranty of 13 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 14 | # GNU General Public License for more details. 15 | # 16 | # You should have received a copy of the GNU General Public License 17 | # along with this program. If not, see . (c) Balint Czeba, Balint Pal Toth 18 | # 19 | # Please cite the following paper if this code was useful for your research: 20 | # 21 | # Bálint Pál Tóth, Bálint Czeba, 22 | # "Convolutional Neural Networks for Large-Scale Bird Song Classification in Noisy Environment", 23 | # In: Working Notes of Conference and Labs of the Evaluation Forum, Évora, Portugália, 2016, p. 8 24 | 25 | from keras.models import Sequential 26 | from keras.layers.core import Dense, Activation, Dropout, Flatten 27 | from keras.layers import BatchNormalization 28 | from keras.optimizers import SGD, RMSprop 29 | from keras.layers.recurrent import LSTM 30 | from keras.callbacks import EarlyStopping, ModelCheckpoint 31 | from keras.layers.convolutional import Convolution2D, MaxPooling2D 32 | 33 | # Strongly AlexNet based convolutional neural network 34 | 35 | model = Sequential() 36 | 37 | # convolutional layer 38 | model.add(Convolution2D(input_shape=(1,200,310), 39 | nb_filter=48*2, 40 | nb_row=16, 41 | nb_col=16, 42 | border_mode='valid', 43 | init='glorot_normal', #glorot_normal lecun_uniform he_uniform 44 | activation='relu', 45 | subsample=(6, 6) 46 | )) 47 | model.add(BatchNormalization()) 48 | model.add(MaxPooling2D(pool_size=(2,2))) 49 | 50 | model.add(Convolution2D(nb_filter=128*2, 51 | nb_row=3, 52 | nb_col=3, 53 | border_mode='valid', 54 | init='lecun_uniform', #glorot_normal lecun_uniform he_uniform 55 | activation='relu' 56 | )) 57 | model.add(BatchNormalization()) 58 | model.add(MaxPooling2D(pool_size=(2,2))) 59 | 60 | model.add(Convolution2D(nb_filter=192*2, 61 | nb_row=3, 62 | nb_col=3, 63 | border_mode='same', 64 | init='lecun_uniform', #glorot_normal lecun_uniform he_uniform 65 | activation='relu' 66 | )) 67 | 68 | model.add(Convolution2D(nb_filter=192*2, 69 | nb_row=3, 70 | nb_col=3, 71 | border_mode='same', 72 | init='lecun_uniform', #glorot_normal lecun_uniform he_uniform 73 | activation='relu' 74 | )) 75 | 76 | model.add(Convolution2D(nb_filter=128*2, 77 | nb_row=3, 78 | nb_col=3, 79 | border_mode='same', 80 | init='lecun_uniform', #glorot_normal lecun_uniform he_uniform 81 | activation='relu' 82 | )) 83 | 84 | model.add(BatchNormalization()) 85 | model.add(MaxPooling2D(pool_size=(2,2))) 86 | 87 | # dense layers 88 | model.add(Flatten()) 89 | model.add(Dropout(0.5)) 90 | model.add(Dense(4096, activation='relu')) 91 | model.add(Dropout(0.5)) 92 | model.add(Dense(4096, activation='relu')) 93 | model.add(Dropout(0.5)) 94 | model.add(Dense(output_dim = output_dim, activation='softmax')) 95 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Trainig scripts for deep convolutional neural network based audio classification in Keras 2 | 3 | The following scripts were created for the BirdCLEF 2016 competition by Bálint Czeba and Bálint Pál Tóth. 4 | 5 | *The LifeCLEF bird identification challenge provides a largescale 6 | testbed for the system-oriented evaluation of bird species identifi- 7 | cation based on audio recordings. One of its main strength is that the 8 | data used for the evaluation is collected through Xeno-Canto, the largest 9 | network of bird sound recordists in the world. This makes the task closer 10 | to the conditions of a real-world application than previous, similar initiatives. 11 | The main novelty of the 2016-th edition of the challenge was the 12 | inclusion of soundscape recordings in addition to the usual xeno-canto 13 | recordings that focus on a single foreground species. This paper reports 14 | the methodology of the conducted evaluation, the overview of the systems 15 | experimented by the 6 participating research groups and a synthetic 16 | analysis of the obtained results.* (More details: http://www.imageclef.org/lifeclef/2016/bird) 17 | 18 | With some tweeks (reading meta-data and modifing network structure / how the spectogram is preprocessed) it is possible to apply it to arbitrary audio classification problems. 19 | 20 | # Citation 21 | 22 | Please cite the following paper if this code was useful for your research: 23 | 24 | *Tóth Bálint Pál, Czeba Bálint, "Convolutional Neural Networks for Large-Scale Bird Song Classification in Noisy Environment", In: Working Notes of Conference and Labs of the Evaluation Forum, Évora, Portugália, 2016, p. 8* 25 | Download from here (PDF): http://ceur-ws.org/Vol-1609/16090560.pdf 26 | 27 | ``` 28 | @article{tothczeba, 29 | author = "B\'{a}lint P\'{a}l T\'{o}th, B\'{a}lint Czeba", 30 | title = "{Convolutional Neural Networks for Large-Scale Bird Song Classification in Noisy Environment}", 31 | booktitle = "{Working Notes of Conference and Labs of the Evaluation Forum}, 32 | pages = "8", 33 | year = "2016", 34 | } 35 | ``` 36 | 37 | # Prerequisites 38 | You will need SOX for wave file resampling and Keras deep learning frameworks and some necessary modules. At the time of writeing you can install them in the following way: 39 | ``` 40 | sudo apt-get install sox 41 | sudo apt-get install python-tk 42 | sudo pip install scipy 43 | sudo pip install matplotlib 44 | sudo pip install sklearn 45 | sudo pip install tensorflow-gpu 46 | sudo pip install keras 47 | ``` 48 | The code is tested under Python 2.7. with TensorFlow (GPU) 1.0.0a0 and Keras 1.1.1. backend, NVidia Titan X 12GB GPU. 49 | 50 | If you use TensorFlow as a backend with Keras 1.x you should set 51 | ``` 52 | "image_dim_ordering": "th", 53 | ``` 54 | in ~/.keras/keras.json configuration file. 55 | 56 | In Keras 2 "image_dim_ordering" is deprecated. If you use TensorFlow + Keras 2.x, you should change the "image_data_format" setting to "channels_first". 57 | 58 | # Directory structure and files 59 | ``` 60 | doall.sh - run this script and it will do everything (you will need plenty of disk space > 100 GB) 61 | preprocess/loadData.py - responsible for preprocessing the data (wavs and XML meta-data) 62 | preprocess/sample_wavs_to_16k.sh - simple script that resamples wave files to 16 kHz with SOX 63 | preprocess/xmltodict.py - XML processing from https://github.com/martinblech/xmltodict 64 | train/trainModel.py - after preprocessing this script trains the neural networks 65 | train/model-AlexNet.py - AlexNet inspired model for audio classification 66 | train/model-BirdClef.py - Another convolutional neural net model for audio classification 67 | train/MAPCallback.py - Script to calculate MAP scores during training the neural nets 68 | train/generateImages.py - Generate images from the preprocessed spectogram for visualization purposes 69 | train/io_utils_mod.py - Functions for loading and saving data to HDF5 70 | train/log.py - Functions for logging purposes 71 | predict/predict.py - Predict after preprocessing and training is done 72 | ``` 73 | 74 | # Training (and download data and preprocess) 75 | 76 | For training you have to simply run 77 | ``` 78 | ./doall.sh 79 | ``` 80 | Be aware that this will download all the data (>50 GB) from http://otmedia.lirmm.fr/LifeCLEF/BirdCLEF2016/ to 81 | ``` 82 | birdclef_data 83 | ``` 84 | directory, unpack it and resample to 16 kHz and preprocess it into HDF5 files. You need cca. 280 GB of free space for the whole process. If you would like to put the data to somewhere else, please modify the doall.sh, preprocess/loadData.py and train/trainModel.py scripts. 85 | 86 | The download process, preprocessing and training takes 4-5 days on an i7 CPU + Titan X GPU. 87 | 88 | # Prediction 89 | 90 | After the preprocessing and training is do simpy run the following script to make predictions on test data: 91 | 92 | ``` 93 | ./predict.sh 94 | ``` 95 | The prediction results will be written in a .csv file in the predict/ directory. 96 | -------------------------------------------------------------------------------- /predict/predict.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | # 3 | # Birdsong classificatione in noisy environment with convolutional neural nets in Keras 4 | # Copyright (C) 2017 Báint Czeba, Bálint Pál Tóth (toth.b@tmit.bme.hu) 5 | # 6 | # This program is free software: you can redistribute it and/or modify 7 | # it under the terms of the GNU General Public License as published by 8 | # the Free Software Foundation, either version 3 of the License, or 9 | # (at your option) any later version. 10 | # 11 | # This program is distributed in the hope that it will be useful, 12 | # but WITHOUT ANY WARRANTY; without even the implied warranty of 13 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 14 | # GNU General Public License for more details. 15 | # 16 | # You should have received a copy of the GNU General Public License 17 | # along with this program. If not, see . (c) Balint Czeba, Balint Pal Toth 18 | # 19 | # Please cite the following paper if this code was useful for your research: 20 | # 21 | # Bálint Pál Tóth, Bálint Czeba, 22 | # "Convolutional Neural Networks for Large-Scale Bird Song Classification in Noisy Environment", 23 | # In: Working Notes of Conference and Labs of the Evaluation Forum, Évora, Portugália, 2016, p. 8 24 | 25 | from scipy import io 26 | from scipy.io import wavfile 27 | import pandas as pd 28 | import numpy as np 29 | np.random.seed(0) 30 | import time 31 | import pickle 32 | import os 33 | import h5py 34 | import sys, getopt 35 | import datetime 36 | from sklearn.metrics import average_precision_score, accuracy_score 37 | sys.path.append('../preprocess/') 38 | import loadData 39 | 40 | # prediction related paths, should be consistent with /preprocess/loadData.py and /train/trainModel.py 41 | PATH_TEST_IN_16KWAVS = '../birdclef_data/test/wav_16khz' 42 | PATH_TEST_IN_XMLPICKLEFILE = '../birdclef_data/test/xml_data.pickle' 43 | modelPath = '../train/model-AlexNet.py' 44 | modelWeightsPath = '../train/modelWeights/best_val_map_999.hdf5' 45 | labelBinarizerPath = "../birdclef_data/labelBinarizer_top999.pickle" 46 | 47 | output_dim = 999 48 | scalerFilePath = None 49 | 50 | if __name__ == "__main__": 51 | argv=sys.argv[1:] 52 | 53 | try: 54 | opts, args = getopt.getopt(argv,"he:p:m:s:",["nbepochs=","hdf5path=","scalerpath"]) 55 | except getopt.GetoptError: 56 | print 'trainModel.py -p ' 57 | sys.exit(2) 58 | for opt, arg in opts: 59 | if opt == '-h': 60 | print 'trainModel.py -p -m ' 61 | sys.exit() 62 | elif opt in ("-p", "--p"): 63 | hdf5path = arg 64 | elif opt in ("-m", "--m"): 65 | modelPath = arg 66 | elif opt in ("-s", "--s"): 67 | scalerFilePath = arg 68 | 69 | # function to convert probabilities to classes 70 | def proba_to_class(a): 71 | classCount = len(a[0]) 72 | to_return = np.empty((0,classCount)) 73 | for row in a: 74 | maxind = np.argmax(row) 75 | to_return = np.vstack((to_return,[1 if i==maxind else 0 for i in range(classCount)])) 76 | return to_return 77 | 78 | # run preprocessing and prediction on one file 79 | def runModelOnWavByPath(model, path): 80 | (tempSpecUnfiltered, tempSpecFiltered) = loadData.audioToFilteredSpectrogram(io.wavfile.read(path)[1], expandByOne=True) 81 | tempList = list() 82 | tempList.append(tempSpecFiltered) 83 | X,_,_,_ = loadData.spectrogramListToT4(tempList, N = 5*62) # the N must be consistent with /preprocess/loadData.py 84 | result = model.predict(X) 85 | 86 | return result 87 | 88 | scaler = None 89 | scaleData = None 90 | if scalerFilePath is not None: 91 | scaler = pickle.load(open(scalerFilePath, 'rb')) 92 | # Can't use the build in scaler.transform because it only supports 2d arrays. 93 | def scaleData(X): 94 | return (X-scaler.mean_)/scaler.scale_ 95 | 96 | # load the meta-data saved into a pickle file 97 | df = pd.read_pickle(PATH_TEST_IN_XMLPICKLEFILE) 98 | df = df.iloc[np.random.permutation(len(df))] 99 | df.reset_index(drop=True, inplace=True) 100 | lb = pickle.load(open(labelBinarizerPath, 'rb')) 101 | 102 | # calculate prediction time 103 | startTime = time.time() 104 | 105 | # Build model and load saved weights 106 | execfile(modelPath) 107 | model.load_weights(modelWeightsPath) 108 | 109 | ap=[] 110 | i=0 111 | 112 | # writeing predictions 113 | resultColumns = lb.inverse_transform(np.diag([1 for i in range(999)])) 114 | resultsFileName = "test_2015_{}_{}.csv".format(os.path.split(modelPath)[-1], datetime.datetime.now().strftime('%Y-%m-%d-%H-%M')) 115 | 116 | # running test with all the data 117 | for i in range(int(df.shape[0]*0.0),df.shape[0]): 118 | print ('{}/{}'.format(i, df.shape[0])) 119 | result = runModelOnWavByPath(model, os.path.join(PATH_TEST_IN_16KWAVS, df.FileName[i])) 120 | 121 | result_avg = np.mean(result, axis=0) 122 | result_avg = result_avg/np.sum(result_avg) 123 | 124 | singleResultDF = pd.DataFrame([result_avg], columns = resultColumns) 125 | singleResultDF['MediaId'] = df.MediaId[i] 126 | with open(resultsFileName, "a") as myfile: 127 | for row in singleResultDF.iterrows(): 128 | for k,v in row[1].iterkv(): 129 | if (k is not 'MediaId'): 130 | resultLine = "{};{};{:.16f}\n".format(row[1].MediaId, k, v) 131 | myfile.write(resultLine) 132 | 133 | elapsed = time.time()-startTime; 134 | print("Execution time: {0} s".format(elapsed)) 135 | -------------------------------------------------------------------------------- /train/trainModel.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | # 3 | # Birdsong classificatione in noisy environment with convolutional neural nets in Keras 4 | # Copyright (C) 2017 Báint Czeba, Bálint Pál Tóth (toth.b@tmit.bme.hu) 5 | # 6 | # This program is free software: you can redistribute it and/or modify 7 | # it under the terms of the GNU General Public License as published by 8 | # the Free Software Foundation, either version 3 of the License, or 9 | # (at your option) any later version. 10 | # 11 | # This program is distributed in the hope that it will be useful, 12 | # but WITHOUT ANY WARRANTY; without even the implied warranty of 13 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 14 | # GNU General Public License for more details. 15 | # 16 | # You should have received a copy of the GNU General Public License 17 | # along with this program. If not, see . (c) Balint Czeba, Balint Pal Toth 18 | # 19 | # Please cite the following paper if this code was useful for your research: 20 | # 21 | # Bálint Pál Tóth, Bálint Czeba, 22 | # "Convolutional Neural Networks for Large-Scale Bird Song Classification in Noisy Environment", 23 | # In: Working Notes of Conference and Labs of the Evaluation Forum, Évora, Portugália, 2016, p. 8 24 | 25 | # this script is responsible for training the neural networks 26 | 27 | from scipy import io 28 | import pandas as pd 29 | import numpy as np 30 | import time 31 | import pickle 32 | import os 33 | import h5py 34 | import sys, getopt 35 | import datetime 36 | from MapCallback import MapCallback 37 | 38 | if __name__ == "__main__": 39 | argv=sys.argv[1:] 40 | 41 | nb_epochs = 20000 # number of epochs, should be high, the end of the learning process is controled by early stoping 42 | es_patience = 100 # patience for early stoping 43 | batchSize = 1350 # batch size for mini-batch training 44 | hdf5path = '../birdclef_data/data_top999_nozero.hdf5' # training data generated by loadData.py 45 | modelPath = './model-AlexNet.py' # filename of the model to use (currently model-birdClef.py or model-AlexNet.py) 46 | logfileName = 'log.xls' 47 | #scalerFilePath = '../birdclef_data/standardScaler_5000.pickle' 48 | scalerFilePath = None 49 | preTrainedModelWeightsPath = None # path and filename to pretrained network: if there is a pretrained network, we can load it and continue to train it 50 | tensorflowBackend = False # set true if Keras has TensorFlow backend - this way we set TF not to allocate all the GPU memory 51 | 52 | if (tensorflowBackend): 53 | import tensorflow as tf 54 | config = tf.ConfigProto() 55 | config.gpu_options.allow_growth=True 56 | sess = tf.Session(config=config) 57 | from keras import backend as K 58 | K.set_session(sess) 59 | 60 | print('nb_epochs: %d, hdf5path: %s, scalerFilePath: %s' % (nb_epochs, hdf5path,scalerFilePath)) 61 | 62 | scaler = None 63 | scaleData = None 64 | # if a scaler file generated by loadData.py is given, than load it and define a scaler function that will be used later 65 | if scalerFilePath is not None: 66 | scaler = pickle.load(open(scalerFilePath, 'rb')) 67 | # Can't use scaler.transform because it only supports 2d arrays. 68 | def scaleData(X): 69 | return (X-scaler.mean_)/scaler.scale_ 70 | 71 | from io_utils_mod import HDF5Matrix 72 | f = h5py.File(hdf5path, 'r') 73 | X = f.get('X') 74 | y = f.get('y') 75 | print("Shape of X: ") 76 | print(X.shape) 77 | dataSetLength = X.shape[0] 78 | output_dim = y.shape[1] #len(y_train[0]) 79 | # test and validation splits 80 | testSplit = 0.01 # 1% 81 | validationSplit = 0.05 # 5% 82 | f.close() 83 | # load training data 84 | X_train = HDF5Matrix(hdf5path, 'X', 0, int(dataSetLength*(1-(testSplit+validationSplit))), normalizer = scaleData) 85 | y_train = HDF5Matrix(hdf5path, 'y', 0, int(dataSetLength*(1-(testSplit+validationSplit)))) 86 | # load validation data 87 | X_validation = HDF5Matrix(hdf5path, 'X', int(dataSetLength*(1-(testSplit+validationSplit)))+1, int(dataSetLength*(1-testSplit)), normalizer = scaleData) 88 | y_validation = HDF5Matrix(hdf5path, 'y', int(dataSetLength*(1-(testSplit+validationSplit)))+1, int(dataSetLength*(1-testSplit))) 89 | # load test data 90 | X_test = HDF5Matrix(hdf5path, 'X', int(dataSetLength*(1-testSplit))+1, dataSetLength, normalizer = scaleData) 91 | y_test = HDF5Matrix(hdf5path, 'y', int(dataSetLength*(1-testSplit))+1, dataSetLength) 92 | 93 | print("Shape of X_train after train-validation-test split:") 94 | print(X_train.shape) 95 | 96 | # store the starting time 97 | startTime = time.time() 98 | 99 | # load model and compile it, we use RMSprop here, other optimizer algorithm should be tested 100 | execfile(modelPath) 101 | model.compile(loss='categorical_crossentropy', optimizer='rmsprop')#, metrics=["accuracy"]) 102 | 103 | # print the model 104 | print("The following model is used: ") 105 | for layer in model.layers: 106 | print("{} output shape: {}".format(layer.name, layer.output_shape)) 107 | 108 | # load pretrained model if it is set 109 | if preTrainedModelWeightsPath is not None: 110 | model.load_weights(preTrainedModelWeightsPath) 111 | print("Reloaded weights from: {}".format(preTrainedModelWeightsPath)) 112 | 113 | # define callback functions 114 | mapcallback = MapCallback() 115 | earlyStopping = EarlyStopping(monitor='val_loss', patience = es_patience) # early stoping 116 | # save best models based on accuracy, loss and MAP metrics 117 | #bestModelFilePath_val_map = './modelWeights/best_val_map_{}_{}.hdf5'.format(output_dim, datetime.datetime.now().strftime('%Y-%m-%d-%M-%S')) 118 | #bestModelFilePath_val_acc = './modelWeights/best_val_acc_{}_{}.hdf5'.format(output_dim, datetime.datetime.now().strftime('%Y-%m-%d-%M-%S')) 119 | #bestModelFilePath_val_loss = './modelWeights/best_val_loss_{}_{}.hdf5'.format(output_dim, datetime.datetime.now().strftime('%Y-%m-%d-%M-%S')) 120 | bestModelFilePath_val_acc = './modelWeights/best_val_acc_{}.hdf5'.format(output_dim) 121 | bestModelFilePath_val_loss = './modelWeights/best_val_loss_{}.hdf5'.format(output_dim) 122 | bestModelFilePath_val_map = './modelWeights/best_val_map_{}.hdf5'.format(output_dim) 123 | checkpointer_val_acc = ModelCheckpoint(filepath = bestModelFilePath_val_acc, verbose = 1, monitor = 'val_acc', save_best_only = True) 124 | checkpointer_val_loss = ModelCheckpoint(filepath = bestModelFilePath_val_loss, verbose = 1, monitor = 'val_loss', save_best_only = True) 125 | checkpointer_val_map = ModelCheckpoint(filepath = bestModelFilePath_val_map, verbose = 1, monitor = 'val_map', mode = 'max', save_best_only = True) 126 | 127 | # training 128 | fitting_result = model.fit(X_train, y_train, nb_epoch = nb_epochs, batch_size = batchSize, callbacks = [earlyStopping, mapcallback, checkpointer_val_acc, checkpointer_val_loss, checkpointer_val_map], shuffle = 'batch', validation_data = (X_validation, y_validation)) 129 | 130 | # calculate the elapsed time 131 | elapsed = time.time()-startTime; 132 | print("Execution time: {0} s".format(elapsed)) 133 | 134 | # convert the output (probabilistics) to classes 135 | def proba_to_class(a): 136 | classCount = len(a[0]) 137 | to_return = np.empty((0,classCount)) 138 | for row in a: 139 | maxind = np.argmax(row) 140 | to_return = np.vstack((to_return, [1 if i == maxind else 0 for i in range(classCount)])) 141 | return to_return 142 | 143 | # calculate metrics on test data with the last model 144 | from sklearn.metrics import average_precision_score, accuracy_score 145 | y_result = model.predict(X_test) 146 | map = average_precision_score( y_test.data[y_test.start: y_test.end], y_result, average='micro') 147 | accuracy = accuracy_score(y_test.data[y_test.start: y_test.end], proba_to_class(y_result)) 148 | print("AveragePrecision: {}".format(map)) 149 | print("Accuracy: {}".format(accuracy)) 150 | 151 | # reload the best model with smallest validation loss and calculate metrics on test data 152 | print("----- Loading best model from: {} -------".format(bestModelFilePath_val_loss)) 153 | model.load_weights(bestModelFilePath_val_loss) 154 | y_result_bm = model.predict(X_test) 155 | map_bm_val_loss = average_precision_score( y_test.data[y_test.start: y_test.end], y_result_bm, average='macro') 156 | accuracy_bm_val_loss = accuracy_score(y_test.data[y_test.start: y_test.end], proba_to_class(y_result_bm)) 157 | print("AveragePrecision: {}".format(map_bm_val_loss)) 158 | print("Accuracy: {}".format(accuracy_bm_val_loss)) 159 | 160 | # reload the best model with highest validation accuracy and calculate metrics on test data 161 | print("----- Loading best model from: {} -------".format(bestModelFilePath_val_acc)) 162 | model.load_weights(bestModelFilePath_val_acc) 163 | y_result_bm = model.predict(X_test) 164 | map_bm_val_acc = average_precision_score( y_test.data[y_test.start: y_test.end], y_result_bm, average='macro') 165 | accuracy_bm_val_acc = accuracy_score(y_test.data[y_test.start: y_test.end], proba_to_class(y_result_bm)) 166 | print("AveragePrecision: {}".format(map_bm_val_acc)) 167 | print("Accuracy: {}".format(accuracy_bm_val_acc)) 168 | 169 | # save the results summery into an excel file 170 | import log 171 | log.logToXLS(logfileName, model, fitting_result, {'execution(s)':elapsed, 'map':map, 'accuracy':accuracy, 'map_bm_val_loss':map_bm_val_loss, 'accuracy_bm_val_loss':accuracy_bm_val_loss,'map_bm_val_acc':map_bm_val_acc, 'accuracy_bm_val_acc':accuracy_bm_val_acc, 'modelPyFile': modelPath}) 172 | -------------------------------------------------------------------------------- /preprocess/xmltodict.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | "Makes working with XML feel like you are working with JSON" 3 | """ Source: https://github.com/martinblech/xmltodict""" 4 | 5 | try: 6 | import defusedexpat as expat 7 | except ImportError: 8 | from xml.parsers import expat 9 | from xml.sax.saxutils import XMLGenerator 10 | from xml.sax.xmlreader import AttributesImpl 11 | try: # pragma no cover 12 | from cStringIO import StringIO 13 | except ImportError: # pragma no cover 14 | try: 15 | from StringIO import StringIO 16 | except ImportError: 17 | from io import StringIO 18 | try: # pragma no cover 19 | from collections import OrderedDict 20 | except ImportError: # pragma no cover 21 | try: 22 | from ordereddict import OrderedDict 23 | except ImportError: 24 | OrderedDict = dict 25 | 26 | try: # pragma no cover 27 | _basestring = basestring 28 | except NameError: # pragma no cover 29 | _basestring = str 30 | try: # pragma no cover 31 | _unicode = unicode 32 | except NameError: # pragma no cover 33 | _unicode = str 34 | 35 | __author__ = 'Martin Blech' 36 | __version__ = '0.10.1' 37 | __license__ = 'MIT' 38 | 39 | 40 | class ParsingInterrupted(Exception): 41 | pass 42 | 43 | 44 | class _DictSAXHandler(object): 45 | def __init__(self, 46 | item_depth=0, 47 | item_callback=lambda *args: True, 48 | xml_attribs=True, 49 | attr_prefix='@', 50 | cdata_key='#text', 51 | force_cdata=False, 52 | cdata_separator='', 53 | postprocessor=None, 54 | dict_constructor=OrderedDict, 55 | strip_whitespace=True, 56 | namespace_separator=':', 57 | namespaces=None, 58 | force_list=None): 59 | self.path = [] 60 | self.stack = [] 61 | self.data = [] 62 | self.item = None 63 | self.item_depth = item_depth 64 | self.xml_attribs = xml_attribs 65 | self.item_callback = item_callback 66 | self.attr_prefix = attr_prefix 67 | self.cdata_key = cdata_key 68 | self.force_cdata = force_cdata 69 | self.cdata_separator = cdata_separator 70 | self.postprocessor = postprocessor 71 | self.dict_constructor = dict_constructor 72 | self.strip_whitespace = strip_whitespace 73 | self.namespace_separator = namespace_separator 74 | self.namespaces = namespaces 75 | self.force_list = force_list 76 | 77 | def _build_name(self, full_name): 78 | if not self.namespaces: 79 | return full_name 80 | i = full_name.rfind(self.namespace_separator) 81 | if i == -1: 82 | return full_name 83 | namespace, name = full_name[:i], full_name[i+1:] 84 | short_namespace = self.namespaces.get(namespace, namespace) 85 | if not short_namespace: 86 | return name 87 | else: 88 | return self.namespace_separator.join((short_namespace, name)) 89 | 90 | def _attrs_to_dict(self, attrs): 91 | if isinstance(attrs, dict): 92 | return attrs 93 | return self.dict_constructor(zip(attrs[0::2], attrs[1::2])) 94 | 95 | def startElement(self, full_name, attrs): 96 | name = self._build_name(full_name) 97 | attrs = self._attrs_to_dict(attrs) 98 | self.path.append((name, attrs or None)) 99 | if len(self.path) > self.item_depth: 100 | self.stack.append((self.item, self.data)) 101 | if self.xml_attribs: 102 | attr_entries = [] 103 | for key, value in attrs.items(): 104 | key = self.attr_prefix+self._build_name(key) 105 | if self.postprocessor: 106 | entry = self.postprocessor(self.path, key, value) 107 | else: 108 | entry = (key, value) 109 | if entry: 110 | attr_entries.append(entry) 111 | attrs = self.dict_constructor(attr_entries) 112 | else: 113 | attrs = None 114 | self.item = attrs or None 115 | self.data = [] 116 | 117 | def endElement(self, full_name): 118 | name = self._build_name(full_name) 119 | if len(self.path) == self.item_depth: 120 | item = self.item 121 | if item is None: 122 | item = (None if not self.data 123 | else self.cdata_separator.join(self.data)) 124 | 125 | should_continue = self.item_callback(self.path, item) 126 | if not should_continue: 127 | raise ParsingInterrupted() 128 | if len(self.stack): 129 | data = (None if not self.data 130 | else self.cdata_separator.join(self.data)) 131 | item = self.item 132 | self.item, self.data = self.stack.pop() 133 | if self.strip_whitespace and data: 134 | data = data.strip() or None 135 | if data and self.force_cdata and item is None: 136 | item = self.dict_constructor() 137 | if item is not None: 138 | if data: 139 | self.push_data(item, self.cdata_key, data) 140 | self.item = self.push_data(self.item, name, item) 141 | else: 142 | self.item = self.push_data(self.item, name, data) 143 | else: 144 | self.item = None 145 | self.data = [] 146 | self.path.pop() 147 | 148 | def characters(self, data): 149 | if not self.data: 150 | self.data = [data] 151 | else: 152 | self.data.append(data) 153 | 154 | def push_data(self, item, key, data): 155 | if self.postprocessor is not None: 156 | result = self.postprocessor(self.path, key, data) 157 | if result is None: 158 | return item 159 | key, data = result 160 | if item is None: 161 | item = self.dict_constructor() 162 | try: 163 | value = item[key] 164 | if isinstance(value, list): 165 | value.append(data) 166 | else: 167 | item[key] = [value, data] 168 | except KeyError: 169 | if self._should_force_list(key, data): 170 | item[key] = [data] 171 | else: 172 | item[key] = data 173 | return item 174 | 175 | def _should_force_list(self, key, value): 176 | if not self.force_list: 177 | return False 178 | try: 179 | return key in self.force_list 180 | except TypeError: 181 | return self.force_list(self.path[:-1], key, value) 182 | 183 | 184 | def parse(xml_input, encoding=None, expat=expat, process_namespaces=False, 185 | namespace_separator=':', **kwargs): 186 | """Parse the given XML input and convert it into a dictionary. 187 | 188 | `xml_input` can either be a `string` or a file-like object. 189 | 190 | If `xml_attribs` is `True`, element attributes are put in the dictionary 191 | among regular child elements, using `@` as a prefix to avoid collisions. If 192 | set to `False`, they are just ignored. 193 | 194 | Simple example:: 195 | 196 | >>> import xmltodict 197 | >>> doc = xmltodict.parse(\"\"\" 198 | ... 199 | ... 1 200 | ... 2 201 | ... 202 | ... \"\"\") 203 | >>> doc['a']['@prop'] 204 | u'x' 205 | >>> doc['a']['b'] 206 | [u'1', u'2'] 207 | 208 | If `item_depth` is `0`, the function returns a dictionary for the root 209 | element (default behavior). Otherwise, it calls `item_callback` every time 210 | an item at the specified depth is found and returns `None` in the end 211 | (streaming mode). 212 | 213 | The callback function receives two parameters: the `path` from the document 214 | root to the item (name-attribs pairs), and the `item` (dict). If the 215 | callback's return value is false-ish, parsing will be stopped with the 216 | :class:`ParsingInterrupted` exception. 217 | 218 | Streaming example:: 219 | 220 | >>> def handle(path, item): 221 | ... print 'path:%s item:%s' % (path, item) 222 | ... return True 223 | ... 224 | >>> xmltodict.parse(\"\"\" 225 | ... 226 | ... 1 227 | ... 2 228 | ... \"\"\", item_depth=2, item_callback=handle) 229 | path:[(u'a', {u'prop': u'x'}), (u'b', None)] item:1 230 | path:[(u'a', {u'prop': u'x'}), (u'b', None)] item:2 231 | 232 | The optional argument `postprocessor` is a function that takes `path`, 233 | `key` and `value` as positional arguments and returns a new `(key, value)` 234 | pair where both `key` and `value` may have changed. Usage example:: 235 | 236 | >>> def postprocessor(path, key, value): 237 | ... try: 238 | ... return key + ':int', int(value) 239 | ... except (ValueError, TypeError): 240 | ... return key, value 241 | >>> xmltodict.parse('12x', 242 | ... postprocessor=postprocessor) 243 | OrderedDict([(u'a', OrderedDict([(u'b:int', [1, 2]), (u'b', u'x')]))]) 244 | 245 | You can pass an alternate version of `expat` (such as `defusedexpat`) by 246 | using the `expat` parameter. E.g: 247 | 248 | >>> import defusedexpat 249 | >>> xmltodict.parse('hello', expat=defusedexpat.pyexpat) 250 | OrderedDict([(u'a', u'hello')]) 251 | 252 | You can use the force_list argument to force lists to be created even 253 | when there is only a single child of a given level of hierarchy. The 254 | force_list argument is a tuple of keys. If the key for a given level 255 | of hierarchy is in the force_list argument, that level of hierarchy 256 | will have a list as a child (even if there is only one sub-element). 257 | The index_keys operation takes precendence over this. This is applied 258 | after any user-supplied postprocessor has already run. 259 | 260 | For example, given this input: 261 | 262 | 263 | host1 264 | Linux 265 | 266 | 267 | em0 268 | 10.0.0.1 269 | 270 | 271 | 272 | 273 | 274 | If called with force_list=('interface',), it will produce 275 | this dictionary: 276 | {'servers': 277 | {'server': 278 | {'name': 'host1', 279 | 'os': 'Linux'}, 280 | 'interfaces': 281 | {'interface': 282 | [ {'name': 'em0', 'ip_address': '10.0.0.1' } ] } } } 283 | 284 | `force_list` can also be a callable that receives `path`, `key` and 285 | `value`. This is helpful in cases where the logic that decides whether 286 | a list should be forced is more complex. 287 | """ 288 | handler = _DictSAXHandler(namespace_separator=namespace_separator, 289 | **kwargs) 290 | if isinstance(xml_input, _unicode): 291 | if not encoding: 292 | encoding = 'utf-8' 293 | xml_input = xml_input.encode(encoding) 294 | if not process_namespaces: 295 | namespace_separator = None 296 | parser = expat.ParserCreate( 297 | encoding, 298 | namespace_separator 299 | ) 300 | try: 301 | parser.ordered_attributes = True 302 | except AttributeError: 303 | # Jython's expat does not support ordered_attributes 304 | pass 305 | parser.StartElementHandler = handler.startElement 306 | parser.EndElementHandler = handler.endElement 307 | parser.CharacterDataHandler = handler.characters 308 | parser.buffer_text = True 309 | try: 310 | parser.ParseFile(xml_input) 311 | except (TypeError, AttributeError): 312 | parser.Parse(xml_input, True) 313 | return handler.item 314 | 315 | 316 | def _emit(key, value, content_handler, 317 | attr_prefix='@', 318 | cdata_key='#text', 319 | depth=0, 320 | preprocessor=None, 321 | pretty=False, 322 | newl='\n', 323 | indent='\t', 324 | full_document=True): 325 | if preprocessor is not None: 326 | result = preprocessor(key, value) 327 | if result is None: 328 | return 329 | key, value = result 330 | if (not hasattr(value, '__iter__') 331 | or isinstance(value, _basestring) 332 | or isinstance(value, dict)): 333 | value = [value] 334 | for index, v in enumerate(value): 335 | if full_document and depth == 0 and index > 0: 336 | raise ValueError('document with multiple roots') 337 | if v is None: 338 | v = OrderedDict() 339 | elif not isinstance(v, dict): 340 | v = _unicode(v) 341 | if isinstance(v, _basestring): 342 | v = OrderedDict(((cdata_key, v),)) 343 | cdata = None 344 | attrs = OrderedDict() 345 | children = [] 346 | for ik, iv in v.items(): 347 | if ik == cdata_key: 348 | cdata = iv 349 | continue 350 | if ik.startswith(attr_prefix): 351 | if not isinstance(iv, _unicode): 352 | iv = _unicode(iv) 353 | attrs[ik[len(attr_prefix):]] = iv 354 | continue 355 | children.append((ik, iv)) 356 | if pretty: 357 | content_handler.ignorableWhitespace(depth * indent) 358 | content_handler.startElement(key, AttributesImpl(attrs)) 359 | if pretty and children: 360 | content_handler.ignorableWhitespace(newl) 361 | for child_key, child_value in children: 362 | _emit(child_key, child_value, content_handler, 363 | attr_prefix, cdata_key, depth+1, preprocessor, 364 | pretty, newl, indent) 365 | if cdata is not None: 366 | content_handler.characters(cdata) 367 | if pretty and children: 368 | content_handler.ignorableWhitespace(depth * indent) 369 | content_handler.endElement(key) 370 | if pretty and depth: 371 | content_handler.ignorableWhitespace(newl) 372 | 373 | 374 | def unparse(input_dict, output=None, encoding='utf-8', full_document=True, 375 | **kwargs): 376 | """Emit an XML document for the given `input_dict` (reverse of `parse`). 377 | 378 | The resulting XML document is returned as a string, but if `output` (a 379 | file-like object) is specified, it is written there instead. 380 | 381 | Dictionary keys prefixed with `attr_prefix` (default=`'@'`) are interpreted 382 | as XML node attributes, whereas keys equal to `cdata_key` 383 | (default=`'#text'`) are treated as character data. 384 | 385 | The `pretty` parameter (default=`False`) enables pretty-printing. In this 386 | mode, lines are terminated with `'\n'` and indented with `'\t'`, but this 387 | can be customized with the `newl` and `indent` parameters. 388 | 389 | """ 390 | if full_document and len(input_dict) != 1: 391 | raise ValueError('Document must have exactly one root.') 392 | must_return = False 393 | if output is None: 394 | output = StringIO() 395 | must_return = True 396 | content_handler = XMLGenerator(output, encoding) 397 | if full_document: 398 | content_handler.startDocument() 399 | for key, value in input_dict.items(): 400 | _emit(key, value, content_handler, full_document=full_document, 401 | **kwargs) 402 | if full_document: 403 | content_handler.endDocument() 404 | if must_return: 405 | value = output.getvalue() 406 | try: # pragma no cover 407 | value = value.decode(encoding) 408 | except AttributeError: # pragma no cover 409 | pass 410 | return value 411 | 412 | if __name__ == '__main__': # pragma: no cover 413 | import sys 414 | import marshal 415 | try: 416 | stdin = sys.stdin.buffer 417 | stdout = sys.stdout.buffer 418 | except AttributeError: 419 | stdin = sys.stdin 420 | stdout = sys.stdout 421 | 422 | (item_depth,) = sys.argv[1:] 423 | item_depth = int(item_depth) 424 | 425 | 426 | def handle_item(path, item): 427 | marshal.dump((path, item), stdout) 428 | return True 429 | 430 | try: 431 | root = parse(stdin, 432 | item_depth=item_depth, 433 | item_callback=handle_item, 434 | dict_constructor=dict) 435 | if item_depth == 0: 436 | handle_item([], root) 437 | except KeyboardInterrupt: 438 | pass 439 | -------------------------------------------------------------------------------- /preprocess/loadData.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | # 3 | # Birdsong classificatione in noisy environment with convolutional neural nets in Keras 4 | # Copyright (C) 2017 Balint Czeba, Balint Pal Toth (toth.b@tmit.bme.hu) 5 | # 6 | # This program is free software: you can redistribute it and/or modify 7 | # it under the terms of the GNU General Public License as published by 8 | # the Free Software Foundation, either version 3 of the License, or 9 | # (at your option) any later version. 10 | # 11 | # This program is distributed in the hope that it will be useful, 12 | # but WITHOUT ANY WARRANTY; without even the implied warranty of 13 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 14 | # GNU General Public License for more details. 15 | # 16 | # You should have received a copy of the GNU General Public License 17 | # along with this program. If not, see . (c) Balint Czeba, Balint Pal Toth 18 | # 19 | # Please cite the following paper if this code was useful for your research: 20 | # 21 | # Tóth Bálint Pál, Czeba Bálint, 22 | # "Convolutional Neural Networks for Large-Scale Bird Song Classification in Noisy Environment", 23 | # In: Working Notes of Conference and Labs of the Evaluation Forum, Évora, Portugália, 2016, p. 8 24 | 25 | from scipy import io 26 | from scipy.io import wavfile 27 | import numpy as np 28 | import os 29 | import matplotlib 30 | import matplotlib.pyplot as plt 31 | from matplotlib import mlab 32 | import xmltodict 33 | import pandas as pd 34 | import h5py 35 | import pickle 36 | import gc 37 | np.random.seed(0) 38 | matplotlib.use('Agg') 39 | plt.ioff() 40 | 41 | # traiing data related paths 42 | PATH_TRAIN_IN_16KWAVS = '../birdclef_data/TrainingSet/wav_16khz' # the location where the 16kHz resampled wavs are located 43 | PATH_TRAIN_IN_XMLFILES = '../birdclef_data/TrainingSet/xml/' # the path where the XML meta-data files are located 44 | PATH_TRAIN_OUT_XMLPICKLEFILE = '../birdclef_data/TrainingSet/xml_data.pickle' # the location and filename where the XML meta-data will be saved 45 | PATH_TRAIN_OUT_HDF5 = '../birdclef_data/' # the path where preprocessed data will be saved in HDF5 format 46 | 47 | # same as above just with test data, the wav related data will be generating in train/predict.py 48 | PATH_TEST_IN_XMLFILES = '../birdclef_data/test/xml2015' 49 | PATH_TEST_OUT_XMLPICKLEFILE = '../birdclef_data/test/xml_data.pickle' 50 | 51 | # this parameter is used for preprocessing 52 | # the number comes from the following equation: np.floor(sampling_frequency/(FFT_length-FFT_overlap))*num_of_seconds 53 | # we use 16kHz sampling rate for the wavs, 512 FFT window length with 256 overlap and we investigate 5 seconds 54 | spectrogramWindowLength = int(5*np.floor(16000/(512-256))); 55 | 56 | # function to load the wave files from dirPath 57 | def loadWavs(dirPath): 58 | data = list() 59 | for path, subdirs, files in os.walk(dirpath): 60 | for name in files: 61 | if (name.endswith('.wav')): 62 | print os.path.join(path, name) 63 | tempFileName = name.split('.')[0] 64 | tempXmlFile = open(os.path.join(xmldirpath, tempFileName + '.xml'), 'rb') 65 | metadata = tempXmlFile.readlines() 66 | tempXmlFile.close() 67 | data.append([io.wavfile.read(os.path.join(path, name))[1], xmltodict.parse(''.join(metadata))['Audio']]) 68 | return data; 69 | 70 | # function to load corresponding XML files into a Pandas dataframe 71 | def readXMLs(xmldirpath): 72 | df = pd.DataFrame() 73 | for path, subdirs, files in os.walk(xmldirpath): 74 | for name in files: 75 | print os.path.join(path, name) 76 | if (name.endswith('.xml')): 77 | print os.path.join(path, name) 78 | tempXmlFile = open(os.path.join(path, name), 'rb') 79 | metadata = tempXmlFile.readlines() 80 | tempXmlFile.close() 81 | df = df.append(pd.DataFrame(xmltodict.parse(''.join(metadata))['Audio'], index=[0])).reset_index(drop=True) 82 | return df 83 | 84 | # function to merge meta-data of the XML file 85 | def mergeOFGS(row): 86 | return "{} {} {} {}".format(row["Order"], row["Family"], row["Genus"], row["Species"]) 87 | 88 | # if 'x' array contains 1, this expands it inthe given directions 89 | # used for the mask applied to the spectogram 90 | def expandOnes(x, directions = [[-1,0], [1,0], [0,-1], [0,1]]): 91 | expand = np.zeros(x.shape) 92 | for i in range(x.shape[0]): 93 | for j in range(x.shape[1]): 94 | if (x[i,j] == 1): 95 | for direction in directions: 96 | cx = i + direction[0] 97 | cy = j + direction[1] 98 | if (0 <= cx < x.shape[0] and 0 <= cy < x.shape[1]): 99 | expand[cx,cy]=1; 100 | return x+expand; 101 | 102 | # function to filter the spectogram based on the energy of the signal 103 | # 104 | # parameters: 105 | # data audio data 106 | # expandByOne if it is True, than the mask of the spectogram will be expanded in every direction 107 | # dropZeroColumnsPercent determines the ratio of 0 values along the frequency axis when a timeslice is dropped 108 | # 109 | # return values: 110 | # spectogram 111 | # filtered spectogram 112 | 113 | def audioToFilteredSpectrogram(data, expandByOne = True, dropZeroColumnsPercent = 0.95): 114 | # calculate the spectogram 115 | tempSpec = np.log10(mlab.specgram(data, NFFT=512, noverlap=256, Fs=16000)[0]) 116 | 117 | # drop higher frequencies 118 | tempSpec = tempSpec[0:200,:] 119 | tempSpecFiltered = np.copy(tempSpec) 120 | 121 | # we analize the spectogram by 20x30 sized cells 122 | # to achieve better accuray the size of this cell should be finetuned 123 | rowBorders = np.ceil(np.linspace(0,tempSpec.shape[0], 20)) 124 | columnBorders = np.hstack((np.ceil(np.arange(0,tempSpec.shape[1], 30)), tempSpec.shape[1])) 125 | rowBorders = [ int(x) for x in rowBorders ] 126 | columnBorders = [ int(x) for x in columnBorders ] 127 | keepCells = np.ones((len(rowBorders)-1, len(columnBorders)-1)) 128 | 129 | # we create a mask for the spectogram: we scan the spectogram with the 20x30 sized 130 | # cell and create 0 mask based on the mean and std of the spectogram calculated for the cells and rows 131 | for i in range(len(rowBorders)-1): 132 | row_mean = np.mean(tempSpec[rowBorders[i]:rowBorders[i+1],:]) 133 | row_std = np.std(tempSpec[rowBorders[i]:rowBorders[i+1],:]) 134 | 135 | for j in range(len(columnBorders)-1): 136 | cell_mean = np.mean(tempSpec[rowBorders[i]:rowBorders[i+1],columnBorders[j]:columnBorders[j+1]]) 137 | cell_max_top10_mean = np.mean(np.sort(tempSpec[rowBorders[i]:rowBorders[i+1],columnBorders[j]:columnBorders[j+1]], axis=None)[-10:]) 138 | 139 | if (cell_mean < 0 or ((cell_max_top10_mean) < (row_mean + row_std)*1.5)): 140 | keepCells[i,j]=0 141 | 142 | # expand by ones (see above) 143 | if expandByOne: 144 | keepCells = expandOnes(keepCells) 145 | 146 | # apply the mask to the spectogram 147 | for i in range(keepCells.shape[0]): 148 | for j in range(keepCells.shape[1]): 149 | if not keepCells[i,j]: 150 | tempSpecFiltered[rowBorders[i]:rowBorders[i+1],columnBorders[j]:columnBorders[j+1]] = 0 151 | 152 | # drop zero columns 153 | # the amount of zero values along axis 0 (frequency) is calculated for every column (timeslice) 154 | # and it is dropped, if the number of zero values is higher than the dropZeroColumnsPercent 155 | # eg. dropZeroColumnsPercent=0.95, than a column (timeslice) is dropped, if more than 95% of the values (frequencies) is 0 156 | tempSpecFilteredBackup = np.copy(tempSpecFiltered) 157 | tempSpecFiltered = np.delete(tempSpecFiltered, np.nonzero((tempSpecFiltered==0).sum(axis=0) > tempSpecFiltered.shape[0]*dropZeroColumnsPercent), axis=1) 158 | 159 | # if every row was 0 than use the backed up spectogram 160 | if tempSpecFiltered.shape[1] == 0: 161 | tempSpecFiltered = tempSpecFilteredBackup 162 | 163 | return tempSpec, tempSpecFiltered; 164 | 165 | # function to return most common classes in the dataset 166 | def getMostCommon(df, N=10): 167 | from collections import Counter 168 | c = Counter(df["ClassId"]) 169 | mostCommon = c.most_common(N) 170 | df_mostCommon = pd.DataFrame() 171 | for item in mostCommon: 172 | df_mostCommon = df_mostCommon.append(df[df["ClassId"] == item[0]], ignore_index=True) 173 | df_mostCommon.reset_index(drop=True) 174 | return df_mostCommon; 175 | 176 | # function to return data inbetween minQuality and maxQuality 177 | def filterByQuality(df, minQuality=0, maxQuality=5): 178 | df_filtered = pd.DataFrame() 179 | for quality in range(minQuality, maxQuality+1): 180 | df_filtered = df_filtered.append(df[df["Quality"] == str(quality)], ignore_index=True) 181 | df_filtered.reset_index(drop=True, inplace=True) 182 | return df_filtered; 183 | 184 | # function to remove samples where background species exist 185 | def removeSamplesWithBackgroundSpecies(df): 186 | return df[df.BackgroundSpecies.isnull()] 187 | 188 | # create onehot encoding for order, family, genus, specie 189 | def getOneHotOFGS(df): 190 | from sklearn.preprocessing import LabelBinarizer 191 | lb = LabelBinarizer() 192 | lb.fit(df["OFGS"]) 193 | return ( lb, lb.transform(df["OFGS"]) ) 194 | 195 | # create onehot encoding for classid 196 | def getOneHotClassId(df): 197 | from sklearn.preprocessing import LabelBinarizer 198 | lb = LabelBinarizer() 199 | lb.fit(df["ClassId"]) 200 | return ( lb, lb.transform(df["ClassId"]) ) 201 | 202 | # bulk processing of wav files in path 203 | # 204 | # parameters: 205 | # path the source path 206 | # filenames the filenames in the path 207 | # dontFilter does not filter the spectogram if it is set to True 208 | # 209 | # return value: 210 | # specotogram data of multiple files 211 | 212 | def wavsToSpectrogramByList(path, filenames, dontFilter=False): 213 | print("wavsToSpectrogramByList...") 214 | data=list() 215 | for filename in filenames: 216 | print('\r Processing {}'.format(os.path.join(path, filename))), 217 | (tempSpecUnfiltered,tempSpecFiltered) = audioToFilteredSpectrogram(io.wavfile.read(os.path.join(path, filename))[1], expandByOne=True) 218 | if (not dontFilter): 219 | data.append(tempSpecFiltered) 220 | else: 221 | data.append(tempSpecUnfiltered) 222 | print("\nwavsToSpectrogramByList finished") 223 | return data; 224 | 225 | # function to create training data from the list generated by wavsToSpectogramByList function 226 | # 227 | # parameters: 228 | # slist the spectogram list generated by wavsToSpectogramByList function 229 | # labels contain the class labels of the corresponding spectograms 230 | # N (1*44100)/(1024-512)=86 231 | # filenames filenames to return 232 | # classIds class IDs to return 233 | # 234 | # return values 235 | # X the constructed input 236 | # y the constructed output 237 | # fn filenames 238 | # cIds class IDs 239 | # 240 | 241 | def spectrogramListToT4(slist, labels=None, N=spectrogramWindowLength, filenames=None, classIds=None): 242 | print("SpectrogramListToT4...") 243 | 244 | rows = len(slist[0]) 245 | X = np.empty((0,1,rows,N)) 246 | y = [] 247 | fn = [] 248 | cIds = [] 249 | 250 | # process all spectograms 251 | for i in range(len(slist)): 252 | print('\r Processing no. %d / %d' % (i, len(slist))) 253 | ranges = np.hstack((np.arange(0, len(slist[i][0]), N), len(slist[i][0]))) 254 | 255 | for j in range(len(ranges)-1): 256 | # variable contains 257 | tempSpec = np.empty((1,rows,N)) 258 | 259 | if (len(slist[i][0]) < N): # if data is shorter than N than fill up with zeros 260 | tempSpec[0] = np.hstack((slist[i],np.zeros((rows, N-len(slist[i][0]))))) 261 | elif (ranges[j+1]-ranges[j] < N): # last element 262 | tempSpec[0] = slist[i][:,-N:] 263 | else: # other part of the spectrum 264 | tempSpec[0] = slist[i][:,ranges[j]:ranges[j+1]] 265 | 266 | X = np.vstack((X,[tempSpec])) 267 | 268 | if labels is not None: 269 | y.append(labels[i]) 270 | if filenames is not None: 271 | fn.append(filenames[i]) 272 | if classIds is not None: 273 | cIds.append(classIds[i]) 274 | 275 | print("SpectrogramListToT4 finished") 276 | return X, y, fn, cIds 277 | 278 | # calculates the standard scaler coefficients of the input data for 0 mean and 1 variance 279 | # 280 | # parameters 281 | # numberOfFiles the number of files to process (we assume that the mean and variance will be similar in case 282 | # of a subset of the training data and we don't have to process the whole database 283 | # wavdirpath the path that contains the wavs (sampled at 16kHz) 284 | # xmlpicklepath the path and filename that contains the XML file for training (xml_data.pickle) 285 | # 286 | # return values 287 | # scaler 288 | # spectogramData 289 | # 290 | 291 | def generateScaler(numberOfFiles=100, wavdirpath=PATH_TRAIN_IN_16KWAVS, xmlpicklepath=PATH_TRAIN_OUT_XMLPICKLEFILE, todirrootpath=PATH_TRAIN_OUT_HDF5): 292 | if not os.path.exists(todirrootpath): 293 | os.makedirs(todirrootpath) 294 | 295 | import pickle 296 | df = pd.read_pickle(xmlpicklepath) # contains the metadata 297 | print("Metadata loaded") 298 | 299 | # Shuffle rows 300 | df = df.iloc[np.random.permutation(len(df))] 301 | df.reset_index(drop=True, inplace=True) 302 | print("Metadata shuffled") 303 | 304 | # Calculate spectograms 305 | spectrogramData = wavsToSpectrogramByList(wavdirpath, df.FileName[:numberOfFiles], dontFilter=False) 306 | print("Spectrograms done.") 307 | 308 | print('Building scaler...') 309 | from sklearn.preprocessing import StandardScaler 310 | scaler = StandardScaler() 311 | # calculate the scaler variables spectogram by spectogram 312 | for sData in spectrogramData: 313 | scaler.partial_fit(sData.reshape(-1,1)) 314 | 315 | # filename where we save the scaler 316 | saveTo = os.path.join(todirrootpath,"standardScaler_{}.pickle".format(numberOfFiles)) 317 | from sklearn.externals import joblib 318 | import pickle 319 | pickle.dump(scaler, open(saveTo, 'wb')) 320 | print('Scaler saved to: {}'.format(saveTo)) 321 | 322 | return scaler, spectrogramData 323 | 324 | # function that constructs training data 325 | # 326 | # parameters: 327 | # N number of most classes to take into account 328 | # wavdirpath path of the wave files (16kHz) 329 | # xmlpicklepath the path and filename that contains the XML file for training (xml_data.pickle) 330 | # todirrootpath path were to save the training data 331 | # 332 | # return values: 333 | # X,y,fn for debuging purposes 334 | # 335 | 336 | def processNMostCommon(N=3, wavdirpath=PATH_TRAIN_IN_16KWAVS, xmlpicklepath=PATH_TRAIN_OUT_XMLPICKLEFILE, todirrootpath=PATH_TRAIN_OUT_HDF5): 337 | global spectrogramWindowLength 338 | 339 | if not os.path.exists(todirrootpath): 340 | os.makedirs(todirrootpath) 341 | 342 | spectrogramHeight = 200 343 | 344 | f = h5py.File(os.path.join(todirrootpath,"data_top{}_nozero.hdf5".format(N)), "w") 345 | dsetX = f.create_dataset('X', (0,1,spectrogramHeight,spectrogramWindowLength), maxshape=(None, 1,spectrogramHeight,spectrogramWindowLength)) 346 | dsety = f.create_dataset('y', (0,N), maxshape=(None,N)) 347 | dsetMediaId = f.create_dataset('MediaId', (0,1), maxshape=(None,1)) 348 | dsetClassId = f.create_dataset('ClassId', (0,1), maxshape=(None,1), dtype=h5py.special_dtype(vlen=unicode)) 349 | 350 | import pickle 351 | df = pd.read_pickle(xmlpicklepath) # read the metadata 352 | 353 | # if we would like to keep recordings with a given quality than we can do it here by uncommenting the next line 354 | #df = filterByQuality(df, 0, 3) 355 | 356 | df["OFGS"] = df.apply(mergeOFGS, axis=1) # merge Order, Family, Genus, Species 357 | df_mc = getMostCommon(df, N) # get N most common classes from the dataset 358 | df = None # let GC free up some memory 359 | print("Metadata loaded") 360 | 361 | # Shuffle rows 362 | df_mc = df_mc.iloc[np.random.permutation(len(df_mc))] 363 | df_mc.reset_index(drop=True, inplace=True) 364 | (lb,binaryLabels) = getOneHotClassId(df_mc) # generate one-hot labels 365 | pickle.dump(lb, open(os.path.join(todirrootpath,"labelBinarizer_top{}.pickle".format(N)), 'wb')) 366 | 367 | # process the selected files of top N classes and save the data into HDF5 368 | fileRanges = np.hstack((np.arange(0, len(df_mc), 30), len(df_mc))) 369 | for i in range(len(fileRanges)-1): 370 | tempSG = wavsToSpectrogramByList(wavdirpath, df_mc.FileName[fileRanges[i]: fileRanges[i+1]], dontFilter=False) 371 | X, y, fn, cIds = spectrogramListToT4(tempSG, \ 372 | binaryLabels[fileRanges[i]: fileRanges[i+1]], \ 373 | filenames = df_mc.MediaId[fileRanges[i]: fileRanges[i+1]].values, N=spectrogramWindowLength, \ 374 | classIds = df_mc.ClassId[fileRanges[i]: fileRanges[i+1]].values) #convert to t4 375 | pre_len = dsetX.shape[0] 376 | add_len = X.shape[0] 377 | dsetX.resize(pre_len+add_len, axis=0) 378 | dsety.resize(pre_len+add_len, axis=0) 379 | dsetMediaId.resize(pre_len + add_len, axis=0) 380 | dsetClassId.resize(pre_len + add_len, axis=0) 381 | dsetX[pre_len:pre_len+add_len,:,:,:] = X 382 | dsety[pre_len:pre_len+add_len,:] = y 383 | dsetMediaId[pre_len:pre_len+add_len,:] = np.transpose([[int(i) for i in fn]]) 384 | dsetClassId[pre_len:pre_len+add_len,:] = np.transpose([[s.encode('utf8') for s in cIds]]) 385 | f.flush() 386 | 387 | f.close 388 | return (X,y,fn) # return last batch for debug purposes 389 | 390 | print("== Generating training data ==") 391 | print("Reading XML files and generating pickle file") 392 | df_xml = readXMLs(PATH_TRAIN_IN_XMLFILES) # read XML files with meta-data 393 | df_xml.to_pickle(PATH_TRAIN_OUT_XMLPICKLEFILE) # save the loaded meta-data into a pickle file with all the informatio 394 | print("Process wav files and save them into HDF5") 395 | (X, y, fn) = processNMostCommon(999, wavdirpath=PATH_TRAIN_IN_16KWAVS, xmlpicklepath=PATH_TRAIN_OUT_XMLPICKLEFILE, todirrootpath=PATH_TRAIN_OUT_HDF5) # processes the most common 999 species (so the whole dataset) 396 | # print("Generating scaler") 397 | # scaler, data = generateScaler(5000, wavdirpath=PATH_TRAIN_IN_16KWAVS, xmlpicklepath=PATH_TRAIN_OUT_XMLPICKLEFILE, todirrootpath=PATH_TRAIN_OUT_HDF5) # calculates and saves the standard scaler based on 5000 wav files 398 | 399 | print("== Generating test data ==") 400 | print("Reading XML files and generating pickle file") 401 | df_xml = readXMLs(PATH_TEST_IN_XMLFILES) # read XML files with meta-data 402 | df_xml.to_pickle(PATH_TEST_OUT_XMLPICKLEFILE) # save the loaded meta-data into a pickle file with all the informatio 403 | 404 | 405 | ''' 406 | # The following code is for plotting some data for inspection 407 | for data_item in data: 408 | print("Processing {}".format(data_item[1]['FileName'])) 409 | tempSpec, tempSpecFiltered = audioToFilteredSpectrogram(data_item[0]) 410 | plt.ioff() 411 | fig = plt.figure(figsize=(19,10)) 412 | plt.hold(False) 413 | ax1=plt.subplot(211) 414 | plt.tight_layout() 415 | plt.pcolormesh(tempSpec, cmap=plt.cm.binary) 416 | plt.subplot(212, sharex=ax1, sharey=ax1) 417 | plt.pcolormesh(tempSpecFiltered, cmap=plt.cm.binary) 418 | fig.savefig('../figures/filter_expand1/' + data_item[1]['FileName'][:-4] + '.png', transparent = True) 419 | plt.close(fig) 420 | ''' 421 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | 622 | --------------------------------------------------------------------------------