I use pytorch to train a model to classify iris, but my acc was about 0.4

user_19198804 0
Viewed 70

I have tried many improvements like increasing epochs, using better loss functions and optimizers, deepening the network and shuffling the dataset, etc, but still to no avail. This problem has been bothering me for a long time, thanks for your help. Below is my code.

load and process dataset(updated)

def Iris_Reader(dataset):
    train_data, test_data, train_label, test_label = train_test_split(dataset.data, dataset.target, test_size=0.4)

    # scaler = StandardScaler()
    # train_data = scaler.fit_transform(train_data)
    # test_data = scaler.transform(test_data)
    
    return torch.FloatTensor(train_data), torch.LongTensor(train_label), torch.FloatTensor(test_data), torch.LongTensor(test_label)

Define the classifier

class Classifier(nn.Module):
    def __init__(self):
        super().__init__()
        
        #4*3*3 network
        self.model = nn.Sequential(
            nn.Linear(4,3),
            nn.ReLU(),

            nn.Linear(3,3),
        )
        
        #SGD
        self.optimiser = torch.optim.SGD(self.parameters(), lr = 0.1)
        
        #MSE LOSS_FUNCTION
        self.loss_fn = nn.CrossEntropyLoss()

        self.counter = 0
        self.progress = []

    def forward(self, input):
        return self.model(input)
    
    def train(self, input, target):
        output = self.forward(input)

        loss = self.loss_fn(output, target)

        self.counter += 1
        self.progress.append(loss.item())

        self.optimiser.zero_grad()
        loss.backward()
        self.optimiser.step()
    # plot loss
    def plot_loss(self):
        plt.figure(dpi=100)
        plt.ylim([0,1.0])
        plt.yticks([0, 0.25, 0.5, 1.0])
        plt.scatter(x = [i for i in range(len(self.progress))], y = self.progress, marker = '.', alpha = 0.2)
        plt.grid('on')
        plt.show()

TRAIN

C = Classifier()
epochs = 10
dataset = datasets.load_iris()

for epoch in range(epochs):
    train_data, train_label, _, _ = Iris_Reader(dataset)
    for i, j in zip(train_data, train_label):
        C.train(i, j)

TEST

score = 0
num = 0
# for epoch in range(epochs):
_, _, test_data, test_label = Iris_Reader(dataset)
for i,j in zip(test_data, test_label):
    output = C.forward(i).detach().argmax()
    if output == j:
        # print(C.forward(i).detach(), j)
        score += 1
    num += 1
print(score, num, round(score/num, 3))

OUTPUT: 53 60 0.883

iris-dataset pytorch neural-network model python
1 Answers

There's a bunch of problems here:

  • First, you seem to shuffle data and labels independently, rendering the dataset useless.
  • Also, you recreate the dataset inside the loop every epoch, wasting the CPU time pointlessly.

Overall, the dataset creation can be shortened to something like this:

def Iris_Reader(dataset):
    train_data, test_data, train_label, test_label = sklearn.model_selection.train_test_split(dataset.data, dataset.target, test_size=0.2)
    return torch.FloatTensor(train_data), torch.LongTensor(train_label), torch.FloatTensor(test_data), torch.LongTensor(test_label)

and should be taken outside the loop.

Next, MSELoss() is suited for regression. For classification, CrossEntropyLoss() is the default choice.

Using sigmoid as activation in an intermediate layer is not the best choice, especially with a short number of epochs. ReLU should converge much better.

Last but not least, your loss chart would look much cleaner if the values were averaged per epoch.

Update: the implementation that ensures the target having the same size as network output, with additional feature scaling:

from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler

def Iris_Reader(dataset):
    label = nn.functional.one_hot(torch.LongTensor(dataset.target), num_classes=3).float()

    train_data, test_data,train_label, test_label = train_test_split(dataset.data, label, test_size=0.2)
    scaler = StandardScaler()
    train_data = scaler.fit_transform(train_data)
    test_data = scaler.transform(test_data)
    return torch.FloatTensor(train_data), train_label, torch.FloatTensor(test_data), test_label

Oh, and you should probably also remove the final Sigmoid() since CrossEntropyLoss() applies logsoftmax anyway.

dx2-66_192801950
Related