PredNet PyTorch Implementation

References

• Paper: Deep Predictive Coding Networks for Video Prediction and Unsupervised Learning
• Original Keras code: coxlab/prednet liscenced under the MIT Liscense.

---

Implementation

Directory

• prednet_pytorch_mnist/

Training

• File: mnist_train_all.py
• Parameters:
  • num_epochs: 6
  • batch_size: 2
  • lr: 0.001
  • nt: 20 (sequence length)
  • n_train_seq: 7000
  • n_val_seq: 1000

Dataset

• File: mnist_data.py
• Parameters:
  • nt: 20 (frames per sequence)
  • image_size: (64, 64)
  • channels: 3 (RGB)

Model Hyperparameters

• loss_mode: 'L_0' or 'L_all' (default: 'L_all')
• peephole: False
• lstm_tied_bias: False
• gating_mode: 'mul' or 'sub' (default: 'mul')
• A_channels & R_channels: (3, 48, 96, 192)

Output

• Best model: prednet-L_all-mul-peepFalse-tbiasFalse-best.pt
• Loss history: original_mnist-prednet-L_all-mul-peepFalse-tbiasFalse-loss_history.jsonl

Results

• Example predicted frames:

License

This project is licensed under the Apache 2.0 License.

Third-Party Notices

This project is derived from the following works (see NOTICE for full attribution):

• coxlab/prednet — MIT License, Copyright (c) 2017 coxlab.