├── Gated Transformer 论文IJCAI版 ├── .idea │ ├── .gitignore │ ├── ADL_Transformer.iml │ ├── inspectionProfiles │ │ └── profiles_settings.xml │ ├── misc.xml │ ├── modules.xml │ └── vcs.xml ├── Gated Transformer 架构图.png ├── dataset_process │ ├── __pycache__ │ │ └── dataset_process.cpython-37.pyc │ └── dataset_process.py ├── font │ └── simsun.ttc ├── gather_figure │ ├── AUSLAN │ │ ├── AUSLAN channel_gather.jpg │ │ └── AUSLAN input_gather.jpg │ ├── ECG │ │ ├── ECG all_sample_gather.jpg │ │ ├── ECG channel_gather.jpg │ │ └── ECG input_gather.jpg │ └── JapaneseVowels │ │ ├── JapaneseVowels Sample_gather.jpg │ │ ├── JapaneseVowels channel_gather.jpg │ │ └── JapaneseVowels input_gather.jpg ├── heatmap_figure_in_test │ └── JapaneseVowels │ │ └── JapaneseVowels accuracy=98.11 0.jpg ├── module │ ├── __pycache__ │ │ ├── encoder.cpython-37.pyc │ │ ├── feedForward.cpython-37.pyc │ │ ├── loss.cpython-37.pyc │ │ ├── multiHeadAttention.cpython-37.pyc │ │ └── transformer.cpython-37.pyc │ ├── encoder.py │ ├── feedForward.py │ ├── loss.py │ ├── multiHeadAttention.py │ └── transformer.py ├── mytest │ ├── DTW_test.py │ ├── HeatMap.py │ ├── HeatMap_DTW.py │ ├── __pycache__ │ │ └── HeatMap_DTW.cpython-37.pyc │ └── kmeans_test.py ├── result_figure │ ├── CharacterTrajectories 96.6% Adam epoch=100 batch=4 lr=0.0001 pe=True mask=True [512,6,6,8,8,0].png │ ├── ECG 86.0% Adam epoch=10 batch=20 lr=0.0001 pe=True mask=True [512,6,6,8,8,0].png │ ├── ECG 86.0% Adam epoch=20 batch=20 lr=0.0001 pe=True mask=True [512,6,6,8,8,0].png │ ├── ECG 87.0% Adam epoch=50 batch=20 lr=0.0001 pe=True mask=True [512,6,6,8,8,0].png │ └── JapaneseVowels 98.11% Adagrad epoch=100 batch=3 lr=0.0001 pe=True mask=True [512,8,8,8,8,0.2].png ├── run.py ├── run_with_saved_model.py └── utils │ ├── TSNE.py │ ├── __pycache__ │ ├── TSNE.cpython-37.pyc │ ├── colorful_line.cpython-37.pyc │ ├── draw_line.cpython-37.pyc │ ├── heatMap.cpython-37.pyc │ ├── random_seed.cpython-37.pyc │ └── visualization.cpython-37.pyc │ ├── colorful_line.py │ ├── draw_line.py │ ├── heatMap.py │ ├── kmeans.py │ ├── random_seed.py │ └── visualization.py └── README.md /Gated Transformer 论文IJCAI版/.idea/.gitignore: -------------------------------------------------------------------------------- 1 | # Default ignored files 2 | /workspace.xml -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/.idea/ADL_Transformer.iml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/.idea/inspectionProfiles/profiles_settings.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 6 | -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/.idea/misc.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/.idea/modules.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/.idea/vcs.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/Gated Transformer 架构图.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/Gated Transformer 架构图.png -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/dataset_process/__pycache__/dataset_process.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/dataset_process/__pycache__/dataset_process.cpython-37.pyc -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/dataset_process/dataset_process.py: -------------------------------------------------------------------------------- 1 | from torch.utils.data import Dataset 2 | import torch 3 | 4 | from scipy.io import loadmat 5 | 6 | 7 | class MyDataset(Dataset): 8 | def __init__(self, 9 | path: str, 10 | dataset: str): 11 | """ 12 | 训练数据集与测试数据集的Dataset对象 13 | :param path: 数据集路径 14 | :param dataset: 区分是获得训练集还是测试集 15 | """ 16 | super(MyDataset, self).__init__() 17 | self.dataset = dataset # 选择获取测试集还是训练集 18 | self.train_len, \ 19 | self.test_len, \ 20 | self.input_len, \ 21 | self.channel_len, \ 22 | self.output_len, \ 23 | self.train_dataset, \ 24 | self.train_label, \ 25 | self.test_dataset, \ 26 | self.test_label, \ 27 | self.max_length_sample_inTest, \ 28 | self.train_dataset_with_no_paddding = self.pre_option(path) 29 | 30 | def __getitem__(self, index): 31 | if self.dataset == 'train': 32 | return self.train_dataset[index], self.train_label[index] - 1 33 | elif self.dataset == 'test': 34 | return self.test_dataset[index], self.test_label[index] - 1 35 | 36 | def __len__(self): 37 | if self.dataset == 'train': 38 | return self.train_len 39 | elif self.dataset == 'test': 40 | return self.test_len 41 | 42 | # 数据预处理 43 | def pre_option(self, path: str): 44 | """ 45 | 数据预处理由于每个样本的时间步维度不同，在此使用最长的时间步作为时间步的维度，使用0进行填充 46 | :param path: 数据集路径 47 | :return: 训练集样本数量，测试集样本数量，时间步维度，通道数，分类数，训练集数据，训练集标签，测试集数据，测试集标签，测试集中时间步最长的样本列表，没有padding的训练集数据 48 | """ 49 | m = loadmat(path) 50 | 51 | # m中是一个字典有4个key 其中最后一个键值对存储的是数据 52 | x1, x2, x3, x4 = m 53 | data = m[x4] 54 | 55 | data00 = data[0][0] 56 | # print('data00.shape', data00.shape) # () data00才到达数据的维度 57 | 58 | index_train = str(data.dtype).find('train\'') 59 | index_trainlabels = str(data.dtype).find('trainlabels') 60 | index_test = str(data.dtype).find('test\'') 61 | index_testlabels = str(data.dtype).find('testlabels') 62 | list = [index_test, index_train, index_testlabels, index_trainlabels] 63 | list = sorted(list) 64 | index_train = list.index(index_train) 65 | index_trainlabels = list.index(index_trainlabels) 66 | index_test = list.index(index_test) 67 | index_testlabels = list.index(index_testlabels) 68 | 69 | # [('trainlabels', 'O'), ('train', 'O'), ('testlabels', 'O'), ('test', 'O')] O 表示数据类型为 numpy.object 70 | train_label = data00[index_trainlabels] 71 | train_data = data00[index_train] 72 | test_label = data00[index_testlabels] 73 | test_data = data00[index_test] 74 | 75 | train_label = train_label.squeeze() 76 | train_data = train_data.squeeze() 77 | test_label = test_label.squeeze() 78 | test_data = test_data.squeeze() 79 | 80 | train_len = train_data.shape[0] 81 | test_len = test_data.shape[0] 82 | output_len = len(tuple(set(train_label))) 83 | 84 | # 时间步最大值 85 | max_lenth = 0 # 93 86 | for item in train_data: 87 | item = torch.as_tensor(item).float() 88 | if item.shape[1] > max_lenth: 89 | max_lenth = item.shape[1] 90 | # max_length_index = train_data.tolist().index(item.tolist()) 91 | 92 | for item in test_data: 93 | item = torch.as_tensor(item).float() 94 | if item.shape[1] > max_lenth: 95 | max_lenth = item.shape[1] 96 | 97 | # 填充Padding 使用0进行填充 98 | # train_data, test_data为numpy.object 类型，不能直接对里面的numpy.ndarray进行处理 99 | train_dataset_with_no_paddding = [] 100 | test_dataset_with_no_paddding = [] 101 | train_dataset = [] 102 | test_dataset = [] 103 | max_length_sample_inTest = [] 104 | for x1 in train_data: 105 | train_dataset_with_no_paddding.append(x1.transpose(-1, -2).tolist()) 106 | x1 = torch.as_tensor(x1).float() 107 | if x1.shape[1] != max_lenth: 108 | padding = torch.zeros(x1.shape[0], max_lenth - x1.shape[1]) 109 | x1 = torch.cat((x1, padding), dim=1) 110 | train_dataset.append(x1) 111 | 112 | for index, x2 in enumerate(test_data): 113 | test_dataset_with_no_paddding.append(x2.transpose(-1, -2).tolist()) 114 | x2 = torch.as_tensor(x2).float() 115 | if x2.shape[1] != max_lenth: 116 | padding = torch.zeros(x2.shape[0], max_lenth - x2.shape[1]) 117 | x2 = torch.cat((x2, padding), dim=1) 118 | else: 119 | max_length_sample_inTest.append(x2.transpose(-1, -2)) 120 | test_dataset.append(x2) 121 | 122 | # 最后维度 [数据条数,时间步数最大值,时间序列维度] 123 | # train_dataset_with_no_paddding = torch.stack(train_dataset_with_no_paddding, dim=0).permute(0, 2, 1) 124 | # test_dataset_with_no_paddding = torch.stack(test_dataset_with_no_paddding, dim=0).permute(0, 2, 1) 125 | train_dataset = torch.stack(train_dataset, dim=0).permute(0, 2, 1) 126 | test_dataset = torch.stack(test_dataset, dim=0).permute(0, 2, 1) 127 | train_label = torch.Tensor(train_label) 128 | test_label = torch.Tensor(test_label) 129 | channel = test_dataset[0].shape[-1] 130 | input = test_dataset[0].shape[-2] 131 | 132 | return train_len, test_len, input, channel, output_len, train_dataset, train_label, test_dataset, test_label, max_length_sample_inTest, train_dataset_with_no_paddding -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/font/simsun.ttc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/font/simsun.ttc -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/gather_figure/AUSLAN/AUSLAN channel_gather.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/gather_figure/AUSLAN/AUSLAN channel_gather.jpg -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/gather_figure/AUSLAN/AUSLAN input_gather.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/gather_figure/AUSLAN/AUSLAN input_gather.jpg -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/gather_figure/ECG/ECG all_sample_gather.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/gather_figure/ECG/ECG all_sample_gather.jpg -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/gather_figure/ECG/ECG channel_gather.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/gather_figure/ECG/ECG channel_gather.jpg -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/gather_figure/ECG/ECG input_gather.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/gather_figure/ECG/ECG input_gather.jpg -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/gather_figure/JapaneseVowels/JapaneseVowels Sample_gather.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/gather_figure/JapaneseVowels/JapaneseVowels Sample_gather.jpg -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/gather_figure/JapaneseVowels/JapaneseVowels channel_gather.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/gather_figure/JapaneseVowels/JapaneseVowels channel_gather.jpg -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/gather_figure/JapaneseVowels/JapaneseVowels input_gather.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/gather_figure/JapaneseVowels/JapaneseVowels input_gather.jpg -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/heatmap_figure_in_test/JapaneseVowels/JapaneseVowels accuracy=98.11 0.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/heatmap_figure_in_test/JapaneseVowels/JapaneseVowels accuracy=98.11 0.jpg -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/module/__pycache__/encoder.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/module/__pycache__/encoder.cpython-37.pyc -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/module/__pycache__/feedForward.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/module/__pycache__/feedForward.cpython-37.pyc -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/module/__pycache__/loss.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/module/__pycache__/loss.cpython-37.pyc -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/module/__pycache__/multiHeadAttention.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/module/__pycache__/multiHeadAttention.cpython-37.pyc -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/module/__pycache__/transformer.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/module/__pycache__/transformer.cpython-37.pyc -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/module/encoder.py: -------------------------------------------------------------------------------- 1 | from torch.nn import Module 2 | import torch 3 | 4 | from module.feedForward import FeedForward 5 | from module.multiHeadAttention import MultiHeadAttention 6 | 7 | class Encoder(Module): 8 | def __init__(self, 9 | d_model: int, 10 | d_hidden: int, 11 | q: int, 12 | v: int, 13 | h: int, 14 | device: str, 15 | mask: bool = False, 16 | dropout: float = 0.1): 17 | super(Encoder, self).__init__() 18 | 19 | self.MHA = MultiHeadAttention(d_model=d_model, q=q, v=v, h=h, mask=mask, device=device, dropout=dropout) 20 | self.feedforward = FeedForward(d_model=d_model, d_hidden=d_hidden) 21 | self.dropout = torch.nn.Dropout(p=dropout) 22 | self.layerNormal_1 = torch.nn.LayerNorm(d_model) 23 | self.layerNormal_2 = torch.nn.LayerNorm(d_model) 24 | 25 | def forward(self, x, stage): 26 | 27 | residual = x 28 | x, score = self.MHA(x, stage) 29 | x = self.dropout(x) 30 | x = self.layerNormal_1(x + residual) 31 | 32 | residual = x 33 | x = self.feedforward(x) 34 | x = self.dropout(x) 35 | x = self.layerNormal_2(x + residual) 36 | 37 | return x, score 38 | -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/module/feedForward.py: -------------------------------------------------------------------------------- 1 | from torch.nn import Module 2 | import torch 3 | import torch.nn.functional as F 4 | 5 | class FeedForward(Module): 6 | def __init__(self, 7 | d_model: int, 8 | d_hidden: int = 512): 9 | super(FeedForward, self).__init__() 10 | 11 | self.linear_1 = torch.nn.Linear(d_model, d_hidden) 12 | self.linear_2 = torch.nn.Linear(d_hidden, d_model) 13 | 14 | def forward(self, x): 15 | 16 | x = self.linear_1(x) 17 | x = F.relu(x) 18 | x = self.linear_2(x) 19 | 20 | return x -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/module/loss.py: -------------------------------------------------------------------------------- 1 | from torch.nn import Module 2 | import torch 3 | from torch.nn import CrossEntropyLoss 4 | 5 | class Myloss(Module): 6 | def __init__(self): 7 | super(Myloss, self).__init__() 8 | self.loss_function = CrossEntropyLoss() 9 | 10 | def forward(self, y_pre, y_true): 11 | y_true = y_true.long() 12 | loss = self.loss_function(y_pre, y_true) 13 | 14 | return loss -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/module/multiHeadAttention.py: -------------------------------------------------------------------------------- 1 | from torch.nn import Module 2 | import torch 3 | import math 4 | import torch.nn.functional as F 5 | 6 | 7 | class MultiHeadAttention(Module): 8 | def __init__(self, 9 | d_model: int, 10 | q: int, 11 | v: int, 12 | h: int, 13 | device: str, 14 | mask: bool=False, 15 | dropout: float = 0.1): 16 | super(MultiHeadAttention, self).__init__() 17 | 18 | self.W_q = torch.nn.Linear(d_model, q * h) 19 | self.W_k = torch.nn.Linear(d_model, q * h) 20 | self.W_v = torch.nn.Linear(d_model, v * h) 21 | 22 | self.W_o = torch.nn.Linear(v * h, d_model) 23 | 24 | self.device = device 25 | self._h = h 26 | self._q = q 27 | 28 | self.mask = mask 29 | self.dropout = torch.nn.Dropout(p=dropout) 30 | self.score = None 31 | 32 | def forward(self, x, stage): 33 | Q = torch.cat(self.W_q(x).chunk(self._h, dim=-1), dim=0) 34 | K = torch.cat(self.W_k(x).chunk(self._h, dim=-1), dim=0) 35 | V = torch.cat(self.W_v(x).chunk(self._h, dim=-1), dim=0) 36 | 37 | score = torch.matmul(Q, K.transpose(-1, -2)) / math.sqrt(self._q) 38 | self.score = score 39 | 40 | if self.mask and stage == 'train': 41 | mask = torch.ones_like(score[0]) 42 | mask = torch.tril(mask, diagonal=0) 43 | score = torch.where(mask > 0, score, torch.Tensor([-2**32+1]).expand_as(score[0]).to(self.device)) 44 | 45 | score = F.softmax(score, dim=-1) 46 | 47 | attention = torch.matmul(score, V) 48 | 49 | attention_heads = torch.cat(attention.chunk(self._h, dim=0), dim=-1) 50 | 51 | self_attention = self.W_o(attention_heads) 52 | 53 | return self_attention, self.score -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/module/transformer.py: -------------------------------------------------------------------------------- 1 | from torch.nn import Module 2 | import torch 3 | from torch.nn import ModuleList 4 | from module.encoder import Encoder 5 | import math 6 | import torch.nn.functional as F 7 | 8 | 9 | class Transformer(Module): 10 | def __init__(self, 11 | d_model: int, 12 | d_input: int, 13 | d_channel: int, 14 | d_output: int, 15 | d_hidden: int, 16 | q: int, 17 | v: int, 18 | h: int, 19 | N: int, 20 | device: str, 21 | dropout: float = 0.1, 22 | pe: bool = False, 23 | mask: bool = False): 24 | super(Transformer, self).__init__() 25 | 26 | self.encoder_list_1 = ModuleList([Encoder(d_model=d_model, 27 | d_hidden=d_hidden, 28 | q=q, 29 | v=v, 30 | h=h, 31 | mask=mask, 32 | dropout=dropout, 33 | device=device) for _ in range(N)]) 34 | 35 | self.encoder_list_2 = ModuleList([Encoder(d_model=d_model, 36 | d_hidden=d_hidden, 37 | q=q, 38 | v=v, 39 | h=h, 40 | dropout=dropout, 41 | device=device) for _ in range(N)]) 42 | 43 | self.embedding_channel = torch.nn.Linear(d_channel, d_model) 44 | self.embedding_input = torch.nn.Linear(d_input, d_model) 45 | 46 | self.gate = torch.nn.Linear(d_model * d_input + d_model * d_channel, 2) 47 | self.output_linear = torch.nn.Linear(d_model * d_input + d_model * d_channel, d_output) 48 | 49 | self.pe = pe 50 | self._d_input = d_input 51 | self._d_model = d_model 52 | 53 | def forward(self, x, stage): 54 | """ 55 | 前向传播 56 | :param x: 输入 57 | :param stage: 用于描述此时是训练集的训练过程还是测试集的测试过程测试过程中均不在加mask机制 58 | :return: 输出，gate之后的二维向量，step-wise encoder中的score矩阵，channel-wise encoder中的score矩阵，step-wise embedding后的三维矩阵，channel-wise embedding后的三维矩阵，gate 59 | """ 60 | # step-wise 61 | # score矩阵为 input，默认加mask 和 pe 62 | encoding_1 = self.embedding_channel(x) 63 | input_to_gather = encoding_1 64 | 65 | if self.pe: 66 | pe = torch.ones_like(encoding_1[0]) 67 | position = torch.arange(0, self._d_input).unsqueeze(-1) 68 | temp = torch.Tensor(range(0, self._d_model, 2)) 69 | temp = temp * -(math.log(10000) / self._d_model) 70 | temp = torch.exp(temp).unsqueeze(0) 71 | temp = torch.matmul(position.float(), temp) # shape:[input, d_model/2] 72 | pe[:, 0::2] = torch.sin(temp) 73 | pe[:, 1::2] = torch.cos(temp) 74 | 75 | encoding_1 = encoding_1 + pe 76 | 77 | for encoder in self.encoder_list_1: 78 | encoding_1, score_input = encoder(encoding_1, stage) 79 | 80 | # channel-wise 81 | # score矩阵为channel 默认不加mask和pe 82 | encoding_2 = self.embedding_input(x.transpose(-1, -2)) 83 | channel_to_gather = encoding_2 84 | 85 | for encoder in self.encoder_list_2: 86 | encoding_2, score_channel = encoder(encoding_2, stage) 87 | 88 | # 三维变二维 89 | encoding_1 = encoding_1.reshape(encoding_1.shape[0], -1) 90 | encoding_2 = encoding_2.reshape(encoding_2.shape[0], -1) 91 | 92 | # gate 93 | gate = F.softmax(self.gate(torch.cat([encoding_1, encoding_2], dim=-1)), dim=-1) 94 | encoding = torch.cat([encoding_1 * gate[:, 0:1], encoding_2 * gate[:, 1:2]], dim=-1) 95 | 96 | # 输出 97 | output = self.output_linear(encoding) 98 | 99 | return output, encoding, score_input, score_channel, input_to_gather, channel_to_gather, gate 100 | -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/mytest/DTW_test.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | 3 | from dtw import dtw 4 | 5 | # x = np.array([2, 0, 1, 1, 2, 4, 2, 1, 2, 0]).reshape(-1, 1) 6 | # y = np.array([1, 1, 2, 4, 2, 1, 2, 0]).reshape(-1, 1) 7 | x = np.array([8, 9, 1, 9, 6, 1, 3, 5]).reshape(-1, 1) 8 | y = np.array([2, 5, 4, 6, 7, 8, 3, 7, 7, 2]).reshape(-1, 1) 9 | 10 | 11 | euclidean_norm = lambda x, y: np.abs(x - y) 12 | 13 | d, cost_matrix, acc_cost_matrix, path = dtw(x, y, dist=euclidean_norm) 14 | 15 | print('d ', d) 16 | print('cost_matrix \r\n', cost_matrix) 17 | print('acc_cost_matrix \r\n', acc_cost_matrix) 18 | print('path ', path) -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/mytest/HeatMap.py: -------------------------------------------------------------------------------- 1 | import matplotlib.pyplot as plt 2 | import seaborn as sns 3 | sns.set() 4 | flights_long = sns.load_dataset("flights") 5 | flights = flights_long.pivot("month", "year", "passengers") 6 | # 绘制x-y-z的热力图，比如年-月-销量的热力图 7 | f, ax = plt.subplots(figsize=(9, 6)) 8 | #绘制热力图，还要将数值写到热力图上 9 | sns.heatmap(flights, annot=True, fmt="d", ax=ax) 10 | #设置坐标字体方向 11 | label_y = ax.get_yticklabels() 12 | plt.setp(label_y, rotation=360, horizontalalignment='right') 13 | label_x = ax.get_xticklabels() 14 | plt.setp(label_x, rotation=45, horizontalalignment='right') 15 | plt.show() -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/mytest/HeatMap_DTW.py: -------------------------------------------------------------------------------- 1 | from torch.utils.data import Dataset 2 | import torch 3 | import numpy as np 4 | from dtw import dtw 5 | import matplotlib.pyplot as plt 6 | import seaborn as sns 7 | import os 8 | 9 | plt.rcParams['font.sans-serif']=['SimHei'] #显示中文标签 10 | plt.rcParams['axes.unicode_minus']=False #这两行需要手动设 11 | 12 | from scipy.io import loadmat 13 | 14 | from torch.utils.data import Dataset 15 | import torch 16 | 17 | from scipy.io import loadmat 18 | 19 | 20 | class MyDataset(Dataset): 21 | def __init__(self, path, dataset): 22 | super(MyDataset, self).__init__() 23 | self.dataset = dataset 24 | self.train_len, \ 25 | self.test_len, \ 26 | self.input_len, \ 27 | self.channel_len, \ 28 | self.output_len, \ 29 | self.train_dataset, \ 30 | self.train_label, \ 31 | self.test_dataset, \ 32 | self.test_label = self.pre_option(path) 33 | 34 | def __getitem__(self, index): 35 | if self.dataset == 'train': 36 | return self.train_dataset[index], self.train_label[index] - 1 37 | elif self.dataset == 'test': 38 | return self.test_dataset[index], self.test_label[index] - 1 39 | 40 | def __len__(self): 41 | if self.dataset == 'train': 42 | return self.train_len 43 | elif self.dataset == 'test': 44 | return self.test_len 45 | 46 | def pre_option(self, path): 47 | m = loadmat(path) 48 | 49 | # m中是一个字典有4个key 其中最后一个键值对存储的是数据 50 | x1, x2, x3, x4 = m 51 | data = m[x4] 52 | 53 | data00 = data[0][0] 54 | # print('data00.shape', data00.shape) # () data00才到达数据的维度 55 | 56 | index_train = str(data.dtype).find('train\'') 57 | index_trainlabels = str(data.dtype).find('trainlabels') 58 | index_test = str(data.dtype).find('test\'') 59 | index_testlabels = str(data.dtype).find('testlabels') 60 | list = [index_test, index_train, index_testlabels, index_trainlabels] 61 | list = sorted(list) 62 | index_train = list.index(index_train) 63 | index_trainlabels = list.index(index_trainlabels) 64 | index_test = list.index(index_test) 65 | index_testlabels = list.index(index_testlabels) 66 | 67 | # [('trainlabels', 'O'), ('train', 'O'), ('testlabels', 'O'), ('test', 'O')] O 表示数据类型为 numpy.object 68 | train_label = data00[index_trainlabels] 69 | train_data = data00[index_train] 70 | test_label = data00[index_testlabels] 71 | test_data = data00[index_test] 72 | 73 | train_label = train_label.squeeze() 74 | train_data = train_data.squeeze() 75 | test_label = test_label.squeeze() 76 | test_data = test_data.squeeze() 77 | 78 | train_len = train_data.shape[0] 79 | test_len = test_data.shape[0] 80 | output_len = len(tuple(set(train_label))) 81 | 82 | # 时间步最大值 83 | max_lenth = 0 # 93 84 | for item in train_data: 85 | item = torch.as_tensor(item).float() 86 | if item.shape[1] > max_lenth: 87 | max_lenth = item.shape[1] 88 | 89 | for item in test_data: 90 | item = torch.as_tensor(item).float() 91 | if item.shape[1] > max_lenth: 92 | max_lenth = item.shape[1] 93 | 94 | # train_data, test_data为numpy.object 类型，不能直接对里面的numpy.ndarray进行处理 95 | train_dataset = [] 96 | test_dataset = [] 97 | for x1 in train_data: 98 | x1 = torch.as_tensor(x1).float() 99 | if x1.shape[1] != max_lenth: 100 | padding = torch.zeros(x1.shape[0], max_lenth - x1.shape[1]) 101 | x1 = torch.cat((x1, padding), dim=1) 102 | train_dataset.append(x1) 103 | 104 | for x2 in test_data: 105 | x2 = torch.as_tensor(x2).float() 106 | if x2.shape[1] != max_lenth: 107 | padding = torch.zeros(x2.shape[0], max_lenth - x2.shape[1]) 108 | x2 = torch.cat((x2, padding), dim=1) 109 | test_dataset.append(x2) 110 | 111 | # 最后维度 [数据条数,时间步数最大值,时间序列维度] 112 | train_dataset = torch.stack(train_dataset, dim=0).permute(0, 2, 1) 113 | test_dataset = torch.stack(test_dataset, dim=0).permute(0, 2, 1) 114 | train_label = torch.Tensor(train_label) 115 | test_label = torch.Tensor(test_label) 116 | channel = test_dataset[0].shape[-1] 117 | input = test_dataset[0].shape[-2] 118 | 119 | return train_len, test_len, input, channel, output_len, train_dataset, train_label, test_dataset, test_label 120 | 121 | def heatMap_channel(matrix, file_name, EPOCH): 122 | test_data = matrix[0].detach().numpy() 123 | euclidean_norm = lambda x, y: np.abs(x - y) 124 | matrix_0 = np.ones((test_data.shape[1], test_data.shape[1])) 125 | matrix_1 = np.ones((test_data.shape[1], test_data.shape[1])) # 相差度 126 | for i in range(test_data.shape[1]): 127 | for j in range(test_data.shape[1]): 128 | x = test_data[:, i] 129 | y = test_data[:, j] 130 | d, cost_matrix, acc_cost_matrix, path = dtw(x, y, dist=euclidean_norm) 131 | matrix_0[i, j] = d 132 | matrix_1[i, j] = np.mean((x - y) ** 2) 133 | 134 | sns.set() 135 | # f, ax = plt.subplots(figsize=(9, 6)) 136 | sns.heatmap(matrix_0, annot=True) 137 | plt.title('CHANNEL DTW') 138 | if os.path.exists(f'../heatmap_figure/{file_name}'): 139 | plt.savefig(f'../heatmap_figure/{file_name}/channel DTW EPOCH:{EPOCH}.jpg') 140 | else: 141 | os.makedirs(f'../heatmap_figure/{file_name}') 142 | plt.savefig(f'../heatmap_figure/{file_name}/channel DTW EPOCH:{EPOCH}.jpg') 143 | sns.heatmap(matrix_1, annot=True) 144 | plt.title('CHANNEL difference') 145 | plt.savefig(f'../heatmap_figure/{file_name}/channel difference EPOCH:{EPOCH}.jpg') 146 | 147 | def heatMap_input(matrix, file_name, EPOCH): 148 | test_data = matrix[0].detach().numpy() 149 | euclidean_norm = lambda x, y: np.abs(x - y) 150 | matrix_0 = np.ones((test_data.shape[0], test_data.shape[0])) # DTW 151 | matrix_1 = np.ones((test_data.shape[0], test_data.shape[0])) # 相差度 152 | for i in range(test_data.shape[0]): 153 | for j in range(test_data.shape[0]): 154 | x = test_data[i, :] 155 | y = test_data[j, :] 156 | d, cost_matrix, acc_cost_matrix, path = dtw(x, y, dist=euclidean_norm) 157 | matrix_0[i, j] = d 158 | matrix_1[i, j] = np.mean((x - y) ** 2) 159 | 160 | sns.set() 161 | # f, ax = plt.subplots(figsize=(9, 6)) 162 | sns.heatmap(matrix_0, annot=True) 163 | plt.title('INPUT DTW') 164 | if os.path.exists(f'../heatmap_figure/{file_name}'): 165 | plt.savefig(f'../heatmap_figure/{file_name}/input DTW EPOCH:{EPOCH}.jpg') 166 | else: 167 | os.makedirs(f'../heatmap_figure/{file_name}') 168 | plt.savefig(f'../heatmap_figure/{file_name}/input DTW EPOCH:{EPOCH}.jpg') 169 | sns.heatmap(matrix_1, annot=True) 170 | plt.title('INPUT difference') 171 | plt.savefig(f'../heatmap_figure/{file_name}/input difference EPOCH:{EPOCH}.jpg') 172 | 173 | def heatMap_score(matrix_input, matrix_channel, file_name, EPOCH): 174 | score_input = matrix_input[0].detach().numpy() 175 | score_channel = matrix_channel[0].detach().numpy() 176 | sns.set() 177 | # f, ax = plt.subplots(figsize=(9, 6)) 178 | sns.heatmap(score_input, annot=True) 179 | plt.title('SCORE INPUT') 180 | if os.path.exists(f'../heatmap_figure/{file_name}'): 181 | plt.savefig(f'../heatmap_figure/{file_name}/score input EPOCH:{EPOCH}.jpg') 182 | else: 183 | os.makedirs(f'../heatmap_figure/{file_name}') 184 | plt.savefig(f'../heatmap_figure/{file_name}/ score input EPOCH:{EPOCH}.jpg') 185 | sns.heatmap(score_channel, annot=True) 186 | plt.title('SCORE CHANNEL') 187 | plt.savefig(f'../heatmap_figure/{file_name}/score channel EPOCH:{EPOCH}.jpg') 188 | 189 | 190 | if __name__ == '__main__': 191 | path = 'E:\\PyCharmWorkSpace\\mtsdata\\JapaneseVowels\\JapaneseVowels.mat' # lenth=270 input=29 channel=12 output=9 192 | 193 | dataset = MyDataset(path, 'train') 194 | train_dataset = dataset.train_dataset 195 | print(train_dataset.shape) 196 | test_data = train_dataset[0].numpy() 197 | print(test_data[12, :]) 198 | # step_3 = train_dataset[0, 3, :].numpy() 199 | # step_12 = train_dataset[0, 12, :].numpy() 200 | # step_15 = train_dataset[0, 15, :].numpy() 201 | # step_25 = train_dataset[0, 25, :].numpy() 202 | # step_21 = train_dataset[0, 21, :].numpy() 203 | # step_27 = train_dataset[0, 27, :].numpy() 204 | # 205 | # # print(step_25.shape) 206 | # print(step_15) 207 | # print(step_25) 208 | # print(step_27) 209 | 210 | euclidean_norm = lambda x, y: np.abs(x - y) 211 | 212 | # d1, cost_matrix, acc_cost_matrix, path = dtw(step_3, step_12, dist=euclidean_norm) 213 | # d2, cost_matrix, acc_cost_matrix, path = dtw(step_15, step_25, dist=euclidean_norm) 214 | # d3, cost_matrix, acc_cost_matrix, path = dtw(step_3, step_21, dist=euclidean_norm) 215 | # d4, cost_matrix, acc_cost_matrix, path = dtw(step_15, step_27, dist=euclidean_norm) 216 | # 217 | # print(d1) 218 | # print(d2) 219 | # print(d3) 220 | # print(d4) 221 | 222 | print(test_data.shape) 223 | # matrix = np.ones((test_data.shape[0], test_data.shape[0])) # DTW 224 | # matrix_1 = np.ones((test_data.shape[0], test_data.shape[0])) # 欧氏距离 225 | # matrix_2 = np.ones((test_data.shape[0], test_data.shape[0])) # 相差度 226 | # matrix_3 = np.ones((test_data.shape[0], test_data.shape[0])) # 相差度 227 | matrix = np.ones((test_data.shape[1], test_data.shape[1])) # DTW 228 | matrix_1 = np.ones((test_data.shape[1], test_data.shape[1])) # 欧氏距离 229 | matrix_2 = np.ones((test_data.shape[1], test_data.shape[1])) # 相差度 230 | matrix_3 = np.ones((test_data.shape[1], test_data.shape[1])) # 点成 231 | for i in range(test_data.shape[1]): 232 | for j in range(test_data.shape[1]): 233 | # x = test_data[i, :] 234 | # y = test_data[j, :] 235 | x = test_data[:, i] 236 | y = test_data[:, j] 237 | d, cost_matrix, acc_cost_matrix, path = dtw(x, y, dist=euclidean_norm) 238 | matrix[i, j] = d 239 | matrix_1[i, j] = np.sum(np.abs(x-y)) 240 | matrix_2[i, j] = np.mean((x - y) ** 2) 241 | matrix_3[i, j] = float(x.dot(y)) 242 | 243 | # print(matrix) 244 | 245 | import numpy as np 246 | import seaborn as sns 247 | import matplotlib.pyplot as plt 248 | sns.set() 249 | # np.random.seed(0) 250 | # uniform_data = np.random.rand(10, 12) 251 | # ax = sns.heatmap(matrix) 252 | # f, ax = plt.subplots(figsize=(9, 6)) 253 | 254 | sns.heatmap(matrix, annot=True) 255 | plt.title('input') 256 | # os.mkdir('../heatmap_figure/lala') 257 | if os.path.exists('../heatmap_figure/lala'): 258 | plt.savefig(f'../heatmap_figure/lala/1.jpg') 259 | else: 260 | os.makedirs('../heatmap_figure/lala') 261 | plt.savefig(f'../heatmap_figure/lala/1.jpg') 262 | plt.show() 263 | 264 | plt.title('channel') 265 | sns.heatmap(matrix_3, annot=True) 266 | plt.savefig('2.jpg') 267 | plt.show() 268 | -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/mytest/__pycache__/HeatMap_DTW.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/mytest/__pycache__/HeatMap_DTW.cpython-37.pyc -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/mytest/kmeans_test.py: -------------------------------------------------------------------------------- 1 | # import cPickle 2 | # X,y = cPickle.load(open('data.pkl','r')) #X和y都是numpy.ndarray类型 3 | # X.shape #输出(1000,2) 4 | # y.shape #输出(1000,)对应每个样本的真实标签 5 | from dataset_process.dataset_process import MyDataset 6 | 7 | # path = 'E:\\PyCharmWorkSpace\\mtsdata\\ECG\\ECG.mat' # lenth=100 input=152 channel=2 output=2 8 | # path = 'E:\\PyCharmWorkSpace\\mtsdata\\JapaneseVowels\\JapaneseVowels.mat' # lenth=270 input=29 channel=12 output=9 9 | path = 'E:\\PyCharmWorkSpace\\mtsdata\\KickvsPunch\\KickvsPunch.mat' # lenth=10 input=841 channel=62 output=2 10 | 11 | 12 | 13 | dataset = MyDataset(path, 'train') 14 | X = dataset.train_dataset 15 | # X = torch.mean(X, dim=1).numpy() 16 | X = X.reshape(X.shape[0], -1).numpy() 17 | y = dataset.train_label.numpy() 18 | 19 | 20 | import numpy as np 21 | import matplotlib.pyplot as plt 22 | from utils.kmeans import KMeans 23 | def draw(X, Y): 24 | clf = KMeans(n_clusters=2, initCent='random' ,max_iter=100) 25 | clf.fit(X) 26 | cents = clf.centroids#质心 27 | labels = clf.labels#样本点被分配到的簇的索引 28 | sse = clf.sse 29 | #画出聚类结果，每一类用一种颜色 30 | colors = ['b','g','r','k','c','m','y','#e24fff','#524C90','#845868'] 31 | n_clusters = 2 32 | for i in range(n_clusters): 33 | index = np.nonzero(labels==i)[0] 34 | x0 = X[index,0] 35 | x1 = X[index,1] 36 | y_i = Y[index] 37 | for j in range(len(x0)): 38 | plt.text(x0[j],x1[j],str(int(y_i[j])),color=colors[i],\ 39 | fontdict={'weight': 'bold', 'size': 9}) 40 | plt.scatter(cents[i,0],cents[i,1],marker='x',color=colors[i],linewidths=12) 41 | plt.title("SSE={:.2f}".format(sse)) 42 | plt.axis([-30,30,-30,30]) 43 | plt.show() -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/result_figure/CharacterTrajectories 96.6% Adam epoch=100 batch=4 lr=0.0001 pe=True mask=True [512,6,6,8,8,0].png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/result_figure/CharacterTrajectories 96.6% Adam epoch=100 batch=4 lr=0.0001 pe=True mask=True [512,6,6,8,8,0].png -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/result_figure/ECG 86.0% Adam epoch=10 batch=20 lr=0.0001 pe=True mask=True [512,6,6,8,8,0].png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/result_figure/ECG 86.0% Adam epoch=10 batch=20 lr=0.0001 pe=True mask=True [512,6,6,8,8,0].png -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/result_figure/ECG 86.0% Adam epoch=20 batch=20 lr=0.0001 pe=True mask=True [512,6,6,8,8,0].png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/result_figure/ECG 86.0% Adam epoch=20 batch=20 lr=0.0001 pe=True mask=True [512,6,6,8,8,0].png -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/result_figure/ECG 87.0% Adam epoch=50 batch=20 lr=0.0001 pe=True mask=True [512,6,6,8,8,0].png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/result_figure/ECG 87.0% Adam epoch=50 batch=20 lr=0.0001 pe=True mask=True [512,6,6,8,8,0].png -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/result_figure/JapaneseVowels 98.11% Adagrad epoch=100 batch=3 lr=0.0001 pe=True mask=True [512,8,8,8,8,0.2].png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/result_figure/JapaneseVowels 98.11% Adagrad epoch=100 batch=3 lr=0.0001 pe=True mask=True [512,8,8,8,8,0.2].png -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/run.py: -------------------------------------------------------------------------------- 1 | # @Time : 2021/01/22 25:16 2 | # @Author : SY.M 3 | # @FileName: run.py 4 | 5 | import torch 6 | from torch.utils.data import DataLoader 7 | from dataset_process.dataset_process import MyDataset 8 | import torch.optim as optim 9 | from time import time 10 | from tqdm import tqdm 11 | import os 12 | 13 | from module.transformer import Transformer 14 | from module.loss import Myloss 15 | from utils.random_seed import setup_seed 16 | from utils.visualization import result_visualization 17 | 18 | # from mytest.gather.main import draw 19 | 20 | setup_seed(30) # 设置随机数种子 21 | reslut_figure_path = 'result_figure' # 结果图像保存路径 22 | 23 | # 数据集路径选择 24 | # path = 'E:\PyCharmWorkSpace\\dataset\\MTS_dataset\\AUSLAN\\AUSLAN.mat' # lenth=1140 input=136 channel=22 output=95 25 | # path = 'E:\PyCharmWorkSpace\\dataset\\MTS_dataset\\CharacterTrajectories\\CharacterTrajectories.mat' 26 | # path = 'E:\PyCharmWorkSpace\\dataset\\MTS_dataset\\CMUsubject16\\CMUsubject16.mat' # lenth=29,29 input=580 channel=62 output=2 27 | # path = 'E:\PyCharmWorkSpace\\dataset\\MTS_dataset\\ECG\\ECG.mat' # lenth=100 input=152 channel=2 output=2 28 | # path = 'E:\PyCharmWorkSpace\\dataset\\MTS_dataset\\JapaneseVowels\\JapaneseVowels.mat' # lenth=270 input=29 channel=12 output=9 29 | # path = 'E:\PyCharmWorkSpace\\dataset\\MTS_dataset\\Libras\\Libras.mat' # lenth=180 input=45 channel=2 output=15 30 | # path = 'E:\PyCharmWorkSpace\\dataset\\MTS_dataset\\UWave\\UWave.mat' # lenth=4278 input=315 channel=3 output=8 31 | # path = 'E:\PyCharmWorkSpace\\dataset\\MTS_dataset\\KickvsPunch\\KickvsPunch.mat' # lenth=10 input=841 channel=62 output=2 32 | # path = 'E:\PyCharmWorkSpace\\dataset\\MTS_dataset\\NetFlow\\NetFlow.mat' # lenth=803 input=997 channel=4 output=只有1和13 33 | # path = 'E:\PyCharmWorkSpace\\dataset\\MTS_dataset\\ArabicDigits\\ArabicDigits.mat' # lenth=6600 input=93 channel=13 output=10 34 | # path = 'E:\PyCharmWorkSpace\\dataset\\MTS_dataset\\PEMS\\PEMS.mat' 35 | # path = 'E:\PyCharmWorkSpace\\dataset\\MTS_dataset\\Wafer\\Wafer.mat' 36 | path = 'E:\PyCharmWorkSpace\dataset\\MTS_dataset\\WalkvsRun\\WalkvsRun.mat' 37 | 38 | test_interval = 5 # 测试间隔单位：epoch 39 | draw_key = 1 # 大于等于draw_key才会保存图像 40 | file_name = path.split('\\')[-1][0:path.split('\\')[-1].index('.')] # 获得文件名字 41 | 42 | # 超参数设置 43 | EPOCH = 100 44 | BATCH_SIZE = 3 45 | LR = 1e-4 46 | DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") # 选择设备 CPU or GPU 47 | print(f'use device: {DEVICE}') 48 | 49 | d_model = 512 50 | d_hidden = 1024 51 | q = 8 52 | v = 8 53 | h = 8 54 | N = 8 55 | dropout = 0.2 56 | pe = True # # 设置的是双塔中 score=pe score=channel默认没有pe 57 | mask = True # 设置的是双塔中 score=input的mask score=channel默认没有mask 58 | # 优化器选择 59 | optimizer_name = 'Adagrad' 60 | 61 | train_dataset = MyDataset(path, 'train') 62 | test_dataset = MyDataset(path, 'test') 63 | train_dataloader = DataLoader(dataset=train_dataset, batch_size=BATCH_SIZE, shuffle=True) 64 | test_dataloader = DataLoader(dataset=test_dataset, batch_size=BATCH_SIZE, shuffle=False) 65 | 66 | DATA_LEN = train_dataset.train_len # 训练集样本数量 67 | d_input = train_dataset.input_len # 时间部数量 68 | d_channel = train_dataset.channel_len # 时间序列维度 69 | d_output = train_dataset.output_len # 分类类别 70 | 71 | # 维度展示 72 | print('data structure: [lines, timesteps, features]') 73 | print(f'train data size: [{DATA_LEN, d_input, d_channel}]') 74 | print(f'mytest data size: [{train_dataset.test_len, d_input, d_channel}]') 75 | print(f'Number of classes: {d_output}') 76 | 77 | # 创建Transformer模型 78 | net = Transformer(d_model=d_model, d_input=d_input, d_channel=d_channel, d_output=d_output, d_hidden=d_hidden, 79 | q=q, v=v, h=h, N=N, dropout=dropout, pe=pe, mask=mask, device=DEVICE).to(DEVICE) 80 | # 创建loss函数此处使用交叉熵损失 81 | loss_function = Myloss() 82 | if optimizer_name == 'Adagrad': 83 | optimizer = optim.Adagrad(net.parameters(), lr=LR) 84 | elif optimizer_name == 'Adam': 85 | optimizer = optim.Adam(net.parameters(), lr=LR) 86 | 87 | # 用于记录准确率变化 88 | correct_on_train = [] 89 | correct_on_test = [] 90 | # 用于记录损失变化 91 | loss_list = [] 92 | time_cost = 0 93 | 94 | 95 | # 测试函数 96 | def test(dataloader, flag='test_set'): 97 | correct = 0 98 | total = 0 99 | with torch.no_grad(): 100 | net.eval() 101 | for x, y in dataloader: 102 | x, y = x.to(DEVICE), y.to(DEVICE) 103 | y_pre, _, _, _, _, _, _ = net(x, 'test') 104 | _, label_index = torch.max(y_pre.data, dim=-1) 105 | total += label_index.shape[0] 106 | correct += (label_index == y.long()).sum().item() 107 | if flag == 'test_set': 108 | correct_on_test.append(round((100 * correct / total), 2)) 109 | elif flag == 'train_set': 110 | correct_on_train.append(round((100 * correct / total), 2)) 111 | print(f'Accuracy on {flag}: %.2f %%' % (100 * correct / total)) 112 | 113 | return round((100 * correct / total), 2) 114 | 115 | 116 | # 训练函数 117 | def train(): 118 | net.train() 119 | max_accuracy = 0 120 | pbar = tqdm(total=EPOCH) 121 | begin = time() 122 | for index in range(EPOCH): 123 | for i, (x, y) in enumerate(train_dataloader): 124 | optimizer.zero_grad() 125 | 126 | y_pre, _, _, _, _, _, _ = net(x.to(DEVICE), 'train') 127 | 128 | loss = loss_function(y_pre, y.to(DEVICE)) 129 | 130 | print(f'Epoch:{index + 1}:\t\tloss:{loss.item()}') 131 | loss_list.append(loss.item()) 132 | 133 | loss.backward() 134 | 135 | optimizer.step() 136 | 137 | if ((index + 1) % test_interval) == 0: 138 | current_accuracy = test(test_dataloader) 139 | test(train_dataloader, 'train_set') 140 | print(f'当前最大准确率\t测试集:{max(correct_on_test)}%\t 训练集:{max(correct_on_train)}%') 141 | 142 | if current_accuracy > max_accuracy: 143 | max_accuracy = current_accuracy 144 | torch.save(net, f'saved_model/{file_name} batch={BATCH_SIZE}.pkl') 145 | 146 | pbar.update() 147 | 148 | os.rename(f'saved_model/{file_name} batch={BATCH_SIZE}.pkl', 149 | f'saved_model/{file_name} {max_accuracy} batch={BATCH_SIZE}.pkl') 150 | 151 | end = time() 152 | time_cost = round((end - begin) / 60, 2) 153 | 154 | # 结果图 155 | result_visualization(loss_list=loss_list, correct_on_test=correct_on_test, correct_on_train=correct_on_train, 156 | test_interval=test_interval, 157 | d_model=d_model, q=q, v=v, h=h, N=N, dropout=dropout, DATA_LEN=DATA_LEN, BATCH_SIZE=BATCH_SIZE, 158 | time_cost=time_cost, EPOCH=EPOCH, draw_key=draw_key, reslut_figure_path=reslut_figure_path, 159 | file_name=file_name, 160 | optimizer_name=optimizer_name, LR=LR, pe=pe, mask=mask) 161 | 162 | 163 | if __name__ == '__main__': 164 | train() 165 | -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/run_with_saved_model.py: -------------------------------------------------------------------------------- 1 | # @Time : 2021/01/22 25:50 2 | # @Author : SY.M 3 | # @FileName: run_with_saved_model.py 4 | 5 | 6 | import torch 7 | print('当前使用的pytorch版本：', torch.__version__) 8 | from utils.random_seed import setup_seed 9 | from torch.utils.data import DataLoader 10 | from dataset_process.dataset_process import MyDataset 11 | from utils.heatMap import heatMap_all 12 | from utils.TSNE import gather_by_tsne 13 | from utils.TSNE import gather_all_by_tsne 14 | import numpy as np 15 | from utils.colorful_line import draw_colorful_line 16 | 17 | setup_seed(30) 18 | DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") 19 | # 数据集维度展示 20 | # ArabicDigits length=6600 input=93 channel=13 output=10 21 | # AUSLAN length=1140 input=136 channel=22 output=95 22 | # CharacterTrajectories 23 | # CMUsubject16 length=29,29 input=580 channel=62 output=2 24 | # ECG length=100 input=152 channel=2 output=2 25 | # JapaneseVowels length=270 input=29 channel=12 output=9 26 | # Libras length=180 input=45 channel=2 output=15 27 | # UWave length=4278 input=315 channel=3 output=8 28 | # KickvsPunch length=10 input=841 channel=62 output=2 29 | # NetFlow length=803 input=997 channel=4 output=只有1和13 需要修改数据集处理的代码 30 | # Wafer length=803 input=997 channel=4 31 | 32 | # 选择要跑的模型 33 | save_model_path = 'saved_model/ECG 91.0 batch=2.pkl' 34 | file_name = save_model_path.split('/')[-1].split(' ')[0] 35 | path = f'E:\PyCharmWorkSpace\\dataset\\MTS_dataset\\{file_name}\\{file_name}.mat' # 拼装数据集路径 36 | 37 | # 绘制HeatMap的命名准备工作 38 | ACCURACY = save_model_path.split('/')[-1].split(' ')[1] # 使用的模型的准确率 39 | BATCH_SIZE = int(save_model_path[save_model_path.find('=')+1:save_model_path.rfind('.')]) # 使用的模型的batch_size 40 | heatMap_or_not = False # 是否绘制Score矩阵的HeatMap图 41 | gather_or_not = False # 是否绘制单个样本的step和channel上的聚类图 42 | gather_all_or_not = True # 是否绘制所有样本在特征提取后的聚类图 43 | 44 | # 加载模型 45 | net = torch.load(save_model_path, map_location=torch.device('cpu')) # map_location 设置使用的设备，可能是因为原来的pkl是在colab上用GPU跑的 46 | # 加载测试集数据 47 | test_dataset = MyDataset(path, 'test') 48 | test_dataloader = DataLoader(dataset=test_dataset, batch_size=BATCH_SIZE, shuffle=False) 49 | 50 | print(f'step为最大值的sample个数:{len(test_dataset.max_length_sample_inTest)}') 51 | if len(test_dataset.max_length_sample_inTest) == 0: 52 | gather_or_not = False 53 | heatMap_or_not = False 54 | print('测试集中没有step为最大值的样本，将不能绘制make sense 的heatmap 和 gather 图可尝试换一个数据集') 55 | 56 | correct = 0 57 | total = 0 58 | with torch.no_grad(): 59 | all_sample_X = [] 60 | all_sample_Y = [] 61 | for x, y in test_dataloader: 62 | x, y = x.to(DEVICE), y.to(DEVICE) 63 | y_pre, encoding, score_input, score_channel, gather_input, gather_channel, gate = net(x.to(DEVICE), 'test') 64 | 65 | all_sample_X.append(encoding) 66 | all_sample_Y.append(y) 67 | if heatMap_or_not: 68 | for index, sample in enumerate(test_dataset.max_length_sample_inTest): 69 | if sample.numpy().tolist() in x.numpy().tolist(): 70 | target_index = x.numpy().tolist().index(sample.numpy().tolist()) 71 | print('正在绘制heatmap图...') 72 | heatMap_all(score_input[target_index], score_channel[target_index], sample, 'heatmap_figure_in_test', file_name, ACCURACY, index) 73 | print('heatmap图绘制完成！') 74 | if gather_or_not: 75 | for index, sample in enumerate(test_dataset.max_length_sample_inTest): 76 | if sample.numpy().tolist() in x.numpy().tolist(): 77 | target_index = x.numpy().tolist().index(sample.numpy().tolist()) 78 | print('正在绘制gather图...') 79 | gather_by_tsne(gather_input[target_index].numpy(), np.arange(gather_input[target_index].shape[0]), index, file_name+' input_gather') 80 | gather_by_tsne(gather_channel[target_index].numpy(), np.arange(gather_channel[target_index].shape[0]), index, file_name+' channel_gather') 81 | print('gather图绘制完成！') 82 | draw_data = x[target_index].transpose(-1, -2)[0].numpy() 83 | draw_colorful_line(draw_data) 84 | gather_or_not = False 85 | 86 | _, label_index = torch.max(y_pre.data, dim=-1) 87 | total += label_index.shape[0] 88 | correct += (label_index == y.long()).sum().item() 89 | 90 | if gather_all_or_not: 91 | all_sample_X = torch.cat(all_sample_X, dim=0).numpy() 92 | all_sample_Y = torch.cat(all_sample_Y, dim=0).numpy() 93 | print('正在绘制gather图...') 94 | gather_all_by_tsne(all_sample_X, all_sample_Y, test_dataset.output_len, file_name+' all_sample_gather') 95 | print('gather图绘制完成！') 96 | 97 | print(f'Accuracy: %.2f %%' % (100 * correct / total)) 98 | 99 | 100 | -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/utils/TSNE.py: -------------------------------------------------------------------------------- 1 | from dataset_process.dataset_process import MyDataset 2 | import torch 3 | import numpy as np 4 | import os 5 | import matplotlib.pyplot as plt 6 | 7 | # path = 'E:\\PyCharmWorkSpace\\mtsdata\\ECG\\ECG.mat' # lenth=100 input=152 channel=2 output=2 8 | # 9 | # dataset = MyDataset(path, 'train') 10 | # X = dataset.train_dataset 11 | # X = torch.mean(X, dim=1).numpy() 12 | # # X = X.reshape(X.shape[0], -1).numpy() 13 | # y = dataset.train_label.numpy() 14 | 15 | from matplotlib import cm 16 | 17 | 18 | def plot_with_labels(lowDWeights, labels, kinds, file_name): 19 | plt.cla() 20 | # 降到二维了，分别给x和y 21 | X, Y = lowDWeights[:, 0], lowDWeights[:, 1] 22 | # 遍历每个点以及对应标签 23 | for x, y, s in zip(X, Y, labels): 24 | c = cm.rainbow(int(255/kinds * s)) # 为了使得颜色有区分度，把0-255颜色区间分为9分,然后把标签映射到一个区间 25 | plt.text(x, y, s, backgroundcolor=c, fontsize=6) 26 | plt.xlim(X.min(), X.max()) 27 | plt.ylim(Y.min(), Y.max()) 28 | 29 | plt.title('Clustering Step-Wise after Embedding') 30 | plt.rcParams['figure.figsize'] = (10.0, 10.0) # 设置figure_size尺寸 31 | 32 | if os.path.exists(f'gather_figure/{file_name.split(" ")[0]}') == False: 33 | os.makedirs(f'gather_figure/{file_name.split(" ")[0]}') 34 | 35 | plt.savefig(f'gather_figure/{file_name.split(" ")[0]}/{file_name}.jpg', dpi=600) 36 | # plt.show() 37 | plt.close() 38 | 39 | def plot_only(lowDWeights, labels, index, file_name): 40 | """ 41 | 绘制聚类图并为标签打上颜色 42 | :param lowDWeights: 将为之后的用于绘制聚类图的数据 43 | :param labels: lowDWeights对应的标签 44 | :param index: 用于命名文件是进行区分防止覆盖 45 | :param file_name: 文件名称和聚类的方式 46 | :return: None 47 | """ 48 | plt.cla() 49 | # 降到二维了，分别给x和y 50 | X, Y = lowDWeights[:, 0], lowDWeights[:, 1] 51 | # 遍历每个点以及对应标签 52 | # 聚类图中自定义的颜色的绘制请在下面for循环中完成 53 | for x, y, s in zip(X, Y, labels): 54 | position = 255 55 | if x < -850: 56 | position = 255 57 | elif 0.5*x - 225 < y: 58 | position = 0 59 | elif x < 1500: 60 | position = 50 61 | else: 62 | position = 100 63 | 64 | # c = cm.rainbow(int(255/9 * s)) # 为了使得颜色有区分度，把0-255颜色区间分为9分,然后把标签映射到一个区间 65 | c = cm.rainbow(position) # 为了使得颜色有区分度，把0-255颜色区间分为9分,然后把标签映射到一个区间 66 | plt.text(x, y, s, backgroundcolor=c, fontsize=6) 67 | plt.xlim(X.min(), X.max()) 68 | plt.ylim(Y.min(), Y.max()) 69 | # plt.title('Clustering Step-Wise after Embedding') 70 | 71 | plt.rcParams['figure.figsize'] = (10.0, 10.0) # 设置figure_size尺寸 72 | 73 | if os.path.exists(f'gather_figure/{file_name.split(" ")[0]}') == False: 74 | os.makedirs(f'gather_figure/{file_name.split(" ")[0]}') 75 | 76 | plt.savefig(f'gather_figure/{file_name.split(" ")[0]}/{file_name} {index}.jpg', dpi=600) 77 | # plt.show() 78 | plt.close() 79 | 80 | 81 | from sklearn.manifold import TSNE 82 | 83 | 84 | def gather_by_tsne(X: np.ndarray, 85 | Y: np.ndarray, 86 | index: int, 87 | file_name: str): 88 | tsne = TSNE(perplexity=30, n_components=2, init='pca', n_iter=4000) # TSNE降维，降到2 89 | low_dim_embs = tsne.fit_transform(X[:, :]) 90 | labels = Y[:] 91 | plot_only(low_dim_embs, labels, index, file_name) 92 | 93 | 94 | def gather_all_by_tsne(X: np.ndarray, 95 | Y: np.ndarray, 96 | kinds: int, 97 | file_name: str): 98 | """ 99 | 对gate之后的二维数据进行聚类 100 | :param X: 聚类数据 2维数据 101 | :param Y: 聚类数据对应标签 102 | :param kinds: 分类数 103 | :param file_name: 用于文件命名 104 | :return: None 105 | """ 106 | tsne = TSNE(perplexity=30, n_components=2, init='pca', n_iter=4000) # TSNE降维，降到2 107 | low_dim_embs = tsne.fit_transform(X[:, :]) 108 | labels = Y[:] 109 | plot_with_labels(low_dim_embs, labels, kinds, file_name) 110 | 111 | 112 | if __name__ == '__main__': 113 | path = 'E:\\PyCharmWorkSpace\\mtsdata\\ECG\\ECG.mat' # lenth=100 input=152 channel=2 output=2 114 | 115 | dataset = MyDataset(path, 'train') 116 | X = dataset.train_dataset 117 | X = torch.mean(X, dim=1).numpy() 118 | # X = X.reshape(X.shape[0], -1).numpy() 119 | Y = dataset.train_label.numpy() 120 | gather_by_tsne(X, Y) 121 | -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/utils/__pycache__/TSNE.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/utils/__pycache__/TSNE.cpython-37.pyc -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/utils/__pycache__/colorful_line.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/utils/__pycache__/colorful_line.cpython-37.pyc -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/utils/__pycache__/draw_line.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/utils/__pycache__/draw_line.cpython-37.pyc -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/utils/__pycache__/heatMap.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/utils/__pycache__/heatMap.cpython-37.pyc -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/utils/__pycache__/random_seed.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/utils/__pycache__/random_seed.cpython-37.pyc -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/utils/__pycache__/visualization.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZZUFaceBookDL/GTN/46aa05e8b609609edff193f3dba5bc0436faba2e/Gated Transformer 论文IJCAI版/utils/__pycache__/visualization.cpython-37.pyc -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/utils/colorful_line.py: -------------------------------------------------------------------------------- 1 | from matplotlib.collections import LineCollection 2 | import numpy as np 3 | import math 4 | import matplotlib.pyplot as plt 5 | 6 | 7 | def draw_colorful_line(draw_data: np.ndarray): 8 | # 定义颜色列表与各个点一一对应维度与点的个数相同 9 | colors = [] 10 | length = len(draw_data) 11 | x = np.arange(length) 12 | # 定义属于各个颜色的点 13 | red = [1, 2, 3] 14 | blue = [4, 5, 6] 15 | cyan = [7, 8, 9] 16 | # 填充颜色列表 17 | for i in range(length): 18 | if i in red: 19 | colors.append('red') 20 | elif i in blue: 21 | colors.append('blue') 22 | elif i in cyan: 23 | colors.append('cyan') 24 | else: 25 | colors.append('purple') 26 | y = draw_data 27 | 28 | points = np.array([x, y]).T.reshape(-1, 1, 2) # shape 152，1，2 29 | # 分片因为颜色不能画在点上而是画在线段上 30 | segments = np.concatenate([points[:-1], points[1:]], axis=1) # shape 151,2,2 31 | lc = LineCollection(segments, color=colors) 32 | ax = plt.axes() 33 | ax.set_xlim(0, length) 34 | ax.set_ylim(min(y), max(y)) 35 | ax.add_collection(lc) 36 | plt.show() 37 | plt.close() 38 | 39 | 40 | if __name__ == '__main__': 41 | pi = 3.1415 42 | 43 | x = np.linspace(0, 4 * pi, 100) 44 | y = [math.cos(xx) for xx in x] 45 | lwidths = abs(x) 46 | color = [] 47 | for i in range(len(y)): 48 | if i < 50: 49 | color.append('#FF0000') 50 | else: 51 | color.append('#000000') 52 | print(color) 53 | # print(x) 54 | # print(y) 55 | print('--------------------------------------') 56 | points = np.array([x, y]).T.reshape(-1, 1, 2) 57 | print(np.array([x, y]).shape) 58 | print(points.shape) 59 | print('--------------------------------------') 60 | print(points[:-1].shape) 61 | print(points[1:].shape) 62 | segments = np.concatenate([points[:-1], points[1:]], axis=1) 63 | print(segments.shape) 64 | lc = LineCollection(segments, linewidths=lwidths, color=color) 65 | 66 | ax = plt.axes() 67 | ax.set_xlim(min(x), max(x)) 68 | ax.set_ylim(min(y), max(y)) 69 | ax.add_collection(lc) 70 | plt.show() 71 | plt.close() 72 | 73 | ''' 74 | fig, a = plt.subplots() 75 | a.add_collection(lc) 76 | a.set_xlim(0, 4*pi) 77 | a.set_ylim(-1.1, 1.1) 78 | fig.show() 79 | ''' 80 | -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/utils/draw_line.py: -------------------------------------------------------------------------------- 1 | import matplotlib.pyplot as plt 2 | 3 | 4 | def draw_line(list_1, list_2, list_3): 5 | plt.style.use('seaborn') 6 | fig = plt.figure() # 创建基础图 7 | ax1 = fig.add_subplot(211) 8 | ax2 = fig.add_subplot(212) 9 | 10 | ax1.plot(list_3, color='red', label='Target_Weight_for_step') 11 | ax1.set_xlabel('epoch') 12 | ax1.set_ylabel('value') 13 | ax1.set_title('One Sample Weight for Step') 14 | 15 | ax2.plot(list_1, color='red', label='Weight_for_step') 16 | ax2.plot(list_2, color='blue', label='Weight_for_channel') 17 | ax2.set_xlabel('epoch') 18 | ax2.set_ylabel('value') 19 | ax2.set_title('Mean Weight Allocation') 20 | 21 | plt.legend(loc='best') 22 | plt.show() 23 | 24 | def draw_heatmap_anylasis(sample): 25 | channel_00 = sample[:, 0].numpy() 26 | channel_11 = sample[:, 1].numpy() 27 | sample = sample.transpose(-1, -2) 28 | channel_0 = sample[0, :].numpy().tolist() 29 | channel_1 = sample[1, :].numpy().tolist() 30 | channel_6 = sample[6, :].numpy().tolist() 31 | channel_4 = sample[4, :].numpy().tolist() 32 | channel_3 = sample[3, :].numpy().tolist() 33 | channel_7 = sample[7, :].numpy().tolist() 34 | channel_11 = sample[11, :].numpy().tolist() 35 | channel_9 = sample[9, :].numpy().tolist() 36 | 37 | ax1 = plt.subplot(511) 38 | ax2 = plt.subplot(512) 39 | ax3 = plt.subplot(513) 40 | ax4 = plt.subplot(514) 41 | ax5 = plt.subplot(515) 42 | ax1.plot(channel_11, color='red', label='channel_11') 43 | ax1.plot(channel_7, label='channel_7') 44 | ax2.plot(channel_11, color='red', label='channel_11') 45 | ax2.plot(channel_6, label='channel_6') 46 | ax3.plot(channel_11, color='red', label='channel_11') 47 | ax3.plot(channel_3, label='channel_3') 48 | ax4.plot(channel_3, color='red', label='channel_3') 49 | ax4.plot(channel_4, label='channel_4') 50 | ax5.plot(channel_9, color='red', label='channel_6') 51 | ax5.plot(channel_0, label='channel_0') 52 | ax1.legend(loc='best') 53 | ax2.legend(loc='best') 54 | ax3.legend(loc='best') 55 | ax4.legend(loc='best') 56 | plt.suptitle('Feature-wise Series Compare') 57 | plt.show() -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/utils/heatMap.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from dtw import dtw 3 | import matplotlib.pyplot as plt 4 | import seaborn as sns 5 | import os 6 | import torch 7 | 8 | def heatMap_all(score_input: torch.Tensor, # 二维数据 9 | score_channel: torch.Tensor, # 二维数据 10 | x: torch.Tensor, # 二维数据 11 | save_root: str, 12 | file_name: str, 13 | accuracy: str, 14 | index: int) -> None: 15 | score_channel = score_channel.detach().numpy() 16 | score_input = score_input.detach().numpy() 17 | draw_data = x.detach().numpy() 18 | 19 | euclidean_norm = lambda x, y: np.abs(x - y) # 用于计算DTW使用的函数,此处是一个计算欧氏距离的函数 20 | 21 | matrix_00 = np.ones((draw_data.shape[1], draw_data.shape[1])) # 用于记录channel之间DTW值的矩阵 22 | # matrix_01 = np.ones((draw_data.shape[1], draw_data.shape[1])) # 用于记录channel之间相差度的矩阵 23 | 24 | # matrix_10 = np.ones((draw_data.shape[0], draw_data.shape[0])) # 用于记录input之间DTW值的矩阵 25 | matrix_11 = np.ones((draw_data.shape[0], draw_data.shape[0])) # 用于记录input之间相差度值的矩阵 26 | 27 | for i in range(draw_data.shape[0]): 28 | for j in range(draw_data.shape[0]): 29 | x = draw_data[i, :] 30 | y = draw_data[j, :] 31 | # d, cost_matrix, acc_cost_matrix, path = dtw(x, y, dist=euclidean_norm) 32 | # matrix_10[i, j] = d 33 | matrix_11[i, j] = np.sqrt(np.sum((x - y) ** 2)) 34 | 35 | draw_data = draw_data.transpose(-1, -2) 36 | for i in range(draw_data.shape[0]): 37 | for j in range(draw_data.shape[0]): 38 | x = draw_data[i, :] 39 | y = draw_data[j, :] 40 | d, cost_matrix, acc_cost_matrix, path = dtw(x, y, dist=euclidean_norm) 41 | matrix_00[i, j] = d 42 | # matrix_01[i, j] = np.mean((x - y) ** 2) 43 | 44 | plt.rcParams['figure.figsize'] = (10.0, 8.0) # 设置figure_size尺寸 45 | plt.subplot(221) 46 | sns.heatmap(score_channel, cmap="YlGnBu", vmin=0) 47 | # plt.title('channel-wise attention') 48 | 49 | plt.subplot(222) 50 | sns.heatmap(matrix_00, cmap="YlGnBu", vmin=0) 51 | # plt.title('channel-wise DTW') 52 | 53 | plt.subplot(223) 54 | sns.heatmap(score_input, cmap="YlGnBu", vmin=0) 55 | # plt.title('step-wise attention') 56 | 57 | plt.subplot(224) 58 | sns.heatmap(matrix_11, cmap="YlGnBu", vmin=0) 59 | # plt.title('step-wise L2 distance') 60 | 61 | # plt.suptitle(f'{file_name.lower()}') 62 | 63 | if os.path.exists(f'{save_root}/{file_name}') == False: 64 | os.makedirs(f'{save_root}/{file_name}') 65 | plt.savefig(f'{save_root}/{file_name}/{file_name} accuracy={accuracy} {index}.jpg', dpi=400) 66 | 67 | # plt.show() 68 | plt.close() 69 | 70 | 71 | if __name__ == '__main__': 72 | matrix = torch.Tensor(range(24)).reshape(2, 3, 4) 73 | print(matrix.shape) 74 | file_name = 'lall' 75 | epcoh = 1 76 | 77 | data_channel = matrix.detach() 78 | data_input = matrix.detach() 79 | 80 | plt.subplot(2, 2, 1) 81 | sns.heatmap(data_channel[0].data.cpu().numpy()) 82 | plt.title("1") 83 | 84 | plt.subplot(2, 2, 2) 85 | sns.heatmap(data_input[0].data.cpu().numpy()) 86 | plt.title("2") 87 | 88 | plt.subplot(2, 2, 3) 89 | sns.heatmap(data_input[0].data.cpu().numpy()) 90 | plt.title("3") 91 | 92 | plt.subplot(2, 2, 4) 93 | sns.heatmap(data_input[0].data.cpu().numpy()) 94 | plt.title("4") 95 | 96 | plt.suptitle("JapaneseVowels Attention Heat Map", fontsize='x-large', fontweight='bold') 97 | # plt.savefig('result_figure/JapaneseVowels Attention Heat Map.png') 98 | plt.show() 99 | -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/utils/kmeans.py: -------------------------------------------------------------------------------- 1 | # coding=utf-8 2 | 3 | ''' 4 | @author: wepon, http://2hwp.com 5 | Reference: 6 | Book: <> 7 | Software: sklearn.cluster.KMeans 8 | ''' 9 | import numpy as np 10 | 11 | 12 | class KMeans(object): 13 | """ 14 | - 参数 15 | n_clusters: 16 | 聚类个数，即k 17 | initCent: 18 | 质心初始化方式，可选"random"或指定一个具体的array,默认random，即随机初始化 19 | max_iter: 20 | 最大迭代次数 21 | """ 22 | 23 | def __init__(self, n_clusters=5, initCent='random', max_iter=300): 24 | if hasattr(initCent, '__array__'): 25 | n_clusters = initCent.shape[0] 26 | self.centroids = np.asarray(initCent, dtype=np.float) 27 | else: 28 | self.centroids = None 29 | 30 | self.n_clusters = n_clusters 31 | self.max_iter = max_iter 32 | self.initCent = initCent 33 | self.clusterAssment = None 34 | self.labels = None 35 | self.sse = None 36 | 37 | # 计算两点的欧式距离 38 | 39 | def _distEclud(self, vecA, vecB): 40 | return np.linalg.norm(vecA - vecB) 41 | 42 | # 随机选取k个质心,必须在数据集的边界内 43 | def _randCent(self, X, k): 44 | n = X.shape[1] # 特征维数 45 | centroids = np.empty((k, n)) # k*n的矩阵，用于存储质心 46 | for j in range(n): # 产生k个质心，一维一维地随机初始化 47 | minJ = min(X[:, j]) 48 | rangeJ = float(max(X[:, j]) - minJ) 49 | centroids[:, j] = (minJ + rangeJ * np.random.rand(k, 1)).flatten() 50 | return centroids 51 | 52 | def fit(self, X): 53 | # 类型检查 54 | if not isinstance(X, np.ndarray): 55 | try: 56 | X = np.asarray(X) 57 | except: 58 | raise TypeError("numpy.ndarray required for X") 59 | 60 | m = X.shape[0] # m代表样本数量 61 | self.clusterAssment = np.empty((m, 2)) # m*2的矩阵，第一列存储样本点所属的族的索引值， 62 | # 第二列存储该点与所属族的质心的平方误差 63 | if self.initCent == 'random': 64 | self.centroids = self._randCent(X, self.n_clusters) 65 | 66 | clusterChanged = True 67 | for _ in range(self.max_iter): 68 | clusterChanged = False 69 | for i in range(m): # 将每个样本点分配到离它最近的质心所属的族 70 | minDist = np.inf; 71 | minIndex = -1 72 | for j in range(self.n_clusters): 73 | distJI = self._distEclud(self.centroids[j, :], X[i, :]) 74 | if distJI < minDist: 75 | minDist = distJI; 76 | minIndex = j 77 | if self.clusterAssment[i, 0] != minIndex: 78 | clusterChanged = True 79 | self.clusterAssment[i, :] = minIndex, minDist ** 2 80 | 81 | if not clusterChanged: # 若所有样本点所属的族都不改变,则已收敛，结束迭代 82 | break 83 | for i in range(self.n_clusters): # 更新质心，即将每个族中的点的均值作为质心 84 | ptsInClust = X[np.nonzero(self.clusterAssment[:, 0] == i)[0]] # 取出属于第i个族的所有点 85 | self.centroids[i, :] = np.mean(ptsInClust, axis=0) 86 | 87 | self.labels = self.clusterAssment[:, 0] 88 | self.sse = sum(self.clusterAssment[:, 1]) 89 | 90 | def predict(self, X): # 根据聚类结果，预测新输入数据所属的族 91 | # 类型检查 92 | if not isinstance(X, np.ndarray): 93 | try: 94 | X = np.asarray(X) 95 | except: 96 | raise TypeError("numpy.ndarray required for X") 97 | 98 | m = X.shape[0] # m代表样本数量 99 | preds = np.empty((m,)) 100 | for i in range(m): # 将每个样本点分配到离它最近的质心所属的族 101 | minDist = np.inf 102 | for j in range(self.n_clusters): 103 | distJI = self._distEclud(self.centroids[j, :], X[i, :]) 104 | if distJI < minDist: 105 | minDist = distJI 106 | preds[i] = j 107 | return preds 108 | 109 | 110 | class biKMeans(object): 111 | def __init__(self, n_clusters=5): 112 | self.n_clusters = n_clusters 113 | self.centroids = None 114 | self.clusterAssment = None 115 | self.labels = None 116 | self.sse = None 117 | 118 | # 计算两点的欧式距离 119 | def _distEclud(self, vecA, vecB): 120 | return np.linalg.norm(vecA - vecB) 121 | 122 | def fit(self, X): 123 | m = X.shape[0] 124 | self.clusterAssment = np.zeros((m, 2)) 125 | centroid0 = np.mean(X, axis=0).tolist() 126 | centList = [centroid0] 127 | for j in range(m): # 计算每个样本点与质心之间初始的平方误差 128 | self.clusterAssment[j, 1] = self._distEclud(np.asarray(centroid0), X[j, :]) ** 2 129 | 130 | while (len(centList) < self.n_clusters): 131 | lowestSSE = np.inf 132 | for i in range(len(centList)): # 尝试划分每一族,选取使得误差最小的那个族进行划分 133 | ptsInCurrCluster = X[np.nonzero(self.clusterAssment[:, 0] == i)[0], :] 134 | clf = KMeans(n_clusters=2) 135 | clf.fit(ptsInCurrCluster) 136 | centroidMat, splitClustAss = clf.centroids, clf.clusterAssment # 划分该族后，所得到的质心、分配结果及误差矩阵 137 | sseSplit = sum(splitClustAss[:, 1]) 138 | sseNotSplit = sum(self.clusterAssment[np.nonzero(self.clusterAssment[:, 0] != i)[0], 1]) 139 | if (sseSplit + sseNotSplit) < lowestSSE: 140 | bestCentToSplit = i 141 | bestNewCents = centroidMat 142 | bestClustAss = splitClustAss.copy() 143 | lowestSSE = sseSplit + sseNotSplit 144 | # 该族被划分成两个子族后,其中一个子族的索引变为原族的索引，另一个子族的索引变为len(centList),然后存入centList 145 | bestClustAss[np.nonzero(bestClustAss[:, 0] == 1)[0], 0] = len(centList) 146 | bestClustAss[np.nonzero(bestClustAss[:, 0] == 0)[0], 0] = bestCentToSplit 147 | centList[bestCentToSplit] = bestNewCents[0, :].tolist() 148 | centList.append(bestNewCents[1, :].tolist()) 149 | self.clusterAssment[np.nonzero(self.clusterAssment[:, 0] == bestCentToSplit)[0], :] = bestClustAss 150 | 151 | self.labels = self.clusterAssment[:, 0] 152 | self.sse = sum(self.clusterAssment[:, 1]) 153 | self.centroids = np.asarray(centList) 154 | 155 | def predict(self, X): # 根据聚类结果，预测新输入数据所属的族 156 | # 类型检查 157 | if not isinstance(X, np.ndarray): 158 | try: 159 | X = np.asarray(X) 160 | except: 161 | raise TypeError("numpy.ndarray required for X") 162 | 163 | m = X.shape[0] # m代表样本数量 164 | preds = np.empty((m,)) 165 | for i in range(m): # 将每个样本点分配到离它最近的质心所属的族 166 | minDist = np.inf 167 | for j in range(self.n_clusters): 168 | distJI = self._distEclud(self.centroids[j, :], X[i, :]) 169 | if distJI < minDist: 170 | minDist = distJI 171 | preds[i] = j 172 | return preds -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/utils/random_seed.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import numpy as np 3 | 4 | 5 | def setup_seed(seed): 6 | torch.manual_seed(seed) 7 | torch.cuda.manual_seed_all(seed) 8 | np.random.seed(seed) 9 | torch.backends.cudnn.deterministic = True 10 | 11 | 12 | -------------------------------------------------------------------------------- /Gated Transformer 论文IJCAI版/utils/visualization.py: -------------------------------------------------------------------------------- 1 | import matplotlib.pyplot as plt 2 | from matplotlib.font_manager import FontProperties as fp # 1、引入FontProperties 3 | import math 4 | 5 | def result_visualization(loss_list: list, 6 | correct_on_test: list, 7 | correct_on_train: list, 8 | test_interval: int, 9 | d_model: int, 10 | q: int, 11 | v: int, 12 | h: int, 13 | N: int, 14 | dropout: float, 15 | DATA_LEN: int, 16 | BATCH_SIZE: int, 17 | time_cost: float, 18 | EPOCH: int, 19 | draw_key: int, 20 | reslut_figure_path: str, 21 | optimizer_name: str, 22 | file_name: str, 23 | LR: float, 24 | pe: bool, 25 | mask: bool): 26 | my_font = fp(fname=r"font/simsun.ttc") # 2、设置字体路径 27 | 28 | # 设置风格 29 | plt.style.use('seaborn') 30 | 31 | fig = plt.figure() # 创建基础图 32 | ax1 = fig.add_subplot(311) # 创建两个子图 33 | ax2 = fig.add_subplot(313) 34 | 35 | ax1.plot(loss_list) # 添加折线 36 | ax2.plot(correct_on_test, color='red', label='on Test Dataset') 37 | ax2.plot(correct_on_train, color='blue', label='on Train Dataset') 38 | 39 | # 设置坐标轴标签和图的标题 40 | ax1.set_xlabel('epoch') 41 | ax1.set_ylabel('loss') 42 | ax2.set_xlabel(f'epoch/{test_interval}') 43 | ax2.set_ylabel('correct') 44 | ax1.set_title('LOSS') 45 | ax2.set_title('CORRECT') 46 | 47 | plt.legend(loc='best') 48 | 49 | # 设置文本 50 | fig.text(x=0.13, y=0.4, s=f'最小loss：{min(loss_list)}' ' ' 51 | f'最小loss对应的epoch数:{math.ceil((loss_list.index(min(loss_list)) + 1) / math.ceil((DATA_LEN / BATCH_SIZE)))}' ' ' 52 | f'最后一轮loss:{loss_list[-1]}' '\n' 53 | f'最大correct：测试集:{max(correct_on_test)}% 训练集:{max(correct_on_train)}%' ' ' 54 | f'最大correct对应的已训练epoch数:{(correct_on_test.index(max(correct_on_test)) + 1) * test_interval}' ' ' 55 | f'最后一轮correct：{correct_on_test[-1]}%' '\n' 56 | f'd_model={d_model} q={q} v={v} h={h} N={N} drop_out={dropout}' '\n' 57 | f'共耗时{round(time_cost, 2)}分钟', FontProperties=my_font) 58 | 59 | # 保存结果图测试不保存图（epoch少于draw_key） 60 | if EPOCH >= draw_key: 61 | plt.savefig( 62 | f'{reslut_figure_path}/{file_name} {max(correct_on_test)}% {optimizer_name} epoch={EPOCH} batch={BATCH_SIZE} lr={LR} pe={pe} mask={mask} [{d_model},{q},{v},{h},{N},{dropout}].png') 63 | 64 | # 展示图 65 | plt.show() 66 | 67 | print('正确率列表', correct_on_test) 68 | 69 | print(f'最小loss：{min(loss_list)}\r\n' 70 | f'最小loss对应的epoch数:{math.ceil((loss_list.index(min(loss_list)) + 1) / math.ceil((DATA_LEN / BATCH_SIZE)))}\r\n' 71 | f'最后一轮loss:{loss_list[-1]}\r\n') 72 | 73 | print(f'最大correct：测试集:{max(correct_on_test)}\t 训练集:{max(correct_on_train)}\r\n' 74 | f'最correct对应的已训练epoch数:{(correct_on_test.index(max(correct_on_test)) + 1) * test_interval}\r\n' 75 | f'最后一轮correct:{correct_on_test[-1]}') 76 | 77 | print(f'共耗时{round(time_cost, 2)}分钟') -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Gated-Transformer-on-MTS 2 | 基于Pytorch，使用改良的Transformer模型应用于多维时间序列的分类任务上 3 | 4 | ## 实验结果 5 | 对比模型选择 Fully Convolutional Networks (FCN) and Residual Net works (ResNet)
6 | 7 | 8 | DataSet|MLP|FCN|ResNet|Encoder|MCNN|t-LeNet|MCDCNN|Time-CNN|TWIESN|Gated Transformer| 9 | -------|---|---|------|-------|----|-------|------|--------|------|-----------------| 10 | ArabicDigits|96.9|99.4|**99.6**|98.1|10.0|10.0|95.9|95.8|85.3|98.8| 11 | AUSLAN|93.3|**97.5**|97.4|93.8|1.1|1.1|85.4|72.6|72.4|**97.5**| 12 | CharacterTrajectories|96.9|**99.0**|**99.0**|97.1|5.4|6.7|93.8|96.0|92.0|97.0| 13 | CMUsubject16|60.0|**100**|99.7|98.3|53.1|51.0|51.4|97.6|89.3|**100**| 14 | ECG|74.8|87.2|86.7|87.2|67.0|67.0|50.0|84.1|73.7|**91.0**| 15 | JapaneseVowels|97.6||**99.3**|99.2|97.6|9.2|23.8|94.4|95.6|96.5|98.7| 16 | Libras|78.0|**96.4**|95.4|78.3|6.7|6.7|65.1|63.7|79.4|88.9| 17 | UWave|90.1|**93.4**|92.6|90.8|12.5|12.5|84.5|8.9|75.4|91.0| 18 | KickvsPunch|61.0|54.0|51.0|61.0|54.0|50.0|56.0|62.0|67.0|**90.0**| 19 | NetFlow|55.0|89.1|62.7|77.7|77.9|72.3|63.0|89.0|94.5|**100**| 20 | PEMS|-|-|-|-|-|-|-|-|-|93.6| 21 | Wafer|89.4|98.2|98.9|98.6|89.4|89.4|65.8|94.8|94.9|**99.1**| 22 | WalkvsRun|70.0|**100**|**100**|**100**|75.0|60.0|45.0|**100.0**|94.4|**100**| 23 | 24 | ## 实验环境 25 | 环境|描述| 26 | ---|---------| 27 | 语言|Python3.7| 28 | 框架|Pytorch1.6| 29 | IDE|Pycharm and Colab| 30 | 设备|CPU and GPU| 31 | 32 | ## 数据集 33 | 多元时间序列数据集, 文件为.mat格式，训练集与测试集在一个文件中，且预先定义为了测试集数据，测试集标签，训练集数据与训练集标签。
34 | 数据集下载使用百度云盘，连接如下：
35 | 链接：https://pan.baidu.com/s/1u2HN6tfygcQvzuEK5XBa2A
36 | 提取码：dxq6
37 | Google drive link:https://drive.google.com/drive/folders/1QFadJOmbOLWMjLrcebZQR_w2fBX7x0Vm?usp=share_link 38 | 39 | UEA and UCR dataset:http://www.timeseriesclassification.com/index.php 40 | 41 | --- 42 | 43 | 数据集维度描述 44 | DataSet|Number of Classes|Size of training Set|Size of testing Set|Max Time series Length|Channel| 45 | -------|-----------------|--------------------|-------------------|----------------------|-------| 46 | ArabicDigits|10|6600|2200|93|13| 47 | AUSLAN|95|1140|1425|136|22| 48 | CharacterTrajectories|20|300|2558|205|3| 49 | CMUsubject16|2|29|29|580|62| 50 | ECG|2|100|100|152|2| 51 | JapaneseVowels|9|270|370|29|12| 52 | Libras|15|180|180|45|2| 53 | UWave|8|200|4278|315|3| 54 | KickvsPunch|2|16|10|841|62| 55 | NetFlow|2|803|534|997|4| 56 | PEMS|7|267|173|144|963| 57 | Wafer|2|298|896|198|6| 58 | WalkvsRun|2|28|16|1918|62| 59 | 60 | ## 数据预处理 61 | 详细数据集处理过程参看 dataset_process.py文件。
62 | - 创建torch.utils.data.Dataset对象，在类中对数据集进行处理，其成员变量定义的有训练集数据，训练集标签，测试集数据，测试集标签等。创建torch.utils.data.DataLoader对象，生成训练过程中的mini-batch与数据集的随机shuffle
63 | - 数据集中不同样本的Time series Length不同，处理时使用所有样本中（测试集与训练集中）**最长**的时间步作为Time series Length，使用**0**进行填充。
64 | - 数据集处理过程中保存未添加Padding的训练集数据与测试集数据，还有测试集中最长时间步的样本列表以供探索模型使用。
65 | - NetFlow数据集中标签为**1和13**，在使用此数据集时要对返回的标签值进行处理。
66 | 67 | ## 模型描述 68 | 69 |

70 | 71 | - 仅使用Encoder：由于是分类任务，模型删去传统Transformer中的decoder，**仅使用Encoder**进行分类 72 | - Two Tower：在不同的Step之间或者不同Channel之间，显然存在着诸多联系，传统Transformer使用Attentino机制用来关注不同的step或channel之间的相关程度，但仅选择一个进行计算。不同于CNN模型处理时间序列，它可以使用二维卷积核同时关注step-wise和channel-wise，在这里我们使**双塔**模型，即同时计算step-wise Attention和channel-wise Attention。 73 | - Gate机制：对于不同的数据集，不同的Attention机制有好有坏，对于双塔的特征提取的结果，简单的方法，是对两个塔的输出尽心简单的拼接，不过在这里，我们使用模型学习两个权重值，为每个塔的输出进行权重的分配，公式如下。
74 | `h = W · Concat(C, S) + b`
75 | `g1, g2 = Softmax(h)`
76 | `y = Concat(C · g1, S · g2)`
77 | - 在step-wise,模型如传统Transformer一样，添加位置编码与mask机制，而在channel-wise，模型舍弃位置编码与mask，因为对于没有时间特性的channel之间，这两个机制没有实际的意义。 78 | 79 | ## 超参描述 80 | 超参|描述| 81 | ----|---| 82 | d_model|模型处理的为时间序列而非自然语言，所以省略了NLP中对词语的编码，仅使用一个线性层映射成d_model维的稠密向量，此外，d_model保证了在每个模块衔接的地方的维度相同| 83 | d_hidden|Position-wise FeedForword 中隐藏层的维度| 84 | d_input|时间序列长度，其实是一个数据集中最长时间步的维度 **固定**的，直接由数据集预处理决定| 85 | d_channel|多元时间序列的时间通道数，即是几维的时间序列 **固定**的，直接由数据集预处理决定| 86 | d_output|分类类别数 **固定**的，直接由数据集预处理决定| 87 | q,v|Multi-Head Attention中线性层映射维度| 88 | h|Multi-Head Attention中头的数量| 89 | N|Encoder栈中Encoder的数量| 90 | dropout|随机失活| 91 | EPOCH|训练迭代次数| 92 | BATCH_SIZE|mini-batch size| 93 | LR|学习率定义为1e-4| 94 | optimizer_name|优化器选择建议**Adagrad**和Adam| 95 | 96 | ## 文件描述 97 | 文件名称|描述| 98 | -------|----| 99 | dataset_process|数据集处理| 100 | font|存储字体，用于结果图中的文字| 101 | gather_figure|聚类结果图| 102 | heatmap_figure_in_test|测试模型时绘制的score矩阵的热力图| 103 | module|模型的各个模块| 104 | mytest|各种测试代码| 105 | reslut_figure|准确率结果图| 106 | saved_model|保存的pkl文件| 107 | utils|工具类文件| 108 | run.py|训练模型| 109 | run_with_saved_model.py|使用训练好的模型（保存为pkl文件）测试结果| 110 | 111 | ## utils工具描述 112 | 简单介绍几个 113 | - random_seed:用于设置**随机种子**，使每一次的实验结果可复现。 114 | - heatMap.py:用于绘制双塔的score矩阵的**热力图**，用来分析channel与channel之间或者step与step之间的相关程度，用于比较的还有**DTW**矩阵和欧氏距离矩阵，用来分析决定权重分配的因素。 115 | - draw_line:用于绘制折线图，一般需要根据需要自定义新的函数进行绘制。 116 | - visualization:用于绘制训练模型的loss变化曲线和accuracy变化曲线，判断是否收敛与过拟合。 117 | - TSNE：**降维聚类算法**并绘制聚类图，用来评估模型特征提取的效果或者时间序列之间的相似性。 118 | 119 | ## Tips 120 | - .pkl文件需要先训练，并在训练结束时进行保存(设置参数为True)，由于github对文件大小的限制，上传文件中不包含训练好的.pkl文件。 121 | - .pkl文件使用pycharm上的1.6版本的pytorch和colab上1.7的pytorch保存，若想load模型直接进行测试，需要测试使用的pytorch版本尽可能高于等于**1.6版本**。 122 | - 根目录文件如saved_model,reslut_figure为保存的默认路径，请勿删除或者修改名称，除非直接在源代码中对路径进行修改。 123 | - 请使用百度云盘提供的数据集，不同的MTS数据集文件格式不同，本数据集处理的是.mat文件。 124 | - utils中的工具类，在绘制彩色曲线和聚类图时，对于图中颜色的划分，由于需求不能泛化，请在函数中自行编写代码定义。 125 | - save model保存的.pkl文件在迭代过程中不断更新，在最后保存最高准确率的模型并命名，命名格式请勿修改，因为在run_with_saved_model.py中，对文件命名中的信息会加以利用，若干绘图结果的命名也会参考其中的信息。 126 | - 优先选择GPU，没有则使用CPU。 127 | 128 | ## 参考 129 | ``` 130 | [Wang et al., 2017] Z. Wang, W. Yan, and T. Oates. Time series classification from scratch with deep neural networks:A strong baseline. In 2017 International Joint Conference on Neural Networks (IJCNN), pages 1578–1585, 2017. 131 | ``` 132 | 133 | ## 本人学识浅薄，代码和文字若有不当之处欢迎批评与指正！ 134 | ## 联系方式：masiyuan007@qq.com 135 | --------------------------------------------------------------------------------