version: v3_3
backbone: backbones/ViT-L-14.pt
seed: 42
epochs: 20
batch_size: 8
test_batch_size: 1
learning_rate: 0.0004
weight_decay: 0.05
dropout: 0.1
jpeg_augment:
- 60
- 100
jpeg_p: 0.1
blur_p: 0.1
blur_augment:
- 0.1
- 2.5
resize_augment:
- 0.25
- 2
resize_p: 0.1
replace_p: 0.1
replace_ratio:
- 0.2
- 0.98
replace_grid: 16
num_workers: 8
train_set: genimage_dr3_rec
test_set: genimage
train_set_size: 50000
test_set_size: null
use_patch_transformer: false
linear_classifier: true
transformer_layers: 0
transformer_heads: 2
tile_layers: 1
tile_heads: 2
default_train_tiles: 16
use_global_image: true
mil_heads: 8
fusion_hidden: 512
v2_l1_cross_shuffle_p: 0.9
v2_l1_replace_ratio_range:
- 0.5
- 0.9
v2_l1_with_replacement_p: 0.5
v2_l2_cross_shuffle_p: 0.9
v2_l2_replace_ratio_range:
- 0.1
- 0.8
v2_l2_with_replacement_p: 0.5
v2_bag_token_shuffle_p: 0.5
v2_cross_shuffle_min_keep_original: 1
v2_use_clip_intermediate_for_mil: false
v3_3_use_pool_gate_fusion: true
v3_3_pool_gate_hidden: 16
weights: weights/epoch_20.pth
eval_amp: false