TGDA: fine-grained image recognition from scratch with teacher-guided data augmentation

These are the checkpoints behind Fine-Grained Image Recognition from Scratch with Teacher-Guided Data Augmentation (arXiv:2507.12157). Students trained from scratch learn from a fine-grained teacher (a ResNet-101 trained with counterfactual attention learning, CAL), which also guides data-aware augmentation of the student's inputs. Code: arkel23/TGDA.

249 checkpoints, one per configuration, each the last epoch of one training seed. Each file is a torch.save dict with config (the full training configuration), model (the state dict), accuracy and epoch, with no optimizer state. File names are the runs' experiment-log names, ending in the serial. Load them with fgir-zoo. The collection groups this repo with the paper.

Layout

Folder	Serials	What	Files
sota_448	301-304	Main state-of-the-art table: ResNet-18/34/50/101 students, TGDA with TrivialAugment, label smoothing and stochastic depth	12
lrnet_128	201-204	Main LRNet table: low-resolution students at 128 px from a 448 px teacher	24
vitfs_448	100, 101	Efficient ViTs and the ViTFS variants, teacher and student at 448 px	54
vitfs_224_t448	102	Teacher at 448 px, student at 224 px	15
vitfs_224	103	Teacher and student at 224 px	9
vitfs_128	109	ViTFS tiny/micro/nano students at 128 px from a 448 px teacher	8
vit_ce_448	107	The serial-101 models trained with cross-entropy only (no teacher)	35
vitfs_transfer	104-106	A TGDA-trained student transferred to two new datasets with cross-entropy	8
ablations_128	212, 214, 221-223, 241-248, 251, 261-263, 271-278, 281	Ablations: TGDA without extra regularization, one- vs two-stage training, augmentation type, teacher type, transfer	73
scratch_vs_pt_448	320, 321, 331	200 epochs from scratch with cross-entropy or KD, and ImageNet-pretrained with KD	9
sweeps_128	0, 1	KD weight and temperature sweeps on CUB	2

manifest.csv lists every file with its dataset, student, teacher, serial, seed, image size, class count, accuracy, SHA-256 and size.

Load a checkpoint and classify an image

import torch
from PIL import Image
from torchvision import transforms
from fgir_zoo import tgda

model = tgda.create_model('sota_448/cub_resnet18_resnet101_cal_tgda_301')
cfg = model.config
tf = transforms.Compose([
    transforms.Resize((cfg.test_resize_size, cfg.test_resize_size),
                      interpolation=transforms.InterpolationMode.BICUBIC),
    transforms.CenterCrop(cfg.student_image_size or cfg.image_size),
    transforms.ToTensor(),
    transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
# a CUB-200-2011 test image, class index 50 (051.Horned_Grebe)
x = tf(Image.open('Horned_Grebe_0050_34561.jpg').convert('RGB')).unsqueeze(0)
with torch.no_grad():
    logits = model(x)
print(logits.argmax(-1).item(), logits.softmax(-1).max().item())  # 50 0.9855

Accuracy of the released checkpoints

Top-1 accuracy (%) stored in each file, on the dataset's test split at the end of training. Each value is one seed; the paper reports its own aggregates, so its tables can differ slightly.

sota_448

Model	aircraft	cars	cub
resnet101 <- resnet101	94.81	95.19	89.25
resnet18 <- resnet101	93.73	94.86	87.49
resnet34 <- resnet101	94.24	95.04	88.40
resnet50 <- resnet101	94.78	95.16	88.95

lrnet_128

Model	aircraft	cars	cub
lrnet14 <- resnet101	89.02	92.74	80.48
lrnet18 <- resnet101	90.67	93.21	82.48
lrnet26 <- resnet101	91.30	93.57	82.65
lrnet50 <- resnet101	91.81	93.74	84.12
resnet101 <- resnet101	88.57	92.19	81.48
resnet18 <- resnet101	86.95	91.18	77.68
resnet34 <- resnet101	85.45	91.03	78.81
resnet50 <- resnet101	90.19	92.96	82.74

vitfs_448

Model	aircraft	cars	cub
efficientvit_m2 <- resnet101	92.50	94.11	85.42
efficientvit_m3 <- resnet101	92.80	94.50	85.48
lcnet_035 <- resnet101	91.42	91.12	78.43
levit_128s <- resnet101	88.66	91.98	81.24
mobilevit_s <- resnet101	92.68	93.79	83.48
mobilevitv2_100 <- resnet101	94.36	94.57	87.59
mobilevitv2_125 <- resnet101	93.64	94.63	87.35
nest_tiny <- resnet101	89.89	93.25	85.48
pvt_v2_b0 <- resnet101	91.84	93.79	86.93
pvt_v2_b1 <- resnet101	92.53	94.29	86.85
swin_tiny_patch4_window7_224 <- resnet101	89.62	93.35	82.40
tiny_vit_5m_224 <- resnet101	93.37	94.43	88.45
vit_t16 <- resnet101	61.00		58.70
vitfs_tiny_patch16_clc_gap_reg4_dinov2_bn <- resnet101	90.58	94.01	85.47
vitfs_tiny_patch16_gap_224 <- resnet101	85.30		82.72
vitfs_tiny_patch16_gap_reg4_dinov2_bn_init <- resnet101	92.20	94.48	86.26
vitfs_tiny_patch16_gap_reg4_dinov2_init <- resnet101	89.32	93.07	83.72
vitfs_tiny_patch16_kvm_gap_reg4_dinov2_bn <- resnet101		94.25	83.91

vitfs_224_t448

Model	aircraft	cars	cub
vit_t16 <- resnet101	55.06	51.29	49.48
vitfs_micro_patch16_gap_reg4_dinov2_bn <- resnet101	88.81	92.41	78.25
vitfs_nano_patch16_gap_reg4_dinov2_bn <- resnet101	85.06	90.44	70.88
vitfs_tiny_patch16_gap_reg4_dinov2_bn_init <- resnet101	90.49	93.91	81.20
vitfs_tiny_patch16_gap_reg4_dinov2_init <- resnet101	86.35	92.71	77.94

vitfs_224

Model	aircraft	cars	cub
vitfs_tiny_patch16_gap_224 <- resnet101	52.63	86.16	75.61
vitfs_tiny_patch16_gap_reg4_dinov2_bn_init <- resnet101	60.13	92.64	79.89
vitfs_tiny_patch16_gap_reg4_dinov2_init <- resnet101	47.76	86.57	76.46

vitfs_128

Model	aircraft	cars	cub
vitfs_micro_patch16_gap_reg4_dinov2_bn <- resnet101	79.21	88.45	66.40
vitfs_nano_patch16_gap_reg4_dinov2_bn <- resnet101	77.47	83.97	61.32
vitfs_tiny_patch16_gap_reg4_dinov2_bn_init <- resnet101	81.10	90.24

vit_ce_448

Model	aircraft	cars	cub
efficientvit_m2	87.49	83.75	73.25
efficientvit_m3	87.46	80.19	72.82
levit_128s	82.27	87.07	69.59
mobilevit_s	88.36	91.20	72.01
mobilevitv2_100	89.11	90.57	78.60
mobilevitv2_125	87.85	90.34	81.50
nest_tiny	82.99	81.23	71.13
pvt_v2_b0	86.41	89.76	75.09
pvt_v2_b1	85.63
swin_tiny_patch4_window7_224		63.26
vit_t16	45.60	26.38	30.81
vitfs_tiny_patch16_gap_224			0.78
vitfs_tiny_patch16_gap_reg4_dinov2_bn_init		88.41	69.92
vitfs_tiny_patch16_gap_reg4_dinov2_init	76.84	81.35	60.30

vitfs_transfer

Model	aircraft	cars	cub
vit_t16	50.23	39.12
vitfs_tiny_patch16_gap_reg4_dinov2_bn_init	87.19	90.70	82.52

ablations_128

Model	aircraft	cars	cub
lrnet14	86.05	89.34	65.10
lrnet14 <- resnet101	85.30	93.06	81.79
resnet101d <- resnet101	88.18	92.54
resnet18	73.72	82.13	56.54
resnet18 <- resnet101	82.03	88.22	75.34
resnet34d <- resnet101	87.46	91.64	78.65

scratch_vs_pt_448

Model	aircraft	cars	cub
resnet101	83.08	84.36	66.41
resnet101 <- resnet101	93.91	94.60	87.63

sweeps_128

Model	cub
resnet18 <- resnet101	80.34

Requirements

• fgir-zoo (pip install git+https://github.com/arkel23/fgir-zoo.git), which pins timm==0.9.12
• torch (checked with 2.5.1)

Citation

@misc{rios_tgda_2025,
  title         = {Fine-Grained Image Recognition from Scratch with Teacher-Guided Data Augmentation},
  author        = {Rios, Edwin Arkel and Mikael, Fernando and Gosal, Oswin and Oyerinde, Femiloye and
                   Liang, Hao-Chun and Lai, Bo-Cheng and Hu, Min-Chun},
  year          = {2025},
  eprint        = {2507.12157},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2507.12157},
  url           = {https://arxiv.org/abs/2507.12157}
}