FST-only training for paper experiments

FST-paper-experiment/FontDiffuserFST_training_phase_2_config.yaml

@@ -0,0 +1,167 @@
+adam_beta1: 0.9
+adam_beta2: 0.999
+adam_epsilon: 1.0e-08
+adam_weight_decay: 0.01
+algorithm_type: dpmsolver++
+batch_size: 1
+beta_scheduler: scaled_linear
+channel_attn: true
+channels_last: false
+character_input: false
+characters: null
+characters_file: null
+ckpt_dir: null
+ckpt_interval: 250
+compile: false
+compute_fid: false
+consistency_loss_weight: 0.1
+content_character: null
+content_encoder_downsample_size: 3
+content_image_path: null
+content_image_size: !!python/tuple
+- 96
+- 96
+content_start_channel: 64
+controlnet: false
+correcting_x0_fn: null
+data_root: .
+dataset_split: train_original
+demo: false
+deterministic: false
+device: cuda:0
+dro_div_weight: 0.0
+dro_lpips_weight: 1.0
+dro_max_timestep_frac: 0.3
+dro_normalise_reward: false
+dro_reward_scale: 1.0
+dro_sharp_weight: 0.0
+dro_ssim_weight: 1.0
+dro_warmup_steps: 0
+dro_weight: 0.1
+drop_prob: 0.1
+enable_attention_slicing: false
+enable_style_transform: false
+enable_xformers: false
+end_line: null
+evaluate: false
+experience_name: FontDiffuserFST_training_phase_2
+export_onnx: false
+fast_sampling: false
+feature_dim: 512
+ffn_dim: 2048
+fp16: false
+freeze_modules: ''
+frequency_filter_type: gaussian
+frequency_low_cutoff: 0.1
+frequency_mid_cutoff: 0.4
+frequency_mid_target: both
+frequency_use_mid_band: true
+fst_ckpt_path: null
+fst_feature_channels: 64,128,256,512,1024
+fst_num_queries: 220
+fst_num_scales: 5
+fst_query_dim: 256
+gradient_accumulation_steps: 2
+ground_truth_dir: null
+grpo_clip_eps: 0.2
+grpo_group_size: 4
+grpo_kl_coeff: 0.01
+grpo_pg_weight: 0.01
+grpo_reward_clip: 5.0
+grpo_sample_steps: 5
+grpo_warmup_steps: 1000
+guidance_scale: 7.5
+guidance_type: classifier-free
+hidden_dim: 256
+identity_adaptive_max_weight: 1.0
+identity_adaptive_min_weight: 0.1
+identity_log_metrics: true
+identity_loss_type: frobenius
+identity_loss_weight: 0.1
+identity_matrix_size: null
+identity_metric_interval: 100
+identity_pair_mode: random
+identity_pooled_reduction: mean
+identity_reg_weight: 0.01
+identity_regularization: orthogonal
+identity_similarity_threshold: 0.8
+instructpix2pix: false
+learning_rate: 5.0e-05
+local_rank: 0
+log_interval: 50
+logging_dir: logs
+lr_scheduler: cosine
+lr_warmup_steps: 2000
+max_grad_norm: 1.0
+max_train_steps: 15000
+method: multistep
+mixed_precision: 'no'
+mode: refinement
+model_type: noise
+mss_base_channels: 64
+mss_num_scales: 5
+nce_layers: 0,1,2,3
+num_consistency_pairs: 3
+num_heads: 8
+num_identity_pairs: 0
+num_inference_steps: 20
+num_neg: 34
+num_workers: 3
+offset_coefficient: 0.3
+onnx_export_dir: null
+onnx_opset_version: 17
+order: 2
+output_dir: outputs/FontArchitect/FST-paper-experiment
+perceptual_coefficient: 0.03
+phase_1: false
+phase_1_ckpt_dir: ckpt/FST-paper-experiment/final
+phase_2: true
+report_to: wandb
+resolution: 96
+resume_from_checkpoint: ''
+save_image: false
+save_image_dir: null
+save_interval: 10
+sc_coefficient: 0.03
+scale_lr: false
+scr_ckpt_path: ckpt/FST-paper-experiments/phase2/checkpoint_step_500/scr.safetensors
+scr_image_size: 96
+seed: 123
+skeleton_distance_method: hybrid
+skeleton_fusion_method: concat
+skeleton_max_distance: 12.0
+skeleton_method: medial_axis
+skeleton_output_mode: dual_channel
+skeleton_sigma: 1.5
+skip_type: time_uniform
+start_line: 1
+style_image_path: null
+style_image_size: !!python/tuple
+- 96
+- 96
+style_images: null
+style_source_same_prob: 0.5
+style_start_channel: 64
+style_transform_coefficient: 0.1
+summary: false
+t_end: null
+t_start: null
+temperature: 0.07
+train_batch_size: 4
+ttf_path: ttf/KaiXinSongA.ttf
+unet_channels: !!python/tuple
+- 64
+- 128
+- 256
+- 512
+use_adaptive_identity_loss: false
+use_dro: false
+use_frequency_decomp: false
+use_fst: true
+use_grpo: false
+use_pooled_identity_loss: false
+use_skeleton_content: false
+use_wandb: true
+val_interval: 100
+wandb_project: fontdiffuser-eval
+wandb_run_name: null

FST-paper-experiment/checkpoint_step_1000/content_encoder.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:747907e2145a118a7ba13cde36d6eb678e4878b57552181a6ae5614b37013f98
+size 4756580

FST-paper-experiment/checkpoint_step_1000/fst_module.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0548923d7a644c340a3ddacd005160db4749b1fa1e73b62bb51f49fb26d03c7a
+size 238575652

FST-paper-experiment/checkpoint_step_1000/fst_projection.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6176668b642f983b5ef896aa3c5ce5f07e67c546b17a1468e38a51480898dbe5
+size 4198552

FST-paper-experiment/checkpoint_step_1000/mss_encoder.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ebe8ef08d74e01053d154912b90ff6ff1440da41dcaacbb653ff46584da665f
+size 25261992

FST-paper-experiment/checkpoint_step_1000/original_style_projection.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd13f64e9982a3b2ce31f89040a6dd9d810f19b80e333e80bb6f384af126be30
+size 4198552

FST-paper-experiment/checkpoint_step_1000/scr.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d9d4a234af4258f939d420fd4aceb7ac6b8303386ad44ac7cc6f0d69ccf94c2
+size 284227968

FST-paper-experiment/checkpoint_step_1000/style_encoder.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc5a32d87b76ecd4a8e05409c4882f9cc399fa4b79a0cb665884de094fa8e51a
+size 82394556

FST-paper-experiment/checkpoint_step_1000/training_state.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:23355b821f3ff7d26dc80d8c21c38f6d887f6e32e3e35c559a3c989ff0db4f31
+size 1349161117

FST-paper-experiment/checkpoint_step_1000/unet.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:41e7225a08b3704bcbf1ae76af2475d512eacc232ffb0d490c43eaa16c72dfc5
+size 314927748

FST-paper-experiment/checkpoint_step_250/content_encoder.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:99e4b3c75877fd9c274e58f90962686f94fc1406a67b38dd134c5234cac605b2
+size 4756580

FST-paper-experiment/checkpoint_step_250/fst_module.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eec039b77276706ec44b55f132671555d2797d089e2dc098d85b0780bfad043a
+size 238575652

FST-paper-experiment/checkpoint_step_250/fst_projection.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:50f8d3fe7b9f5fb5c838757655ab3032598c33ae10cac879db84e2291e2c080f
+size 4198552

FST-paper-experiment/checkpoint_step_250/mss_encoder.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5cba98dc202813956834d70de2884b590f3add70352196e4164264b475db9f19
+size 25261992

FST-paper-experiment/checkpoint_step_250/original_style_projection.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c72fa10dad557c208feda50e7ec43d8bda15db2d848c632e428924ea212abc1
+size 4198552

FST-paper-experiment/checkpoint_step_250/scr.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dcba70bca1370a3d1cc9aa1d38cb65c4430be2efe2d71c61fd8d7cfa9ae021b5
+size 284227968

FST-paper-experiment/checkpoint_step_250/style_encoder.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1775a42b2683f8f954c7e502c42b809962f40de7b5567c4c2c75d4b557ce3f99
+size 82394556

FST-paper-experiment/checkpoint_step_250/training_state.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5a13f32ace78a11e6f12b3859d8d595c22cd03b0f0324f8d9fdb6c2b47210be
+size 1349161117

FST-paper-experiment/checkpoint_step_250/unet.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2739f053aba119820a66271c5e3a27ccaa9e4d3d29edb87cc29f8299de049567
+size 314927748

FST-paper-experiment/checkpoint_step_500/content_encoder.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0822b5ff787565cf9f94ce03f470fe2c177742758eff1828ca1ff218d4b58b1e
+size 4756580

FST-paper-experiment/checkpoint_step_500/fst_module.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac9ae9b07229a69d37d0f3727e43f0da6b60a2c66d8858e9c0515f6f5c076b51
+size 238575652

FST-paper-experiment/checkpoint_step_500/fst_projection.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4695237575ca12799dfb2bbd4d6540b625213f894fc222c2d64d5b9c50f06938
+size 4198552

FST-paper-experiment/checkpoint_step_500/mss_encoder.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6eaf71c0e6cb967368668326a1a79664a56d293327bf16e2428167c7bcad6e66
+size 25261992

FST-paper-experiment/checkpoint_step_500/original_style_projection.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd2c21e5915abaf8bc7fe888e87928842b90844473718a92e17a346ef6ee09c1
+size 4198552

FST-paper-experiment/checkpoint_step_500/scr.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f843fd29d3d9a58a5d759ccd73fac41194e02c6e97f0795f7402fecc50ae30b
+size 284227968

FST-paper-experiment/checkpoint_step_500/style_encoder.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cdaca9ab10f1efcc355acbc7e41b7a45bb6d6fff1db142f4cf647decb323035a
+size 82394556

FST-paper-experiment/checkpoint_step_500/training_state.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69392bd06c5aff1214adfcc1363aed9ef9ff7b2cf7e20a8359b0f67ba61fe63b
+size 1349161117

FST-paper-experiment/checkpoint_step_500/unet.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6f8c7bd942ace00d2f548623bba0e352608325288a5aabe4cd1a1a7edf859ae
+size 314927748