因为我观察了baseline错误分类数据的频域发现，他们的振幅所处的频率相同，因此我想用振幅比较大对应的频率来分。我选择这些数据中振幅大于0.1的振幅对应的频率。需要改进的地方可能是应该采取所有训练集中大于0.1振幅的频率。
import matplotlib.pyplot as plt import numpy as np import pandas as pd import torch import torch.fft as fft df = pd.read_csv('train.csv') df=df.drop(['ID'],axis=1) nmp=df.to_numpy() feature=nmp[:-20,:-1] label=nmp[:-20,-1]#（210，240） feature=torch.fft.fft(torch.Tensor(feature)) feature=torch.abs(feature)/240*2 feature=feature[:,[0, 1, 60, 180, 239, 59, 120, 181, 61, 179, 119, 121, 58, 182, 57, 183, 62, 178]] test_feature=nmp[-20:,:-1] test_label=nmp[-20:,-1]#（210，240）
test_feature=torch.fft.fft(torch.Tensor(test_feature)) test_feature=torch.abs(test_feature)/240*2 test_feature=test_feature[:,[0, 1, 60, 180, 239, 59, 120, 181, 61, 179, 119, 121, 58, 182, 57, 183, 62, 178]] from torch import nn import torch loss=nn.MSELoss() feature=torch.Tensor(feature) label=torch.Tensor(label) label=label.reshape(-1,1)
test_feature=torch.Tensor(test_feature) test_label=torch.Tensor(test_label) test_label=test_label.reshape(-1,1)
network=nn.Sequential(nn.Linear(18,5),nn.Sigmoid(),nn.Linear(5,1),nn.Sigmoid()) import torch.optim as optim optimizer = optim.Adam(network.parameters(), lr=0.004) for epoch in range(10000): optimizer.zero_grad() out=network(feature) l=loss(out,label) l.backward() optimizer.step() Y = torch.ge(out, 0.5).float() acc=Y.eq(label).float().sum()/len(label) out=network(test_feature) Y = torch.ge(out, 0.5).float() test_acc=Y.eq(test_label).float().sum()/len(test_label) print(epoch,l,acc,test_acc) #if test_acc==0.50 and acc>0.93: if acc>0.83 and test_acc==0.85: break
df = pd.read_csv('test.csv') df=df.drop(['ID'],axis=1) nmp=df.to_numpy() feature=nmp[:,:] feature=torch.fft.fft(torch.Tensor(feature)) feature=torch.abs(feature)/240*2 feature=torch.Tensor(feature[:,[0, 1, 60, 180, 239, 59, 120, 181, 61, 179, 119, 121, 58, 182, 57, 183, 62, 178]]) out=network(feature) out=out.detach().numpy() out=out>0.5 out=out.astype(np.int) out=pd.DataFrame(out) out.columns = ['CLASS'] w=[] for k in range(out.shape[0]): w.append(k+210) out['ID']=np.reshape(w,(-1,1)) out[['ID','CLASS']].to_csv('out.csv',index=False) 标签：fft,torch,程序,feature,label,准确率,0.81,test,out From： https://www.cnblogs.com/hahaah/p/16944503.html