Evaluate paper_dsprites_cnn extrapolation seed 2
Browse files
checkpoints/paper_dsprites_cnn/dsprites/extrapolation/cnn/seed_02/evaluation/status.yaml
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
status: completed
|
| 2 |
+
updated_at: '2026-07-27T01:53:11.582343+00:00'
|
| 3 |
+
experiment: paper_dsprites_cnn
|
| 4 |
+
dataset: dsprites
|
| 5 |
+
variant: extrapolation
|
| 6 |
+
model: cnn
|
| 7 |
+
seed: 2
|
| 8 |
+
slurm_job_id: '1229229'
|
| 9 |
+
slurm_array_job_id: '1229229'
|
| 10 |
+
slurm_array_task_id: '11'
|
checkpoints/paper_dsprites_cnn/dsprites/extrapolation/cnn/seed_02/evaluation/test_metrics.json
ADDED
|
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"checkpoint_model": "cnn",
|
| 3 |
+
"checkpoints": [
|
| 4 |
+
"/net/scratch/mx99/idgb/results/paper_dsprites_cnn/dsprites/extrapolation/cnn/seed_02/last.pt"
|
| 5 |
+
],
|
| 6 |
+
"created_at": "2026-07-27T01:53:11.579025+00:00",
|
| 7 |
+
"dataset": "dsprites",
|
| 8 |
+
"device": "cuda",
|
| 9 |
+
"elapsed_seconds": 14.412404144997708,
|
| 10 |
+
"experiment": "paper_dsprites_cnn",
|
| 11 |
+
"max_examples": null,
|
| 12 |
+
"metrics": {
|
| 13 |
+
"examples": 497280,
|
| 14 |
+
"factor_names": [
|
| 15 |
+
"shape",
|
| 16 |
+
"scale",
|
| 17 |
+
"orientation",
|
| 18 |
+
"x-position",
|
| 19 |
+
"y-position"
|
| 20 |
+
],
|
| 21 |
+
"factor_variance": [
|
| 22 |
+
0.16666666666666666,
|
| 23 |
+
0.11666666666666659,
|
| 24 |
+
0.08760683760683764,
|
| 25 |
+
0.08870967741935493,
|
| 26 |
+
0.08870967741935493
|
| 27 |
+
],
|
| 28 |
+
"mse": 0.03467760499751134,
|
| 29 |
+
"mse_per_factor": {
|
| 30 |
+
"orientation": 0.1265565141867177,
|
| 31 |
+
"scale": 0.008429279441742788,
|
| 32 |
+
"shape": 0.035347416815184905,
|
| 33 |
+
"x-position": 0.001484351611298606,
|
| 34 |
+
"y-position": 0.001570462932612655
|
| 35 |
+
},
|
| 36 |
+
"rsquared": 0.6473264265930907,
|
| 37 |
+
"rsquared_clipped": 0.6473264265930907,
|
| 38 |
+
"rsquared_clipped_per_factor": {
|
| 39 |
+
"orientation": 0.0,
|
| 40 |
+
"scale": 0.9277490333564904,
|
| 41 |
+
"shape": 0.7879154991088906,
|
| 42 |
+
"x-position": 0.9832673091089975,
|
| 43 |
+
"y-position": 0.9822965996687301
|
| 44 |
+
},
|
| 45 |
+
"rsquared_per_factor": {
|
| 46 |
+
"orientation": -0.4445963082776552,
|
| 47 |
+
"scale": 0.9277490333564904,
|
| 48 |
+
"shape": 0.7879154991088906,
|
| 49 |
+
"x-position": 0.9832673091089975,
|
| 50 |
+
"y-position": 0.9822965996687301
|
| 51 |
+
}
|
| 52 |
+
},
|
| 53 |
+
"model": "cnn",
|
| 54 |
+
"provenance": {
|
| 55 |
+
"config_path": "/home/mx99/comp-gen-bench/InDomainGeneralizationBenchmark/experiments/configs/paper/dsprites_cnn.yaml",
|
| 56 |
+
"slurm_array_job_id": "1229229",
|
| 57 |
+
"slurm_array_task_id": "11",
|
| 58 |
+
"slurm_job_id": "1229229"
|
| 59 |
+
},
|
| 60 |
+
"schema_version": 1,
|
| 61 |
+
"seed": 2,
|
| 62 |
+
"test_mode": "test",
|
| 63 |
+
"variant": "extrapolation"
|
| 64 |
+
}
|