Arch: resnet50_pt
Bs trn: 128
Bs val: 128
Hidden dim: 256
Dataset: celebA
Resample class: 
Slice with: rep
Rep cluster method: gmm
Num anchor: 32
Num positive: 32
Num negative: 32
Num negative easy: 0
Weight anc by loss: False
Weight pos by loss: False
Weight neg by loss: False
Anc loss temp: 0.5
Pos loss temp: 0.5
Neg loss temp: 0.5
Data wide pos: False
Target sample ratio: 1
Balance targets: False
Additional negatives: False
Hard negative factor: 0
Full contrastive: False
Train encoder: False
No projection head: False
Projection dim: 128
Batch factor: None
Temperature: 0.05
Single pos: False
Supervised linear scale up: False
Supervised update delay: 0
Contrastive weight: 0.5
Classifier update interval: 8
Optim: sgd
Max epoch: 5
Lr: 0.0001
Momentum: 0.9
Weight decay: 0.1
Weight decay c: 0.1
Stopping window: 30
Load encoder: 
Freeze encoder: False
Finetune epochs: 0
Clip grad norm: False
Lr scheduler classifier: 
Lr scheduler: 
Grad clip grad norm: False
Erm: False
Erm only: False
Pretrained spurious path: 
Max epoch s: 1
Bs trn s: 32
Lr s: 0.001
Momentum s: 0.9
Weight decay s: 0.0005
Slice temp: 10
Log loss interval: 10
Checkpoint interval: 50
Grad checkpoint interval: 50
Log visual interval: 100
Log grad visual interval: 50
Verbose: True
Seed: 38
Replicate: 0
No cuda: False
Resume: False
New slice: False
Num workers: 12
Evaluate: False
Data cmap: hsv
Test cmap: 
P correlation: 0.9
P corr by class: None
Train classes: ['blond', 'nonblond']
Train class ratios: None
Test shift: random
Flipped: False
Q: 0.7
Pretrained bmodel: False
Cosine: False
Exp: stage_one_erm
Supervised contrast: True
Prioritize spurious pos: False
Contrastive type: cnc
Compute auroc: False
Model type: resnet50_pt_cnc
Criterion: cross_entropy
Pretrained: False
Max grad norm: 1.0
Adam epsilon: 1e-08
Warmup steps: 0
Max grad norm s: 1.0
Adam epsilon s: 1e-08
Warmup steps s: 0
Grad max grad norm: 1.0
Grad adam epsilon: 1e-08
Grad warmup steps: 0
Device: cuda
Img file type: .png
Display image: False
Image path: ./images/celebA/celebA/config/contrastive_umaps
Log interval: 1
Log path: ./logs/celebA/config
Results path: ./results/celebA/config
Model path: ./model/celebA/config
Loss factor: 1
Supersample labels: False
Subsample labels: False
Weigh slice samples by loss: True
Val split: 0.2
Spurious train split: 0.2
Subsample groups: False
Train method: sc
Max robust acc: -1
Max robust epoch: -1
Max robust group acc: (None, None)
Root dir: ./datasets/data/CelebA/
Target name: Blond_Hair
Confounder names: ['Male']
Image mean: 0.449
Image std: 0.226
Augment data: False
Task: celebA
Num classes: 2
Experiment configs: config
Experiment name: cnc-celebA-sw=re-na=32-np=32-nn=32-nne=0-tsr=1-t=0.05-bf=None-cw=0.5-sud=0-me=5-bst=128-o=sgd-lr=0.0001-mo=0.9-wd=0.1-wdc=0.1-spur-me=1-bst=32-lr=0.001-mo=0.9-wd=0.0005-sts=0.2-s=38-r=0
Mi resampled: None

Loading checkpoints for train split:
[-1 -1 -1 ... -1 -1 -1]
<class 'numpy.ndarray'>
[0 1 2 3] [71629 66874 22880  1387]
Loading checkpoints for val split:
[-1 -1 -1 ... -1  1 -1]
<class 'numpy.ndarray'>
[0 1 2 3] [8535 8276 2874  182]
Loading checkpoints for test split:
[-1 -1 -1 ... -1 -1  1]
<class 'numpy.ndarray'>
[0 1 2 3] [9767 7535 2480  180]
Train dataset:
    Blond_Hair = 0, Male = 0 : n = 71629
    Blond_Hair = 0, Male = 1 : n = 66874
    Blond_Hair = 1, Male = 0 : n = 22880
    Blond_Hair = 1, Male = 1 : n = 1387
Val dataset:
    Blond_Hair = 0, Male = 0 : n = 8535
    Blond_Hair = 0, Male = 1 : n = 8276
    Blond_Hair = 1, Male = 0 : n = 2874
    Blond_Hair = 1, Male = 1 : n = 182
Test dataset:
    Blond_Hair = 0, Male = 0 : n = 9767
    Blond_Hair = 0, Male = 1 : n = 7535
    Blond_Hair = 1, Male = 0 : n = 2480
    Blond_Hair = 1, Male = 1 : n = 180
Pretrained model loaded from 
Epoch:   1 | Train Loss: 0.000 | Train Acc: 84.967 | Val Loss: 0.003 | Val Acc: 84.618
Training:
Accuracies by groups:
0, 0  acc: 71527 / 71629 =  99.858
0, 1  acc: 66741 / 66874 =  99.801
1, 0  acc:    30 / 22880 =   0.131
1, 1  acc:     3 /  1387 =   0.216
--------------------------------------
Average acc: 138301 / 162770 =  84.967
Robust  acc:    30 / 22880 =   0.131
--------------------------------------
Validating:
Accuracies by groups:
0, 0  acc:  8535 /  8535 = 100.000
0, 1  acc:  8276 /  8276 = 100.000
1, 0  acc:     0 /  2874 =   0.000
1, 1  acc:     0 /   182 =   0.000
------------------------------------
Average acc: 16811 / 19867 =  84.618
Robust  acc:     0 /  2874 =   0.000
------------------------------------
Save biased model at epoch 0
replace: True
Checkpoint saved at ./model/celebA/config/stage_one_erm_model_b_epoch0_seed38.pt
New max average-worst acc gap: 84.61770775658127
bias model - Saving best checkpoint at epoch 0
replace: True
Checkpoint saved at ./model/celebA/config/stage_one_erm_model_b_worst_avg_gap_best_epoch0_seed38.pt
-------------------------------------------
Avg Test Loss: 0.003 | Avg Test Acc: 86.675
Robust Acc: 0.000 | Best Acc: 100.000
-------------------------------------
Training, Epoch 0:
Accuracies by groups:
0, 0  acc:  9767 /  9767 = 100.000
0, 1  acc:  7535 /  7535 = 100.000
1, 0  acc:     0 /  2480 =   0.000
1, 1  acc:     0 /   180 =   0.000
------------------------------------
Average acc: 17302 / 19962 =  86.675
Robust  acc:     0 /  2480 =   0.000
------------------------------------
Accuracies by groups:
0, 0  acc:  9767 /  9767 = 100.000
0, 1  acc:  7535 /  7535 = 100.000
1, 0  acc:     0 /  2480 =   0.000
1, 1  acc:     0 /   180 =   0.000
------------------------------------
Average acc: 17302 / 19962 =  86.675
Robust  acc:     0 /  2480 =   0.000
------------------------------------
Testing:
Accuracies by groups:
0, 0  acc:  9767 /  9767 = 100.000
0, 1  acc:  7535 /  7535 = 100.000
1, 0  acc:     0 /  2480 =   0.000
1, 1  acc:     0 /   180 =   0.000
------------------------------------
Average acc: 17302 / 19962 =  86.675
Robust  acc:     0 /  2480 =   0.000
------------------------------------
Epoch:   2 | Train Loss: 0.001 | Train Acc: 86.673 | Val Loss: 0.002 | Val Acc: 91.056
Training:
Accuracies by groups:
0, 0  acc: 71408 / 71629 =  99.691
0, 1  acc: 66872 / 66874 =  99.997
1, 0  acc:  2789 / 22880 =  12.190
1, 1  acc:     9 /  1387 =   0.649
--------------------------------------
Average acc: 141078 / 162770 =  86.673
Robust  acc:     9 /  1387 =   0.649
--------------------------------------
Validating:
Accuracies by groups:
0, 0  acc:  8316 /  8535 =  97.434
0, 1  acc:  8276 /  8276 = 100.000
1, 0  acc:  1493 /  2874 =  51.949
1, 1  acc:     5 /   182 =   2.747
------------------------------------
Average acc: 18090 / 19867 =  91.056
Robust  acc:     5 /   182 =   2.747
------------------------------------
Save biased model at epoch 1
replace: True
Checkpoint saved at ./model/celebA/config/stage_one_erm_model_b_epoch1_seed38.pt
New max average-worst acc gap: 88.30826645544519
bias model - Saving best checkpoint at epoch 1
replace: True
Checkpoint saved at ./model/celebA/config/stage_one_erm_model_b_worst_avg_gap_best_epoch1_seed38.pt
-------------------------------------------
Avg Test Loss: 0.001 | Avg Test Acc: 91.905
Robust Acc: 4.444 | Best Acc: 100.000
-------------------------------------
Training, Epoch 1:
Accuracies by groups:
0, 0  acc:  9617 /  9767 =  98.464
0, 1  acc:  7535 /  7535 = 100.000
1, 0  acc:  1186 /  2480 =  47.823
1, 1  acc:     8 /   180 =   4.444
------------------------------------
Average acc: 18346 / 19962 =  91.905
Robust  acc:     8 /   180 =   4.444
------------------------------------
Accuracies by groups:
0, 0  acc:  9617 /  9767 =  98.464
0, 1  acc:  7535 /  7535 = 100.000
1, 0  acc:  1186 /  2480 =  47.823
1, 1  acc:     8 /   180 =   4.444
------------------------------------
Average acc: 18346 / 19962 =  91.905
Robust  acc:     8 /   180 =   4.444
------------------------------------
Testing:
Accuracies by groups:
0, 0  acc:  9617 /  9767 =  98.464
0, 1  acc:  7535 /  7535 = 100.000
1, 0  acc:  1186 /  2480 =  47.823
1, 1  acc:     8 /   180 =   4.444
------------------------------------
Average acc: 18346 / 19962 =  91.905
Robust  acc:     8 /   180 =   4.444
------------------------------------
Epoch:   3 | Train Loss: 0.000 | Train Acc: 92.984 | Val Loss: 0.001 | Val Acc: 93.904
Training:
Accuracies by groups:
0, 0  acc: 69385 / 71629 =  96.867
0, 1  acc: 66763 / 66874 =  99.834
1, 0  acc: 15049 / 22880 =  65.774
1, 1  acc:   153 /  1387 =  11.031
--------------------------------------
Average acc: 151350 / 162770 =  92.984
Robust  acc:   153 /  1387 =  11.031
--------------------------------------
Validating:
Accuracies by groups:
0, 0  acc:  8090 /  8535 =  94.786
0, 1  acc:  8255 /  8276 =  99.746
1, 0  acc:  2286 /  2874 =  79.541
1, 1  acc:    25 /   182 =  13.736
------------------------------------
Average acc: 18656 / 19867 =  93.904
Robust  acc:    25 /   182 =  13.736
------------------------------------
Save biased model at epoch 2
replace: True
Checkpoint saved at ./model/celebA/config/stage_one_erm_model_b_epoch2_seed38.pt
-------------------------------------------
Avg Test Loss: 0.001 | Avg Test Acc: 94.284
Robust Acc: 21.111 | Best Acc: 99.854
-------------------------------------
Training, Epoch 2:
Accuracies by groups:
0, 0  acc:  9399 /  9767 =  96.232
0, 1  acc:  7524 /  7535 =  99.854
1, 0  acc:  1860 /  2480 =  75.000
1, 1  acc:    38 /   180 =  21.111
------------------------------------
Average acc: 18821 / 19962 =  94.284
Robust  acc:    38 /   180 =  21.111
------------------------------------
Accuracies by groups:
0, 0  acc:  9399 /  9767 =  96.232
0, 1  acc:  7524 /  7535 =  99.854
1, 0  acc:  1860 /  2480 =  75.000
1, 1  acc:    38 /   180 =  21.111
------------------------------------
Average acc: 18821 / 19962 =  94.284
Robust  acc:    38 /   180 =  21.111
------------------------------------
Testing:
Accuracies by groups:
0, 0  acc:  9399 /  9767 =  96.232
0, 1  acc:  7524 /  7535 =  99.854
1, 0  acc:  1860 /  2480 =  75.000
1, 1  acc:    38 /   180 =  21.111
------------------------------------
Average acc: 18821 / 19962 =  94.284
Robust  acc:    38 /   180 =  21.111
------------------------------------
Epoch:   4 | Train Loss: 0.000 | Train Acc: 94.167 | Val Loss: 0.001 | Val Acc: 94.342
Training:
Accuracies by groups:
0, 0  acc: 68844 / 71629 =  96.112
0, 1  acc: 66625 / 66874 =  99.628
1, 0  acc: 17502 / 22880 =  76.495
1, 1  acc:   305 /  1387 =  21.990
--------------------------------------
Average acc: 153276 / 162770 =  94.167
Robust  acc:   305 /  1387 =  21.990
--------------------------------------
Validating:
Accuracies by groups:
0, 0  acc:  8101 /  8535 =  94.915
0, 1  acc:  8251 /  8276 =  99.698
1, 0  acc:  2358 /  2874 =  82.046
1, 1  acc:    33 /   182 =  18.132
------------------------------------
Average acc: 18743 / 19867 =  94.342
Robust  acc:    33 /   182 =  18.132
------------------------------------
Save biased model at epoch 3
replace: True
Checkpoint saved at ./model/celebA/config/stage_one_erm_model_b_epoch3_seed38.pt
-------------------------------------------
Avg Test Loss: 0.001 | Avg Test Acc: 94.925
Robust Acc: 26.111 | Best Acc: 99.748
-------------------------------------
Training, Epoch 3:
Accuracies by groups:
0, 0  acc:  9425 /  9767 =  96.498
0, 1  acc:  7516 /  7535 =  99.748
1, 0  acc:  1961 /  2480 =  79.073
1, 1  acc:    47 /   180 =  26.111
------------------------------------
Average acc: 18949 / 19962 =  94.925
Robust  acc:    47 /   180 =  26.111
------------------------------------
Accuracies by groups:
0, 0  acc:  9425 /  9767 =  96.498
0, 1  acc:  7516 /  7535 =  99.748
1, 0  acc:  1961 /  2480 =  79.073
1, 1  acc:    47 /   180 =  26.111
------------------------------------
Average acc: 18949 / 19962 =  94.925
Robust  acc:    47 /   180 =  26.111
------------------------------------
Testing:
Accuracies by groups:
0, 0  acc:  9425 /  9767 =  96.498
0, 1  acc:  7516 /  7535 =  99.748
1, 0  acc:  1961 /  2480 =  79.073
1, 1  acc:    47 /   180 =  26.111
------------------------------------
Average acc: 18949 / 19962 =  94.925
Robust  acc:    47 /   180 =  26.111
------------------------------------
Epoch:   5 | Train Loss: 0.000 | Train Acc: 94.632 | Val Loss: 0.001 | Val Acc: 94.720
Training:
Accuracies by groups:
0, 0  acc: 68776 / 71629 =  96.017
0, 1  acc: 66579 / 66874 =  99.559
1, 0  acc: 18289 / 22880 =  79.934
1, 1  acc:   389 /  1387 =  28.046
--------------------------------------
Average acc: 154033 / 162770 =  94.632
Robust  acc:   389 /  1387 =  28.046
--------------------------------------
Validating:
Accuracies by groups:
0, 0  acc:  8162 /  8535 =  95.630
0, 1  acc:  8252 /  8276 =  99.710
1, 0  acc:  2367 /  2874 =  82.359
1, 1  acc:    37 /   182 =  20.330
------------------------------------
Average acc: 18818 / 19867 =  94.720
Robust  acc:    37 /   182 =  20.330
------------------------------------
Save biased model at epoch 4
replace: True
Checkpoint saved at ./model/celebA/config/stage_one_erm_model_b_epoch4_seed38.pt
-------------------------------------------
Avg Test Loss: 0.001 | Avg Test Acc: 95.191
Robust Acc: 30.000 | Best Acc: 99.748
-------------------------------------
Training, Epoch 4:
Accuracies by groups:
0, 0  acc:  9469 /  9767 =  96.949
0, 1  acc:  7516 /  7535 =  99.748
1, 0  acc:  1963 /  2480 =  79.153
1, 1  acc:    54 /   180 =  30.000
------------------------------------
Average acc: 19002 / 19962 =  95.191
Robust  acc:    54 /   180 =  30.000
------------------------------------
Accuracies by groups:
0, 0  acc:  9469 /  9767 =  96.949
0, 1  acc:  7516 /  7535 =  99.748
1, 0  acc:  1963 /  2480 =  79.153
1, 1  acc:    54 /   180 =  30.000
------------------------------------
Average acc: 19002 / 19962 =  95.191
Robust  acc:    54 /   180 =  30.000
------------------------------------
Testing:
Accuracies by groups:
0, 0  acc:  9469 /  9767 =  96.949
0, 1  acc:  7516 /  7535 =  99.748
1, 0  acc:  1963 /  2480 =  79.153
1, 1  acc:    54 /   180 =  30.000
------------------------------------
Average acc: 19002 / 19962 =  95.191
Robust  acc:    54 /   180 =  30.000
------------------------------------
replace: True
Checkpoint saved at ./model/celebA/config/bias-end_seed38.pt
training biased model is done
