BrepMFR Model Weights

Model checkpoints for BrepMFR, which does machining feature recognition in B-rep models through deep learning and domain adaptation.

Recommended: trial1_finetune_ss/best-epoch32-f1_0.9354.ckpt has the best face-level macro F1 on 157 real parts (0.8886).

val F1 is model-level multi-label F1, which is what the filenames record. face F1 is per-face macro F1. These two metrics rank checkpoints differently. The "Local path" column is where the BrepMFR scripts expect each file. See the repo README.

Files

HuggingFace file	Local path used by scripts	Metric	Notes
brepformer/best-epoch43-f1_0.8638.ckpt	results/brepformer/best-epoch=43-val/f1=0.8638.ckpt	val F1 0.8638	Legacy architecture (BatchNorm, single FFN); won't load into the current BrepClassifier
brepformer/best-epoch97-f1_0.8482.ckpt	results/brepformer/best-epoch=97-val/f1=0.8482.ckpt	val F1 0.8482	Legacy architecture (BatchNorm, single FFN); won't load into the current BrepClassifier
brepformer/best-epoch99-f1_0.8466.ckpt	results/brepformer/best-epoch=99-val/f1=0.8466.ckpt	val F1 0.8466	Legacy architecture (BatchNorm, single FFN); won't load into the current BrepClassifier
brepsimplifier/f1_0.9415.ckpt	brepformer/brepsimplifier/model_checkpoint/f1=0.9415.ckpt	val F1 0.9415	Same file as trial1_ss1500 epoch 57; path BrepSimplifier expects
face_seg_heavy/best-epoch43-f1_0.8626.ckpt	results/face_seg_heavy/best-epoch=43-val/f1=0.8626.ckpt	val F1 0.8626	Legacy architecture (BatchNorm, single FFN); won't load into the current BrepClassifier
face_seg_heavy/best-epoch46-f1_0.8608.ckpt	results/face_seg_heavy/best-epoch=46-val/f1=0.8608.ckpt	val F1 0.8608	Legacy architecture (BatchNorm, single FFN); won't load into the current BrepClassifier
fix4_onehot_def/bestface-epoch81-face_f1_0.8714.ckpt	results/fix4_onehot_def/bestface-epoch=81-val/face_f1=0.8714.ckpt	face F1 0.8714
fix8_uv_def/bestface-epoch60-face_f1_0.8577.ckpt	results/fix8_uv_def/bestface-epoch=60-val/face_f1=0.8577.ckpt	face F1 0.8577
fix8ctl_legacy_def/bestface-epoch00-face_f1_0.8653.ckpt	results/fix8ctl_legacy_def/bestface-epoch=00-val/face_f1=0.8653.ckpt	face F1 0.8653
pipe_classifier/best-epoch64-f1_0.5517.ckpt	results/pipe_classifier/best-epoch=64-val/f1=0.5517.ckpt	val F1 0.5517	Different model class (pipe classifier)
t6_base/bestface-epoch01-face_f1_0.1255.ckpt	results/t6_base/bestface-epoch=01-val/face_f1=0.1255.ckpt	face F1 0.1255	Collapsed run, kept for the record
t6_f/bestface-epoch00-face_f1_0.1255.ckpt	results/t6_f/bestface-epoch=00-val/face_f1=0.1255.ckpt	face F1 0.1255	Collapsed run, kept for the record
t6_init_def/bestface-epoch00-face_f1_0.6958.ckpt	results/t6_init_def/bestface-epoch=00-val/face_f1=0.6958.ckpt	face F1 0.6958
t6_init_def_wf/bestface-epoch61-face_f1_0.8448.ckpt	results/t6_init_def_wf/bestface-epoch=61-val/face_f1=0.8448.ckpt	face F1 0.8448
t6_init_ft/bestface-epoch52-face_f1_0.8212.ckpt	results/t6_init_ft/bestface-epoch=52-val/face_f1=0.8212.ckpt	face F1 0.8212
t6_init_ft_wf/bestface-epoch47-face_f1_0.8378.ckpt	results/t6_init_ft_wf/bestface-epoch=47-val/face_f1=0.8378.ckpt	face F1 0.8378
t6_w/bestface-epoch00-face_f1_0.1982.ckpt	results/t6_w/bestface-epoch=00-val/face_f1=0.1982.ckpt	face F1 0.1982	Collapsed run, kept for the record
t6_wf/bestface-epoch00-face_f1_0.1942.ckpt	results/t6_wf/bestface-epoch=00-val/face_f1=0.1942.ckpt	face F1 0.1942	Collapsed run, kept for the record
t6a_init_t5/bestface-epoch74-face_f1_0.8676.ckpt	results/t6a_init_t5/bestface-epoch=74-val/face_f1=0.8676.ckpt	face F1 0.8676
t6a_init_t5_wf/bestface-epoch62-face_f1_0.8642.ckpt	results/t6a_init_t5_wf/bestface-epoch=62-val/face_f1=0.8642.ckpt	face F1 0.8642
t6a_init_t5_wf/bestface-epoch85-face_f1_0.8621.ckpt	results/t6a_init_t5_wf/bestface-epoch=85-val/face_f1=0.8621.ckpt	face F1 0.8621
t6actl_init_t5/bestface-epoch60-face_f1_0.8877.ckpt	results/t6actl_init_t5/bestface-epoch=60-val/face_f1=0.8877.ckpt	face F1 0.8877
t6actl_init_t5/bestface-epoch61-face_f1_0.8873.ckpt	results/t6actl_init_t5/bestface-epoch=61-val/face_f1=0.8873.ckpt	face F1 0.8873
t6ctl_base/bestface-epoch01-face_f1_0.1001.ckpt	results/t6ctl_base/bestface-epoch=01-val/face_f1=0.1001.ckpt	face F1 0.1001	Collapsed run, kept for the record
t6ctl_init_ft/bestface-epoch00-face_f1_0.8637.ckpt	results/t6ctl_init_ft/bestface-epoch=00-val/face_f1=0.8637.ckpt	face F1 0.8637
t6ctl_wf/bestface-epoch02-face_f1_0.0887.ckpt	results/t6ctl_wf/bestface-epoch=02-val/face_f1=0.0887.ckpt	face F1 0.0887	Collapsed run, kept for the record
trial1_finetune_ss/best-epoch26-f1_0.9356.ckpt	results/trial1_finetune_ss/best-epoch=26-val/f1=0.9356.ckpt	val F1 0.9356
trial1_finetune_ss/best-epoch32-f1_0.9354.ckpt	results/trial1_finetune_ss/best-epoch=32-val/f1=0.9354.ckpt	val F1 0.9354	Recommended — best face macro F1 on real parts (0.8886)
trial1_finetune_ss/last.ckpt	results/trial1_finetune_ss/last.ckpt	—
trial1_ss/best-epoch41-f1_0.9476.ckpt	results/trial1_ss/best-epoch=41-val/f1=0.9476.ckpt	val F1 0.9476
trial1_ss/last.ckpt	results/trial1_ss/last.ckpt	—
trial1_ss1000/best-epoch28-f1_0.9146.ckpt	results/trial1_ss1000/best-epoch=28-val/f1=0.9146.ckpt	val F1 0.9146
trial1_ss1500/best-epoch43-f1_0.9425.ckpt	results/trial1_ss1500/best-epoch=43-val/f1=0.9425.ckpt	val F1 0.9425
trial1_ss1500/best-epoch46-f1_0.9455.ckpt	results/trial1_ss1500/best-epoch=46-val/f1=0.9455.ckpt	val F1 0.9455
trial1_ss1500/best-epoch57-f1_0.9415.ckpt	results/trial1_ss1500/best-epoch=57-val/f1=0.9415.ckpt	val F1 0.9415	defeature_v2.py default
trial1_ss1500/last.ckpt	results/trial1_ss1500/last.ckpt	—
trial2/best-epoch111-f1_0.8558.ckpt	results/trial2/best-epoch=111-val/f1=0.8558.ckpt	val F1 0.8558	Legacy architecture (BatchNorm, single FFN); won't load into the current BrepClassifier
trial2_finetune_ss/best-epoch28-f1_0.9289.ckpt	results/trial2_finetune_ss/best-epoch=28-val/f1=0.9289.ckpt	val F1 0.9289
trial2_finetune_ss/last.ckpt	results/trial2_finetune_ss/last.ckpt	—
trial3/best-epoch41-f1_0.8799.ckpt	results/trial3/best-epoch=41-val/f1=0.8799.ckpt	val F1 0.8799	Legacy architecture (BatchNorm, single FFN); won't load into the current BrepClassifier
trial3/best-epoch42-f1_0.8795.ckpt	results/trial3/best-epoch=42-val/f1=0.8795.ckpt	val F1 0.8795	Legacy architecture (BatchNorm, single FFN); won't load into the current BrepClassifier
trial4/best-epoch26-f1_0.8839.ckpt	results/trial4/best-epoch=26-val/f1=0.8839.ckpt	val F1 0.8839	Legacy architecture (BatchNorm, single FFN); won't load into the current BrepClassifier
trial5/best-epoch39-f1_0.8832.ckpt	results/trial5/best-epoch=39-val/f1=0.8832.ckpt	val F1 0.8832	27-class MFTRCAD
trial5/best-epoch43-f1_0.8823.ckpt	results/trial5/best-epoch=43-val/f1=0.8823.ckpt	val F1 0.8823	27-class MFTRCAD

Source Code

https://github.com/cad-results/BrepMFR