mne. filter_data, notch_filter\nX2 = X[[,,,, ,,,, ,,,, ,,,]] - X[[,,,, ,,,, ,,,, ,,,]]\nX2 = notch_filter(X2, , , n_jobs=-, verbose=)\nX2 = filter_data(X2, , , , n_jobs=-, verbose=) \n==> to kHz\n==> Frequencies to notch Hz = ?? ( idea about used )\n==> After Difference, applied notch frequency filters ??\n (nn.KLDivLoss):\n ():\n (WeightedKLDivWithLogitsLoss, self).__init__(size_average=, reduce=, reduction=)\n self.register_buffer(, weight)\n\n ():\n \n \n batch_size = .size()\n log_prob = F.log_softmax(, )\n element_loss = (WeightedKLDivWithLogitsLoss, self).forward(log_prob, target)\n\n sample_loss = torch.(element_loss, dim=)\n sample_weight = torch.(target * self.weight, dim=)\n\n weighted_loss = sample_loss*sample_weight\n \n avg_loss = torch.(weighted_loss) / batch_size\n\n avg_loss \noptimizer = optim.Adam(model_cnn.parameters(), lr=*, betas = (,),eps = *, weight_decay=*)\n==> lr = *\n==> eps = *\n==> betas = (, )\nbatch_size = \n (nn.Sequential):\n ():\n (_DenseLayer, self).__init__()\n batch_norm:\n self.add_module(, nn.BatchNorm1d(num_input_features)),\n \n self.add_module(, nn.ELU()),\n self.add_module(, nn.Conv1d(num_input_features, bn_size * growth_rate, kernel_size=, stride=, bias=conv_bias)),\n batch_norm:\n self.add_module(, nn.BatchNorm1d(bn_size * growth_rate)),\n \n self.add_module(, nn.ELU()),\n self.add_module(, nn.Conv1d(bn_size * growth_rate, growth_rate, kernel_size=, stride=, padding=, bias=conv_bias)),\n \n self.drop_rate = drop_rate\n\n ():\n \n \n new_features = (_DenseLayer, self).forward(x)\n \n \n self.drop_rate > :\n new_features = F.dropout(new_features, p=self.drop_rate, training=self.training)\n torch.cat([x, new_features], )\n\n\n (nn.Sequential):\n ():\n (_DenseBlock, self).__init__()\n i (num_layers):\n layer = _DenseLayer(num_input_features + i * growth_rate, growth_rate, bn_size, drop_rate, conv_bias, batch_norm)\n self.add_module( % (i + ), layer)\n\n\n (nn.Sequential):\n ():\n (_Transition, self).__init__()\n batch_norm:\n self.add_module(, nn.BatchNorm1d(num_input_features))\n \n self.add_module(, nn.ELU())\n self.add_module(, nn.Conv1d(num_input_features, num_output_features, kernel_size=, stride=, bias=conv_bias))\n self.add_module(, nn.AvgPool1d(kernel_size=, stride=))\n\n\n (nn.Module):\n ():\n\n (DenseNetEnconder, self).__init__()\n\n \n first_conv = OrderedDict([(, nn.Conv1d(in_channels, num_init_features, kernel_size=, stride=, padding=, bias=conv_bias))])\n \n \n\n \n \n \n \n\n batch_norm:\n first_conv[] = nn.BatchNorm1d(num_init_features)\n \n first_conv[] = nn.ELU()\n first_conv[] = nn.MaxPool1d(kernel_size=, stride=, padding=)\n\n self.densenet = nn.Sequential(first_conv)\n\n num_features = num_init_features\n i, num_layers (block_config):\n block = _DenseBlock(num_layers=num_layers, num_input_features=num_features,\n bn_size=bn_size, growth_rate=growth_rate, drop_rate=drop_rate, conv_bias=conv_bias, batch_norm=batch_norm)\n self.densenet.add_module( % (i + ), block)\n num_features = num_features + num_layers * growth_rate\n i != (block_config) - :\n trans = _Transition(num_input_features=num_features, num_output_features=num_features // , conv_bias=conv_bias, batch_norm=batch_norm)\n self.densenet.add_module( % (i + ), trans)\n num_features = num_features // \n\n \n batch_norm:\n self.densenet.add_module(.((block_config) + ), nn.BatchNorm1d(num_features))\n \n\n self.densenet.add_module(.((block_config) + ), nn.ReLU())\n self.densenet.add_module(.((block_config) + ), nn.AvgPool1d(kernel_size=, stride=)) \n\n self.num_features = num_features\n\n \n m self.modules():\n (m, nn.Conv1d):\n nn.init.kaiming_normal_(m.weight.data)\n (m, nn.BatchNorm1d):\n m.weight.data.fill_()\n m.bias.data.zero_()\n (m, nn.Linear):\n m.bias.data.zero_()\n\n ():\n features = self.densenet(x)\n \n \n features.view(features.size(), -)\n\n\n (nn.Module):\n \n \n ():\n\n (DenseNetClassifier, self).__init__()\n\n self.features = DenseNetEnconder(growth_rate=growth_rate, block_config=block_config, in_channels=in_channels,\n num_init_features=num_init_features, bn_size=bn_size, drop_rate=drop_rate,\n conv_bias=conv_bias, batch_norm=batch_norm)\n\n \n self.classifier = nn.Sequential(\n nn.Dropout(p=drop_fc),\n nn.Linear(self.features.num_features, num_classes)\n )\n\n \n m self.modules():\n (m, nn.Conv1d):\n nn.init.kaiming_normal_(m.weight.data)\n (m, nn.BatchNorm1d):\n m.weight.data.fill_()\n m.bias.data.zero_()\n (m, nn.Linear):\n m.bias.data.zero_()\n\n ():\n features = self.features(x)\n out = self.classifier(features)\n out, features\n mne. filter_data, notch_filter\nX2 = X[[,,,, ,,,, ,,,, ,,,]] - X[[,,,, ,,,, ,,,, ,,,]]\nX2 = notch_filter(X2, , , n_jobs=-, verbose=)\nX2 = filter_data(X2, , , , n_jobs=-, verbose=) \n==> to kHz\n==> Frequencies to notch Hz = ?? ( idea about used )\n==> After Difference, applied notch frequency filters ??\n (nn.KLDivLoss):\n ():\n (WeightedKLDivWithLogitsLoss, self).__init__(size_average=, reduce=, reduction=)\n self.register_buffer(, weight)\n\n ():\n \n \n batch_size = .size()\n log_prob = F.log_softmax(, )\n element_loss = (WeightedKLDivWithLogitsLoss, self).forward(log_prob, target)\n\n sample_loss = torch.(element_loss, dim=)\n sample_weight = torch.(target * self.weight, dim=)\n\n weighted_loss = sample_loss*sample_weight\n \n avg_loss = torch.(weighted_loss) / batch_size\n\n avg_loss \noptimizer = optim.Adam(model_cnn.parameters(), lr=*, betas = (,),eps = *, weight_decay=*)\n==> lr = *\n==> eps = *\n==> betas = (, )\nbatch_size = \n (nn.Sequential):\n ():\n (_DenseLayer, self).__init__()\n batch_norm:\n self.add_module(, nn.BatchNorm1d(num_input_features)),\n \n self.add_module(, nn.ELU()),\n self.add_module(, nn.Conv1d(num_input_features, bn_size * growth_rate, kernel_size=, stride=, bias=conv_bias)),\n batch_norm:\n self.add_module(, nn.BatchNorm1d(bn_size * growth_rate)),\n \n self.add_module(, nn.ELU()),\n self.add_module(, nn.Conv1d(bn_size * growth_rate, growth_rate, kernel_size=, stride=, padding=, bias=conv_bias)),\n \n self.drop_rate = drop_rate\n\n ():\n \n \n new_features = (_DenseLayer, self).forward(x)\n \n \n self.drop_rate > :\n new_features = F.dropout(new_features, p=self.drop_rate, training=self.training)\n torch.cat([x, new_features], )\n\n\n (nn.Sequential):\n ():\n (_DenseBlock, self).__init__()\n i (num_layers):\n layer = _DenseLayer(num_input_features + i * growth_rate, growth_rate, bn_size, drop_rate, conv_bias, batch_norm)\n self.add_module( % (i + ), layer)\n\n\n (nn.Sequential):\n ():\n (_Transition, self).__init__()\n batch_norm:\n self.add_module(, nn.BatchNorm1d(num_input_features))\n \n self.add_module(, nn.ELU())\n self.add_module(, nn.Conv1d(num_input_features, num_output_features, kernel_size=, stride=, bias=conv_bias))\n self.add_module(, nn.AvgPool1d(kernel_size=, stride=))\n\n\n (nn.Module):\n ():\n\n (DenseNetEnconder, self).__init__()\n\n \n first_conv = OrderedDict([(, nn.Conv1d(in_channels, num_init_features, kernel_size=, stride=, padding=, bias=conv_bias))])\n \n \n\n \n \n \n \n\n batch_norm:\n first_conv[] = nn.BatchNorm1d(num_init_features)\n \n first_conv[] = nn.ELU()\n first_conv[] = nn.MaxPool1d(kernel_size=, stride=, padding=)\n\n self.densenet = nn.Sequential(first_conv)\n\n num_features = num_init_features\n i, num_layers (block_config):\n block = _DenseBlock(num_layers=num_layers, num_input_features=num_features,\n bn_size=bn_size, growth_rate=growth_rate, drop_rate=drop_rate, conv_bias=conv_bias, batch_norm=batch_norm)\n self.densenet.add_module( % (i + ), block)\n num_features = num_features + num_layers * growth_rate\n i != (block_config) - :\n trans = _Transition(num_input_features=num_features, num_output_features=num_features // , conv_bias=conv_bias, batch_norm=batch_norm)\n self.densenet.add_module( % (i + ), trans)\n num_features = num_features // \n\n \n batch_norm:\n self.densenet.add_module(.((block_config) + ), nn.BatchNorm1d(num_features))\n \n\n self.densenet.add_module(.((block_config) + ), nn.ReLU())\n self.densenet.add_module(.((block_config) + ), nn.AvgPool1d(kernel_size=, stride=)) \n\n self.num_features = num_features\n\n \n m self.modules():\n (m, nn.Conv1d):\n nn.init.kaiming_normal_(m.weight.data)\n (m, nn.BatchNorm1d):\n m.weight.data.fill_()\n m.bias.data.zero_()\n (m, nn.Linear):\n m.bias.data.zero_()\n\n ():\n features = self.densenet(x)\n \n \n features.view(features.size(), -)\n\n\n (nn.Module):\n \n \n ():\n\n (DenseNetClassifier, self).__init__()\n\n self.features = DenseNetEnconder(growth_rate=growth_rate, block_config=block_config, in_channels=in_channels,\n num_init_features=num_init_features, bn_size=bn_size, drop_rate=drop_rate,\n conv_bias=conv_bias, batch_norm=batch_norm)\n\n \n self.classifier = nn.Sequential(\n nn.Dropout(p=drop_fc),\n nn.Linear(self.features.num_features, num_classes)\n )\n\n \n m self.modules():\n (m, nn.Conv1d):\n nn.init.kaiming_normal_(m.weight.data)\n (m, nn.BatchNorm1d):\n m.weight.data.fill_()\n m.bias.data.zero_()\n (m, nn.Linear):\n m.bias.data.zero_()\n\n ():\n features = self.features(x)\n out = self.classifier(features)\n out, features\n@seshurajup wow! you are quick. Upvoted!
", "rawMarkdown": "seshurajup wow! you are quick. Upvoted!", "votes": null }, { "id": "2623980", "postDate": "01/28/2024 14:05:14", "content": "Some study material for the future 😋
\nupvoted
", "rawMarkdown": "Some study material for the future 😋\n\nupvoted", "votes": null }, { "id": "2624003", "postDate": "01/28/2024 14:26:48", "content": "@renyiwei i like this paper (exact match with competition data), so informative and now trying to train the code.
", "rawMarkdown": "renyiwei i like this paper (exact match with competition data), so informative and now trying to train the code.", "votes": null }, { "id": "2624745", "postDate": "01/29/2024 00:15:11", "content": "Regarding:
\nX2 = notch_filter(X2, 200, 60, =-1, =)\nX2 = filter_data(X2, 200, 0.5, 40, =-1, =) \nThe notch filter is at 60 Hz which is the AC \"Utility frequency\"; wiki says: \"in large parts of the world this is 50 Hz, although in the Americas and parts of Asia it is typically 60 Hz.\" This signal can be relatively strong, even after differencing the channels.
\nThe next filter selects the band of frequencies from 0.5 to 40 Hz; the 0.5 highpass gets rid of slow drifts in the waveform.
@seshurajup wow! you are quick. Upvoted!
", "votes": null, "replies": [ { "id": 2624003, "author_name": "seshurajup", "author_url": "", "post_date": "01/28/2024 14:26:48", "content": "@renyiwei i like this paper (exact match with competition data), so informative and now trying to train the code.
", "votes": null, "replies": [] } ] }, { "id": 2623980, "author_name": "gabrielfreddi", "author_url": "", "post_date": "01/28/2024 14:05:14", "content": "Some study material for the future 😋
\nupvoted
", "votes": null, "replies": [] }, { "id": 2624745, "author_name": "dan3dewey", "author_url": "", "post_date": "01/29/2024 00:15:11", "content": "Regarding:
\nX2 = notch_filter(X2, 200, 60, =-1, =)\nX2 = filter_data(X2, 200, 0.5, 40, =-1, =) \nThe notch filter is at 60 Hz which is the AC \"Utility frequency\"; wiki says: \"in large parts of the world this is 50 Hz, although in the Americas and parts of Asia it is typically 60 Hz.\" This signal can be relatively strong, even after differencing the channels.
\nThe next filter selects the band of frequencies from 0.5 to 40 Hz; the 0.5 highpass gets rid of slow drifts in the waveform.