r2s-v10-b128-5247da7-f66c07e7-47646604
cancelled —
Hypothesis
With batch 32 -> 128 with learning rate 5e-4 -> 1e-3 (square-root rule), 60k steps, so four times the samples, and everything else as in v9, the render-to-program model transcribes the 339 validation renders with lower mean 72 px pixel error than v9 trained on the same machine and seed. Pass requires the paired 95% bootstrap interval on the per-icon reduction against that v9 rerun to exclude zero. Its greedy latency on the RTX 4080 is recorded; a pass that is too slow for the browser demo is a teacher for distillation, not the demo model.
State transitions
- running2026-10-13T18:02:53Z
- cancelled2026-10-13T18:54:30Zjob time limit too short for batch 128 (no resume); rerun in slurm 1825792
Run record
Verbatim from runs/r2s-v10-b128-5247da7-f66c07e7-47646604/run.yaml, the record committed before launch.
schema_version
2
run_id
r2s-v10-b128-5247da7-f66c07e7-47646604
state
cancelled
planned_at
2026-10-13T18:02:53Z
hypothesis
With batch 32 -> 128 with learning rate 5e-4 -> 1e-3 (square-root rule), 60k steps, so four times the samples, and everything else as in v9, the render-to-program model transcribes the 339 validation renders with lower mean 72 px pixel error than v9 trained on the same machine and seed. Pass requires the paired 95% bootstrap interval on the per-icon reduction against that v9 rerun to exclude zero. Its greedy latency on the RTX 4080 is recorded; a pass that is too slow for the browser demo is a teacher for distillation, not the demo model.
parent_run
r2s-full-v9-colour
git_commit
5247da7ce9cb85d7952d2cf5833ef2daa3a416a1
dirty_patch_sha256
79f30c7a2834c8d8…config
configs/render2svg/v10-b128.yaml
config_sha256
f66c07e76a348eaf…config_resolved
model
image_size
144
d_model
256
heads
8
encoder_layers
2
decoder_layers
6
feedforward
1024
dropout
0.1
metric
true
fourier
10
order
path
training
steps
60000
batch_size
128
learning_rate
0.001
weight_decay
0.01
warmup_steps
1000
seed
7001
eval_every
2000
eval_icons
64
train_icons
None
evaluate_on_train
false
augment_variants
0
augment_seed
9001
augment_mirror
0.5
augment_max_shift
48
augment_colour
0.1
augment_online
true
compose_probability
0.5
compose_parts
4
augment_original
0.1
loader_workers
16
extra_training
bf16
true
final_eval_icons
339
clip_icons
32
dataset
pilot_config
configs/learning/openmoji-g1-geometric-gate-v16.yaml
cache_sha256
476466042da98d72…train_icons
2681
augmented_variants
0
extra_training
None
augmented_sha256
None
evaluated_on
primary/validation
seed
7001
determinism
seeded; cuDNN and SDPA kernels not forced deterministic
model_parameters
9025826
command
.venv/bin/python -m mojidiff.learning.render2svg --config configs/render2svg/v10-b128.yaml
outputs
run_dir
runs/r2s-v10-b128-5247da7-f66c07e7-47646604
checkpoints
/data/fs201459/cp98556/predictable/cache/runs/r2s-v10-b128-5247da7-f66c07e7-47646604
notes
cancelled_at
2026-10-13T18:54:30Z
cancellation
reason
Stopped by the operator agent at step 14,000 of 60,000 (eval pixel error 0.078 on 64 icons): batch 128 runs at about 4.4 steps/s on an H100, so the run needed about 3.8 h against its job limit of 3.5 h (Slurm 1824774), which cannot be raised while running, and there is no resume. Rerun unchanged in its own single-GPU job (Slurm 1825792) under a new run id.
rerun_job
slurm 1825792