Model Achieving 100% accuracy if I call train model function 3 to 4 times

1

I am new to pytorch and experimenting few things, I got to know about torch.manual_seed() , and I was using it this way

torch.manual_seed(52)
def train_model(model,X_train,y_train,X_test,y_test,optimizer):
    for epoch in range(0,500):
        model.train()
        y_pred=model(X_train).squeeze()
        y_pred_prob=torch.round(torch.sigmoid(y_pred))
        loss_data=loss(y_pred,y_train)
        acc = accuracy_fn(y_true=y_train, 
                      y_pred=y_pred_prob) 
        optimizer.zero_grad()
        loss_data.backward()
        optimizer.step()
        model.eval()
        with torch.inference_mode():
            # 1. Forward pass
            test_logits = model(X_test).squeeze() 
            test_pred = torch.round(torch.sigmoid(test_logits))
            # 2. Caculate loss/accuracy
            test_loss = loss(test_logits,
                                y_test)
            test_acc = accuracy_fn(y_true=y_test,
                                y_pred=test_pred)

        # Print out what's happening every 10 epochs
        if epoch % 10 == 0:
            print(f"Epoch: {epoch} | Loss: {loss_data:.5f}, Accuracy: {acc:.2f}% | Test loss: {test_loss:.5f}, Test acc: {test_acc:.2f}%")
    return model

The loss that I am using is loss=nn.BCEWithLogitsLoss() and the optimizer is optim=torch.optim.SGD(model_non_linear.parameters(),lr=0.02)

I created model in two types one by using class and other by using nn.sequential as show here

model_seq_non_linear=nn.Sequential(
        nn.Linear(in_features=2,out_features=8),
        nn.ReLU(),
        nn.Linear(in_features=8,out_features=5),
        nn.ReLU(),
        nn.Linear(in_features=5,out_features=3),
        nn.ReLU(),
        nn.Linear(in_features=3,out_features=1)
)
model_seq_non_linear.to(device)

class CirclePredictionNonLinear(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.layer1=nn.Linear(in_features=2,out_features=8)
        self.layer2=nn.Linear(in_features=8,out_features=5)
        self.layer3=nn.Linear(in_features=5,out_features=3)
        self.layer4=nn.Linear(in_features=3,out_features=1)
        self.relu=nn.ReLU()
    def forward(self,x):
        return self.layer4(self.relu(self.layer3(self.relu(self.layer2(self.relu(self.layer1(x)))))))

the problem is when I call the train model function 2 to 3 times, I start getting 100% accuracy

train_model(model_non_linear,X_train,y_train,X_test,y_test,optim)

image
image773×737 128 KB

I feel it has to do something with torch.manual_seed , because when I was not using it my training results will be same no matter how many times I call the function but now it seems if I call it , it remembers the previous results and train on it somehow .

The dataset is this

from sklearn.datasets import make_circles 
X,y=make_circles(n_samples=1000,noise=0.08,shuffle=True,factor=0.3,random_state=30)
import pandas as pd
circles = pd.DataFrame({"X1": X[:, 0],
    "X2": X[:, 1],
    "label": y
})
X_data=torch.from_numpy(X).type(torch.float32)
y_data=torch.from_numpy(y).type(torch.float32)
X_train, X_test, y_train, y_test = train_test_split(X_data, 
                                                    y_data, 
                                                    test_size=0.2, # 20% test, 80% train
                                                    random_state=42) # make the random split reproducible

len(X_train), len(X_test), len(y_train), len(y_test)
X_train, y_train = X_train.to(device), y_train.to(device)
X_test, y_test = X_test.to(device), y_test.to(device)

Thank You for helping

2

I don’t know what the actual issue and if you are concerned about the achieved 100% accuracy, the seed, or multiple calls into train_model.
The 100% accuracy could point to a small and easy to learn dataset.
E.g. look at this very simple example:

data = torch.randn(64, 1)
target = (data > 0.).float()

model = nn.Linear(1, 1)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-1)
criterion = nn.BCEWithLogitsLoss()

for epoch in range(50):
    optimizer.zero_grad()
    output = model(data)
    loss = criterion(output, target)
    loss.backward()
    optimizer.step()
    pred = (output > 0.).float()
    acc = (pred == target).float().mean()
    print(f"epoch: {epoch}, loss: {loss:.3f}, acc: {acc}")

# epoch: 0, loss: 1.001, acc: 0.28125
# epoch: 1, loss: 0.940, acc: 0.28125
# epoch: 2, loss: 0.884, acc: 0.28125
# epoch: 3, loss: 0.832, acc: 0.171875
# epoch: 4, loss: 0.785, acc: 0.125
# ...
# epoch: 44, loss: 0.245, acc: 1.0
# epoch: 45, loss: 0.242, acc: 1.0
# epoch: 46, loss: 0.239, acc: 1.0
# epoch: 47, loss: 0.237, acc: 1.0
# epoch: 48, loss: 0.235, acc: 1.0
# epoch: 49, loss: 0.232, acc: 1.0

Setting the seed will make sure the pseudorandom number generator will return a defined sequence of random values.
Calling train_model multiple times will continue training the model for 500 epochs each.

1 Like
3

image
image883×816 142 KB

First call performance results

I am sorry, i included too many factors, my main issue was that since the train model function is training the model for 500 epochs , why is it retaining the weights values of the previous training call .
I will include the call results of all 5 in 5 different replies as it seems I am not allowed to attach more than 1 media results at a time

4

This is the second call results

train_model(model_seq_non_linear,X_train,y_train,X_test,y_test,optim_seq) 

the output 
Output exceeds the size limit. Open the full output data in a text editorEpoch: 0 | Loss: 0.68657, Accuracy: 66.62% | Test loss: 0.68570, Test acc: 66.50%
Epoch: 10 | Loss: 0.68640, Accuracy: 67.25% | Test loss: 0.68548, Test acc: 67.50%
Epoch: 20 | Loss: 0.68622, Accuracy: 67.00% | Test loss: 0.68525, Test acc: 66.00%
Epoch: 30 | Loss: 0.68604, Accuracy: 66.62% | Test loss: 0.68502, Test acc: 67.00%
Epoch: 40 | Loss: 0.68586, Accuracy: 65.50% | Test loss: 0.68478, Test acc: 66.50%
Epoch: 50 | Loss: 0.68567, Accuracy: 65.50% | Test loss: 0.68454, Test acc: 66.50%
Epoch: 60 | Loss: 0.68547, Accuracy: 64.75% | Test loss: 0.68430, Test acc: 66.00%
Epoch: 70 | Loss: 0.68527, Accuracy: 64.50% | Test loss: 0.68406, Test acc: 66.50%
Epoch: 80 | Loss: 0.68507, Accuracy: 63.88% | Test loss: 0.68382, Test acc: 65.50%
Epoch: 90 | Loss: 0.68486, Accuracy: 63.88% | Test loss: 0.68359, Test acc: 66.50%
Epoch: 100 | Loss: 0.68466, Accuracy: 63.88% | Test loss: 0.68335, Test acc: 67.00%
Epoch: 110 | Loss: 0.68445, Accuracy: 63.75% | Test loss: 0.68312, Test acc: 67.00%
Epoch: 120 | Loss: 0.68424, Accuracy: 63.62% | Test loss: 0.68289, Test acc: 67.00%
Epoch: 130 | Loss: 0.68402, Accuracy: 63.62% | Test loss: 0.68266, Test acc: 67.00%
Epoch: 140 | Loss: 0.68380, Accuracy: 63.62% | Test loss: 0.68242, Test acc: 67.50%
Epoch: 150 | Loss: 0.68358, Accuracy: 63.62% | Test loss: 0.68218, Test acc: 68.00%
Epoch: 160 | Loss: 0.68336, Accuracy: 63.75% | Test loss: 0.68194, Test acc: 68.00%
Epoch: 170 | Loss: 0.68312, Accuracy: 64.00% | Test loss: 0.68169, Test acc: 68.00%
Epoch: 180 | Loss: 0.68289, Accuracy: 64.12% | Test loss: 0.68144, Test acc: 68.00%
Epoch: 190 | Loss: 0.68265, Accuracy: 64.50% | Test loss: 0.68118, Test acc: 68.00%
Epoch: 200 | Loss: 0.68241, Accuracy: 64.50% | Test loss: 0.68093, Test acc: 68.00%
Epoch: 210 | Loss: 0.68216, Accuracy: 64.62% | Test loss: 0.68066, Test acc: 68.50%
Epoch: 220 | Loss: 0.68190, Accuracy: 64.62% | Test loss: 0.68039, Test acc: 68.50%
Epoch: 230 | Loss: 0.68164, Accuracy: 65.00% | Test loss: 0.68011, Test acc: 68.00%
Epoch: 240 | Loss: 0.68138, Accuracy: 65.25% | Test loss: 0.67982, Test acc: 68.00%
...
Epoch: 460 | Loss: 0.67342, Accuracy: 69.38% | Test loss: 0.67123, Test acc: 72.00%
Epoch: 470 | Loss: 0.67293, Accuracy: 69.62% | Test loss: 0.67070, Test acc: 72.00%
Epoch: 480 | Loss: 0.67242, Accuracy: 69.62% | Test loss: 0.67016, Test acc: 72.50%
Epoch: 490 | Loss: 0.67190, Accuracy: 69.75% | Test loss: 0.66961, Test acc: 72.50%
CirclePredictionNonLinear(
  (layer1): Linear(in_features=2, out_features=8, bias=True)
  (layer2): Linear(in_features=8, out_features=5, bias=True)
  (layer3): Linear(in_features=5, out_features=3, bias=True)
  (layer4): Linear(in_features=3, out_features=1, bias=True)
  (relu): ReLU()
)

The third time I call the same line

Output exceeds the size limit. Open the full output data in a text editorEpoch: 0 | Loss: 0.67136, Accuracy: 70.12% | Test loss: 0.66903, Test acc: 72.50%
Epoch: 10 | Loss: 0.67080, Accuracy: 70.38% | Test loss: 0.66844, Test acc: 72.50%
Epoch: 20 | Loss: 0.67023, Accuracy: 70.50% | Test loss: 0.66784, Test acc: 72.50%
Epoch: 30 | Loss: 0.66964, Accuracy: 70.50% | Test loss: 0.66721, Test acc: 72.50%
Epoch: 40 | Loss: 0.66903, Accuracy: 70.50% | Test loss: 0.66657, Test acc: 72.50%
Epoch: 50 | Loss: 0.66840, Accuracy: 70.62% | Test loss: 0.66590, Test acc: 72.50%
Epoch: 60 | Loss: 0.66775, Accuracy: 71.00% | Test loss: 0.66522, Test acc: 73.00%
Epoch: 70 | Loss: 0.66708, Accuracy: 71.12% | Test loss: 0.66451, Test acc: 73.00%
Epoch: 80 | Loss: 0.66639, Accuracy: 71.38% | Test loss: 0.66377, Test acc: 73.50%
Epoch: 90 | Loss: 0.66567, Accuracy: 71.50% | Test loss: 0.66302, Test acc: 74.00%
Epoch: 100 | Loss: 0.66493, Accuracy: 71.75% | Test loss: 0.66225, Test acc: 74.00%
Epoch: 110 | Loss: 0.66417, Accuracy: 72.12% | Test loss: 0.66146, Test acc: 74.00%
Epoch: 120 | Loss: 0.66338, Accuracy: 72.38% | Test loss: 0.66064, Test acc: 74.00%
Epoch: 130 | Loss: 0.66256, Accuracy: 72.62% | Test loss: 0.65979, Test acc: 74.50%
Epoch: 140 | Loss: 0.66172, Accuracy: 72.75% | Test loss: 0.65891, Test acc: 75.50%
Epoch: 150 | Loss: 0.66086, Accuracy: 72.75% | Test loss: 0.65800, Test acc: 75.50%
Epoch: 160 | Loss: 0.65996, Accuracy: 73.00% | Test loss: 0.65706, Test acc: 75.50%
Epoch: 170 | Loss: 0.65904, Accuracy: 73.12% | Test loss: 0.65609, Test acc: 75.50%
Epoch: 180 | Loss: 0.65809, Accuracy: 73.25% | Test loss: 0.65509, Test acc: 75.50%
Epoch: 190 | Loss: 0.65711, Accuracy: 73.38% | Test loss: 0.65405, Test acc: 75.50%
Epoch: 200 | Loss: 0.65610, Accuracy: 73.62% | Test loss: 0.65300, Test acc: 76.00%
Epoch: 210 | Loss: 0.65505, Accuracy: 73.88% | Test loss: 0.65191, Test acc: 76.50%
Epoch: 220 | Loss: 0.65398, Accuracy: 74.00% | Test loss: 0.65079, Test acc: 76.50%
Epoch: 230 | Loss: 0.65286, Accuracy: 74.25% | Test loss: 0.64963, Test acc: 76.50%
Epoch: 240 | Loss: 0.65170, Accuracy: 74.25% | Test loss: 0.64844, Test acc: 76.50%
...
Epoch: 460 | Loss: 0.61325, Accuracy: 78.88% | Test loss: 0.60937, Test acc: 80.00%
Epoch: 470 | Loss: 0.61072, Accuracy: 79.00% | Test loss: 0.60683, Test acc: 80.00%
Epoch: 480 | Loss: 0.60810, Accuracy: 79.62% | Test loss: 0.60423, Test acc: 80.00%
Epoch: 490 | Loss: 0.60540, Accuracy: 80.00% | Test loss: 0.60157, Test acc: 80.00%
CirclePredictionNonLinear(
  (layer1): Linear(in_features=2, out_features=8, bias=True)
  (layer2): Linear(in_features=8, out_features=5, bias=True)
  (layer3): Linear(in_features=5, out_features=3, bias=True)
  (layer4): Linear(in_features=3, out_features=1, bias=True)
  (relu): ReLU()
)

When I call it fourth time

Output exceeds the size limit. Open the full output data in a text editorEpoch: 0 | Loss: 0.60262, Accuracy: 80.12% | Test loss: 0.59881, Test acc: 80.00%
Epoch: 10 | Loss: 0.59973, Accuracy: 80.38% | Test loss: 0.59598, Test acc: 80.00%
Epoch: 20 | Loss: 0.59673, Accuracy: 80.75% | Test loss: 0.59307, Test acc: 80.00%
Epoch: 30 | Loss: 0.59366, Accuracy: 80.88% | Test loss: 0.59006, Test acc: 81.00%
Epoch: 40 | Loss: 0.59051, Accuracy: 81.00% | Test loss: 0.58695, Test acc: 81.00%
Epoch: 50 | Loss: 0.58726, Accuracy: 81.25% | Test loss: 0.58374, Test acc: 80.50%
Epoch: 60 | Loss: 0.58390, Accuracy: 81.00% | Test loss: 0.58047, Test acc: 80.50%
Epoch: 70 | Loss: 0.58043, Accuracy: 81.25% | Test loss: 0.57710, Test acc: 81.00%
Epoch: 80 | Loss: 0.57684, Accuracy: 81.62% | Test loss: 0.57362, Test acc: 81.00%
Epoch: 90 | Loss: 0.57312, Accuracy: 81.62% | Test loss: 0.57002, Test acc: 81.50%
Epoch: 100 | Loss: 0.56927, Accuracy: 81.75% | Test loss: 0.56630, Test acc: 82.50%
Epoch: 110 | Loss: 0.56530, Accuracy: 82.38% | Test loss: 0.56247, Test acc: 83.50%
Epoch: 120 | Loss: 0.56117, Accuracy: 82.88% | Test loss: 0.55849, Test acc: 83.50%
Epoch: 130 | Loss: 0.55691, Accuracy: 83.25% | Test loss: 0.55437, Test acc: 83.50%
Epoch: 140 | Loss: 0.55255, Accuracy: 83.12% | Test loss: 0.55018, Test acc: 83.50%
Epoch: 150 | Loss: 0.54809, Accuracy: 83.62% | Test loss: 0.54587, Test acc: 83.50%
Epoch: 160 | Loss: 0.54350, Accuracy: 83.88% | Test loss: 0.54145, Test acc: 83.50%
Epoch: 170 | Loss: 0.53880, Accuracy: 84.00% | Test loss: 0.53693, Test acc: 84.00%
Epoch: 180 | Loss: 0.53392, Accuracy: 84.25% | Test loss: 0.53226, Test acc: 84.00%
Epoch: 190 | Loss: 0.52891, Accuracy: 84.38% | Test loss: 0.52745, Test acc: 84.00%
Epoch: 200 | Loss: 0.52379, Accuracy: 84.50% | Test loss: 0.52258, Test acc: 84.00%
Epoch: 210 | Loss: 0.51856, Accuracy: 85.00% | Test loss: 0.51765, Test acc: 84.00%
Epoch: 220 | Loss: 0.51325, Accuracy: 85.38% | Test loss: 0.51263, Test acc: 84.50%
Epoch: 230 | Loss: 0.50789, Accuracy: 85.75% | Test loss: 0.50757, Test acc: 84.50%
Epoch: 240 | Loss: 0.50241, Accuracy: 86.12% | Test loss: 0.50248, Test acc: 84.50%
...
Epoch: 460 | Loss: 0.36920, Accuracy: 91.00% | Test loss: 0.38700, Test acc: 89.00%
Epoch: 470 | Loss: 0.36351, Accuracy: 91.25% | Test loss: 0.38228, Test acc: 89.00%
Epoch: 480 | Loss: 0.35778, Accuracy: 91.38% | Test loss: 0.37786, Test acc: 89.00%
Epoch: 490 | Loss: 0.35200, Accuracy: 91.62% | Test loss: 0.37374, Test acc: 89.00%
CirclePredictionNonLinear(
  (layer1): Linear(in_features=2, out_features=8, bias=True)
  (layer2): Linear(in_features=8, out_features=5, bias=True)
  (layer3): Linear(in_features=5, out_features=3, bias=True)
  (layer4): Linear(in_features=3, out_features=1, bias=True)
  (relu): ReLU()
)

And the last call

Output exceeds the size limit. Open the full output data in a text editorEpoch: 0 | Loss: 0.34634, Accuracy: 91.62% | Test loss: 0.36953, Test acc: 90.00%
Epoch: 10 | Loss: 0.34066, Accuracy: 91.88% | Test loss: 0.36541, Test acc: 90.00%
Epoch: 20 | Loss: 0.33518, Accuracy: 92.00% | Test loss: 0.36125, Test acc: 90.00%
Epoch: 30 | Loss: 0.32981, Accuracy: 92.12% | Test loss: 0.35709, Test acc: 90.00%
Epoch: 40 | Loss: 0.32450, Accuracy: 92.12% | Test loss: 0.35301, Test acc: 90.00%
Epoch: 50 | Loss: 0.31925, Accuracy: 92.25% | Test loss: 0.34907, Test acc: 90.00%
Epoch: 60 | Loss: 0.31406, Accuracy: 92.25% | Test loss: 0.34522, Test acc: 90.00%
Epoch: 70 | Loss: 0.30882, Accuracy: 92.38% | Test loss: 0.34156, Test acc: 90.00%
Epoch: 80 | Loss: 0.30361, Accuracy: 92.75% | Test loss: 0.33790, Test acc: 90.00%
Epoch: 90 | Loss: 0.29847, Accuracy: 92.88% | Test loss: 0.33419, Test acc: 90.00%
Epoch: 100 | Loss: 0.29340, Accuracy: 93.12% | Test loss: 0.33047, Test acc: 90.00%
Epoch: 110 | Loss: 0.28836, Accuracy: 93.12% | Test loss: 0.32668, Test acc: 90.00%
Epoch: 120 | Loss: 0.28344, Accuracy: 93.12% | Test loss: 0.32291, Test acc: 90.00%
Epoch: 130 | Loss: 0.27847, Accuracy: 93.50% | Test loss: 0.31906, Test acc: 90.00%
Epoch: 140 | Loss: 0.27358, Accuracy: 93.50% | Test loss: 0.31508, Test acc: 90.00%
Epoch: 150 | Loss: 0.26866, Accuracy: 93.50% | Test loss: 0.31101, Test acc: 90.50%
Epoch: 160 | Loss: 0.26372, Accuracy: 93.62% | Test loss: 0.30680, Test acc: 91.00%
Epoch: 170 | Loss: 0.25888, Accuracy: 93.88% | Test loss: 0.30235, Test acc: 91.00%
Epoch: 180 | Loss: 0.25390, Accuracy: 94.00% | Test loss: 0.29788, Test acc: 91.00%
Epoch: 190 | Loss: 0.24888, Accuracy: 94.00% | Test loss: 0.29310, Test acc: 91.00%
Epoch: 200 | Loss: 0.24392, Accuracy: 94.12% | Test loss: 0.28803, Test acc: 91.00%
Epoch: 210 | Loss: 0.23902, Accuracy: 94.25% | Test loss: 0.28296, Test acc: 91.00%
Epoch: 220 | Loss: 0.23402, Accuracy: 94.50% | Test loss: 0.27763, Test acc: 91.00%
Epoch: 230 | Loss: 0.22910, Accuracy: 94.75% | Test loss: 0.27222, Test acc: 91.00%
Epoch: 240 | Loss: 0.22411, Accuracy: 94.88% | Test loss: 0.26646, Test acc: 91.50%
...
Epoch: 460 | Loss: 0.11658, Accuracy: 99.62% | Test loss: 0.13017, Test acc: 100.00%
Epoch: 470 | Loss: 0.11262, Accuracy: 99.88% | Test loss: 0.12518, Test acc: 100.00%
Epoch: 480 | Loss: 0.10882, Accuracy: 99.88% | Test loss: 0.12047, Test acc: 100.00%
Epoch: 490 | Loss: 0.10512, Accuracy: 99.88% | Test loss: 0.11594, Test acc: 100.00%
CirclePredictionNonLinear(
  (layer1): Linear(in_features=2, out_features=8, bias=True)
  (layer2): Linear(in_features=8, out_features=5, bias=True)
  (layer3): Linear(in_features=5, out_features=3, bias=True)
  (layer4): Linear(in_features=3, out_features=1, bias=True)
  (relu): ReLU()
)

Now that it has achieved 100% let’s call once more

Output exceeds the size limit. Open the full output data in a text editorEpoch: 0 | Loss: 0.10161, Accuracy: 99.88% | Test loss: 0.11168, Test acc: 100.00%
Epoch: 10 | Loss: 0.09823, Accuracy: 99.88% | Test loss: 0.10765, Test acc: 100.00%
Epoch: 20 | Loss: 0.09499, Accuracy: 99.88% | Test loss: 0.10386, Test acc: 100.00%
Epoch: 30 | Loss: 0.09188, Accuracy: 99.88% | Test loss: 0.10025, Test acc: 100.00%
Epoch: 40 | Loss: 0.08889, Accuracy: 99.88% | Test loss: 0.09676, Test acc: 100.00%
Epoch: 50 | Loss: 0.08603, Accuracy: 99.88% | Test loss: 0.09340, Test acc: 100.00%
Epoch: 60 | Loss: 0.08329, Accuracy: 99.88% | Test loss: 0.09023, Test acc: 100.00%
Epoch: 70 | Loss: 0.08070, Accuracy: 99.88% | Test loss: 0.08724, Test acc: 100.00%
Epoch: 80 | Loss: 0.07821, Accuracy: 99.88% | Test loss: 0.08435, Test acc: 100.00%
Epoch: 90 | Loss: 0.07582, Accuracy: 99.88% | Test loss: 0.08151, Test acc: 100.00%
Epoch: 100 | Loss: 0.07352, Accuracy: 99.88% | Test loss: 0.07884, Test acc: 100.00%
Epoch: 110 | Loss: 0.07132, Accuracy: 99.88% | Test loss: 0.07631, Test acc: 100.00%
Epoch: 120 | Loss: 0.06922, Accuracy: 99.88% | Test loss: 0.07385, Test acc: 100.00%
Epoch: 130 | Loss: 0.06723, Accuracy: 99.88% | Test loss: 0.07157, Test acc: 100.00%
Epoch: 140 | Loss: 0.06532, Accuracy: 99.88% | Test loss: 0.06941, Test acc: 100.00%
Epoch: 150 | Loss: 0.06350, Accuracy: 99.88% | Test loss: 0.06734, Test acc: 100.00%
Epoch: 160 | Loss: 0.06177, Accuracy: 99.88% | Test loss: 0.06538, Test acc: 100.00%
Epoch: 170 | Loss: 0.06011, Accuracy: 99.88% | Test loss: 0.06351, Test acc: 100.00%
Epoch: 180 | Loss: 0.05853, Accuracy: 100.00% | Test loss: 0.06173, Test acc: 100.00%
Epoch: 190 | Loss: 0.05701, Accuracy: 100.00% | Test loss: 0.06003, Test acc: 100.00%
Epoch: 200 | Loss: 0.05555, Accuracy: 100.00% | Test loss: 0.05842, Test acc: 100.00%
Epoch: 210 | Loss: 0.05416, Accuracy: 100.00% | Test loss: 0.05687, Test acc: 100.00%
Epoch: 220 | Loss: 0.05282, Accuracy: 100.00% | Test loss: 0.05540, Test acc: 100.00%
Epoch: 230 | Loss: 0.05154, Accuracy: 100.00% | Test loss: 0.05400, Test acc: 100.00%
Epoch: 240 | Loss: 0.05031, Accuracy: 100.00% | Test loss: 0.05264, Test acc: 100.00%
...
Epoch: 460 | Loss: 0.03187, Accuracy: 100.00% | Test loss: 0.03296, Test acc: 100.00%
Epoch: 470 | Loss: 0.03130, Accuracy: 100.00% | Test loss: 0.03237, Test acc: 100.00%
Epoch: 480 | Loss: 0.03076, Accuracy: 100.00% | Test loss: 0.03181, Test acc: 100.00%
Epoch: 490 | Loss: 0.03023, Accuracy: 100.00% | Test loss: 0.03126, Test acc: 100.00%
CirclePredictionNonLinear(
  (layer1): Linear(in_features=2, out_features=8, bias=True)
  (layer2): Linear(in_features=8, out_features=5, bias=True)
  (layer3): Linear(in_features=5, out_features=3, bias=True)
  (layer4): Linear(in_features=3, out_features=1, bias=True)
  (relu): ReLU()
)

You can observe that the accuracy at 0th epoch is continuosly increasing and achieves 100% directly for the last one , now the data can’t be too simple as the very first inital call it was barely in 50s

5

Nevermind, I got the issue :slightly_smiling_face: , I was asking dumb question , I thought everytime I call the train_model function it will start with random weights but it wont, because I am not initialising my model everytime I call train_model function, hence when I called the train_model function 4 times I trained the model for 2000 epochs.

Thanks a lot for answering also I apologize for this question

6

Yes, that’s exactly the reason.
You are passing the model to the train_model function and are returning it afterwards.
The model’s parameters will be trained and kept which is why you are continuing the training.

No need to apologize as this discussion board is a good place to ask questions and learn more about PyTorch :wink:

2 Likes