Commit Graph

124 Commits

Author SHA1 Message Date
James Betker
5e8b52f34c Misc changes 2020-07-10 09:45:48 -06:00
James Betker
5f2c722a10 SRG2 revival
Big update to SRG2 architecture to pull in a lot of things that have been learned:
- Use group norm instead of batch norm
- Initialize the weights on the transformations low like is done in RRDB rather than using the scalar. Models live or die by their early stages, and this ones early stage is pretty weak
- Transform multiplexer to use u-net like architecture.
- Just use one set of configuration variables instead of a list - flat networks performed fine in this regard.
2020-07-09 17:34:51 -06:00
James Betker
b2507be13c Fix up pixgan loss and pixdisc 2020-07-08 21:27:48 -06:00
James Betker
26a4a66d1c Bug fixes and new gan mechanism
- Removed a bunch of unnecessary image loggers. These were just consuming space and never being viewed
- Got rid of support of artificial var_ref support. The new pixdisc is what i wanted to implement then - it's much better.
- Add pixgan GAN mechanism. This is purpose-built for the pixdisc. It is intended to promote a healthy discriminator
- Megabatchfactor was applied twice on metrics, fixed that

Adds pix_gan (untested) which swaps a portion of the fake and real image with each other, then expects the discriminator
to properly discriminate the swapped regions.
2020-07-08 17:40:26 -06:00
James Betker
8a4eb8241d SRG3 work
Operates on top of a pre-trained SpineNET backbone (trained on CoCo 2017 with RetinaNet)

This variant is extremely shallow.
2020-07-07 13:46:40 -06:00
James Betker
0acad81035 More SRG2 adjustments.. 2020-07-06 22:40:40 -06:00
James Betker
086b2f0570 More bugs 2020-07-06 22:28:07 -06:00
James Betker
d4d4f85fc0 Bug fixes 2020-07-06 22:25:40 -06:00
James Betker
3c31bea1ac SRG2 architectural changes 2020-07-06 22:22:29 -06:00
James Betker
9a1c3241f5 Switch discriminator to groupnorm 2020-07-06 20:59:59 -06:00
James Betker
6beefa6d0c PixDisc - Add two more levels of losses coming from this gen at higher resolutions 2020-07-06 11:15:52 -06:00
James Betker
2636d3b620 Fix assertion error 2020-07-06 09:23:53 -06:00
James Betker
8f92c0a088 Interpolate attention well before softmax 2020-07-06 09:18:30 -06:00
James Betker
72f90cabf8 More pixdisc fixes 2020-07-05 22:03:16 -06:00
James Betker
a47a5dca43 Fix pixdisc bug 2020-07-05 21:57:52 -06:00
James Betker
d0957bd7d4 Alter weight initialization for transformation blocks 2020-07-05 17:32:46 -06:00
James Betker
16d1bf6dd7 Replace ConvBnRelus in SRG2 with Silus 2020-07-05 17:29:20 -06:00
James Betker
10f7e49214 Add ConvBnSilu to replace ConvBnRelu
Relu produced good performance gains over LeakyRelu, but
GAN performance degraded significantly. Try SiLU as an alternative
to see if it's the leaky-ness we are looking for or the smooth activation
curvature.
2020-07-05 13:39:08 -06:00
James Betker
9934e5d082 Move SRG1 to identical to new 2020-07-05 08:49:34 -06:00
James Betker
416538f31c SRG1 conjoined except ConvBnRelu 2020-07-05 08:44:17 -06:00
James Betker
c58c2b09ca Back to remove all biases (looks like a ConvBnRelu made its way in..) 2020-07-04 22:41:02 -06:00
James Betker
86cda86e94 Re-add biases, also add new init
A/B testing where we lost our GAN competitiveness.
2020-07-04 22:24:42 -06:00
James Betker
b03741f30e Remove all biases from generator
Continuing to investigate loss of GAN competitiveness, this is a big difference
between "old" SRG1 and "new".
2020-07-04 22:19:55 -06:00
James Betker
726e946e79 Turn BN off in SRG1
This wont work well but just testing if GAN performance comes back
2020-07-04 14:51:27 -06:00
James Betker
0ee39d419b OrderedDict not needed 2020-07-04 14:09:27 -06:00
James Betker
9048105b72 Break out SRG1 as separate network
Something strange is going on. These networks do not respond to
discriminator gradients properly anymore. SRG1 did, however so
reverting back to last known good state to figure out why.
2020-07-04 13:28:50 -06:00
James Betker
510b2f887d Remove RDB from srg2
Doesnt seem to work so great.
2020-07-03 22:31:20 -06:00
James Betker
703dec4472 Add SpineNet & integrate with SRG
New version of SRG uses SpineNet for a switch backbone.
2020-07-03 12:07:31 -06:00
James Betker
3ed7a2b9ab Move ConvBnRelu/Lelu to arch_util 2020-07-03 12:06:38 -06:00
James Betker
e9ee67ff10 Integrate RDB into SRG
The last RDB for each cluster is switched.
2020-07-01 17:19:55 -06:00
James Betker
6ac6c95177 Fix scaling bug 2020-07-01 16:42:27 -06:00
James Betker
30653181ba Experiment: get rid of post_switch_conv 2020-07-01 16:30:40 -06:00
James Betker
17191de836 Experiment: bring initialize_weights back again
Something really strange going on here..
2020-07-01 15:58:13 -06:00
James Betker
d1d573de07 Experiment: new init and post-switch-conv 2020-07-01 15:25:54 -06:00
James Betker
480d1299d7 Remove RRDB with switching
This idea never really panned out, removing it.
2020-07-01 12:08:32 -06:00
James Betker
e2398ac83c Experiment: revert initialization changes 2020-07-01 12:08:09 -06:00
James Betker
78276afcaa Experiment: Back to lelu 2020-07-01 11:43:25 -06:00
James Betker
b945021c90 SRG v2 - Move to Relu, rely on Module-based initialization 2020-07-01 11:33:32 -06:00
James Betker
604763be68 NSG r7
Converts the switching trunk to a VGG-style network to make it more comparable
to SRG architectures.
2020-07-01 09:54:29 -06:00
James Betker
87f1e9c56f Invert ResGen2 to operate in LR space 2020-06-30 20:57:40 -06:00
James Betker
e07d8abafb NSG rev 6
- Disable style passthrough
- Process multiplexers starting at base resolution
2020-06-30 20:47:26 -06:00
James Betker
3ce1a1878d NSG improvements (r5)
- Get rid of forwards(), it makes numeric_stability.py not work properly.
- Do stability auditing across layers.
- Upsample last instead of first, work in much higher dimensionality for transforms.
2020-06-30 16:59:57 -06:00
James Betker
75f148022d Even more NSG improvements (r4) 2020-06-30 13:52:47 -06:00
James Betker
773753073f More NSG improvements (v3)
Move to a fully fixup residual network for the switch (no
batch norms). Fix a bunch of other small bugs. Add in a
temporary latent feed-forward from the bottom of the
switch. Fix several initialization issues.
2020-06-29 20:26:51 -06:00
James Betker
4b82d0815d NSG improvements
- Just use resnet blocks for the multiplexer trunk of the generator
- Every block initializes itself, rather than everything at the end
- Cleans up some messy parts of the architecture, including unnecessary
  kernel sizes and places where BN is not used properly.
2020-06-29 10:09:51 -06:00
James Betker
978036e7b3 Add NestedSwitchGenerator
An evolution of SwitchedResidualGenerator, this variant nests attention
modules upon themselves to extend the representative capacity of the
model significantly.
2020-06-28 21:22:05 -06:00
James Betker
407224eba1 Re-work SwitchedResgen2
Got rid of the converged multiplexer bases but kept the configurable architecture. The
new multiplexers look a lot like the old one.

Took some queues from the transformer architecture: translate image to a higher filter-space
and stay there for the duration of the models computation. Also perform convs after each
switch to allow the model to anneal issues that arise.
2020-06-25 18:17:05 -06:00
James Betker
42a10b34ce Re-enable batch norm on switch processing blocks
Found out that batch norm is causing the switches to init really poorly -
not using a significant number of transforms. Might be a great time to
re-consider using the attention norm, but for now just re-enable it.
2020-06-24 21:15:17 -06:00
James Betker
4001db1ede Add ConfigurableSwitchComputer 2020-06-24 19:49:37 -06:00
James Betker
83c3b8b982 Add parameterized noise injection into resgen 2020-06-23 10:16:02 -06:00