Hi again

I’ve finally tested it and you’re totally right. I get very high training accuracies (89 %) and low validation accuracies (38%).

I think my problem is that I want the network to give me a classification each 5 time samples. Which means that the same class is classified many times in a row for long time series. I can see that the first 10-11 classifications attempts are wrong but then the model learns to classify the correct class and for the remaining part of the time series the class is correct.

However, it resets and start over when a new time series is used. I calculate loss and update optimizer each time it gives a classification. Since it give many classifications for each time series the model quickly learn the single time series but forget everything else.

```
loss = criterion(output, y)
# Reset gradient
optimizer.zero_grad()
loss.backward()
optimizer.step()
```

Anything to combat this? Does it make sense to simply call model() and from that calculate the predicted class and not update optimizer and calculate loss?

```
output = model(nbatch,batch_size)
```

And then maybe calculate loss and update optimizer at fewer intervals (like every 100 or 1000?)