Denoise_AutoEncoder

The implement of layer-wise training denoise autoencoder in pytorch.

Compress the 5-minute K line in a day (48 dimensions) of stock in A-share market into a vector of 5 dimensions, through 3 pairs of encoder-decoder layer with denoising.

Requirements

• Matplotlib
• Pandas
• Pytorch
• Numpy

Network

AutoEncoder_1 (
  (encoder1): Linear (48 -> 20)
  (decoder1): Linear (20 -> 48)
)  

AutoEncoder_2 (
  (encoder1): Linear (48 -> 20)
  (encoder2): Linear (20 -> 10)
  (decoder2): Linear (10 -> 20)
  (decoder1): Linear (20 -> 48)
)  

AutoEncoder_3 (
  (encoder1): Linear (48 -> 20)
  (encoder2): Linear (20 -> 10)
  (encoder3): Linear (10 -> 5)
  (decoder3): Linear (5 -> 10)
  (decoder2): Linear (10 -> 20)
  (decoder1): Linear (20 -> 48)
)  

Each layers of above network is simple full connection.

Usage

1. Run train_net_1.py to train the AutoEncoder_1.
2. Run train_net_2.py, which copies the parameters of encoder1/decoder1 from AutoEncoder_1 as the fixed parameters for AutoEncoder2 and only train the encoder2/decoder2.
3. Run train_net_3.py, which copies the parameters of encoder1/decoder1 and encoder2/decoder2 from AutoEncoder_2 as the fixed parameters for AutoEncoder3 and only train the encoder3/decoder3.
4. Run make_encoder.py , truncate the AutoEncoder_3 and retain the front part of it as an compression encoder which converts a 48-dimensional vector into a 5-dimension vector.
5. Run encode_vision.py to show the trend chart of stocks recovered from compressed encoding.

Result

Loss sequence on train and test set of each training stage.

The original 5-minute K line sequnce and the recovered sequence from compressed vector.