Baseline.md

Baselines Configurations

We have provided code to reimplement baselines to ablate predictor, decoder, and encoder correpondingly.

The baselines are configured as the following:

method	mod	graph	gconv_unit_type	encoder	decoder
CVP (default)	cvp	fact_gc	n2e2n	traj	comb_late
No-Factor [23]	noFactor	fact_fc	-	traj	noFactor
No-Edge	cvp	fact_gc	noEdge	traj	comb_late
Early-Feat	cvp	fact_gc	n2e2n	traj	comb_early
Mid-Feat	cvp	fact_gc	n2e2n	traj	comb_mid
Pixel	cvp	fact_gc	n2e2n	traj	cPix
No-Z	cvp	fact_gc	n2e2n	noZ	comb_late
FP [6]	cvp	fact_gc	n2e2n	fp	comb_late
LP [6]	lp	fact_gc	n2e2n	lp	comb_late

To run baseline, just set the non-default flag (in bold) to the corresponding one, since the default configuration is set to train our CVP.

Anyway, the remaining part kindly provides straightforward command which gives you the same config as the table shows:

Entity Predictor

• No-Factor: Generating Videos with Scene Dynamics, Vondrick et al. in NeurIPS'16

python train.py --gpu ${GPU_ID} --mod noFactor  --graph fact_fc --decoder noFactor

• No-Edge:

python train.py --gpu ${GPU_ID} --gconv_unit_type noEdge

Frame Decoder

• Early-Feat

python train.py --decoder comb_early

• Mid-Feat

python train.py --decoder comb_mid

• Late-Feat (ours)

python train.py --decoder comb_late

• Pixel

python train.py --decoder cPix

Latent Representations and Encoder

The --encoder could be traj(ours), noZ, fp, or lp.

• No-Z baseline

python train.py --encoder noZ

• FP (Fix-Prior): Stochastic Video Generation with a Learned Prior, Denton et al. in ICML'18

python train.py --encoder fp

• LP (Learned-Prior): Stochastic Video Generation with a Learned Prior, Denton et al. in ICML'18

python train.py --mod lp --encoder lp

Testing

All models could directly be evaluated by running:

python test.py --checkpoint ${PATH_TO_MODEL} --test_mod best_100