This project implements a generative adversarial network to predict future frames of video, as detailed in "Deep Multi-Scale Video Prediction Beyond Mean Square Error" by Mathieu, Couprie & LeCun. Their official code (using Torch) can be found here.
Adversarial generation uses two networks – a generator and a discriminator – to improve the sharpness of generated images. Given the past four frames of video, the generator learns to generate accurate predictions for the next frame. Given either a generated or a real-world image, the discriminator learns to correctly classify between generated and real. The two networks "compete," with the generator attempting to fool the discriminator into classifying its output as real. This forces the generator to create frames that are very similar to what real frames in the domain might look like.
I trained and tested my network on a dataset of frame sequences from Ms. Pac-Man. To compare adversarial training vs. non-adversarial, I trained an adversarial network for 500,000 steps on both the generator and discriminator, and I trained a non-adversarial network for 1,000,000 steps (as the non-adversarial network runs about twice as fast). Training took around 24 hours for each network, using a GTX 980TI GPU.
In the following examples, I ran the networks recursively for 64 frames. (i.e. The input to generate the first frame was [input1, input2, input3, input4], the input to generate the second frame was [input2, input3, input4, generated1], etc.). As the networks are not fed actions from the original game, they cannot predict much of the true motion (such as in which direction Ms. Pac-Man will turn). Thus, the goal is not to line up perfectly with the ground truth images, but to maintain a crisp and likely representation of the world.
The following example exhibits how quickly the non-adversarial network becomes fuzzy and loses definition of the sprites. The adversarial network exhibits this behavior to an extent, but is much better at maintaining sharp representations of at least some sprites throughout the sequence:
This example shows how the adversarial network is able to keep a sharp representation of Ms. Pac-Man around multiple turns, while the non-adversarial network fails to do so:
While the adversarial network is clearly superior in terms of sharpness and consistency over time, the non-adversarial network does generate some fun/spectacular failures:
Using the error measurements outlined in the paper (Peak Signal to Noise Ratio and Sharp Difference) did not show significant difference between adversarial and non-adversarial training. I believe this is because sequential frames from the Ms. Pac-Man dataset have no motion in the majority of pixels, while the original paper was trained on real-world video where there is motion in much of the frame. Despite this, it is clear that adversarial training produces a qualitative improvement in the sharpness of the generated frames, especially over long time spans. You can view the loss and error statistics by running tensorboard --logdir=./Results/Summaries/ from the root of this project.
Clone or download this repository.
Prepare your data:
Data/ in the root of this project for default behavior. Otherwise, you will need to specify your data location using the options outlined in parts 3 and 4.Process training data:
python process_data from the Code/ directory with the following options:
-n/--num_clips= <# clips to process for training> (Default = 5000000)
-t/--train_dir= <Directory of full training frames>
-c/--clips_dir= <Save directory for processed clips>
(I suggest making this a hidden dir so the filesystem doesn't freeze
with so many files. DON'T `ls` THIS DIRECTORY!)
-o/--overwrite (Overwrites the previous data in clips_dir)
-H/--help (prints usage)Train/Test:
-l option. (e.g. python avg_runner.py -l ./Models/Adversarial/model.ckpt-500000).python avg_runner.py from the Code/ directory with the following options:
-l/--load_path= <Relative/path/to/saved/model>
-t/--test_dir= <Directory of test images>
-r--recursions= <# recursive predictions to make on test>
-a/--adversarial= <{t/f}> (Whether to use adversarial training. Default=True)
-n/--name= <Subdirectory of ../Data/Save/*/ in which to save output of this run>
-O/--overwrite (Overwrites all previous data for the model with this save name)
-T/--test_only (Only runs a test step -- no training)
-H/--help (Prints usage)
--stats_freq= <How often to print loss/train error stats, in # steps>
--summary_freq= <How often to save loss/error summaries, in # steps>
--img_save_freq= <How often to save generated images, in # steps>
--test_freq= <How often to test the model on test data, in # steps>
--model_save_freq= <How often to save the model, in # steps>Why don't you train on patches larger then 32x32? Why not train on the whole image?
Memory usage. Since the discriminator has fully-connected layers after the convolutions, the output of the last convolution must be flattened to connect to the first fully-connected layer. The size of this output is dependent on the input image size, and blows up really quickly (e.g. For an input size of 64x64, going from 128 feature maps to a fully connected layer with 512 nodes, you need a connection with 64 64 128 * 512 = 268,435,456 weights). Because of this, training on patches larger than 32x32 causes an out-of-memory error (at least on my machine).
Luckily, you only need the discriminator for training, and the generator network is fully convolutional, so you can test the weights you trained on 32x32 images over images of any size (which is why I'm able to do generations for the entire Ms. Pac-Man board).