Make SynDiff run on current PyTorch, and fix the validation/checkpoint/inference paths - #67

Open

mertcanozdemir wants to merge 21 commits into

icon-lab:mainfrom

mertcanozdemir:fix/torch2-compat

Commits on Aug 17, 2026

compat: fall back to native ops when CUDA extensions cannot be built
mertcanozdemir
committed
5c63136
fix: remove leftover anomaly-detection call from the training loop
mertcanozdemir
committed
5328fed
fix: only strip the DDP 'module.' prefix when it is actually present
mertcanozdemir
committed
35cca82
compat: load the resume checkpoint with weights_only=False
mertcanozdemir
committed
1380ecc
feat: allow single-process and CPU-only training runs
mertcanozdemir
committed
27b96e7
fix: size the validation metric buffers for the final epoch
mertcanozdemir
committed
5cf5375
fix: validate the 2->1 direction with gen_diffusive_2
mertcanozdemir
committed
82e2afe
compat: fall back to CPU in test.py when CUDA is unavailable
mertcanozdemir
committed
d226a61
fix: store synthesised volumes at the cropped image size
mertcanozdemir
committed
f186404
fix: drop the dead args.num_channels assignment
mertcanozdemir
committed
329ee84
fix: pad to exactly 256 and report unusable input clearly
mertcanozdemir
committed
7d4e137
perf: run the generators under no_grad during discriminator updates
mertcanozdemir
committed
6997685
fix: stop wrapping the translation networks in DataParallel
mertcanozdemir
committed
020ff2f
fix: make the minibatch-stddev grouping divide the batch
mertcanozdemir
committed
f36c55d
test: add a CPU test suite for the diffusion, model and data paths
mertcanozdemir
committed
e1cca76
chore: add requirements.txt, ignore .DS_Store and local envs
mertcanozdemir
committed
e1dd4a7
docs: correct the dependency, dataset and inference notes
mertcanozdemir
committed
0009353
chore: remove two unreferenced modules
mertcanozdemir
committed
39252bf
refactor: move the shared diffusion code into diffusion.py
mertcanozdemir
committed
6416389
fix: honour --ngf and --image_size outside the diffusive networks
mertcanozdemir
committed
f26c207
docs: describe what --ngf and --image_size now cover
mertcanozdemir
committed
37f23c0

Navigation Menu

Uh oh!

Make SynDiff run on current PyTorch, and fix the validation/checkpoint/inference paths - #67

Make SynDiff run on current PyTorch, and fix the validation/checkpoint/inference paths#67
mertcanozdemir wants to merge 21 commits into
icon-lab:mainfrom
mertcanozdemir:fix/torch2-compat

Commits on Aug 17, 2026

compat: fall back to native ops when CUDA extensions cannot be built

fix: remove leftover anomaly-detection call from the training loop

fix: only strip the DDP 'module.' prefix when it is actually present

compat: load the resume checkpoint with weights_only=False

feat: allow single-process and CPU-only training runs

fix: size the validation metric buffers for the final epoch

fix: validate the 2->1 direction with gen_diffusive_2

compat: fall back to CPU in test.py when CUDA is unavailable

fix: store synthesised volumes at the cropped image size

fix: drop the dead args.num_channels assignment

fix: pad to exactly 256 and report unusable input clearly

perf: run the generators under no_grad during discriminator updates

fix: stop wrapping the translation networks in DataParallel

fix: make the minibatch-stddev grouping divide the batch

test: add a CPU test suite for the diffusion, model and data paths

chore: add requirements.txt, ignore .DS_Store and local envs

docs: correct the dependency, dataset and inference notes

chore: remove two unreferenced modules

refactor: move the shared diffusion code into diffusion.py

fix: honour --ngf and --image_size outside the diffusive networks

docs: describe what --ngf and --image_size now cover