txt2video

Generating videos, conditioned on text with GANs. Honours thesis. This implementation contains the following paper implementations:

1. To Create What You Tell
2. TGAN
3. TGANv2

With modifications to the last two to condition on text. Text is encoded with a Bi-LSTM which has been pretrained to generate the next token - which from memmory is the same methodlogy as "To Create What You Tell".

Additionally to capture motion in the discriminator more effectively, non-local blocks are utilised (self attention).

Conditional information is introduced similar to StackGAN++. Relativisitc losses are used.

For discriminator we compare the pairs:

• ${(x_r, c_r), (x_f, c_r)}$
• ${(x_r, c_f), (x_f, c_r)}$

For generator we only compare first pair above.

• x_r is real video
• x_f is fake video
• c_r is caption correctly associated to video
• c_f is caption not associated to video

Standard GAN loss is preferred due to 1 discrim step to 1 generator step.

Alternatively I did experiment with non-relativisitc loss, via the following intiution:

• $(x_r, c_r)$ => [should be associated]
• $(x_f, c_r)$ => [should not be]
• $(x_r, c_f)$ => [should not be]
• $(x_f, c_f)$ => not used

The last could optionally be used to learn but doesn't seem to be necessary (at least emprically)

Datasets & Results

Three datasets are used.

1. Synthetic MNIST for moving digits
2. MSR Video to Text (MSRVDC) dataset
3. Custom dataset with videos scraped from reddit

Synthetic MNIST

MNIST with generated data from txt2vid/data/synthetic/generate.py

TCWYT Baseline:

From top to bottom:

'<start> digit 9 is left and right<end>'

'<start> digit 8 is right and left<end>'

'<start> digit 8 is bottom and top<end>'

'<start> digit 4 is top and bottom<end>'

MSRVDC

TCWYT (Conditional)

Bottom is ground truth for both of below

<start> a woman is saying about how to make vegetable tofu <unk> <end>'

<start> the person is cooking <end>'

TGANv2 (Unconditional)

TGANv2 (Conditional) + My Modifications

'<start> the man poured preserves over the chicken<end>'

'<start> a person is dicing and onion<end>'

'<start> a woman is peeling a large shrimp in a glass bowl of water<end>'

reddit-videos

See https://github.com/miguelmartin75/reddit-videos

Didn't end up training on this dataset :/

Details & References

Please see thesis.pdf for more details, references, etc.