PredRNN implementation using Tensorflow.

This is an implementation of PredRNN using Tensorflow. The implementation is based on:

• PredRNN: recurrent neural networks for predictive learning using spatiotemporal LSTMs
• PredRNN: A Recurrent Neural Network for Spatiotemporal Predictive Learning
• Official implementation in Pytorch
• kami93/PredRNN

Specifically, this repo implements the second paper: (stacked) spatial temporal cell and reversed scheduled sampling. There is an implementation for decouple loss; however, due to a bug with Keras's add_loss() function, enabling the loss will raise error during runtime.

Accuracy

Performance (MSE) on Moving MNIST dataset (results produced by running scripts/moving_mnist_all_batch.sh 5)_

	Run #1	Run #2	Run #3	Run #4	Run #5	Mean	Std
ConvLSTM	0.03618	0.00951	0.04391	0.00956	0.03485	0.026802	0.01613769407
PredRNN without Scheduled Sampling	0.00705	0.00728	0.00703	0.00723	0.00733	0.007184	0.0001363084737
PredRNN with Linear Scheduled Sampling	0.00798	0.00741	0.01012	0.01468	0.00822	0.009682	0.00297355343

Performance (MSE) on Moving MNIST dataset with different scheduled sampling strategies.

	Run #1	Run #2	Run #3	Run #4	Run #5	Mean	Std
PredRNN with Linear Scheduled Sampling	0.01254	0.01095	0.01231	0.01032	0.01071	0.011366	0.0009958564154
PredRNN with Expo Scheduled Sampling	0.00816	0.00868	0.01382	0.01027	0.01094	0.010374	0.002234810954
PredRNN with Sigmoid Scheduled Sampling	0.00938	0.00999	0.00797	0.01007	0.00775	0.009032	0.001105404903

Installation

The repository can be installed as module using either pip or poetry.

Development

The repo uses poetry to manage dependencies. To install all the development dependencies, use poetry install.

Examples

In the folder examples, there are some notebooks that create PredRNN model and use it on Moving MNIST dataset.

These examples can be opened as Jupyter notebooks using Jupytext. These development dependencies will be installed along with required dependencies with poetry install.

Since these examples are resource-intensive and time-consuming, it is recommended to run these examples in GPU clusters. In the scripts folder, there are some scripts to run these in IU's Carbornate GPU clusters (you will authorization to use these clusters, of course!). Specifically,

• sbatch scripts/moving_mnist.sbatch will schedule a batch job to run example examples/moving_mnist_predrnn.py. The output notebook will be examples/moving_mnist_predrnn_<job_id>.ipynb and the model will be saved as saved_models/moving_mnist_predrnn_<job_id>.
• Similarly for sbatch scripts/moving_mnist_ss.sbatch.
• scripts/moving_mnist_all_batch.sh <nb_of_runs> will submit a series of jobs (3 * <nb_of_runs>) to the GPU clusters. There will be <nb_of_runs> jobs for each combination of PredRNN/ConvLSTM and with/without scheduled sampling. The goal of this script is to show whether PredRNN is better than ConvLSTM at predicting future frames for Moving MNIST dataset.