[MXNET-1210 ] Gluon Audio

example/gluon/urban_sounds/README.md

@@ -0,0 +1,22 @@
+# Urban Sounds classification in MXNet
+
+Urban Sounds Dataset:
+## Description
+  The dataset contains 8732 wav files which are audio samples(<= 4s)) of street sounds like engine_idling, car_horn, children_playing, dog_barking and so on.
+  The task is to classify these audio samples into one of the 10 labels.
+
+To be able to run this example:
+
+1. Download the dataset(train.zip, test.zip) required for this example from the location:
+**https://drive.google.com/drive/folders/0By0bAi7hOBAFUHVXd1JCN3MwTEU**
+
+
+2. Extract both the zip archives into the **current directory** - after unzipping you would get 2 new folders namely,\
+   **Train** and **Test** and two csv files - **train.csv**, **test.csv**
+
+3. Apache MXNet is installed on the machine. For instructions, go to the link: **https://mxnet.incubator.apache.org/install/**
+
+4. Librosa is installed. To install, use the commands
+   `pip install librosa`,
+   For more details, refer here:
+   **https://librosa.github.io/librosa/install.html**

example/gluon/urban_sounds/model.py

@@ -0,0 +1,34 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+"""
+    This module builds a model an MLP with a configurable output layer( number of units in the last layer).
+    Users can pass any number of units in the last layer. SInce this dataset has 10 labels,
+    the default value of num_labels = 10
+"""
+import mxnet as mx
+from mxnet import gluon
+
+# Defining a neural network with number of labels
+def get_net(num_labels=10):
+    net = gluon.nn.Sequential()
+    with net.name_scope():
+        net.add(gluon.nn.Dense(256, activation="relu")) # 1st layer (256 nodes)
+        net.add(gluon.nn.Dense(256, activation="relu")) # 2nd hidden layer
+        net.add(gluon.nn.Dense(num_labels))
+    net.collect_params().initialize(mx.init.Normal(1.))
+    return net

example/gluon/urban_sounds/predict.py

@@ -0,0 +1,91 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+""" Prediction module for Urban Sounds Classification
+"""
+import os
+import warnings
+import mxnet as mx
+from mxnet import nd
+from mxnet.gluon.contrib.data.audio.transforms import MFCC
+from model import get_net
+
+def predict(prediction_dir='./Test'):
+    """The function is used to run predictions on the audio files in the directory `pred_directory`.
+
+    Parameters
+    ----------
+    net: 
+        The model that has been trained.
+    prediction_dir: string, default ./Test
+        The directory that contains the audio files on which predictions are to be made
+
+    """
+
+    try:
+        import librosa
+    except ImportError:
+        warnings.warn("Librosa is not installed! please run the following command pip install librosa.")
+        return
+
+    if not os.path.exists(prediction_dir):
+        warnings.warn("The directory on which predictions are to be made is not found!")
+        return
+
+    if len(os.listdir(prediction_dir)) == 0:
+        warnings.warn("The directory on which predictions are to be made is empty! Exiting...")
+        return
+
+    # Loading synsets
+    if not os.path.exists('./synset.txt'):
+        warnings.warn("The synset or labels for the dataset do not exist. Please run the training script first.")
+        return
+
+    with open("./synset.txt", "r") as f:
+        synset = [l.rstrip() for l in f]
+    net = get_net(len(synset))
+    print("Trying to load the model with the saved parameters...")
+    if not os.path.exists("./net.params"):
+        warnings.warn("The model does not have any saved parameters... Cannot proceed! Train the model first")
+        return
+
+    net.load_parameters("./net.params")
+    file_names = os.listdir(prediction_dir)
+    full_file_names = [os.path.join(prediction_dir, item) for item in file_names]
+    mfcc = MFCC()
+    print("\nStarting predictions for audio files in ", prediction_dir, " ....\n")
+    for filename in full_file_names:
+        # Argument kaiser_fast to res_type is faster than 'kaiser_best'. To reduce the load time, passing kaiser_fast.
+        X1, _ = librosa.load(filename, res_type='kaiser_fast')
+        transformed_test_data = mfcc(mx.nd.array(X1))
+        output = net(transformed_test_data.reshape((1, -1)))
+        prediction = nd.argmax(output, axis=1)
+        print(filename, " -> ", synset[(int)(prediction.asscalar())])
+
+
+if __name__ == '__main__':
+    try:
+        import argparse
+        parser = argparse.ArgumentParser(description="Urban Sounds clsssification example - MXNet")
+        parser.add_argument('--pred', '-p', help="Enter the folder path that contains your audio files", type=str)
+        args = parser.parse_args()
+        pred_dir = args.pred
+
+    except ImportError:
+        warnings.warn("Argparse module not installed! passing default arguments.")
+        pred_dir = './Test'
+    predict(prediction_dir=pred_dir)
+    print("Urban sounds classification Prediction DONE!")

example/gluon/urban_sounds/train.py

@@ -0,0 +1,165 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""The module to run training on the Urban sounds dataset
+"""
+import os
+import time
+import warnings
+import mxnet as mx
+from mxnet import gluon, nd, autograd
+from mxnet.gluon.contrib.data.audio.datasets import AudioFolderDataset
+from mxnet.gluon.contrib.data.audio.transforms import MFCC
+import model
+
+def evaluate_accuracy(data_iterator, net):
+    """Function to evaluate accuracy of any data iterator passed to it as an argument"""
+    acc = mx.metric.Accuracy()
+    for _, (data, label) in enumerate(data_iterator):
+        output = net(data)
+        predictions = nd.argmax(output, axis=1)
+        predictions = predictions.reshape((-1, 1))
+        acc.update(preds=predictions, labels=label)
+    return acc.get()[1]
+
+
+def train(train_dir=None, train_csv=None, epochs=30, batch_size=32):
+    """The function responsible for running the training the model."""
+    try:
+        import librosa
+    except ImportError:
+        warnings.warn("The dependency librosa is not installed. Cannot continue")
+        return
+    if not train_dir or not os.path.exists(train_dir) or not train_csv:
+        warnings.warn("No train directory could be found ")
+        return
+    # Make a dataset from the local folder containing Audio data
+    print("\nMaking an Audio Dataset...\n")
+    tick = time.time()
+    aud_dataset = AudioFolderDataset(train_dir, train_csv=train_csv, file_format='.wav', skip_rows=1)
+    tock = time.time()
+
+    print("Loading the dataset took ", (tock-tick), " seconds.")
+    print("\n=======================================\n")
+    print("Number of output classes = ", len(aud_dataset.synsets))
+    print("\nThe labels are : \n")
+    print(aud_dataset.synsets)
+    # Get the model to train
+    net = model.get_net(len(aud_dataset.synsets))
+    print("\nNeural Network = \n")
+    print(net)
+    print("\nModel - Neural Network Generated!\n")
+    print("=======================================\n")
+
+    #Define the loss - Softmax CE Loss
+    softmax_loss = gluon.loss.SoftmaxCELoss(from_logits=False, sparse_label=True)
+    print("Loss function initialized!\n")
+    print("=======================================\n")
+
+    #Define the trainer with the optimizer
+    trainer = gluon.Trainer(net.collect_params(), 'adadelta')
+    print("Optimizer - Trainer function initialized!\n")
+    print("=======================================\n")
+    print("Loading the dataset to the Gluon's OOTB Dataloader...")
+
+    #Getting the data loader out of the AudioDataset and passing the transform
+    aud_transform = MFCC()
+    tick = time.time()
+
+    audio_train_loader = gluon.data.DataLoader(aud_dataset.transform_first(aud_transform), batch_size=32, shuffle=True)
+    tock = time.time()
+    print("Time taken to load data and apply transform here is ", (tock-tick), " seconds.")
+    print("=======================================\n")
+
+
+    print("Starting the training....\n")
+    # Training loop
+    tick = time.time()
+    batch_size = batch_size
+    num_examples = len(aud_dataset)
+
+    for e in range(epochs):
+        cumulative_loss = 0
+        for _, (data, label) in enumerate(audio_train_loader):
+            with autograd.record():
+                output = net(data)
+                loss = softmax_loss(output, label)
+            loss.backward()
+
+            trainer.step(batch_size)
+            cumulative_loss += mx.nd.sum(loss).asscalar()
+
+        if e%5 == 0:
+            train_accuracy = evaluate_accuracy(audio_train_loader, net)
+            print("Epoch %s. Loss: %s Train accuracy : %s " % (e, cumulative_loss/num_examples, train_accuracy))
+            print("\n------------------------------\n")
+
+    train_accuracy = evaluate_accuracy(audio_train_loader, net)
+    tock = time.time()
+    print("\nFinal training accuracy: ", train_accuracy)
+
+    print("Training the sound classification for ", epochs, " epochs, MLP model took ", (tock-tick), " seconds")
+    print("====================== END ======================\n")
+
+    print("Trying to save the model parameters here...")
+    net.save_parameters("./net.params")
+    print("Saved the model parameters in current directory.")
+
+
+if __name__ == '__main__':
+
+    try:
+        import argparse
+        parser = argparse.ArgumentParser(description="Urban Sounds clsssification example - MXNet")
+        parser.add_argument('--train', '-t', help="Enter the folder path that contains your audio files", type=str)
+        parser.add_argument('--csv', '-c', help="Enter the filename of the csv that contains filename\
+        to label mapping", type=str)
+        parser.add_argument('--epochs', '-e', help="Enter the number of epochs \
+        you would want to run the training for.", type=int)
+        parser.add_argument('--batch_size', '-b', help="Enter the batch_size of data", type=int)
+        args = parser.parse_args()
+
+        if args:
+            if args.train:
+                training_dir = args.train
+            else:
+                training_dir = './Train'
+
+            if args.csv:
+                training_csv = args.csv
+            else:
+                training_csv = './train.csv'
+
+            if args.epochs:
+                eps = args.epochs
+            else:
+                eps = 30
+
+            if args.batch_size:
+                batch_sz = args.batch_size
+            else:
+                batch_sz = 32
+
+    except ImportError as er:
+        warnings.warn("Argument parsing module could not be imported \
+        Passing default arguments.")
+        training_dir = './Train'
+        training_csv = './train.csv'
+        eps = 30
+        batch_sz = 32
+
+    train(train_dir=training_dir, train_csv=training_csv, epochs=eps, batch_size=batch_sz)
+    print("Urban sounds classification Training DONE!")

python/mxnet/gluon/contrib/data/__init__.py

@@ -22,3 +22,5 @@
 from . import text
 
 from .sampler import *
+
+from . import audio

python/mxnet/gluon/contrib/data/audio/__init__.py

@@ -0,0 +1,20 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+# coding: utf-8
+# pylint: disable=wildcard-import
+"""Audio utilities."""