Hi Dr. Denton,
Thank you very much for releasing your code. I have had fun using SVG to generate images.
Just one quick question, I noticed that the FP code has this:
|
# plot some stuff |
|
frame_predictor.eval() |
|
encoder.eval() |
|
decoder.eval() |
|
posterior.eval() |
|
x = next(testing_batch_generator) |
|
plot(x, epoch) |
|
plot_rec(x, epoch) |
whereas LP has this:
|
# plot some stuff |
|
frame_predictor.eval() |
|
#encoder.eval() |
|
#decoder.eval() |
|
posterior.eval() |
|
prior.eval() |
|
|
|
x = next(testing_batch_generator) |
|
plot(x, epoch) |
|
plot_rec(x, epoch) |
Given that the encoder and decoder have batch normalization layers and are used during inference time, I believe LP should also have eval() enabled? Is my understanding correct?
Hi Dr. Denton,
Thank you very much for releasing your code. I have had fun using SVG to generate images.
Just one quick question, I noticed that the FP code has this:
svg/train_svg_fp.py
Lines 319 to 326 in 3f19f0b
whereas LP has this:
svg/train_svg_lp.py
Lines 359 to 368 in 3f19f0b
Given that the encoder and decoder have batch normalization layers and are used during inference time, I believe LP should also have eval() enabled? Is my understanding correct?