RNN prediction results do not match training results

saschalippert · July 30, 2019, 4:44pm

Hi,

my sine approximation during training works fine, while the prediction (which is feed by previous predictions, with a head start of 20 original values) does not approximate.

I guess there is a bug/misunderstanding in my prediction code A hint would be great.

gist.github.com

https://gist.github.com/saschalippert/36b72313afff86e00f3e10254fb4ff25

simplesine.py

from collections import deque
import torch
from torch import nn
import numpy as np
import matplotlib.pyplot as plt
from torch.utils.data import TensorDataset, DataLoader


class RNN(nn.Module):
    def __init__(self, input_size, output_size, hidden_size, n_layers):

This file has been truncated. show original

[Album] Imgur

Thanks

saschalippert · July 31, 2019, 5:22pm

Further testing showed that feeding the network with the predicted values (which have a slight deviation from the target values) cause to spiral the future values out of control. Using a more complex model and training for more episodes seems to solve the issue.