Repository for frame prediction on the MovingMNIST dataset using seq2seq ConvLSTM following either of these guides:
Make sure you have the following libraries installed!
python=3.6.8
torch=1.1.0
torchvision=0.3.0
pytorch-lightning=0.7.1
matplotlib=3.1.3
tensorboard=1.15.0a20190708
-
Install the above libraries
-
Clone this repo
git clone https://github.com/holmdk/Video-Prediction-using-PyTorch.git
cd ./Video-Prediction-using-PyTorch
- Run main.py
python main.py
- Navigate to http://localhost:6006/ for visualizing results
The first row displays our predictions, the second row the ground truth and the third row the absolute error on a pixel-level. The first 8 columns are the input, followed by output in the final 8 columns. This matches the output from the Tensorboard logging.
After some iterations, we notice that our model is actually generating images of all zeros! This is a common issue people using ConvLSTM reports, however, do not be discouraged! Simply keep training the model, and you should start to see actual and plausible future predictions.
Now, we are actually starting to see actual predictions, however blurry they might be.
- Add video of predictions by model
- Implement other video prediction methods (feel free to contribute!)
- SVG
- PredRNN+
- E3D
- MIM