Skip to content

Fork of code and models accompanying "Deep Predictive Coding Networks for Video Prediction and Unsupervised Learning"

License

Notifications You must be signed in to change notification settings

mschart/prednet_python3

 
 

Repository files navigation

This is a fork of PredNet developed by Bill Lotter, Gabriel Kreiman, and David Cox which has been modified to use Python 3.6 and the most recent stable versions of Tensorflow and Keras. Worth noting that TF is only compatible with CUDA up to 9.0 and the associated CUDNN release.

Below is the original readme from Bill Lotter's repository.

prednet

Code and models accompanying Deep Predictive Coding Networks for Video Prediction and Unsupervised Learning by Bill Lotter, Gabriel Kreiman, and David Cox.

The PredNet is a deep recurrent convolutional neural network that is inspired by the neuroscience concept of predictive coding (Rao and Ballard, 1999; Friston, 2005). Check out example prediction videos here.

The architecture is implemented as a custom layer1 in Keras. Code and model data is now compatible with Keras 2.0. Specifically, it has been tested on Keras 2.0.6 with Theano 0.9.0, Tensorflow 1.2.1, and Python 2.7 (for your convenience, we have added an environment.yml file for setting up your python environment). For previous versions of the code compatible with Keras 1.2.1, use fbcdc18. To convert old PredNet model files and weights for Keras 2.0 compatibility, see convert_model_to_keras2 in keras_utils.py.

KITTI Demo

Code is included for training the PredNet on the raw KITTI dataset. We include code for downloading and processing the data, as well as training and evaluating the model. The preprocessed data and can also be downloaded directly using download_data.sh and the trained weights by running download_models.sh. The model download will include the original weights trained for t+1 prediction, the fine-tuned weights trained to extrapolate predictions for multiple timesteps, and the "Lall" weights trained with an 0.1 loss weight on upper layers (see paper for details).

Steps

  1. Download/process data

    python process_kitti.py

    This will scrape the KITTI website to download the raw data from the city, residential, and road categories (~165 GB) and then process the images (cropping, downsampling). Alternatively, the processed data (~3 GB) can be directly downloaded by executing download_data.sh

  2. Train model

    python kitti_train.py

    This will train a PredNet model for t+1 prediction. See Keras FAQ on how to run using a GPU. To download pre-trained weights, run download_models.sh

  3. Evaluate model

    python kitti_evaluate.py

    This will output the mean-squared error for predictions as well as make plots comparing predictions to ground-truth.

Feature Extraction

Extracting the intermediate features for a given layer in the PredNet can be done using the appropriate output_mode argument. For example, to extract the hidden state of the LSTM (the "Representation" units) in the lowest layer, use output_mode = 'R0'. More details can be found in the PredNet docstring.

Multi-Step Prediction

The PredNet argument extrap_start_time can be used to force multi-step prediction. Starting at this time step, the prediction from the previous time step will be treated as the actual input. For example, if the model is run on a sequence of 15 timesteps with extrap_start_time = 10, the last output will correspond to a t+5 prediction. In the paper, we train in this setting starting from the original t+1 trained weights (see kitti_extrap_finetune.py), and the resulting fine-tuned weights are included in download_models.sh. Note that when training with extrapolation, the "errors" are no longer tied to ground truth, so the loss should be calculated on the pixel predictions themselves. This can be done by using output_mode = 'prediction', as illustrated in kitti_extrap_finetune.py.

Additional Notes

When training on a new dataset, the image size has to be divisible by 2^(nb of layers - 1) because of the cyclical 2x2 max-pooling and upsampling operations.


1 Note on implementation: PredNet inherits from the Recurrent layer class, i.e. it has an internal state and a step function. Given the top-down then bottom-up update sequence, it must currently be implemented in Keras as essentially a 'super' layer where all layers in the PredNet are in one PredNet 'layer'. This is less than ideal, but it seems like the most efficient way as of now. We welcome suggestions if anyone thinks of a better implementation.

About

Fork of code and models accompanying "Deep Predictive Coding Networks for Video Prediction and Unsupervised Learning"

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 98.7%
  • Shell 1.3%