This repository contains instructions and code for a release version of the work on hierarchical state space models for VDSM: Unsupervised Video Disentanglement with State-Space Modeling and Deep Mixtures of Experts (CVPR '21, in print - Vowels, Camgoz, Bowden). A preprint can be found here: https://arxiv.org/abs/2103.07292
Note that these are random samples from the test set, so there may be inactive GIFs if the action is neutral.
Pendulum Swing Generation (60 timesteps)
MovingMNIST [3] Generation (16 timesteps)
Animated Sprites [1] Generation (8 timesteps)
Animated Sprites Action Transfer (first column is target action):
MUG-FED [2] Generation (20 timesteps)
MUG-FED Action Transfer (first column is target action):
python 3.7.6
numpy 1.19.2
pandas 1.0.5
torch 1.6.0
torchvision 0.7.0
pyro 1.4.0
sklearn 0.23.2
imageio 2.9.0
scipy 1.5.3
pytz 2020.1
natsort 7.0.1
datetime
antialiased_cnns
nonechucks
os argparse
N.B. Currently, VDSM is designed to train on GPU.
MUG: Go to https://mug.ee.auth.gr/fed/ and obtain a license.
Use OpenFace 2.0, https://github.com/TadasBaltrusaitis/OpenFace, to crop and align the images (no mask) to 112x112x3 (VDSM does further downscaling to 64x64).
Save the images into a npz with ndarray of shape [num_sequences_in_dataset] such that each sequence has a shape of [seq_len, 112, 112, 3].
Save the labels into a npz array with shape [num_sequences_in_dataset, 2] where the first column is identity label and the second column is the action label (and where both have been categorically encoded e.g. action {0,1...8}).
Put these files in VDSM_release/data/MUG-FED/
SPRITES:
For sprites, additional packages may be required (check create_lpc_dataset.py which is adapted from https://github.com/yatindandi/Disentangled-Sequential-Autoencoder).
cd ./data/sprites
./data/sprites/create_lpc_dataset.py
This will download the sprites data into train and test folders.
PENDULUM:
The code used to create this data is provided in ./data/pendulum_data_creation.ipynb
cd VDSM_release
MUG:
python3 main.py --RUN release_test --rnn_layers 3 --rnn_dim 512 --bs 20 --seq_len 20 --epochs 250 --bs_per_epoch 50 \
--num_test_ids 10 --dataset_name MUG-FED --model_save_interval 50 \
--train_VDSMSeq False --train_VDSMEncDec True --model_test_interval 10 \
--anneal_start_dynamics 0.1 --anneal_end_dynamics 0.6 --anneal_frac_dynamics 1 --lr_VDSMEncDec 0.001 --lr_resume 0.0008 --z_dim 30 --n_e_w 15 \
--dynamics_dim 50 --test_temp_id 10 --temp_id_end 10 --temp_id_start 1 --temp_id_frac 2 --anneal_end_id 1 --anneal_start_id 0.1 \
--anneal_frac_id 3 --anneal_start_t 30.0 --anneal_mid_t1 0.2 --anneal_mid_t2 0.4 --anneal_end_t 1 --anneal_t_midfrac1 0.5 --anneal_t_midfrac2 0.8 \
--rnn_dropout 0.2
Sprites:
python3 main.py --RUN release_test_sprites --rnn_layers 3 --rnn_dim 512 --bs 20 --seq_len 16 --epochs 200 --bs_per_epoch 50 \
--num_test_ids 12 --dataset_name sprites --model_save_interval 50 \
--train_VDSMSeq False --train_VDSMEncDec True --model_test_interval 10 \
--anneal_start_dynamics 0.1 --anneal_end_dynamics 0.6 --anneal_frac_dynamics 1 --lr_VDSMEncDec 0.001 --lr_resume 0.0008 --z_dim 30 --n_e_w 40 \
--dynamics_dim 50 --test_temp_id 10 --temp_id_end 10 --temp_id_start 1 --temp_id_frac 2 --anneal_end_id 1 --anneal_start_id 0.1 \
--anneal_frac_id 3 --anneal_start_t 30.0 --anneal_mid_t1 0.2 --anneal_mid_t2 0.4 --anneal_end_t 1 --anneal_t_midfrac1 0.5 --anneal_t_midfrac2 0.8 \
--rnn_dropout 0.2
If you encounter NaN during training, you can try resuming from where you last checkpointed at (this is taken care of automatically).
You need to manually delete the model.pth files in the models folder if you want to start training from scratch.
If you get NaN again, you can add e.g. --lr_resume 0.0008 argument to reduce the learning rate.
To train the sequence network, you have to specify the {}_VDSM_EncDec.pth number to use as the pretrained model. (Usually it will be (epochs-1) for the number of epochs used during pretraining)
MUG:
python3 main.py --RUN release_test --rnn_layers 3 --rnn_dim 512 --bs 20 --seq_len 20 --epochs 200 --bs_per_epoch 2 \
--num_test_ids 6 --dataset_name MUG-FED --model_save_interval 50 --pretrained_model_VDSMEncDec 249\
--train_VDSMSeq True --train_VDSMEncDec False --model_test_interval 10 \
--anneal_start_dynamics 0.1 --anneal_end_dynamics 1.0 --anneal_frac_dynamics 1 --lr_VDSMSeq 0.001 --z_dim 30 --n_e_w 15 \
--dynamics_dim 50 --test_temp_id 10.0 --temp_id_end 10.0 --temp_id_start 10.0 --temp_id_frac 1 --anneal_end_id 1.0 --anneal_start_id .1 \
--anneal_frac_id 3 --anneal_start_t 0.1 --anneal_mid_t1 0.4 --anneal_mid_t2 0.4 --anneal_end_t 1 --anneal_t_midfrac1 0.5 --anneal_t_midfrac2 0.8 \
--rnn_dropout 0.2
SPRITES Note that the Sprites dataset (due in part to its synthetic nature), has limited variability to separate identity and pose. To improve the consistency of results, we have added a feature in dataset.py which pairs up two random actions (from sprites with the same identity). This produces enough variation across pose to consistently disentangle it from identity.
python3 main.py --RUN release_test_sprites --rnn_layers 3 --rnn_dim 512 --bs 20 --seq_len 8 --epochs 200 --bs_per_epoch 50 \
--num_test_ids 12 --dataset_name sprites --model_save_interval 50 --pretrained_model_VDSMEncDec 199\
--train_VDSMSeq True --train_VDSMEncDec False --model_test_interval 10 \
--anneal_start_dynamics 0.1 --anneal_end_dynamics 1.0 --anneal_frac_dynamics 1 --lr_VDSMSeq 0.001 --z_dim 30 --n_e_w 40 \
--dynamics_dim 50 --test_temp_id 10.0 --temp_id_end 10.0 --temp_id_start 10.0 --temp_id_frac 1 --anneal_end_id 1.0 --anneal_start_id .1 \
--anneal_frac_id 3 --anneal_start_t 0.1 --anneal_mid_t1 0.4 --anneal_mid_t2 0.4 --anneal_end_t 1 --anneal_t_midfrac1 0.5 --anneal_t_midfrac2 0.8 \
--rnn_dropout 0.2
Testing: just set both train_VDSMSeq and train_VDSMEncDec to False.
For FID we used Evan: https://github.com/raahii/evan
and the resnext-101-kinetics.pth model and code available from https://github.com/kenshohara/video-classification-3d-cnn-pytorch
(you may need to download this model and put it in the hidden folder .evan/.
To reproduce the results, you may need to replace the prepare_inception_model function in the evan package with:
def prepare_inception_model(weight_dir: Path = CACHE_DIR, device: torch.device = torch.device("cpu")):
filename = "resnext-101-kinetics.pth"
weight_path = weight_dir / filename
model = resnet.resnet101(num_classes=400, shortcut_type='B', cardinality=32,
sample_size=112, sample_duration=16,
last_fc=False)
model_data = torch.load(str(weight_path), map_location="cpu")
fixed_model_data = OrderedDict()
for key, value in model_data["state_dict"].items():
new_key = key.replace("module.", "")
fixed_model_data[new_key] = value
model.load_state_dict(fixed_model_data, strict=False)
model = model.to(device)
model.eval()
return modelYou may also need to replace the code for resnet.resnet101 model with the model in https://github.com/kenshohara/video-classification-3d-cnn-pytorch
For this we adapted the code from # see https://github.com/yatindandi/Disentangled-Sequential-Autoencoder/blob/master/classifier.py
It can be found in classifier_MUG.ipynb and classifier_sprites.ipynb.
To use these, you need to export the embeddings learned by the model using functions like these:
def extract_id_etc(self, test_images): # test images is shape [bs, seq_len, 3, 64, 64]
num_individuals, num_timepoints, pixels = (test_images.shape[0],
test_images.shape[1],
self.imsize ** 2 * self.nc)
unrav = test_images.view(num_individuals * num_timepoints, self.nc, self.imsize, self.imsize)
loc, _, ID, ID_scale = self.VDSM_Seq.image_enc.forward(unrav.to(self.dev))
loc = loc.view(num_individuals, num_timepoints, -1)
ID, _ = self.id_layers(ID, ID_scale)
ID = torch.mean(ID.view(num_individuals, num_timepoints, -1), 1).unsqueeze(1)
ID = ID * self.temp_id_end
ID_exp = torch.exp(ID)
ID = ID_exp / ID_exp.sum(-1).unsqueeze(-1)
seq = loc.permute(1, 0, 2)
return ID, seq, num_individuals, num_timepoints, pixels
def generate_sequences(self, test_images): # test images is shape [bs, seq_len, 3, 64, 64]
ID, seq, num_individuals, num_timepoints, pixels = self.extract_id_etc(test_images)
futures = self.VDSM_Seq.test_sequence(seq, num_timepoints).permute(1, 0, 2)
recon_gen = torch.zeros(num_individuals, num_timepoints, pixels, device=test_images.device)
for ind in range(num_individuals):
recon_gen[ind] = self.VDSM_Seq.image_dec.forward(futures[ind], ID[ind, 0].unsqueeze(1))
recon_gen = recon_gen.view(num_individuals, num_timepoints, self.nc, self.imsize, self.imsize)
return recon_gen
def generate_codes(self, test_images, labels, p): # test images is shape [bs, seq_len, 3, 64, 64]
ID, seq, num_individuals, num_timepoints, pixels = self.extract_id_etc(test_images)
dynamics = self.VDSM_Seq.return_dynamics(seq).detach().cpu().numpy()
recon_gen = self.generate_sequences(test_images).detach().cpu().numpy()
np.savez(os.path.join(self.code_path, 'id_{}.npz'.format(p)), ID.detach().cpu().numpy())
if labels is not None:
np.savez(os.path.join(self.code_path, 'labels_{}.npz'.format(p)), labels)
np.savez(os.path.join(self.code_path, 'dynamics_{}.npz'.format(p)), dynamics)
np.savez(os.path.join(self.code_path, 'test_images_{}.npz'.format(p)), test_images.detach().cpu().numpy())
np.savez(os.path.join(self.code_path, 'recon_images_{}.npz'.format(p)), recon_gen) These have not yet be integrated into the release code (see TODO below).
- Consolidate repeated functions.
- Create test_harness.py and consolidate with train_test.py
- Add pendulum dataset
- Add moving MNIST dataset
- Adapt for CPU training
- Implement Inter and Intra-Entropy for sprites
- Implement test harness
[1] http://gaurav.munjal.us/Universal-LPC-Spritesheet-Character-Generator/
[2] N. Aifanti, C. Papachristou, and A. Delopoulos. The MUG facial expression database.Proc. 11th Int. Workshop on Image Analysis for Multimedia Interactive Services, 2010.
[3] Srivastava, E. Mansimov, and R. Salakhutdinov. Un-supervised learning of video representations using LSTMs. arXiv:1502.04681v3, 2016.