Can you tell me what's wrong with my LSTM model?

Recently, I’ve been making a model to predict the driver’s intension using the data sampled at 10HZ.
I try to predict the driver’s intention by remembering the data up to 3 seconds ago.
So I set window_size for 3 seconds.
Input data has a 25 features so input data’s shape is (Batchsize) * (timeseq total 30) * (feature total 25)
I’ll predict one of 0,1 and 2 as a result of output
So I thought this problem was a multi-class classification model, so I organized the model as follows.

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.lstm = nn.LSTM(input_size = 25, hidden_size = 30, num_layers = 1,
                            batch_first = True)
        self.linear = nn.Linear(30,10)
        self.linear1 = nn.Linear(10,3)
        self.batch_norm1 = nn.BatchNorm1d(10)
        self.drop = nn.Dropout(p =0.5)
        
        
    def forward(self, input_seq):
        output_seq, _ = self.lstm(input_seq)
        last_output = output_seq[:,-1]
        class_predictions = self.drop(last_output)
        class_predictions = self.linear(class_predictions)
        class_predictions = self.batch_norm1(class_predictions)
        class_predictions = F.relu(class_predictions)
        class_predictions = self.linear1(class_predictions)
        class_predictions = F.softmax(class_predictions,dim=1)
        return class_predictions
        

I used relu function, dropout, and BatchNomalize because the test result and train result were very different.
I think it’s because of overfitting, but I don’t know how to solve it.
Finally, I used weightedrandomsampler to resolve class imbalances in the data used in the model.
I’ll attach the whole code.

import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import pandas as pd
from torch.utils.data import DataLoader, Dataset
from torchvision import transforms, utils
from torchviz import make_dot
from sklearn.preprocessing import MinMaxScaler
from torch.autograd import Variable
from torch.utils.tensorboard import SummaryWriter
from torch.utils.data import WeightedRandomSampler

if torch.cuda.is_available():
    DEVICE = torch.device('cuda')
else:
    DEVICE = torch.device('cpu')


print('Using Pytorch Version: ',torch.__version__,
      'Device: ',DEVICE)
EPOCHS = 1000
batch_size = [64,128,256,512,1024]
learning_rates = [0.001,0.0001]


class FeatureDataset(Dataset):
    def __init__(self, file_name):
        
        _x, _y = [], []
        file_out = pd.read_csv(file_name)
        x = file_out.iloc[:,0:-1].values
        y = file_out.iloc[:,-1].values
        mMscaler = MinMaxScaler()
        mMscaler.fit(x)
        x = mMscaler.fit_transform(x)

    
        for i in range(30, len(y)-29):
            _x.append(x[i-30:i,:])
            _y.append(y[i+29])
        
        ax = np.array(_x)
        ay = np.array(_y)
        
        self.x = torch.tensor(ax, dtype=torch.float32)
        self.y = torch.tensor(ay)
        
    def __len__(self):
        return len(self.y)
    
    def __getitem__(self,idx):
        return self.x[idx], self.y[idx]
    
class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.lstm = nn.LSTM(input_size = 25, hidden_size = 30, num_layers = 1,
                            batch_first = True)
        self.linear = nn.Linear(30,10)
        self.linear1 = nn.Linear(10,3)
        self.batch_norm1 = nn.BatchNorm1d(10)
        self.drop = nn.Dropout(p =0.5)
        
        
    def forward(self, input_seq):
        output_seq, _ = self.lstm(input_seq)
        last_output = output_seq[:,-1]
        class_predictions = self.drop(last_output)
        class_predictions = self.linear(class_predictions)
        class_predictions = self.batch_norm1(class_predictions)
        class_predictions = F.relu(class_predictions)
        class_predictions = self.linear1(class_predictions)
        class_predictions = F.softmax(class_predictions,dim=1)
        return class_predictions
        
arr = FeatureDataset('Untitled 5.csv')
arr2 = FeatureDataset('real_test.csv')

print(arr.x.shape)
class_weights = [1/256668, 1/46829, 1/45261]
sample_weights = [0] * len(arr)

for idx, (data, label) in enumerate(arr):
    class_weight = class_weights[label]
    sample_weights[idx] = class_weight
    
sampler = WeightedRandomSampler(sample_weights, num_samples = len(sample_weights), replacement = True)
train_loader = torch.utils.data.DataLoader(arr,
                                                   batch_size = 32, sampler = sampler,
                                          shuffle = False)
for trainx,trainy in train_loader:
    print(trainx.size, trainx.shape)
    break


for BATCH_SIZE in batch_size:
    for learning_rate in learning_rates:
        model = Model().to(DEVICE)
        optimizer = torch.optim.Adam(model.parameters(), lr = learning_rate)
        criterion = nn.CrossEntropyLoss()
        
        train_loader = torch.utils.data.DataLoader(arr,
                                                   batch_size = BATCH_SIZE, sampler = sampler,
                                          shuffle = False)
   
        print("BATCH_SIZE : ",BATCH_SIZE, "lr = ",learning_rate)
        for epoch in range(EPOCHS):
            cnt_0 = 0
            cnt_1 = 0
            cnt_2 = 0
            train_loss = 0
            train_correct = 0
            model.train()
            for batch_idx,(train_x, train_y) in enumerate(train_loader):
                train_x = train_x.to(DEVICE)
                train_y = train_y.to(DEVICE)
                optimizer.zero_grad()
                output = model(train_x)
                loss = criterion(output, train_y)
                loss.backward()
                optimizer.step()
                prediction = output.max(1, keepdim = True)[1]
                for i in range(len(prediction)):
                    pre = int(prediction[i])
                    if pre is 0:
                        cnt_0 +=1
                    elif pre is 1:
                        cnt_1 +=1
                    elif pre is 2:
                        cnt_2 +=1
                train_correct += prediction.eq(train_y.view_as(prediction)).sum().item()
                train_loss += loss.item()
            
                
            train_loss /= (len(train_loader.dataset) / BATCH_SIZE)
            train_accuracy = 100. * train_correct / len(train_loader.dataset)
            print(cnt_0, '  ', cnt_1, '  ', cnt_2)
            
          
            print("[Train EPOCH: {}], \tTrain Loss: {:.4f}, \tTrain Accuracy: {:.2f} % \n".format(
                        epoch, train_loss, train_accuracy))
            
            
        
        
            test_loader = torch.utils.data.DataLoader(arr2, batch_size = BATCH_SIZE,
                                              shuffle = False)
            
          
           
            test_loss = 0
            test_correct = 0
            model.eval()
            cnt_0 = 0
            cnt_1 = 0
            cnt_2 = 0
            with torch.no_grad():
                for test_x, test_y in test_loader:
                    test_x = test_x.to(DEVICE)
                    test_y = test_y.to(DEVICE)
                    output = model(test_x)
                    loss = criterion(output, test_y)
                    test_loss += criterion(output, test_y).item()
                    prediction = output.max(1, keepdim = True)[1]
                    for i in range(len(prediction)):
                        pre = int(prediction[i])
                        if pre is 0:
                            cnt_0 +=1
                        elif pre is 1:
                            cnt_1 +=1
                        elif pre is 2:
                            cnt_2 +=1
                    test_correct += prediction.eq(test_y.view_as(prediction)).sum().item()
                  
               
                print(cnt_0, '  ', cnt_1, '  ', cnt_2)
                test_loss /= (len(test_loader.dataset) / BATCH_SIZE)
                test_accuracy = 100. * test_correct / len(test_loader.dataset)
                print("[EPOCH: {}], \tTest Loss: {:.4f}, \tTest Accuracy: {:.2f} % \n".format(
                            epoch, test_loss, test_accuracy))
                
                print('-----------------------------------')
       

Your answers always help me a lot. It’s my first time writing a question, so I ask for your understanding even if you don’t understand.

1 Like

It’s my first time to read someone else’s post on this forum.

1 Like

I’ve encountered similar problem recently, when using a model with BatchNorm layers.

For me, trying InstanceNorm instead of BatchNorm solved the problem…

also I found some silimar issue in community. This may be helpful.

Thank you.
I hope this method works

BatchNorm doesnt work mostly in recurrent networks, try LayerNorm or InstanceNorm

my model contains batchn