add audio set, change normalization

dcase2020_task2/data_sets/audio_set.py

@@ -6,7 +6,6 @@ from dcase2020_task2.data_sets import BaseDataSet, CLASS_MAP, INVERSE_CLASS_MAP,
 import librosa
 import numpy as np
 from dcase2020_task2.data_sets import MCMDataSet
-import pickle
 
 
 class AudioSet(BaseDataSet):
@@ -31,6 +30,7 @@ class AudioSet(BaseDataSet):
         self.power = power
         self.fmin = fmin
         self.hop_all = hop_all
+        self.normalize_raw = normalize_raw
 
         kwargs = {
             'data_root': self.data_root,
@@ -38,13 +38,13 @@ class AudioSet(BaseDataSet):
             'num_mel': self.num_mel,
             'n_fft': self.n_fft,
             'hop_size': self.hop_size,
-            'power': power,
-            'normalize': normalize_raw,
-            'fmin': fmin,
-            'hop_all': hop_all
+            'power': self.power,
+            'normalize': self.normalize_raw,
+            'fmin': self.fmin,
+            'hop_all': self.hop_all
         }
 
-        class_names = sorted([class_name for class_name in os.listdir(data_root) if os.path.isdir(os.path.join(data_root, class_name))])[:30]
+        class_names = sorted([class_name for class_name in os.listdir(data_root) if os.path.isdir(os.path.join(data_root, class_name))])
 
         training_sets = []
         data_arrays = []
@@ -88,6 +88,7 @@ class AudioSetClassSubset(torch.utils.data.Dataset):
             normalize=True,
             fmin=0,
             hop_all=False,
+            max_file_per_class=2,
             max_file_length=350
     ):
 
@@ -101,13 +102,14 @@ class AudioSetClassSubset(torch.utils.data.Dataset):
         self.fmin = fmin
         self.hop_all = hop_all
         self.class_name = class_name
+        self.max_file_per_class = max_file_per_class
         self.max_file_length = max_file_length
 
         files = glob.glob(os.path.join(data_root, class_name, '*.wav'))
 
         assert len(files) > 0
 
-        files = sorted(files)
+        files = sorted(files)[:max_file_per_class]
         self.files = files
 
         self.meta_data = self.__load_meta_data__(files)
@@ -171,6 +173,7 @@ class AudioSetClassSubset(torch.utils.data.Dataset):
         x, sr = librosa.load(file, sr=16000, mono=True)
         if len(x) > (self.max_file_length + 1 * self.hop_size) + self.n_fft:
             x = x[:(self.max_file_length + 1) * self.hop_size + self.n_fft]
+
         if self.normalize:
             x = (x - x.mean()) / x.std()
 
@@ -201,6 +204,7 @@ class AudioSetClassSubset(torch.utils.data.Dataset):
             'file_ids': os.sep.join(os.path.normpath(file_path).split(os.sep)[-4:])
         }
 
+
 if __name__ == '__main__':
     mcmc = MCMDataSet(0, 0)
     a = audio_set = AudioSet(

dcase2020_task2/data_sets/complement_dataset.py

@@ -4,6 +4,16 @@ from dcase2020_task2.data_sets import BaseDataSet, CLASS_MAP, INVERSE_CLASS_MAP,
 from dcase2020_task2.data_sets import MachineDataSet
 import numpy as np
 
+valid_types = {
+    0: [1, 2, 5],
+    1: [0, 2, 5],
+    2: [0, 1, 5],
+    5: [0, 1, 2],
+    3: [4],
+    4: [3],
+}
+
+
 class ComplementMCMDataSet(BaseDataSet):
 
     def __init__(
@@ -18,9 +28,11 @@ class ComplementMCMDataSet(BaseDataSet):
             power=1.0,
             fmin=0,
             normalize_raw=False,
-            normalize=None,
             hop_all=False
     ):
+
+        assert type(machine_type) == int and type(machine_id) == int
+
         self.data_root = data_root
         self.context = context
         self.num_mel = num_mel
@@ -28,10 +40,8 @@ class ComplementMCMDataSet(BaseDataSet):
         self.hop_size = hop_size
         self.power = power
         self.fmin = fmin
-        self.normalize = normalize
         self.hop_all = hop_all
-
-        assert type(machine_type) == int and type(machine_id) == int
+        self.normalize_raw = normalize_raw
 
         kwargs = {
             'data_root': self.data_root,
@@ -39,77 +49,28 @@ class ComplementMCMDataSet(BaseDataSet):
             'num_mel': self.num_mel,
             'n_fft': self.n_fft,
             'hop_size': self.hop_size,
-            'power': power,
-            'normalize': normalize_raw,
-            'fmin': fmin,
-            'hop_all': hop_all
+            'power': self.power,
+            'normalize': self.normalize_raw,
+            'fmin': self.fmin,
+            'hop_all': self.hop_all
         }
 
-        if machine_id == -1:
-            training_sets = []
-            validation_sets = []
-            data = []
-            for id_ in ALL_ID_MAP[machine_type]:
-                training_sets.append(MachineDataSet(machine_type, id_, mode='training', **kwargs))
-                validation_sets.append(MachineDataSet(machine_type, id_, mode='validation', **kwargs))
-                data.append(training_sets[-1].data)
-
-            if normalize is None:
-                data = np.concatenate(data, axis=-1)
-                mean = data.mean(axis=1, keepdims=True)
-                std = data.std(axis=1, keepdims=True)
-            else:
-                assert type(normalize) == tuple
-                assert len(normalize) == 2
-                mean, std = normalize
-
-            for training_set, validation_set in zip(training_sets, validation_sets):
-                training_set.data = (training_set.data - mean) / std
-                validation_set.data = (validation_set.data - mean) / std
-
-            del data
-        else:
-            training_set = MachineDataSet(machine_type, machine_id, mode='training', **kwargs)
-            validation_set = MachineDataSet(machine_type, machine_id, mode='validation', **kwargs)
-            if normalize is None:
-                mean = training_set.data.mean(axis=1, keepdims=True)
-                std = training_set.data.std(axis=1, keepdims=True)
-                training_set.data = (training_set.data - mean) / std
-                validation_set.data = (validation_set.data - mean) / std
-            else:
-                assert type(normalize) == tuple
-                assert len(normalize) == 2
-                mean, std = normalize
-                training_set.data = (training_set.data - mean) / std
-                validation_set.data = (validation_set.data - mean) / std
-
         training_sets = []
-        # validation_sets = []
-        valid_types = {
-            0: [1, 2, 5],
-            1: [0, 2, 5],
-            2: [0, 1, 5],
-            5: [0, 1, 2],
-            3: [4],
-            4: [3],
-        }
+
+        data = []
         for type_ in ALL_ID_MAP:
-            if type_ in valid_types[machine_type]:
-                for id_ in ALL_ID_MAP[type_]:
-                    #if type_ != machine_type: #or (id_ != machine_id and machine_id != -1):
+            for id_ in ALL_ID_MAP[type_]:
+                if type_ != machine_type or (id_ != machine_id and machine_id != -1):
                     t = MachineDataSet(type_, id_, mode='training', **kwargs)
-                    t.data = (t.data - mean) / std
+                    data.append(t.data)
                     training_sets.append(t)
+        data = np.concatenate(data, axis=-1)
 
-                    # don't load validation set ...
-                    # v = MachineDataSet(type_, id_, mode='validation', **kwargs)
-                    # v.data = (v.data - mean) / std
-                    # validation_sets.append(v)
+        self.mean = data.mean(axis=1, keepdims=True)
+        self.std = data.std(axis=1, keepdims=True)
+        del data
 
         self.training_set = torch.utils.data.ConcatDataset(training_sets)
-        # self.validation_set = torch.utils.data.ConcatDataset(validation_sets)
-        self.mean = mean
-        self.std = std
 
     @property
     def observation_shape(self) -> tuple:

dcase2020_task2/data_sets/mcm_dataset.py

@@ -31,6 +31,7 @@ class MCMDataSet(BaseDataSet):
         self.power = power
         self.fmin = fmin
         self.hop_all = hop_all
+        self.normalize_raw = normalize_raw
 
         assert type(machine_type) == int and type(machine_id) == int
 
@@ -40,10 +41,10 @@ class MCMDataSet(BaseDataSet):
             'num_mel': self.num_mel,
             'n_fft': self.n_fft,
             'hop_size': self.hop_size,
-            'power': power,
-            'normalize': normalize_raw,
-            'fmin': fmin,
-            'hop_all': hop_all
+            'power': self.power,
+            'normalize': self.normalize_raw,
+            'fmin': self.fmin,
+            'hop_all': self.hop_all
         }
 
         if machine_id == -1:

dcase2020_task2/experiments/classification_experiment.py

@@ -11,7 +11,8 @@ from sacred import SETTINGS
 SETTINGS['CAPTURE_MODE'] = 'sys'
 from datetime import datetime
 
-from dcase2020_task2.data_sets import AudioSet
+from dcase2020_task2.data_sets import AudioSet, ComplementMCMDataSet
+
 
 class ClassificationExperiment(BaseExperiment, pl.LightningModule):
 
@@ -31,12 +32,25 @@ class ClassificationExperiment(BaseExperiment, pl.LightningModule):
         # will be set before each epoch
         self.normal_data_set = self.objects['data_set']
 
-        self.abnormal_data_set = AudioSet(
+        self.abnormal_data_set = ComplementMCMDataSet(
+            self.objects['machine_type'],
+            self.objects['machine_id'],
             **self.objects['fetaure_settings']
         )
 
-        self.mean = torch.from_numpy((self.normal_data_set.mean + self.abnormal_data_set.mean) / 2)
-        self.std = torch.from_numpy((self.normal_data_set.std + self.abnormal_data_set.std) / 2)
+        if self.objects.get('normalize') == 'normal':
+            self.mean = torch.from_numpy(self.normal_data_set.mean)
+            self.std = torch.from_numpy(self.normal_data_set.std)
+        elif self.objects.get('normalize') == 'abnormal':
+            self.mean = torch.from_numpy(self.abnormal_data_set.mean)
+            self.std = torch.from_numpy(self.abnormal_data_set.std)
+        elif self.objects.get('normalize') == 'average':
+            self.mean = torch.from_numpy((self.normal_data_set.mean + self.abnormal_data_set.mean) / 2)
+            self.std = torch.from_numpy((self.normal_data_set.std + self.abnormal_data_set.std) / 2)
+        else:
+            print('No normalization.')
+            self.mean = torch.zeros(self.normal_data_set.mean.shape)
+            self.std = torch.ones(self.normal_data_set.std.shape)
 
         self.inf_data_loader = self.get_inf_data_loader(
             torch.utils.data.DataLoader(
@@ -165,8 +179,8 @@ def configuration():
     context = 32
 
     model_class = 'dcase2020_task2.models.CNN'
-    hidden_size = 64
-    num_hidden = 3
+    hidden_size = 256
+    num_hidden = 4
     dropout_probability = 0.0
 
     epochs = 100
@@ -176,7 +190,6 @@ def configuration():
     if debug:
         num_workers = 0
         epochs = 100
-        hop_all = True
     else:
         num_workers = 4
 
@@ -186,6 +199,7 @@ def configuration():
     weight_decay = 0
 
     normalize_raw = True
+    normalize = None
 
     # TODO: change default descriptor
     descriptor = "ClassificationExperiment_Model:[{}_{}_{}_{}]_Training:[{}_{}_{}_{}]_Features:[{}_{}_{}_{}_{}_{}_{}]_{}".format(
@@ -247,6 +261,7 @@ def configuration():
         'kwargs': {
             'hidden_size': hidden_size,
             'num_hidden': num_hidden,
+            'base_channels': hidden_size,
             'dropout_probability': dropout_probability,
             'batch_norm': False
         }

dcase2020_task2/models/classifier.py

@@ -17,7 +17,7 @@ class FCNN(torch.nn.Module):
             num_outputs=1,
             activation='relu',
             batch_norm=False,
-            dropout_probability=0.1
+            **kwargs
     ):
         super().__init__()
 
@@ -30,11 +30,9 @@ class FCNN(torch.nn.Module):
             layers.append(torch.nn.Linear(i, o))
             if batch_norm:
                 layers.append(torch.nn.BatchNorm1d(o))
-            layers.append(torch.nn.Dropout(p=dropout_probability))
             layers.append(activation_fn())
 
         _ = layers.pop()
-        _ = layers.pop()
 
         if batch_norm:
             _ = layers.pop()
@@ -59,7 +57,7 @@ class CNN(torch.nn.Module):
             num_outputs=1,
             activation='relu',
             batch_norm=False,
-            dropout_probability=0.1
+            **kwargs
     ):
         super().__init__()
 
@@ -72,8 +70,6 @@ class CNN(torch.nn.Module):
             layers.append(torch.nn.Conv2d(i, o, kernel_size=(3, 3), stride=2, padding=(1, 1)))
             if batch_norm:
                 layers.append(torch.nn.BatchNorm2d(o))
-            if dropout_probability > 0:
-                layers.append(torch.nn.Dropout2d(dropout_probability))
             layers.append(activation_fn())
 
         _ = layers.pop()

dcase2020_task2/models/cp_resnet.py

+# coding: utf-8
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.utils.checkpoint import checkpoint_sequential
+
+from librosa.filters import mel as librosa_mel_fn
+
+
+def initialize_weights(module):
+    if isinstance(module, nn.Conv2d):
+        nn.init.kaiming_normal_(module.weight.data, mode='fan_in', nonlinearity="relu")
+        # nn.init.kaiming_normal_(module.weight.data, mode='fan_out')
+    elif isinstance(module, nn.BatchNorm2d):
+        module.weight.data.fill_(1)
+        module.bias.data.zero_()
+    elif isinstance(module, nn.Linear):
+        module.bias.data.zero_()
+
+
+layer_index_total = 0
+
+
+def initialize_weights_fixup(module):
+    if isinstance(module, AttentionAvg):
+        print("AttentionAvg init..")
+        module.forw_conv[0].weight.data.zero_()
+        module.atten[0].bias.data.zero_()
+        nn.init.kaiming_normal_(module.atten[0].weight.data, mode='fan_in', nonlinearity="sigmoid")
+    if isinstance(module, BasicBlock):
+        # He init, rescaled by Fixup multiplier
+        b = module
+        n = b.conv1.kernel_size[0] * b.conv1.kernel_size[1] * b.conv1.out_channels
+        print(b.layer_index, math.sqrt(2. / n), layer_index_total ** (-0.5))
+        b.conv1.weight.data.normal_(0, (layer_index_total ** (-0.5)) * math.sqrt(2. / n))
+        b.conv2.weight.data.zero_()
+        if b.shortcut._modules.get('conv') is not None:
+            convShortcut = b.shortcut._modules.get('conv')
+            n = convShortcut.kernel_size[0] * convShortcut.kernel_size[1] * convShortcut.out_channels
+            convShortcut.weight.data.normal_(0, math.sqrt(2. / n))
+    if isinstance(module, nn.Conv2d):
+        pass
+        # nn.init.kaiming_normal_(module.weight.data, mode='fan_in', nonlinearity="relu")
+        # nn.init.kaiming_normal_(module.weight.data, mode='fan_out')
+    elif isinstance(module, nn.BatchNorm2d):
+        module.weight.data.fill_(1)
+        module.bias.data.zero_()
+    elif isinstance(module, nn.Linear):
+        module.bias.data.zero_()
+
+
+first_RUN = True
+
+
+def calc_padding(kernal):
+    try:
+        return kernal // 3
+    except TypeError:
+        return [k // 3 for k in kernal]
+
+
+class AttentionAvg(nn.Module):
+
+    def __init__(self, in_channels, out_channels, sum_all=True):
+        super(AttentionAvg, self).__init__()
+        self.sum_dims = [2, 3]
+        if sum_all:
+            self.sum_dims = [1, 2, 3]
+        self.forw_conv = nn.Sequential(
+            nn.Conv2d(
+                in_channels,
+                out_channels,
+                kernel_size=1,
+                stride=1,
+                padding=0,
+                bias=False)
+        )
+        self.atten = nn.Sequential(
+            nn.Conv2d(
+                in_channels,
+                out_channels,
+                kernel_size=1,
+                stride=1,
+                padding=0,
+                bias=True),
+            nn.Sigmoid()
+        )
+
+    def forward(self, x):
+        a1 = self.forw_conv(x)
+        atten = self.atten(x)
+        num = atten.size(2) * atten.size(3)
+        asum = atten.sum(dim=self.sum_dims, keepdim=True) + 1e-8
+        return a1 * atten * num / asum
+
+
+class BasicBlock(nn.Module):
+    expansion = 1
+
+    def __init__(self, in_channels, out_channels, stride, k1=3, k2=3):
+        super(BasicBlock, self).__init__()
+        global layer_index_total
+        self.layer_index = layer_index_total
+        layer_index_total = layer_index_total + 1
+        self.conv1 = nn.Conv2d(
+            in_channels,
+            out_channels,
+            kernel_size=k1,
+            stride=stride,  # downsample with first conv
+            padding=calc_padding(k1),
+            bias=False
+        )
+        self.conv2 = nn.Conv2d(
+            out_channels,
+            out_channels,
+            kernel_size=k2,
+            stride=1,
+            padding=calc_padding(k2),
+            bias=False
+        )
+
+        self.shortcut = nn.Sequential()
+        if in_channels != out_channels:
+            self.shortcut.add_module(
+                'conv',
+                nn.Conv2d(
+                    in_channels,
+                    out_channels,
+                    kernel_size=1,
+                    stride=stride,  # downsample
+                    padding=0,
+                    bias=False)
+            )
+
+    def forward(self, x):
+        y = F.relu(self.conv1(x), inplace=True)
+        y = self.conv2(y)
+        y += self.shortcut(x)
+        y = F.relu(y, inplace=True)  # apply ReLU after addition
+        return y
+
+
+class BottleneckBlock(nn.Module):
+    expansion = 4
+
+    def __init__(self, in_channels, out_channels, stride):
+        super(BottleneckBlock, self).__init__()
+
+        bottleneck_channels = out_channels // self.expansion
+
+        self.conv1 = nn.Conv2d(
+            in_channels,
+            bottleneck_channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=False)
+
+        self.conv2 = nn.Conv2d(
+            bottleneck_channels,
+            bottleneck_channels,
+            kernel_size=3,
+            stride=stride,  # downsample with 3x3 conv
+            padding=1,
+            bias=False)
+
+        self.conv3 = nn.Conv2d(
+            bottleneck_channels,
+            out_channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=False)
+
+        self.shortcut = nn.Sequential()  # identity
+        if in_channels != out_channels:
+            self.shortcut.add_module(
+                'conv',
+                nn.Conv2d(
+                    in_channels,
+                    out_channels,
+                    kernel_size=1,
+                    stride=stride,  # downsample
+                    padding=0,
+                    bias=False))
+
+    def forward(self, x):
+        y = F.relu(self.conv1(x), inplace=True)
+        y = F.relu(self.conv2(y), inplace=True)
+        y = self.conv3(y)  # not apply ReLU
+        y += self.shortcut(x)
+        y = F.relu(y, inplace=True)  # apply ReLU after addition
+        return y
+
+
+class Network(nn.Module):
+    def __init__(self, config):
+        super(Network, self).__init__()
+
+        input_shape = config['input_shape']
+        n_classes = config['n_classes']
+
+        base_channels = config['base_channels']
+        block_type = config['block_type']
+        depth = config['depth']
+        self.pooling_padding = config.get("pooling_padding", 0) or 0
+        self.use_raw_spectograms = config.get("use_raw_spectograms") or False
+        self.apply_softmax = config.get("apply_softmax") or False
+
+        assert block_type in ['basic', 'bottleneck']
+        if self.use_raw_spectograms:
+            mel_basis = librosa_mel_fn(
+                22050, 2048, 256)
+            mel_basis = torch.from_numpy(mel_basis).float()
+            self.register_buffer('mel_basis', mel_basis)
+        if block_type == 'basic':
+            block = BasicBlock
+            n_blocks_per_stage = (depth - 2) // 6
+            assert n_blocks_per_stage * 6 + 2 == depth
+        else:
+            block = BottleneckBlock
+            n_blocks_per_stage = (depth - 2) // 9
+            assert n_blocks_per_stage * 9 + 2 == depth
+        n_blocks_per_stage = [n_blocks_per_stage, n_blocks_per_stage, n_blocks_per_stage]
+
+        if config.get("n_blocks_per_stage") is not None:
+            print("n_blocks_per_stage is specified ignoring the depth param, nc=" + str(config.get("n_channels")))
+            n_blocks_per_stage = config.get("n_blocks_per_stage")
+
+        n_channels = config.get("n_channels")
+        if n_channels is None:
+            n_channels = [
+                base_channels,
+                base_channels * 2 * block.expansion,
+                base_channels * 4 * block.expansion
+            ]
+        if config.get("grow_a_lot"):
+            n_channels[2] = base_channels * 8 * block.expansion
+
+        self.in_c = nn.Sequential(nn.Conv2d(
+            input_shape[1],
+            n_channels[0],
+            kernel_size=5,
+            stride=2,
+            padding=1,
+            bias=False),
+            nn.ReLU(True)
+        )
+        self.stage1 = self._make_stage(
+            n_channels[0], n_channels[0], n_blocks_per_stage[0], block, stride=1, maxpool=config['stage1']['maxpool'],
+            k1s=config['stage1']['k1s'], k2s=config['stage1']['k2s'])
+        self.stage2 = self._make_stage(
+            n_channels[0], n_channels[1], n_blocks_per_stage[1], block, stride=1, maxpool=config['stage2']['maxpool'],
+            k1s=config['stage2']['k1s'], k2s=config['stage2']['k2s'])
+        self.stage3 = self._make_stage(