Add files using upload-large-folder tool

04_11_2025/38/basicsr_options.yaml

@@ -0,0 +1,238 @@
+# GENERATE TIME: Tue Nov  4 14:00:39 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 8
+    pin_memory: true
+    persistent_workers: true
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 16
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 360
+  num_heads:
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  mlp_ratio: 2
+  resi_connection: 1conv
+  primary_head: x4
+  head_num_feat: 256
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: ./runs/02_11_2025/34/models/net_g_20000.pth
+  strict_load_g: true
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: false
+  mode: max-autotune
+  dynamic: true
+  fullgraph: false
+  backend: null
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.0002
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.99
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  eagle_pixel_x2_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x2
+  l1_pixel_x2_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  eagle_pixel_x4_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x4
+  l1_pixel_x4_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+val:
+  val_freq: 5000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 10
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: '38'
+name: '38'

04_11_2025/38/train_38_20251104_140039.log

@@ -0,0 +1,577 @@
+2025-11-04 14:00:39,859 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 14:00:39,859 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 8
+      pin_memory: True
+      persistent_workers: True
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 16
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 360
+    num_heads: [12, 12, 12, 12, 12, 12]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    primary_head: x4
+    head_num_feat: 256
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: ./runs/02_11_2025/34/models/net_g_20000.pth
+    strict_load_g: True
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38/visualization
+  ]
+  compile:[
+    enabled: False
+    mode: max-autotune
+    dynamic: True
+    fullgraph: False
+    backend: None
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.0002
+      weight_decay: 0
+      betas: [0.9, 0.99]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    eagle_pixel_x2_opt:[
+      type: Eagle_Loss
+      loss_weight: 5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x2
+    ]
+    l1_pixel_x2_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    eagle_pixel_x4_opt:[
+      type: Eagle_Loss
+      loss_weight: 5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x4
+    ]
+    l1_pixel_x4_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+  ]
+  val:[
+    val_freq: 5000
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 10
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 38
+  name: 38
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 14:00:41,572 INFO: Use wandb logger with id=yupmcnie; project=Swin2SR-Latent-SR.
+2025-11-04 14:00:55,139 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 14:00:55,140 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 8
+	World size (gpu number): 6
+	Steps per epoch: 101219
+	Configured training steps: 125000
+	Approximate epochs to cover: 2.
+2025-11-04 14:00:55,144 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 14:00:55,144 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 14:00:55,146 INFO: Enabled find_unused_parameters=True for multi-head training overrides.
+2025-11-04 14:00:55,621 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 14:01:00,856 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 54,917,584
+2025-11-04 14:01:00,857 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)

04_11_2025/38_archived_20251104_065727/basicsr_options.yaml

@@ -0,0 +1,238 @@
+# GENERATE TIME: Tue Nov  4 06:51:38 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 8
+    pin_memory: true
+    persistent_workers: true
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 16
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 360
+  num_heads:
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  mlp_ratio: 2
+  resi_connection: 1conv
+  primary_head: x4
+  head_num_feat: 256
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: ./runs/02_11_2025/34/models/net_g_20000.pth
+  strict_load_g: true
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: false
+  mode: max-autotune
+  dynamic: true
+  fullgraph: false
+  backend: null
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.0002
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.99
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  eagle_pixel_x2_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x2
+  l1_pixel_x2_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  eagle_pixel_x4_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x4
+  l1_pixel_x4_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+val:
+  val_freq: 5000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 10
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: '38'
+name: '38'

04_11_2025/38_archived_20251104_065727/train_38_20251104_065138.log

@@ -0,0 +1,598 @@
+2025-11-04 06:51:38,640 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 06:51:38,640 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 8
+      pin_memory: True
+      persistent_workers: True
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 16
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 360
+    num_heads: [12, 12, 12, 12, 12, 12]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    primary_head: x4
+    head_num_feat: 256
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: ./runs/02_11_2025/34/models/net_g_20000.pth
+    strict_load_g: True
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38/visualization
+  ]
+  compile:[
+    enabled: False
+    mode: max-autotune
+    dynamic: True
+    fullgraph: False
+    backend: None
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.0002
+      weight_decay: 0
+      betas: [0.9, 0.99]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    eagle_pixel_x2_opt:[
+      type: Eagle_Loss
+      loss_weight: 5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x2
+    ]
+    l1_pixel_x2_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    eagle_pixel_x4_opt:[
+      type: Eagle_Loss
+      loss_weight: 5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x4
+    ]
+    l1_pixel_x4_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+  ]
+  val:[
+    val_freq: 5000
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 10
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 38
+  name: 38
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 06:51:40,278 INFO: Use wandb logger with id=ji9u56mi; project=Swin2SR-Latent-SR.
+2025-11-04 06:51:54,954 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 06:51:54,955 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 8
+	World size (gpu number): 6
+	Steps per epoch: 101219
+	Configured training steps: 125000
+	Approximate epochs to cover: 2.
+2025-11-04 06:51:54,959 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 06:51:54,959 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 06:51:54,960 INFO: Enabled find_unused_parameters=True for multi-head training overrides.
+2025-11-04 06:51:55,415 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 06:51:57,438 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 54,917,584
+2025-11-04 06:51:57,439 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 06:51:57,604 INFO: Loading SwinIRMultiHead from ./runs/02_11_2025/34/models/net_g_20000.pth [key=params].
+2025-11-04 06:51:57,666 INFO: Use EMA with decay: 0.999
+2025-11-04 06:51:58,234 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 06:51:58,422 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 06:51:58,423 INFO: Loading SwinIRMultiHead from ./runs/02_11_2025/34/models/net_g_20000.pth [key=params].
+2025-11-04 06:51:58,471 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 06:51:58,472 INFO: Initialized eagle_pixel_x2_opt in pixel space (w=5e-05).
+2025-11-04 06:51:58,472 INFO: Loss [L1Loss] is created.
+2025-11-04 06:51:58,473 INFO: Initialized l1_pixel_x2_opt in pixel space (w=10.0).
+2025-11-04 06:51:58,473 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 06:51:58,473 INFO: Initialized fft_frequency_x2_opt in pixel space (w=1.0).
+2025-11-04 06:51:58,473 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 06:51:58,474 INFO: Initialized eagle_pixel_x4_opt in pixel space (w=5e-05).
+2025-11-04 06:51:58,474 INFO: Loss [L1Loss] is created.
+2025-11-04 06:51:58,474 INFO: Initialized l1_pixel_x4_opt in pixel space (w=10.0).
+2025-11-04 06:51:58,474 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 06:51:58,475 INFO: Initialized fft_frequency_x4_opt in pixel space (w=1.0).
+2025-11-04 06:51:58,476 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 06:51:58,476 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 06:53:09,920 INFO: Start training from epoch: 0, step: 0
+2025-11-04 06:53:12,036 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae

04_11_2025/38_archived_20251104_140039/basicsr_options.yaml

@@ -0,0 +1,238 @@
+# GENERATE TIME: Tue Nov  4 06:57:27 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 8
+    pin_memory: true
+    persistent_workers: true
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 16
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 360
+  num_heads:
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  mlp_ratio: 2
+  resi_connection: 1conv
+  primary_head: x4
+  head_num_feat: 256
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: ./runs/02_11_2025/34/models/net_g_20000.pth
+  strict_load_g: true
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: false
+  mode: max-autotune
+  dynamic: true
+  fullgraph: false
+  backend: null
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.0002
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.99
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  eagle_pixel_x2_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x2
+  l1_pixel_x2_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  eagle_pixel_x4_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x4
+  l1_pixel_x4_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+val:
+  val_freq: 5000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 10
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: '38'
+name: '38'

04_11_2025/38_continue/basicsr_options.yaml

@@ -0,0 +1,245 @@
+# GENERATE TIME: Tue Nov  4 16:48:19 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 8
+    pin_memory: true
+    persistent_workers: true
+    prefetch_mode: cuda
+    latent_dtype: bf16
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+    latent_dtype: bf16
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 16
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 360
+  num_heads:
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  mlp_ratio: 2
+  resi_connection: 1conv
+  primary_head: x4
+  head_num_feat: 256
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+  strict_load_g: true
+  resume_state: null
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: true
+  mode: auto
+  dynamic: true
+  fullgraph: false
+  backend: inductor
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.00025
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.99
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  eagle_pixel_x2_opt:
+    type: Eagle_Loss
+    loss_weight: 2.5e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x2
+  l1_pixel_x2_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  eagle_pixel_x4_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x4
+  l1_pixel_x4_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+val:
+  val_freq: 100
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 10
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: 38_continue
+name: 38_continue

04_11_2025/38_continue/train_38_continue_20251104_164819.log

@@ -0,0 +1,615 @@
+2025-11-04 16:48:19,402 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 16:48:19,402 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 8
+      pin_memory: True
+      persistent_workers: True
+      prefetch_mode: cuda
+      latent_dtype: bf16
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+      latent_dtype: bf16
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 16
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 360
+    num_heads: [12, 12, 12, 12, 12, 12]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    primary_head: x4
+    head_num_feat: 256
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+    strict_load_g: True
+    resume_state: None
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/visualization
+  ]
+  compile:[
+    enabled: True
+    mode: auto
+    dynamic: True
+    fullgraph: False
+    backend: inductor
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.00025
+      weight_decay: 0
+      betas: [0.9, 0.99]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    eagle_pixel_x2_opt:[
+      type: Eagle_Loss
+      loss_weight: 2.5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x2
+    ]
+    l1_pixel_x2_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    eagle_pixel_x4_opt:[
+      type: Eagle_Loss
+      loss_weight: 5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x4
+    ]
+    l1_pixel_x4_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+  ]
+  val:[
+    val_freq: 100
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 10
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 38_continue
+  name: 38_continue
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 16:48:21,128 INFO: Use wandb logger with id=owuvmq6d; project=Swin2SR-Latent-SR.
+2025-11-04 16:48:33,914 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 16:48:33,915 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 8
+	World size (gpu number): 6
+	Steps per epoch: 101219
+	Configured training steps: 125000
+	Approximate epochs to cover: 2.
+2025-11-04 16:48:33,918 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 16:48:33,918 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 16:48:33,920 WARNING: Forced find_unused_parameters=True for multi-head training because DDP requires it.
+2025-11-04 16:48:34,388 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 16:48:36,476 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 54,917,584
+2025-11-04 16:48:36,477 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 16:48:36,606 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 16:48:36,657 WARNING: torch.compile failed for net_g; running in eager mode. Unrecognized mode=auto, should be one of: default, reduce-overhead, max-autotune-no-cudagraphs, max-autotune
+2025-11-04 16:48:36,659 INFO: Use EMA with decay: 0.999
+2025-11-04 16:48:37,079 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 16:48:37,243 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 16:48:37,244 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 16:48:37,295 INFO: Torch.compile disabled for EMA network; validation runs in eager mode.
+2025-11-04 16:48:37,297 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 16:48:37,298 INFO: Initialized eagle_pixel_x2_opt in pixel space (w=2.5e-05).
+2025-11-04 16:48:37,298 INFO: Loss [L1Loss] is created.
+2025-11-04 16:48:37,299 INFO: Initialized l1_pixel_x2_opt in pixel space (w=10.0).
+2025-11-04 16:48:37,300 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 16:48:37,301 INFO: Initialized fft_frequency_x2_opt in pixel space (w=1.0).
+2025-11-04 16:48:37,303 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 16:48:37,304 INFO: Initialized eagle_pixel_x4_opt in pixel space (w=5e-05).
+2025-11-04 16:48:37,305 INFO: Loss [L1Loss] is created.
+2025-11-04 16:48:37,306 INFO: Initialized l1_pixel_x4_opt in pixel space (w=10.0).
+2025-11-04 16:48:37,307 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 16:48:37,307 INFO: Initialized fft_frequency_x4_opt in pixel space (w=1.0).
+2025-11-04 16:48:37,310 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 16:48:37,310 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 16:48:37,310 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 16:49:52,970 INFO: Use cuda prefetch dataloader
+2025-11-04 16:49:52,972 INFO: Start training from epoch: 0, step: 0
+2025-11-04 16:49:54,587 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae
+2025-11-04 16:51:53,555 INFO: [38_co..][epoch:  0, step:     100, lr:(2.500e-04,)] [eta: 1 day, 12:17:04, time (data): 1.206 (0.012)] eagle_pixel_x2_opt: 4.0349e+00 l1_pixel_x2_opt: 3.5620e-02 fft_frequency_x2_opt: 3.2362e-02 eagle_pixel_x4_opt: 6.1451e+00 l1_pixel_x4_opt: 5.1324e-02 fft_frequency_x4_opt: 4.3936e-02 
+2025-11-04 16:54:22,576 INFO: Validation val_x2
+	 # l1_latent: 1.4049	Best: 1.4049 @ 100 iter
+	 # pixel_psnr_pt: 32.3299	Best: 32.3299 @ 100 iter
+

04_11_2025/38_continue_archived_20251104_150011/basicsr_options.yaml

@@ -0,0 +1,239 @@
+# GENERATE TIME: Tue Nov  4 14:08:56 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 8
+    pin_memory: true
+    persistent_workers: true
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 16
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 360
+  num_heads:
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  mlp_ratio: 2
+  resi_connection: 1conv
+  primary_head: x4
+  head_num_feat: 256
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+  strict_load_g: true
+  resume_state: null
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: false
+  mode: max-autotune
+  dynamic: true
+  fullgraph: false
+  backend: null
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.00025
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.99
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  eagle_pixel_x2_opt:
+    type: Eagle_Loss
+    loss_weight: 2.5e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x2
+  l1_pixel_x2_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  eagle_pixel_x4_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x4
+  l1_pixel_x4_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+val:
+  val_freq: 5000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 10
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: 38_continue
+name: 38_continue

04_11_2025/38_continue_archived_20251104_150011/train_38_continue_20251104_140856.log

@@ -0,0 +1,621 @@
+2025-11-04 14:08:56,280 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 14:08:56,280 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 8
+      pin_memory: True
+      persistent_workers: True
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 16
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 360
+    num_heads: [12, 12, 12, 12, 12, 12]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    primary_head: x4
+    head_num_feat: 256
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+    strict_load_g: True
+    resume_state: None
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/visualization
+  ]
+  compile:[
+    enabled: False
+    mode: max-autotune
+    dynamic: True
+    fullgraph: False
+    backend: None
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.00025
+      weight_decay: 0
+      betas: [0.9, 0.99]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    eagle_pixel_x2_opt:[
+      type: Eagle_Loss
+      loss_weight: 2.5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x2
+    ]
+    l1_pixel_x2_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    eagle_pixel_x4_opt:[
+      type: Eagle_Loss
+      loss_weight: 5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x4
+    ]
+    l1_pixel_x4_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+  ]
+  val:[
+    val_freq: 5000
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 10
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 38_continue
+  name: 38_continue
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 14:08:57,925 INFO: Use wandb logger with id=1wgd5xhu; project=Swin2SR-Latent-SR.
+2025-11-04 14:09:12,328 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 14:09:12,329 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 8
+	World size (gpu number): 6
+	Steps per epoch: 101219
+	Configured training steps: 125000
+	Approximate epochs to cover: 2.
+2025-11-04 14:09:12,333 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 14:09:12,333 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 14:09:12,335 INFO: Enabled find_unused_parameters=True for multi-head training overrides.
+2025-11-04 14:09:12,837 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 14:09:15,401 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 54,917,584
+2025-11-04 14:09:15,402 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 14:09:16,316 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 14:09:16,370 INFO: Use EMA with decay: 0.999
+2025-11-04 14:09:16,820 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 14:09:17,005 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 14:09:17,006 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 14:09:17,059 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 14:09:17,060 INFO: Initialized eagle_pixel_x2_opt in pixel space (w=2.5e-05).
+2025-11-04 14:09:17,061 INFO: Loss [L1Loss] is created.
+2025-11-04 14:09:17,062 INFO: Initialized l1_pixel_x2_opt in pixel space (w=10.0).
+2025-11-04 14:09:17,062 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 14:09:17,063 INFO: Initialized fft_frequency_x2_opt in pixel space (w=1.0).
+2025-11-04 14:09:17,064 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 14:09:17,065 INFO: Initialized eagle_pixel_x4_opt in pixel space (w=5e-05).
+2025-11-04 14:09:17,066 INFO: Loss [L1Loss] is created.
+2025-11-04 14:09:17,067 INFO: Initialized l1_pixel_x4_opt in pixel space (w=10.0).
+2025-11-04 14:09:17,068 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 14:09:17,069 INFO: Initialized fft_frequency_x4_opt in pixel space (w=1.0).
+2025-11-04 14:09:17,071 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 14:09:17,072 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 14:10:50,927 INFO: Start training from epoch: 0, step: 0
+2025-11-04 14:10:52,970 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae
+2025-11-04 14:13:04,206 INFO: [38_co..][epoch:  0, step:     100, lr:(2.500e-04,)] [eta: 1 day, 16:52:32, time (data): 1.333 (0.014)] eagle_pixel_x2_opt: 4.0478e+00 l1_pixel_x2_opt: 3.5212e-02 fft_frequency_x2_opt: 3.2364e-02 eagle_pixel_x4_opt: 6.1842e+00 l1_pixel_x4_opt: 5.1870e-02 fft_frequency_x4_opt: 4.4050e-02 
+2025-11-04 14:15:02,551 INFO: [38_co..][epoch:  0, step:     200, lr:(2.500e-04,)] [eta: 1 day, 16:56:02, time (data): 1.258 (0.007)] eagle_pixel_x2_opt: 4.6370e+00 l1_pixel_x2_opt: 3.7053e-02 fft_frequency_x2_opt: 3.3912e-02 eagle_pixel_x4_opt: 7.4606e+00 l1_pixel_x4_opt: 5.7464e-02 fft_frequency_x4_opt: 4.8318e-02 
+2025-11-04 14:17:01,089 INFO: [38_co..][epoch:  0, step:     300, lr:(2.500e-04,)] [eta: 1 day, 16:57:13, time (data): 1.185 (0.000)] eagle_pixel_x2_opt: 4.4128e+00 l1_pixel_x2_opt: 3.5012e-02 fft_frequency_x2_opt: 3.2407e-02 eagle_pixel_x4_opt: 6.8814e+00 l1_pixel_x4_opt: 5.5703e-02 fft_frequency_x4_opt: 4.5544e-02 
+2025-11-04 14:19:00,053 INFO: [38_co..][epoch:  0, step:     400, lr:(2.500e-04,)] [eta: 1 day, 16:59:03, time (data): 1.188 (0.000)] eagle_pixel_x2_opt: 4.7516e+00 l1_pixel_x2_opt: 3.6611e-02 fft_frequency_x2_opt: 3.4111e-02 eagle_pixel_x4_opt: 7.0534e+00 l1_pixel_x4_opt: 5.5126e-02 fft_frequency_x4_opt: 4.6474e-02 
+2025-11-04 14:20:59,104 INFO: [38_co..][epoch:  0, step:     500, lr:(2.500e-04,)] [eta: 1 day, 16:59:43, time (data): 1.191 (0.000)] eagle_pixel_x2_opt: 4.3646e+00 l1_pixel_x2_opt: 3.3084e-02 fft_frequency_x2_opt: 3.0503e-02 eagle_pixel_x4_opt: 6.5304e+00 l1_pixel_x4_opt: 5.0555e-02 fft_frequency_x4_opt: 4.2839e-02 
+2025-11-04 14:22:57,787 INFO: [38_co..][epoch:  0, step:     600, lr:(2.500e-04,)] [eta: 1 day, 16:58:13, time (data): 1.189 (0.000)] eagle_pixel_x2_opt: 4.4383e+00 l1_pixel_x2_opt: 3.3806e-02 fft_frequency_x2_opt: 3.1516e-02 eagle_pixel_x4_opt: 6.5080e+00 l1_pixel_x4_opt: 5.2778e-02 fft_frequency_x4_opt: 4.3677e-02 
+2025-11-04 14:24:56,514 INFO: [38_co..][epoch:  0, step:     700, lr:(2.500e-04,)] [eta: 1 day, 16:56:43, time (data): 1.187 (0.000)] eagle_pixel_x2_opt: 4.0987e+00 l1_pixel_x2_opt: 3.5029e-02 fft_frequency_x2_opt: 3.1578e-02 eagle_pixel_x4_opt: 6.2405e+00 l1_pixel_x4_opt: 5.2738e-02 fft_frequency_x4_opt: 4.3796e-02 
+2025-11-04 14:26:54,625 INFO: [38_co..][epoch:  0, step:     800, lr:(2.500e-04,)] [eta: 1 day, 16:53:31, time (data): 1.184 (0.000)] eagle_pixel_x2_opt: 4.4201e+00 l1_pixel_x2_opt: 3.7193e-02 fft_frequency_x2_opt: 3.4797e-02 eagle_pixel_x4_opt: 6.7722e+00 l1_pixel_x4_opt: 5.5728e-02 fft_frequency_x4_opt: 4.7275e-02 
+2025-11-04 14:28:52,317 INFO: [38_co..][epoch:  0, step:     900, lr:(2.500e-04,)] [eta: 1 day, 16:49:37, time (data): 1.177 (0.000)] eagle_pixel_x2_opt: 4.2041e+00 l1_pixel_x2_opt: 3.5301e-02 fft_frequency_x2_opt: 3.2906e-02 eagle_pixel_x4_opt: 6.3120e+00 l1_pixel_x4_opt: 5.5019e-02 fft_frequency_x4_opt: 4.5312e-02 
+2025-11-04 14:30:49,817 INFO: [38_co..][epoch:  0, step:   1,000, lr:(2.500e-04,)] [eta: 1 day, 16:45:43, time (data): 1.176 (0.000)] eagle_pixel_x2_opt: 3.4468e+00 l1_pixel_x2_opt: 2.9318e-02 fft_frequency_x2_opt: 2.7578e-02 eagle_pixel_x4_opt: 5.1700e+00 l1_pixel_x4_opt: 4.7038e-02 fft_frequency_x4_opt: 3.7929e-02 
+2025-11-04 14:32:47,723 INFO: [38_co..][epoch:  0, step:   1,100, lr:(2.500e-04,)] [eta: 1 day, 16:42:55, time (data): 1.179 (0.000)] eagle_pixel_x2_opt: 3.2528e+00 l1_pixel_x2_opt: 2.9392e-02 fft_frequency_x2_opt: 2.6386e-02 eagle_pixel_x4_opt: 4.8193e+00 l1_pixel_x4_opt: 4.5272e-02 fft_frequency_x4_opt: 3.6452e-02 
+2025-11-04 14:34:46,044 INFO: [38_co..][epoch:  0, step:   1,200, lr:(2.500e-04,)] [eta: 1 day, 16:40:59, time (data): 1.181 (0.000)] eagle_pixel_x2_opt: 3.6126e+00 l1_pixel_x2_opt: 3.3141e-02 fft_frequency_x2_opt: 2.8999e-02 eagle_pixel_x4_opt: 5.5263e+00 l1_pixel_x4_opt: 5.0832e-02 fft_frequency_x4_opt: 4.0940e-02 
+2025-11-04 14:36:43,841 INFO: [38_co..][epoch:  0, step:   1,300, lr:(2.500e-04,)] [eta: 1 day, 16:38:12, time (data): 1.178 (0.000)] eagle_pixel_x2_opt: 4.1173e+00 l1_pixel_x2_opt: 3.1617e-02 fft_frequency_x2_opt: 2.8865e-02 eagle_pixel_x4_opt: 6.5544e+00 l1_pixel_x4_opt: 5.0085e-02 fft_frequency_x4_opt: 4.0899e-02 
+2025-11-04 14:38:41,427 INFO: [38_co..][epoch:  0, step:   1,400, lr:(2.500e-04,)] [eta: 1 day, 16:35:14, time (data): 1.177 (0.000)] eagle_pixel_x2_opt: 3.9227e+00 l1_pixel_x2_opt: 3.4529e-02 fft_frequency_x2_opt: 3.1000e-02 eagle_pixel_x4_opt: 6.0294e+00 l1_pixel_x4_opt: 5.3412e-02 fft_frequency_x4_opt: 4.3645e-02 
+2025-11-04 14:40:40,035 INFO: [38_co..][epoch:  0, step:   1,500, lr:(2.500e-04,)] [eta: 1 day, 16:33:48, time (data): 1.187 (0.000)] eagle_pixel_x2_opt: 4.4719e+00 l1_pixel_x2_opt: 3.7513e-02 fft_frequency_x2_opt: 3.4572e-02 eagle_pixel_x4_opt: 6.9855e+00 l1_pixel_x4_opt: 5.7395e-02 fft_frequency_x4_opt: 4.7756e-02 
+2025-11-04 14:42:38,395 INFO: [38_co..][epoch:  0, step:   1,600, lr:(2.500e-04,)] [eta: 1 day, 16:31:59, time (data): 1.185 (0.000)] eagle_pixel_x2_opt: 3.8830e+00 l1_pixel_x2_opt: 3.2413e-02 fft_frequency_x2_opt: 2.9470e-02 eagle_pixel_x4_opt: 5.6264e+00 l1_pixel_x4_opt: 5.0683e-02 fft_frequency_x4_opt: 4.0469e-02 
+2025-11-04 14:44:36,497 INFO: [38_co..][epoch:  0, step:   1,700, lr:(2.500e-04,)] [eta: 1 day, 16:29:50, time (data): 1.181 (0.000)] eagle_pixel_x2_opt: 4.5973e+00 l1_pixel_x2_opt: 3.6244e-02 fft_frequency_x2_opt: 3.2904e-02 eagle_pixel_x4_opt: 7.1571e+00 l1_pixel_x4_opt: 5.4271e-02 fft_frequency_x4_opt: 4.5142e-02 
+2025-11-04 14:46:34,393 INFO: [38_co..][epoch:  0, step:   1,800, lr:(2.500e-04,)] [eta: 1 day, 16:27:28, time (data): 1.180 (0.000)] eagle_pixel_x2_opt: 3.7386e+00 l1_pixel_x2_opt: 3.0823e-02 fft_frequency_x2_opt: 2.7784e-02 eagle_pixel_x4_opt: 5.8103e+00 l1_pixel_x4_opt: 4.8842e-02 fft_frequency_x4_opt: 3.9464e-02 
+2025-11-04 14:48:32,143 INFO: [38_co..][epoch:  0, step:   1,900, lr:(2.500e-04,)] [eta: 1 day, 16:24:59, time (data): 1.178 (0.000)] eagle_pixel_x2_opt: 3.5256e+00 l1_pixel_x2_opt: 3.0935e-02 fft_frequency_x2_opt: 2.6623e-02 eagle_pixel_x4_opt: 5.4437e+00 l1_pixel_x4_opt: 4.7190e-02 fft_frequency_x4_opt: 3.7659e-02 
+2025-11-04 14:50:29,860 INFO: [38_co..][epoch:  0, step:   2,000, lr:(2.500e-04,)] [eta: 1 day, 16:22:32, time (data): 1.177 (0.000)] eagle_pixel_x2_opt: 4.7641e+00 l1_pixel_x2_opt: 3.4657e-02 fft_frequency_x2_opt: 3.2652e-02 eagle_pixel_x4_opt: 7.6179e+00 l1_pixel_x4_opt: 5.4961e-02 fft_frequency_x4_opt: 4.7113e-02 
+2025-11-04 14:52:28,560 INFO: [38_co..][epoch:  0, step:   2,100, lr:(2.500e-04,)] [eta: 1 day, 16:21:04, time (data): 1.188 (0.000)] eagle_pixel_x2_opt: 3.7760e+00 l1_pixel_x2_opt: 2.8969e-02 fft_frequency_x2_opt: 2.8778e-02 eagle_pixel_x4_opt: 6.1315e+00 l1_pixel_x4_opt: 4.2914e-02 fft_frequency_x4_opt: 3.9965e-02 
+2025-11-04 14:54:28,277 INFO: [38_co..][epoch:  0, step:   2,200, lr:(2.500e-04,)] [eta: 1 day, 16:20:31, time (data): 1.193 (0.000)] eagle_pixel_x2_opt: 4.9532e+00 l1_pixel_x2_opt: 3.8613e-02 fft_frequency_x2_opt: 3.8673e-02 eagle_pixel_x4_opt: 7.2591e+00 l1_pixel_x4_opt: 5.6348e-02 fft_frequency_x4_opt: 5.1948e-02 

04_11_2025/38_continue_archived_20251104_152426/basicsr_options.yaml

@@ -0,0 +1,239 @@
+# GENERATE TIME: Tue Nov  4 15:00:11 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 8
+    pin_memory: true
+    persistent_workers: true
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 16
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 360
+  num_heads:
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  mlp_ratio: 2
+  resi_connection: 1conv
+  primary_head: x4
+  head_num_feat: 256
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+  strict_load_g: true
+  resume_state: null
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: false
+  mode: max-autotune
+  dynamic: true
+  fullgraph: false
+  backend: null
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.00025
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.99
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  eagle_pixel_x2_opt:
+    type: Eagle_Loss
+    loss_weight: 2.5e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x2
+  l1_pixel_x2_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  eagle_pixel_x4_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x4
+  l1_pixel_x4_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+val:
+  val_freq: 5000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 10
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: 38_continue
+name: 38_continue

04_11_2025/38_continue_archived_20251104_152426/train_38_continue_20251104_150011.log

@@ -0,0 +1,600 @@
+2025-11-04 15:00:11,246 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 15:00:11,247 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 8
+      pin_memory: True
+      persistent_workers: True
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 16
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 360
+    num_heads: [12, 12, 12, 12, 12, 12]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    primary_head: x4
+    head_num_feat: 256
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+    strict_load_g: True
+    resume_state: None
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/visualization
+  ]
+  compile:[
+    enabled: False
+    mode: max-autotune
+    dynamic: True
+    fullgraph: False
+    backend: None
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.00025
+      weight_decay: 0
+      betas: [0.9, 0.99]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    eagle_pixel_x2_opt:[
+      type: Eagle_Loss
+      loss_weight: 2.5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x2
+    ]
+    l1_pixel_x2_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    eagle_pixel_x4_opt:[
+      type: Eagle_Loss
+      loss_weight: 5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x4
+    ]
+    l1_pixel_x4_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+  ]
+  val:[
+    val_freq: 5000
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 10
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 38_continue
+  name: 38_continue
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 15:00:13,050 INFO: Use wandb logger with id=08ecm9q0; project=Swin2SR-Latent-SR.
+2025-11-04 15:00:27,238 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 15:00:27,240 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 8
+	World size (gpu number): 6
+	Steps per epoch: 101219
+	Configured training steps: 125000
+	Approximate epochs to cover: 2.
+2025-11-04 15:00:27,243 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 15:00:27,243 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 15:00:27,244 INFO: Enabled find_unused_parameters=True for multi-head training overrides.
+2025-11-04 15:00:27,722 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 15:00:29,704 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 54,917,584
+2025-11-04 15:00:29,705 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 15:00:31,019 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 15:00:31,072 INFO: Use EMA with decay: 0.999
+2025-11-04 15:00:31,482 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 15:00:31,711 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 15:00:31,712 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 15:00:31,772 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 15:00:31,773 INFO: Initialized eagle_pixel_x2_opt in pixel space (w=2.5e-05).
+2025-11-04 15:00:31,773 INFO: Loss [L1Loss] is created.
+2025-11-04 15:00:31,774 INFO: Initialized l1_pixel_x2_opt in pixel space (w=10.0).
+2025-11-04 15:00:31,774 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 15:00:31,774 INFO: Initialized fft_frequency_x2_opt in pixel space (w=1.0).
+2025-11-04 15:00:31,775 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 15:00:31,775 INFO: Initialized eagle_pixel_x4_opt in pixel space (w=5e-05).
+2025-11-04 15:00:31,775 INFO: Loss [L1Loss] is created.
+2025-11-04 15:00:31,775 INFO: Initialized l1_pixel_x4_opt in pixel space (w=10.0).
+2025-11-04 15:00:31,776 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 15:00:31,776 INFO: Initialized fft_frequency_x4_opt in pixel space (w=1.0).
+2025-11-04 15:00:31,778 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 15:00:31,778 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 15:01:51,001 INFO: Start training from epoch: 0, step: 0
+2025-11-04 15:01:52,913 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae
+2025-11-04 15:04:01,417 INFO: [38_co..][epoch:  0, step:     100, lr:(2.500e-04,)] [eta: 1 day, 16:08:36, time (data): 1.304 (0.013)] eagle_pixel_x2_opt: 4.0388e+00 l1_pixel_x2_opt: 3.5721e-02 fft_frequency_x2_opt: 3.2387e-02 eagle_pixel_x4_opt: 6.2283e+00 l1_pixel_x4_opt: 5.1576e-02 fft_frequency_x4_opt: 4.4294e-02 

04_11_2025/38_continue_archived_20251104_152934/basicsr_options.yaml

@@ -0,0 +1,242 @@
+# GENERATE TIME: Tue Nov  4 15:24:26 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 8
+    pin_memory: true
+    persistent_workers: true
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 16
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 360
+  num_heads:
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  mlp_ratio: 2
+  resi_connection: 1conv
+  primary_head: x4
+  head_num_feat: 256
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+  strict_load_g: true
+  resume_state: null
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: false
+  mode: max-autotune
+  dynamic: true
+  fullgraph: false
+  backend: null
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.00025
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.99
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  eagle_pixel_x2_opt:
+    type: Eagle_Loss
+    loss_weight: 2.5e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x2
+  l1_pixel_x2_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  eagle_pixel_x4_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x4
+  l1_pixel_x4_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+val:
+  val_freq: 5000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 10
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: 38_continue
+name: 38_continue

04_11_2025/38_continue_archived_20251104_152934/train_38_continue_20251104_152426.log

@@ -0,0 +1,603 @@
+2025-11-04 15:24:26,900 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 15:24:26,900 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 8
+      pin_memory: True
+      persistent_workers: True
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 16
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 360
+    num_heads: [12, 12, 12, 12, 12, 12]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    primary_head: x4
+    head_num_feat: 256
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+    strict_load_g: True
+    resume_state: None
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/visualization
+  ]
+  compile:[
+    enabled: False
+    mode: max-autotune
+    dynamic: True
+    fullgraph: False
+    backend: None
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.00025
+      weight_decay: 0
+      betas: [0.9, 0.99]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    eagle_pixel_x2_opt:[
+      type: Eagle_Loss
+      loss_weight: 2.5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x2
+    ]
+    l1_pixel_x2_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    eagle_pixel_x4_opt:[
+      type: Eagle_Loss
+      loss_weight: 5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x4
+    ]
+    l1_pixel_x4_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+  ]
+  val:[
+    val_freq: 5000
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 10
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 38_continue
+  name: 38_continue
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 15:24:28,543 INFO: Use wandb logger with id=bcpzdw2b; project=Swin2SR-Latent-SR.
+2025-11-04 15:24:42,032 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 15:24:42,033 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 8
+	World size (gpu number): 6
+	Steps per epoch: 101219
+	Configured training steps: 125000
+	Approximate epochs to cover: 2.
+2025-11-04 15:24:42,037 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 15:24:42,037 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 15:24:42,039 INFO: Multi-head training overrides active with find_unused_parameters=False; skipping automatic enablement.
+2025-11-04 15:24:42,512 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 15:24:44,788 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 54,917,584
+2025-11-04 15:24:44,789 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 15:24:44,917 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 15:24:44,968 INFO: Use EMA with decay: 0.999
+2025-11-04 15:24:45,373 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 15:24:45,544 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 15:24:45,545 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 15:24:45,594 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 15:24:45,595 INFO: Initialized eagle_pixel_x2_opt in pixel space (w=2.5e-05).
+2025-11-04 15:24:45,596 INFO: Loss [L1Loss] is created.
+2025-11-04 15:24:45,596 INFO: Initialized l1_pixel_x2_opt in pixel space (w=10.0).
+2025-11-04 15:24:45,597 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 15:24:45,598 INFO: Initialized fft_frequency_x2_opt in pixel space (w=1.0).
+2025-11-04 15:24:45,599 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 15:24:45,599 INFO: Initialized eagle_pixel_x4_opt in pixel space (w=5e-05).
+2025-11-04 15:24:45,601 INFO: Loss [L1Loss] is created.
+2025-11-04 15:24:45,602 INFO: Initialized l1_pixel_x4_opt in pixel space (w=10.0).
+2025-11-04 15:24:45,604 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 15:24:45,605 INFO: Initialized fft_frequency_x4_opt in pixel space (w=1.0).
+2025-11-04 15:24:45,607 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 15:24:45,607 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 15:24:45,608 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 15:26:03,370 INFO: Start training from epoch: 0, step: 0

04_11_2025/38_continue_archived_20251104_153443/basicsr_options.yaml

@@ -0,0 +1,242 @@
+# GENERATE TIME: Tue Nov  4 15:29:35 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 8
+    pin_memory: true
+    persistent_workers: true
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 16
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 360
+  num_heads:
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  mlp_ratio: 2
+  resi_connection: 1conv
+  primary_head: x4
+  head_num_feat: 256
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+  strict_load_g: true
+  resume_state: null
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: false
+  mode: max-autotune
+  dynamic: true
+  fullgraph: false
+  backend: null
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.00025
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.99
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  eagle_pixel_x2_opt:
+    type: Eagle_Loss
+    loss_weight: 2.5e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x2
+  l1_pixel_x2_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  eagle_pixel_x4_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x4
+  l1_pixel_x4_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+val:
+  val_freq: 5000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 10
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: 38_continue
+name: 38_continue

04_11_2025/38_continue_archived_20251104_153443/train_38_continue_20251104_152935.log

@@ -0,0 +1,604 @@
+2025-11-04 15:29:35,015 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 15:29:35,015 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 8
+      pin_memory: True
+      persistent_workers: True
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 16
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 360
+    num_heads: [12, 12, 12, 12, 12, 12]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    primary_head: x4
+    head_num_feat: 256
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+    strict_load_g: True
+    resume_state: None
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/visualization
+  ]
+  compile:[
+    enabled: False
+    mode: max-autotune
+    dynamic: True
+    fullgraph: False
+    backend: None
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.00025
+      weight_decay: 0
+      betas: [0.9, 0.99]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    eagle_pixel_x2_opt:[
+      type: Eagle_Loss
+      loss_weight: 2.5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x2
+    ]
+    l1_pixel_x2_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    eagle_pixel_x4_opt:[
+      type: Eagle_Loss
+      loss_weight: 5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x4
+    ]
+    l1_pixel_x4_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+  ]
+  val:[
+    val_freq: 5000
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 10
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 38_continue
+  name: 38_continue
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 15:29:36,858 INFO: Use wandb logger with id=1sjgogd9; project=Swin2SR-Latent-SR.
+2025-11-04 15:29:49,356 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 15:29:49,357 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 8
+	World size (gpu number): 6
+	Steps per epoch: 101219
+	Configured training steps: 125000
+	Approximate epochs to cover: 2.
+2025-11-04 15:29:49,361 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 15:29:49,361 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 15:29:49,361 INFO: Multi-head training overrides active with find_unused_parameters=False; skipping automatic enablement.
+2025-11-04 15:29:49,799 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 15:29:51,965 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 54,917,584
+2025-11-04 15:29:51,966 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 15:29:52,117 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 15:29:52,179 INFO: Use EMA with decay: 0.999
+2025-11-04 15:29:52,885 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 15:29:53,108 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 15:29:53,109 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 15:29:53,169 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 15:29:53,170 INFO: Initialized eagle_pixel_x2_opt in pixel space (w=2.5e-05).
+2025-11-04 15:29:53,171 INFO: Loss [L1Loss] is created.
+2025-11-04 15:29:53,172 INFO: Initialized l1_pixel_x2_opt in pixel space (w=10.0).
+2025-11-04 15:29:53,173 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 15:29:53,174 INFO: Initialized fft_frequency_x2_opt in pixel space (w=1.0).
+2025-11-04 15:29:53,176 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 15:29:53,177 INFO: Initialized eagle_pixel_x4_opt in pixel space (w=5e-05).
+2025-11-04 15:29:53,178 INFO: Loss [L1Loss] is created.
+2025-11-04 15:29:53,179 INFO: Initialized l1_pixel_x4_opt in pixel space (w=10.0).
+2025-11-04 15:29:53,180 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 15:29:53,181 INFO: Initialized fft_frequency_x4_opt in pixel space (w=1.0).
+2025-11-04 15:29:53,182 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 15:29:53,183 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 15:29:53,184 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 15:31:07,745 INFO: Start training from epoch: 0, step: 0
+2025-11-04 15:31:10,129 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae

04_11_2025/38_continue_archived_20251104_153917/basicsr_options.yaml

@@ -0,0 +1,242 @@
+# GENERATE TIME: Tue Nov  4 15:34:43 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 8
+    pin_memory: true
+    persistent_workers: true
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 16
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 360
+  num_heads:
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  mlp_ratio: 2
+  resi_connection: 1conv
+  primary_head: x4
+  head_num_feat: 256
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+  strict_load_g: true
+  resume_state: null
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: false
+  mode: max-autotune
+  dynamic: true
+  fullgraph: false
+  backend: null
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.00025
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.99
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  eagle_pixel_x2_opt:
+    type: Eagle_Loss
+    loss_weight: 2.5e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x2
+  l1_pixel_x2_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  eagle_pixel_x4_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x4
+  l1_pixel_x4_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+val:
+  val_freq: 5000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 10
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: 38_continue
+name: 38_continue

04_11_2025/38_continue_archived_20251104_153917/train_38_continue_20251104_153443.log

@@ -0,0 +1,604 @@
+2025-11-04 15:34:43,226 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 15:34:43,226 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 8
+      pin_memory: True
+      persistent_workers: True
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 16
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 360
+    num_heads: [12, 12, 12, 12, 12, 12]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    primary_head: x4
+    head_num_feat: 256
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+    strict_load_g: True
+    resume_state: None
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/visualization
+  ]
+  compile:[
+    enabled: False
+    mode: max-autotune
+    dynamic: True
+    fullgraph: False
+    backend: None
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.00025
+      weight_decay: 0
+      betas: [0.9, 0.99]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    eagle_pixel_x2_opt:[
+      type: Eagle_Loss
+      loss_weight: 2.5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x2
+    ]
+    l1_pixel_x2_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    eagle_pixel_x4_opt:[
+      type: Eagle_Loss
+      loss_weight: 5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x4
+    ]
+    l1_pixel_x4_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+  ]
+  val:[
+    val_freq: 5000
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 10
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 38_continue
+  name: 38_continue
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 15:34:45,074 INFO: Use wandb logger with id=rx9h74fm; project=Swin2SR-Latent-SR.
+2025-11-04 15:34:58,938 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 15:34:58,939 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 8
+	World size (gpu number): 6
+	Steps per epoch: 101219
+	Configured training steps: 125000
+	Approximate epochs to cover: 2.
+2025-11-04 15:34:58,943 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 15:34:58,944 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 15:34:58,945 INFO: Multi-head training overrides active with find_unused_parameters=False; skipping automatic enablement.
+2025-11-04 15:34:59,419 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 15:35:01,537 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 54,917,584
+2025-11-04 15:35:01,538 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 15:35:01,700 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 15:35:01,763 INFO: Use EMA with decay: 0.999
+2025-11-04 15:35:02,357 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 15:35:02,544 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 15:35:02,545 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 15:35:02,595 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 15:35:02,596 INFO: Initialized eagle_pixel_x2_opt in pixel space (w=2.5e-05).
+2025-11-04 15:35:02,598 INFO: Loss [L1Loss] is created.
+2025-11-04 15:35:02,598 INFO: Initialized l1_pixel_x2_opt in pixel space (w=10.0).
+2025-11-04 15:35:02,598 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 15:35:02,599 INFO: Initialized fft_frequency_x2_opt in pixel space (w=1.0).
+2025-11-04 15:35:02,600 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 15:35:02,601 INFO: Initialized eagle_pixel_x4_opt in pixel space (w=5e-05).
+2025-11-04 15:35:02,602 INFO: Loss [L1Loss] is created.
+2025-11-04 15:35:02,603 INFO: Initialized l1_pixel_x4_opt in pixel space (w=10.0).
+2025-11-04 15:35:02,604 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 15:35:02,605 INFO: Initialized fft_frequency_x4_opt in pixel space (w=1.0).
+2025-11-04 15:35:02,607 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 15:35:02,607 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 15:35:02,608 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 15:36:18,357 INFO: Start training from epoch: 0, step: 0
+2025-11-04 15:36:19,967 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae

04_11_2025/38_continue_archived_20251104_155714/basicsr_options.yaml

@@ -0,0 +1,242 @@
+# GENERATE TIME: Tue Nov  4 15:39:17 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 8
+    pin_memory: true
+    persistent_workers: true
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 16
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 360
+  num_heads:
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  mlp_ratio: 2
+  resi_connection: 1conv
+  primary_head: x4
+  head_num_feat: 256
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+  strict_load_g: true
+  resume_state: null
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: false
+  mode: max-autotune
+  dynamic: true
+  fullgraph: false
+  backend: null
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.00025
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.99
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  eagle_pixel_x2_opt:
+    type: Eagle_Loss
+    loss_weight: 2.5e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x2
+  l1_pixel_x2_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  eagle_pixel_x4_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x4
+  l1_pixel_x4_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+val:
+  val_freq: 5000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 10
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: 38_continue
+name: 38_continue

04_11_2025/38_continue_archived_20251104_155714/train_38_continue_20251104_153917.log

@@ -0,0 +1,606 @@
+2025-11-04 15:39:17,694 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 15:39:17,695 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 8
+      pin_memory: True
+      persistent_workers: True
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 16
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 360
+    num_heads: [12, 12, 12, 12, 12, 12]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    primary_head: x4
+    head_num_feat: 256
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+    strict_load_g: True
+    resume_state: None
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/visualization
+  ]
+  compile:[
+    enabled: False
+    mode: max-autotune
+    dynamic: True
+    fullgraph: False
+    backend: None
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.00025
+      weight_decay: 0
+      betas: [0.9, 0.99]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    eagle_pixel_x2_opt:[
+      type: Eagle_Loss
+      loss_weight: 2.5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x2
+    ]
+    l1_pixel_x2_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    eagle_pixel_x4_opt:[
+      type: Eagle_Loss
+      loss_weight: 5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x4
+    ]
+    l1_pixel_x4_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+  ]
+  val:[
+    val_freq: 5000
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 10
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 38_continue
+  name: 38_continue
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 15:39:19,337 INFO: Use wandb logger with id=xpnahptq; project=Swin2SR-Latent-SR.
+2025-11-04 15:39:32,213 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 15:39:32,213 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 8
+	World size (gpu number): 6
+	Steps per epoch: 101219
+	Configured training steps: 125000
+	Approximate epochs to cover: 2.
+2025-11-04 15:39:32,217 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 15:39:32,217 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 15:39:32,219 WARNING: Forced find_unused_parameters=True for multi-head training because DDP requires it.
+2025-11-04 15:39:32,666 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 15:39:34,767 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 54,917,584
+2025-11-04 15:39:34,768 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 15:39:34,894 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 15:39:34,948 INFO: Use EMA with decay: 0.999
+2025-11-04 15:39:35,354 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 15:39:35,534 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 15:39:35,535 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 15:39:35,586 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 15:39:35,587 INFO: Initialized eagle_pixel_x2_opt in pixel space (w=2.5e-05).
+2025-11-04 15:39:35,588 INFO: Loss [L1Loss] is created.
+2025-11-04 15:39:35,588 INFO: Initialized l1_pixel_x2_opt in pixel space (w=10.0).
+2025-11-04 15:39:35,589 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 15:39:35,589 INFO: Initialized fft_frequency_x2_opt in pixel space (w=1.0).
+2025-11-04 15:39:35,590 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 15:39:35,591 INFO: Initialized eagle_pixel_x4_opt in pixel space (w=5e-05).
+2025-11-04 15:39:35,592 INFO: Loss [L1Loss] is created.
+2025-11-04 15:39:35,593 INFO: Initialized l1_pixel_x4_opt in pixel space (w=10.0).
+2025-11-04 15:39:35,594 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 15:39:35,595 INFO: Initialized fft_frequency_x4_opt in pixel space (w=1.0).
+2025-11-04 15:39:35,597 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 15:39:35,598 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 15:39:35,598 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 15:40:53,193 INFO: Start training from epoch: 0, step: 0
+2025-11-04 15:40:55,784 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae
+2025-11-04 15:42:57,789 INFO: [38_co..][epoch:  0, step:     100, lr:(2.500e-04,)] [eta: 1 day, 12:26:11, time (data): 1.246 (0.021)] eagle_pixel_x2_opt: 3.9566e+00 l1_pixel_x2_opt: 3.5668e-02 fft_frequency_x2_opt: 3.1990e-02 eagle_pixel_x4_opt: 6.1017e+00 l1_pixel_x4_opt: 5.1498e-02 fft_frequency_x4_opt: 4.3687e-02 
+2025-11-04 15:44:43,505 INFO: [38_co..][epoch:  0, step:     200, lr:(2.500e-04,)] [eta: 1 day, 12:31:37, time (data): 1.152 (0.011)] eagle_pixel_x2_opt: 4.8343e+00 l1_pixel_x2_opt: 3.7020e-02 fft_frequency_x2_opt: 3.4819e-02 eagle_pixel_x4_opt: 7.5603e+00 l1_pixel_x4_opt: 5.7494e-02 fft_frequency_x4_opt: 4.8756e-02 

04_11_2025/38_continue_archived_20251104_160331/basicsr_options.yaml

@@ -0,0 +1,245 @@
+# GENERATE TIME: Tue Nov  4 15:57:14 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 8
+    pin_memory: true
+    persistent_workers: true
+    prefetch_mode: cuda
+    latent_dtype: bf16
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+    latent_dtype: bf16
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 16
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 360
+  num_heads:
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  mlp_ratio: 2
+  resi_connection: 1conv
+  primary_head: x4
+  head_num_feat: 256
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+  strict_load_g: true
+  resume_state: null
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: true
+  mode: max-autotune
+  dynamic: true
+  fullgraph: false
+  backend: null
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.00025
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.99
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  eagle_pixel_x2_opt:
+    type: Eagle_Loss
+    loss_weight: 2.5e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x2
+  l1_pixel_x2_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  eagle_pixel_x4_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x4
+  l1_pixel_x4_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+val:
+  val_freq: 5000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 10
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: 38_continue
+name: 38_continue

04_11_2025/38_continue_archived_20251104_160331/train_38_continue_20251104_155714.log

@@ -0,0 +1,609 @@
+2025-11-04 15:57:14,197 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 15:57:14,197 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 8
+      pin_memory: True
+      persistent_workers: True
+      prefetch_mode: cuda
+      latent_dtype: bf16
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+      latent_dtype: bf16
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 16
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 360
+    num_heads: [12, 12, 12, 12, 12, 12]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    primary_head: x4
+    head_num_feat: 256
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+    strict_load_g: True
+    resume_state: None
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/visualization
+  ]
+  compile:[
+    enabled: True
+    mode: max-autotune
+    dynamic: True
+    fullgraph: False
+    backend: None
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.00025
+      weight_decay: 0
+      betas: [0.9, 0.99]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    eagle_pixel_x2_opt:[
+      type: Eagle_Loss
+      loss_weight: 2.5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x2
+    ]
+    l1_pixel_x2_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    eagle_pixel_x4_opt:[
+      type: Eagle_Loss
+      loss_weight: 5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x4
+    ]
+    l1_pixel_x4_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+  ]
+  val:[
+    val_freq: 5000
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 10
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 38_continue
+  name: 38_continue
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 15:57:16,097 INFO: Use wandb logger with id=eqilou43; project=Swin2SR-Latent-SR.
+2025-11-04 15:57:28,833 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 15:57:28,834 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 8
+	World size (gpu number): 6
+	Steps per epoch: 101219
+	Configured training steps: 125000
+	Approximate epochs to cover: 2.
+2025-11-04 15:57:28,837 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 15:57:28,838 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 15:57:28,839 WARNING: Forced find_unused_parameters=True for multi-head training because DDP requires it.
+2025-11-04 15:57:29,288 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 15:57:31,229 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 54,917,584
+2025-11-04 15:57:31,230 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 15:57:31,378 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 15:57:31,584 INFO: Use EMA with decay: 0.999
+2025-11-04 15:57:31,981 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 15:57:32,163 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 15:57:32,164 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 15:57:32,212 INFO: Torch.compile disabled for EMA network; validation runs in eager mode.
+2025-11-04 15:57:32,214 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 15:57:32,215 INFO: Initialized eagle_pixel_x2_opt in pixel space (w=2.5e-05).
+2025-11-04 15:57:32,216 INFO: Loss [L1Loss] is created.
+2025-11-04 15:57:32,217 INFO: Initialized l1_pixel_x2_opt in pixel space (w=10.0).
+2025-11-04 15:57:32,218 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 15:57:32,219 INFO: Initialized fft_frequency_x2_opt in pixel space (w=1.0).
+2025-11-04 15:57:32,219 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 15:57:32,220 INFO: Initialized eagle_pixel_x4_opt in pixel space (w=5e-05).
+2025-11-04 15:57:32,221 INFO: Loss [L1Loss] is created.
+2025-11-04 15:57:32,222 INFO: Initialized l1_pixel_x4_opt in pixel space (w=10.0).
+2025-11-04 15:57:32,223 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 15:57:32,224 INFO: Initialized fft_frequency_x4_opt in pixel space (w=1.0).
+2025-11-04 15:57:32,226 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 15:57:32,226 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 15:57:32,227 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 15:58:51,609 INFO: Use cuda prefetch dataloader
+2025-11-04 15:58:51,611 INFO: Start training from epoch: 0, step: 0
+2025-11-04 15:59:18,974 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae

04_11_2025/38_continue_archived_20251104_161131/basicsr_options.yaml

@@ -0,0 +1,245 @@
+# GENERATE TIME: Tue Nov  4 16:03:31 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 8
+    pin_memory: true
+    persistent_workers: true
+    prefetch_mode: cuda
+    latent_dtype: bf16
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+    latent_dtype: bf16
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 16
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 360
+  num_heads:
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  mlp_ratio: 2
+  resi_connection: 1conv
+  primary_head: x4
+  head_num_feat: 256
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+  strict_load_g: true
+  resume_state: null
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: true
+  mode: max-autotune
+  dynamic: true
+  fullgraph: false
+  backend: inductor
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.00025
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.99
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  eagle_pixel_x2_opt:
+    type: Eagle_Loss
+    loss_weight: 2.5e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x2
+  l1_pixel_x2_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  eagle_pixel_x4_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x4
+  l1_pixel_x4_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+val:
+  val_freq: 5000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 10
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: 38_continue
+name: 38_continue

04_11_2025/38_continue_archived_20251104_161131/train_38_continue_20251104_160331.log

@@ -0,0 +1,609 @@
+2025-11-04 16:03:31,157 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 16:03:31,157 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 8
+      pin_memory: True
+      persistent_workers: True
+      prefetch_mode: cuda
+      latent_dtype: bf16
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+      latent_dtype: bf16
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 16
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 360
+    num_heads: [12, 12, 12, 12, 12, 12]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    primary_head: x4
+    head_num_feat: 256
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+    strict_load_g: True
+    resume_state: None
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/visualization
+  ]
+  compile:[
+    enabled: True
+    mode: max-autotune
+    dynamic: True
+    fullgraph: False
+    backend: inductor
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.00025
+      weight_decay: 0
+      betas: [0.9, 0.99]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    eagle_pixel_x2_opt:[
+      type: Eagle_Loss
+      loss_weight: 2.5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x2
+    ]
+    l1_pixel_x2_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    eagle_pixel_x4_opt:[
+      type: Eagle_Loss
+      loss_weight: 5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x4
+    ]
+    l1_pixel_x4_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+  ]
+  val:[
+    val_freq: 5000
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 10
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 38_continue
+  name: 38_continue
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 16:03:32,941 INFO: Use wandb logger with id=49ewh2jg; project=Swin2SR-Latent-SR.
+2025-11-04 16:03:45,608 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 16:03:45,609 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 8
+	World size (gpu number): 6
+	Steps per epoch: 101219
+	Configured training steps: 125000
+	Approximate epochs to cover: 2.
+2025-11-04 16:03:45,613 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 16:03:45,613 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 16:03:45,614 WARNING: Forced find_unused_parameters=True for multi-head training because DDP requires it.
+2025-11-04 16:03:46,082 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 16:03:48,154 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 54,917,584
+2025-11-04 16:03:48,155 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 16:03:48,311 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 16:03:48,415 INFO: Use EMA with decay: 0.999
+2025-11-04 16:03:49,131 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 16:03:49,354 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 16:03:49,355 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 16:03:49,414 INFO: Torch.compile disabled for EMA network; validation runs in eager mode.
+2025-11-04 16:03:49,417 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 16:03:49,418 INFO: Initialized eagle_pixel_x2_opt in pixel space (w=2.5e-05).
+2025-11-04 16:03:49,418 INFO: Loss [L1Loss] is created.
+2025-11-04 16:03:49,419 INFO: Initialized l1_pixel_x2_opt in pixel space (w=10.0).
+2025-11-04 16:03:49,420 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 16:03:49,421 INFO: Initialized fft_frequency_x2_opt in pixel space (w=1.0).
+2025-11-04 16:03:49,422 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 16:03:49,423 INFO: Initialized eagle_pixel_x4_opt in pixel space (w=5e-05).
+2025-11-04 16:03:49,424 INFO: Loss [L1Loss] is created.
+2025-11-04 16:03:49,425 INFO: Initialized l1_pixel_x4_opt in pixel space (w=10.0).
+2025-11-04 16:03:49,427 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 16:03:49,427 INFO: Initialized fft_frequency_x4_opt in pixel space (w=1.0).
+2025-11-04 16:03:49,428 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 16:03:49,428 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 16:03:49,429 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 16:05:09,824 INFO: Use cuda prefetch dataloader
+2025-11-04 16:05:09,826 INFO: Start training from epoch: 0, step: 0
+2025-11-04 16:07:48,145 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae

04_11_2025/38_continue_archived_20251104_162054/basicsr_options.yaml

@@ -0,0 +1,245 @@
+# GENERATE TIME: Tue Nov  4 16:11:31 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 8
+    pin_memory: true
+    persistent_workers: true
+    prefetch_mode: cuda
+    latent_dtype: bf16
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+    latent_dtype: bf16
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 16
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 360
+  num_heads:
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  mlp_ratio: 2
+  resi_connection: 1conv
+  primary_head: x4
+  head_num_feat: 256
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+  strict_load_g: true
+  resume_state: null
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: true
+  mode: max-autotune
+  dynamic: true
+  fullgraph: false
+  backend: inductor
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.00025
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.99
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  eagle_pixel_x2_opt:
+    type: Eagle_Loss
+    loss_weight: 2.5e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x2
+  l1_pixel_x2_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  eagle_pixel_x4_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x4
+  l1_pixel_x4_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+val:
+  val_freq: 5000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 10
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: 38_continue
+name: 38_continue

04_11_2025/38_continue_archived_20251104_162054/train_38_continue_20251104_161131.log

@@ -0,0 +1,609 @@
+2025-11-04 16:11:31,980 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 16:11:31,981 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 8
+      pin_memory: True
+      persistent_workers: True
+      prefetch_mode: cuda
+      latent_dtype: bf16
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+      latent_dtype: bf16
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 16
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 360
+    num_heads: [12, 12, 12, 12, 12, 12]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    primary_head: x4
+    head_num_feat: 256
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+    strict_load_g: True
+    resume_state: None
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/visualization
+  ]
+  compile:[
+    enabled: True
+    mode: max-autotune
+    dynamic: True
+    fullgraph: False
+    backend: inductor
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.00025
+      weight_decay: 0
+      betas: [0.9, 0.99]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    eagle_pixel_x2_opt:[
+      type: Eagle_Loss
+      loss_weight: 2.5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x2
+    ]
+    l1_pixel_x2_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    eagle_pixel_x4_opt:[
+      type: Eagle_Loss
+      loss_weight: 5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x4
+    ]
+    l1_pixel_x4_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+  ]
+  val:[
+    val_freq: 5000
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 10
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 38_continue
+  name: 38_continue
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 16:11:33,704 INFO: Use wandb logger with id=uqf4znim; project=Swin2SR-Latent-SR.
+2025-11-04 16:11:46,029 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 16:11:46,030 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 8
+	World size (gpu number): 6
+	Steps per epoch: 101219
+	Configured training steps: 125000
+	Approximate epochs to cover: 2.
+2025-11-04 16:11:46,033 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 16:11:46,034 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 16:11:46,035 WARNING: Forced find_unused_parameters=True for multi-head training because DDP requires it.
+2025-11-04 16:11:46,484 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 16:11:48,541 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 54,917,584
+2025-11-04 16:11:48,542 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 16:11:48,754 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 16:11:48,862 INFO: Use EMA with decay: 0.999
+2025-11-04 16:11:49,388 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 16:11:49,576 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 16:11:49,577 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 16:11:49,626 INFO: Torch.compile disabled for EMA network; validation runs in eager mode.
+2025-11-04 16:11:49,628 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 16:11:49,629 INFO: Initialized eagle_pixel_x2_opt in pixel space (w=2.5e-05).
+2025-11-04 16:11:49,630 INFO: Loss [L1Loss] is created.
+2025-11-04 16:11:49,631 INFO: Initialized l1_pixel_x2_opt in pixel space (w=10.0).
+2025-11-04 16:11:49,632 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 16:11:49,633 INFO: Initialized fft_frequency_x2_opt in pixel space (w=1.0).
+2025-11-04 16:11:49,634 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 16:11:49,635 INFO: Initialized eagle_pixel_x4_opt in pixel space (w=5e-05).
+2025-11-04 16:11:49,636 INFO: Loss [L1Loss] is created.
+2025-11-04 16:11:49,637 INFO: Initialized l1_pixel_x4_opt in pixel space (w=10.0).
+2025-11-04 16:11:49,637 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 16:11:49,638 INFO: Initialized fft_frequency_x4_opt in pixel space (w=1.0).
+2025-11-04 16:11:49,640 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 16:11:49,640 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 16:11:49,641 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 16:13:09,221 INFO: Use cuda prefetch dataloader
+2025-11-04 16:13:09,222 INFO: Start training from epoch: 0, step: 0
+2025-11-04 16:14:02,100 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae

04_11_2025/38_continue_archived_20251104_164245/basicsr_options.yaml

@@ -0,0 +1,245 @@
+# GENERATE TIME: Tue Nov  4 16:20:54 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 8
+    pin_memory: true
+    persistent_workers: true
+    prefetch_mode: cuda
+    latent_dtype: bf16
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+    latent_dtype: bf16
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 16
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 360
+  num_heads:
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  mlp_ratio: 2
+  resi_connection: 1conv
+  primary_head: x4
+  head_num_feat: 256
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+  strict_load_g: true
+  resume_state: null
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: false
+  mode: auto
+  dynamic: true
+  fullgraph: false
+  backend: inductor
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.00025
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.99
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  eagle_pixel_x2_opt:
+    type: Eagle_Loss
+    loss_weight: 2.5e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x2
+  l1_pixel_x2_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  eagle_pixel_x4_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x4
+  l1_pixel_x4_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+val:
+  val_freq: 1000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 10
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: 38_continue
+name: 38_continue

04_11_2025/38_continue_archived_20251104_164245/train_38_continue_20251104_162054.log

@@ -0,0 +1,618 @@
+2025-11-04 16:20:54,476 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 16:20:54,477 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 8
+      pin_memory: True
+      persistent_workers: True
+      prefetch_mode: cuda
+      latent_dtype: bf16
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+      latent_dtype: bf16
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 16
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 360
+    num_heads: [12, 12, 12, 12, 12, 12]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    primary_head: x4
+    head_num_feat: 256
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+    strict_load_g: True
+    resume_state: None
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/visualization
+  ]
+  compile:[
+    enabled: False
+    mode: auto
+    dynamic: True
+    fullgraph: False
+    backend: inductor
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.00025
+      weight_decay: 0
+      betas: [0.9, 0.99]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    eagle_pixel_x2_opt:[
+      type: Eagle_Loss
+      loss_weight: 2.5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x2
+    ]
+    l1_pixel_x2_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    eagle_pixel_x4_opt:[
+      type: Eagle_Loss
+      loss_weight: 5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x4
+    ]
+    l1_pixel_x4_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+  ]
+  val:[
+    val_freq: 1000
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 10
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 38_continue
+  name: 38_continue
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 16:20:56,095 INFO: Use wandb logger with id=3k8afxcw; project=Swin2SR-Latent-SR.
+2025-11-04 16:21:10,703 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 16:21:10,704 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 8
+	World size (gpu number): 6
+	Steps per epoch: 101219
+	Configured training steps: 125000
+	Approximate epochs to cover: 2.
+2025-11-04 16:21:10,708 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 16:21:10,708 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 16:21:10,709 WARNING: Forced find_unused_parameters=True for multi-head training because DDP requires it.
+2025-11-04 16:21:11,178 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 16:21:13,052 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 54,917,584
+2025-11-04 16:21:13,053 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 16:21:13,182 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 16:21:13,234 INFO: Use EMA with decay: 0.999
+2025-11-04 16:21:21,047 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 16:21:21,231 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 16:21:21,232 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 16:21:21,281 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 16:21:21,282 INFO: Initialized eagle_pixel_x2_opt in pixel space (w=2.5e-05).
+2025-11-04 16:21:21,282 INFO: Loss [L1Loss] is created.
+2025-11-04 16:21:21,283 INFO: Initialized l1_pixel_x2_opt in pixel space (w=10.0).
+2025-11-04 16:21:21,283 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 16:21:21,284 INFO: Initialized fft_frequency_x2_opt in pixel space (w=1.0).
+2025-11-04 16:21:21,285 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 16:21:21,285 INFO: Initialized eagle_pixel_x4_opt in pixel space (w=5e-05).
+2025-11-04 16:21:21,287 INFO: Loss [L1Loss] is created.
+2025-11-04 16:21:21,288 INFO: Initialized l1_pixel_x4_opt in pixel space (w=10.0).
+2025-11-04 16:21:21,289 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 16:21:21,290 INFO: Initialized fft_frequency_x4_opt in pixel space (w=1.0).
+2025-11-04 16:21:21,292 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 16:21:21,293 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 16:21:21,295 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 16:22:39,592 INFO: Use cuda prefetch dataloader
+2025-11-04 16:22:39,593 INFO: Start training from epoch: 0, step: 0
+2025-11-04 16:22:41,135 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae
+2025-11-04 16:24:35,305 INFO: [38_co..][epoch:  0, step:     100, lr:(2.500e-04,)] [eta: 1 day, 11:35:39, time (data): 1.157 (0.011)] eagle_pixel_x2_opt: 3.9866e+00 l1_pixel_x2_opt: 3.5468e-02 fft_frequency_x2_opt: 3.1928e-02 eagle_pixel_x4_opt: 6.1575e+00 l1_pixel_x4_opt: 5.1574e-02 fft_frequency_x4_opt: 4.3926e-02 
+2025-11-04 16:26:19,446 INFO: [38_co..][epoch:  0, step:     200, lr:(2.500e-04,)] [eta: 1 day, 11:49:57, time (data): 1.099 (0.006)] eagle_pixel_x2_opt: 4.7869e+00 l1_pixel_x2_opt: 3.6885e-02 fft_frequency_x2_opt: 3.4497e-02 eagle_pixel_x4_opt: 7.4621e+00 l1_pixel_x4_opt: 5.7528e-02 fft_frequency_x4_opt: 4.8347e-02 
+2025-11-04 16:28:04,736 INFO: [38_co..][epoch:  0, step:     300, lr:(2.500e-04,)] [eta: 1 day, 12:01:31, time (data): 1.053 (0.000)] eagle_pixel_x2_opt: 4.3357e+00 l1_pixel_x2_opt: 3.5033e-02 fft_frequency_x2_opt: 3.1801e-02 eagle_pixel_x4_opt: 7.0501e+00 l1_pixel_x4_opt: 5.5709e-02 fft_frequency_x4_opt: 4.5875e-02 
+2025-11-04 16:29:50,163 INFO: [38_co..][epoch:  0, step:     400, lr:(2.500e-04,)] [eta: 1 day, 12:07:10, time (data): 1.054 (0.000)] eagle_pixel_x2_opt: 4.7769e+00 l1_pixel_x2_opt: 3.6266e-02 fft_frequency_x2_opt: 3.4324e-02 eagle_pixel_x4_opt: 7.3675e+00 l1_pixel_x4_opt: 5.5401e-02 fft_frequency_x4_opt: 4.7421e-02 
+2025-11-04 16:31:33,816 INFO: [38_co..][epoch:  0, step:     500, lr:(2.500e-04,)] [eta: 1 day, 12:02:30, time (data): 1.036 (0.000)] eagle_pixel_x2_opt: 4.2649e+00 l1_pixel_x2_opt: 3.3435e-02 fft_frequency_x2_opt: 3.0485e-02 eagle_pixel_x4_opt: 6.4789e+00 l1_pixel_x4_opt: 5.1200e-02 fft_frequency_x4_opt: 4.2710e-02 
+2025-11-04 16:33:16,785 INFO: [38_co..][epoch:  0, step:     600, lr:(2.500e-04,)] [eta: 1 day, 11:56:27, time (data): 1.033 (0.000)] eagle_pixel_x2_opt: 4.5858e+00 l1_pixel_x2_opt: 3.3850e-02 fft_frequency_x2_opt: 3.2226e-02 eagle_pixel_x4_opt: 6.8922e+00 l1_pixel_x4_opt: 5.2851e-02 fft_frequency_x4_opt: 4.4684e-02 
+2025-11-04 16:35:00,279 INFO: [38_co..][epoch:  0, step:     700, lr:(2.500e-04,)] [eta: 1 day, 11:53:12, time (data): 1.035 (0.000)] eagle_pixel_x2_opt: 7.2760e+00 l1_pixel_x2_opt: 6.4827e-02 fft_frequency_x2_opt: 4.9087e-02 eagle_pixel_x4_opt: 8.4385e+00 l1_pixel_x4_opt: 1.0337e-01 fft_frequency_x4_opt: 6.1042e-02 
+2025-11-04 16:36:42,821 INFO: [38_co..][epoch:  0, step:     800, lr:(2.500e-04,)] [eta: 1 day, 11:47:51, time (data): 1.030 (0.000)] eagle_pixel_x2_opt: 4.5191e+00 l1_pixel_x2_opt: 3.7611e-02 fft_frequency_x2_opt: 3.5145e-02 eagle_pixel_x4_opt: 7.1436e+00 l1_pixel_x4_opt: 5.6583e-02 fft_frequency_x4_opt: 4.8310e-02 
+2025-11-04 16:38:25,933 INFO: [38_co..][epoch:  0, step:     900, lr:(2.500e-04,)] [eta: 1 day, 11:44:38, time (data): 1.031 (0.000)] eagle_pixel_x2_opt: 4.2590e+00 l1_pixel_x2_opt: 3.5996e-02 fft_frequency_x2_opt: 3.3157e-02 eagle_pixel_x4_opt: 6.5662e+00 l1_pixel_x4_opt: 5.5788e-02 fft_frequency_x4_opt: 4.6527e-02 
+2025-11-04 16:40:08,357 INFO: [38_co..][epoch:  0, step:   1,000, lr:(2.500e-04,)] [eta: 1 day, 11:40:18, time (data): 1.028 (0.000)] eagle_pixel_x2_opt: 3.7622e+00 l1_pixel_x2_opt: 3.2101e-02 fft_frequency_x2_opt: 2.9383e-02 eagle_pixel_x4_opt: 6.1037e+00 l1_pixel_x4_opt: 5.2774e-02 fft_frequency_x4_opt: 4.2371e-02 

04_11_2025/38_continue_archived_20251104_164819/basicsr_options.yaml

@@ -0,0 +1,245 @@
+# GENERATE TIME: Tue Nov  4 16:42:45 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 8
+    pin_memory: true
+    persistent_workers: true
+    prefetch_mode: cuda
+    latent_dtype: bf16
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+    latent_dtype: bf16
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 16
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 360
+  num_heads:
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  - 12
+  mlp_ratio: 2
+  resi_connection: 1conv
+  primary_head: x4
+  head_num_feat: 256
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+  strict_load_g: true
+  resume_state: null
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: true
+  mode: auto
+  dynamic: true
+  fullgraph: false
+  backend: inductor
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.00025
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.99
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  eagle_pixel_x2_opt:
+    type: Eagle_Loss
+    loss_weight: 2.5e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x2
+  l1_pixel_x2_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  eagle_pixel_x4_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x4
+  l1_pixel_x4_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+val:
+  val_freq: 1000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 10
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: 38_continue
+name: 38_continue

04_11_2025/38_continue_archived_20251104_164819/train_38_continue_20251104_164245.log

@@ -0,0 +1,611 @@
+2025-11-04 16:42:45,862 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 16:42:45,862 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 8
+      pin_memory: True
+      persistent_workers: True
+      prefetch_mode: cuda
+      latent_dtype: bf16
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+      latent_dtype: bf16
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 16
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 360
+    num_heads: [12, 12, 12, 12, 12, 12]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    primary_head: x4
+    head_num_feat: 256
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth
+    strict_load_g: True
+    resume_state: None
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/38_continue/visualization
+  ]
+  compile:[
+    enabled: True
+    mode: auto
+    dynamic: True
+    fullgraph: False
+    backend: inductor
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.00025
+      weight_decay: 0
+      betas: [0.9, 0.99]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    eagle_pixel_x2_opt:[
+      type: Eagle_Loss
+      loss_weight: 2.5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x2
+    ]
+    l1_pixel_x2_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    eagle_pixel_x4_opt:[
+      type: Eagle_Loss
+      loss_weight: 5e-05
+      reduction: mean
+      space: pixel
+      patch_size: 3
+      cutoff: 0.5
+      target: x4
+    ]
+    l1_pixel_x4_opt:[
+      type: L1Loss
+      loss_weight: 10.0
+      reduction: mean
+      space: pixel
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 1.0
+      reduction: mean
+      space: pixel
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+  ]
+  val:[
+    val_freq: 1000
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 10
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 38_continue
+  name: 38_continue
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 16:42:47,659 INFO: Use wandb logger with id=elsbc4qx; project=Swin2SR-Latent-SR.
+2025-11-04 16:43:01,597 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 16:43:01,598 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 8
+	World size (gpu number): 6
+	Steps per epoch: 101219
+	Configured training steps: 125000
+	Approximate epochs to cover: 2.
+2025-11-04 16:43:01,602 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 16:43:01,602 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 16:43:01,604 WARNING: Forced find_unused_parameters=True for multi-head training because DDP requires it.
+2025-11-04 16:43:02,075 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 16:43:04,296 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 54,917,584
+2025-11-04 16:43:04,297 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=360, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=360, input_resolution=(32, 32), num_heads=12, window_size=16, shift_size=8, mlp_ratio=2.0
+            (norm1): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=360, window_size=(16, 16), num_heads=12
+              (qkv): Linear(in_features=360, out_features=1080, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=360, out_features=360, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=360, out_features=720, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=720, out_features=360, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(360, 360, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(360, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(256, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 16:43:04,446 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 16:43:04,506 WARNING: torch.compile failed for net_g; running in eager mode. Unrecognized mode=auto, should be one of: default, reduce-overhead, max-autotune-no-cudagraphs, max-autotune
+2025-11-04 16:43:04,508 INFO: Use EMA with decay: 0.999
+2025-11-04 16:43:04,999 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 16:43:05,220 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 16:43:05,221 INFO: Loading SwinIRMultiHead from runs/04_11_2025/38_archived_20251104_140039/models/net_g_15000.pth [key=params].
+2025-11-04 16:43:05,278 INFO: Torch.compile disabled for EMA network; validation runs in eager mode.
+2025-11-04 16:43:05,280 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 16:43:05,280 INFO: Initialized eagle_pixel_x2_opt in pixel space (w=2.5e-05).
+2025-11-04 16:43:05,281 INFO: Loss [L1Loss] is created.
+2025-11-04 16:43:05,282 INFO: Initialized l1_pixel_x2_opt in pixel space (w=10.0).
+2025-11-04 16:43:05,283 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 16:43:05,284 INFO: Initialized fft_frequency_x2_opt in pixel space (w=1.0).
+2025-11-04 16:43:05,285 INFO: Loss [Eagle_Loss] is created.
+2025-11-04 16:43:05,286 INFO: Initialized eagle_pixel_x4_opt in pixel space (w=5e-05).
+2025-11-04 16:43:05,287 INFO: Loss [L1Loss] is created.
+2025-11-04 16:43:05,288 INFO: Initialized l1_pixel_x4_opt in pixel space (w=10.0).
+2025-11-04 16:43:05,289 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 16:43:05,290 INFO: Initialized fft_frequency_x4_opt in pixel space (w=1.0).
+2025-11-04 16:43:05,292 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 16:43:05,292 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 16:43:05,293 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 16:44:21,025 INFO: Use cuda prefetch dataloader
+2025-11-04 16:44:21,026 INFO: Start training from epoch: 0, step: 0
+2025-11-04 16:44:22,761 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae
+2025-11-04 16:46:22,069 INFO: [38_co..][epoch:  0, step:     100, lr:(2.500e-04,)] [eta: 1 day, 11:32:10, time (data): 1.210 (0.013)] eagle_pixel_x2_opt: 4.0121e+00 l1_pixel_x2_opt: 3.5740e-02 fft_frequency_x2_opt: 3.2103e-02 eagle_pixel_x4_opt: 6.1645e+00 l1_pixel_x4_opt: 5.1576e-02 fft_frequency_x4_opt: 4.3915e-02 

04_11_2025/39/basicsr_options.yaml

@@ -0,0 +1,260 @@
+# GENERATE TIME: Tue Nov  4 21:31:42 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: true
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 10
+    pin_memory: true
+    persistent_workers: true
+    prefetch_mode: cuda
+    latent_dtype: bf16
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+    latent_dtype: bf16
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 8
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 180
+  num_heads:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  mlp_ratio: 2
+  resi_connection: 1conv
+  head_num_feat: 128
+  primary_head: x4
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: runs/04_11_2025/39_archived_20251104_212958/models/net_g_5000.pth
+  strict_load_g: true
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: true
+  mode: auto
+  dynamic: true
+  fullgraph: false
+  backend: inductor
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.0005
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.995
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  l1_latent_x2_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x2_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x2
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x2_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x2
+    kernel_size: 5
+    sigma: 1.0
+  l1_latent_x4_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x4_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x4
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x4_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x4
+    kernel_size: 5
+    sigma: 1.0
+val:
+  val_freq: 5000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 3
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: '39'
+name: '39'

04_11_2025/39_archived_20251104_171025/basicsr_options.yaml

@@ -0,0 +1,260 @@
+# GENERATE TIME: Tue Nov  4 17:04:38 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 8
+    pin_memory: true
+    persistent_workers: true
+    prefetch_mode: cuda
+    latent_dtype: bf16
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+    latent_dtype: bf16
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 8
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 180
+  num_heads:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  mlp_ratio: 2
+  resi_connection: 1conv
+  head_num_feat: 128
+  primary_head: x4
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth
+  strict_load_g: true
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: true
+  mode: auto
+  dynamic: true
+  fullgraph: false
+  backend: inductor
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.0002
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.995
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  l1_latent_x2_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x2_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x2
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x2_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x2
+    kernel_size: 5
+    sigma: 1.0
+  l1_latent_x4_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x4_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x4
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x4_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x4
+    kernel_size: 5
+    sigma: 1.0
+val:
+  val_freq: 500
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 3
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: '39'
+name: '39'

04_11_2025/39_archived_20251104_171025/train_39_20251104_170438.log

@@ -0,0 +1,633 @@
+2025-11-04 17:04:38,209 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 17:04:38,209 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 8
+      pin_memory: True
+      persistent_workers: True
+      prefetch_mode: cuda
+      latent_dtype: bf16
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+      latent_dtype: bf16
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 8
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 180
+    num_heads: [6, 6, 6, 6, 6, 6]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    head_num_feat: 128
+    primary_head: x4
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth
+    strict_load_g: True
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/visualization
+  ]
+  compile:[
+    enabled: True
+    mode: auto
+    dynamic: True
+    fullgraph: False
+    backend: inductor
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.0002
+      weight_decay: 0
+      betas: [0.9, 0.995]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    l1_latent_x2_opt:[
+      type: L1Loss
+      loss_weight: 1.0
+      reduction: mean
+      space: latent
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: latent
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    aux_downsample_x2_opt:[
+      type: DownsampleConsistencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: pixel
+      target: x2
+      down_factor: 2
+      mode: bicubic
+    ]
+    hf_pixel_x2_opt:[
+      type: HighFrequencyL1Loss
+      loss_weight: 0.05
+      reduction: mean
+      space: pixel
+      target: x2
+      kernel_size: 5
+      sigma: 1.0
+    ]
+    l1_latent_x4_opt:[
+      type: L1Loss
+      loss_weight: 1.0
+      reduction: mean
+      space: latent
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: latent
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    aux_downsample_x4_opt:[
+      type: DownsampleConsistencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: pixel
+      target: x4
+      down_factor: 2
+      mode: bicubic
+    ]
+    hf_pixel_x4_opt:[
+      type: HighFrequencyL1Loss
+      loss_weight: 0.05
+      reduction: mean
+      space: pixel
+      target: x4
+      kernel_size: 5
+      sigma: 1.0
+    ]
+  ]
+  val:[
+    val_freq: 500
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 3
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 39
+  name: 39
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 17:04:40,013 INFO: Use wandb logger with id=ao6kzb5j; project=Swin2SR-Latent-SR.
+2025-11-04 17:04:53,578 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 17:04:53,579 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 8
+	World size (gpu number): 6
+	Steps per epoch: 101219
+	Configured training steps: 125000
+	Approximate epochs to cover: 2.
+2025-11-04 17:04:53,582 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 17:04:53,583 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 17:04:53,584 WARNING: Forced find_unused_parameters=True for multi-head training because DDP requires it.
+2025-11-04 17:04:53,713 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 17:04:55,700 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 13,743,240
+2025-11-04 17:04:55,701 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=180, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=180, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(180, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(128, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(180, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(128, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 17:04:56,149 INFO: Loading SwinIRMultiHead from ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth [key=params].
+2025-11-04 17:04:56,172 WARNING: torch.compile failed for net_g; running in eager mode. Unrecognized mode=auto, should be one of: default, reduce-overhead, max-autotune-no-cudagraphs, max-autotune
+2025-11-04 17:04:56,174 INFO: Use EMA with decay: 0.999
+2025-11-04 17:04:56,279 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 17:04:56,340 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 17:04:56,341 INFO: Loading SwinIRMultiHead from ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth [key=params].
+2025-11-04 17:04:56,363 INFO: Torch.compile disabled for EMA network; validation runs in eager mode.
+2025-11-04 17:04:56,365 INFO: Loss [L1Loss] is created.
+2025-11-04 17:04:56,365 INFO: Initialized l1_latent_x2_opt in latent space (w=1.0).
+2025-11-04 17:04:56,366 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 17:04:56,367 INFO: Initialized fft_frequency_x2_opt in latent space (w=0.1).
+2025-11-04 17:04:56,368 INFO: Loss [DownsampleConsistencyLoss] is created.
+2025-11-04 17:04:56,369 INFO: Initialized aux_downsample_x2_opt in pixel space (w=0.1).
+2025-11-04 17:04:56,370 INFO: Loss [HighFrequencyL1Loss] is created.
+2025-11-04 17:04:56,371 INFO: Initialized hf_pixel_x2_opt in pixel space (w=0.05).
+2025-11-04 17:04:56,372 INFO: Loss [L1Loss] is created.
+2025-11-04 17:04:56,372 INFO: Initialized l1_latent_x4_opt in latent space (w=1.0).
+2025-11-04 17:04:56,372 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 17:04:56,373 INFO: Initialized fft_frequency_x4_opt in latent space (w=0.1).
+2025-11-04 17:04:56,374 INFO: Loss [DownsampleConsistencyLoss] is created.
+2025-11-04 17:04:56,375 INFO: Initialized aux_downsample_x4_opt in pixel space (w=0.1).
+2025-11-04 17:04:56,375 INFO: Loss [HighFrequencyL1Loss] is created.
+2025-11-04 17:04:56,376 INFO: Initialized hf_pixel_x4_opt in pixel space (w=0.05).
+2025-11-04 17:04:56,378 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 17:04:56,378 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 17:04:56,379 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 17:06:14,307 INFO: Use cuda prefetch dataloader
+2025-11-04 17:06:14,308 INFO: Start training from epoch: 0, step: 0
+2025-11-04 17:06:16,834 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae
+2025-11-04 17:08:06,469 INFO: [39..][epoch:  0, step:     100, lr:(2.000e-04,)] [eta: 1 day, 9:57:41, time (data): 1.122 (0.013)] l1_latent_x2_opt: 7.2474e-01 fft_frequency_x2_opt: 5.2474e-01 aux_downsample_x2_opt: 5.1147e-02 hf_pixel_x2_opt: 3.6085e-02 l1_latent_x4_opt: 8.1167e-01 fft_frequency_x4_opt: 6.0807e-01 aux_downsample_x4_opt: 6.0569e-02 hf_pixel_x4_opt: 3.5472e-02 
+2025-11-04 17:09:44,360 INFO: [39..][epoch:  0, step:     200, lr:(2.000e-04,)] [eta: 1 day, 9:56:04, time (data): 1.050 (0.007)] l1_latent_x2_opt: 7.0972e-01 fft_frequency_x2_opt: 5.1855e-01 aux_downsample_x2_opt: 5.6912e-02 hf_pixel_x2_opt: 3.7748e-02 l1_latent_x4_opt: 8.0876e-01 fft_frequency_x4_opt: 6.1263e-01 aux_downsample_x4_opt: 6.9721e-02 hf_pixel_x4_opt: 3.9873e-02 

04_11_2025/39_archived_20251104_171250/basicsr_options.yaml

@@ -0,0 +1,260 @@
+# GENERATE TIME: Tue Nov  4 17:10:25 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 32
+    pin_memory: true
+    persistent_workers: true
+    prefetch_mode: cuda
+    latent_dtype: bf16
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+    latent_dtype: bf16
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 8
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 180
+  num_heads:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  mlp_ratio: 2
+  resi_connection: 1conv
+  head_num_feat: 128
+  primary_head: x4
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth
+  strict_load_g: true
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: true
+  mode: auto
+  dynamic: true
+  fullgraph: false
+  backend: inductor
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.0005
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.995
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  l1_latent_x2_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x2_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x2
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x2_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x2
+    kernel_size: 5
+    sigma: 1.0
+  l1_latent_x4_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x4_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x4
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x4_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x4
+    kernel_size: 5
+    sigma: 1.0
+val:
+  val_freq: 500
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 3
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: '39'
+name: '39'

04_11_2025/39_archived_20251104_171250/train_39_20251104_171025.log

@@ -0,0 +1,631 @@
+2025-11-04 17:10:25,033 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 17:10:25,033 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 32
+      pin_memory: True
+      persistent_workers: True
+      prefetch_mode: cuda
+      latent_dtype: bf16
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+      latent_dtype: bf16
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 8
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 180
+    num_heads: [6, 6, 6, 6, 6, 6]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    head_num_feat: 128
+    primary_head: x4
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth
+    strict_load_g: True
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/visualization
+  ]
+  compile:[
+    enabled: True
+    mode: auto
+    dynamic: True
+    fullgraph: False
+    backend: inductor
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.0005
+      weight_decay: 0
+      betas: [0.9, 0.995]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    l1_latent_x2_opt:[
+      type: L1Loss
+      loss_weight: 1.0
+      reduction: mean
+      space: latent
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: latent
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    aux_downsample_x2_opt:[
+      type: DownsampleConsistencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: pixel
+      target: x2
+      down_factor: 2
+      mode: bicubic
+    ]
+    hf_pixel_x2_opt:[
+      type: HighFrequencyL1Loss
+      loss_weight: 0.05
+      reduction: mean
+      space: pixel
+      target: x2
+      kernel_size: 5
+      sigma: 1.0
+    ]
+    l1_latent_x4_opt:[
+      type: L1Loss
+      loss_weight: 1.0
+      reduction: mean
+      space: latent
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: latent
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    aux_downsample_x4_opt:[
+      type: DownsampleConsistencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: pixel
+      target: x4
+      down_factor: 2
+      mode: bicubic
+    ]
+    hf_pixel_x4_opt:[
+      type: HighFrequencyL1Loss
+      loss_weight: 0.05
+      reduction: mean
+      space: pixel
+      target: x4
+      kernel_size: 5
+      sigma: 1.0
+    ]
+  ]
+  val:[
+    val_freq: 500
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 3
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 39
+  name: 39
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 17:10:27,012 INFO: Use wandb logger with id=gvww3tbe; project=Swin2SR-Latent-SR.
+2025-11-04 17:10:40,716 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 17:10:40,717 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 32
+	World size (gpu number): 6
+	Steps per epoch: 25305
+	Configured training steps: 125000
+	Approximate epochs to cover: 5.
+2025-11-04 17:10:40,720 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 17:10:40,721 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 17:10:40,722 WARNING: Forced find_unused_parameters=True for multi-head training because DDP requires it.
+2025-11-04 17:10:40,851 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 17:10:42,926 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 13,743,240
+2025-11-04 17:10:42,927 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=180, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=180, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(180, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(128, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(180, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(128, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 17:10:42,978 INFO: Loading SwinIRMultiHead from ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth [key=params].
+2025-11-04 17:10:43,000 WARNING: torch.compile failed for net_g; running in eager mode. Unrecognized mode=auto, should be one of: default, reduce-overhead, max-autotune-no-cudagraphs, max-autotune
+2025-11-04 17:10:43,002 INFO: Use EMA with decay: 0.999
+2025-11-04 17:10:43,111 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 17:10:43,170 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 17:10:43,171 INFO: Loading SwinIRMultiHead from ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth [key=params].
+2025-11-04 17:10:43,192 INFO: Torch.compile disabled for EMA network; validation runs in eager mode.
+2025-11-04 17:10:43,194 INFO: Loss [L1Loss] is created.
+2025-11-04 17:10:43,194 INFO: Initialized l1_latent_x2_opt in latent space (w=1.0).
+2025-11-04 17:10:43,196 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 17:10:43,197 INFO: Initialized fft_frequency_x2_opt in latent space (w=0.1).
+2025-11-04 17:10:43,198 INFO: Loss [DownsampleConsistencyLoss] is created.
+2025-11-04 17:10:43,199 INFO: Initialized aux_downsample_x2_opt in pixel space (w=0.1).
+2025-11-04 17:10:43,200 INFO: Loss [HighFrequencyL1Loss] is created.
+2025-11-04 17:10:43,201 INFO: Initialized hf_pixel_x2_opt in pixel space (w=0.05).
+2025-11-04 17:10:43,202 INFO: Loss [L1Loss] is created.
+2025-11-04 17:10:43,202 INFO: Initialized l1_latent_x4_opt in latent space (w=1.0).
+2025-11-04 17:10:43,203 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 17:10:43,204 INFO: Initialized fft_frequency_x4_opt in latent space (w=0.1).
+2025-11-04 17:10:43,205 INFO: Loss [DownsampleConsistencyLoss] is created.
+2025-11-04 17:10:43,205 INFO: Initialized aux_downsample_x4_opt in pixel space (w=0.1).
+2025-11-04 17:10:43,206 INFO: Loss [HighFrequencyL1Loss] is created.
+2025-11-04 17:10:43,207 INFO: Initialized hf_pixel_x4_opt in pixel space (w=0.05).
+2025-11-04 17:10:43,209 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 17:10:43,209 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 17:10:43,210 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 17:11:59,782 INFO: Use cuda prefetch dataloader
+2025-11-04 17:11:59,783 INFO: Start training from epoch: 0, step: 0
+2025-11-04 17:12:01,526 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae

04_11_2025/39_archived_20251104_171656/basicsr_options.yaml

@@ -0,0 +1,260 @@
+# GENERATE TIME: Tue Nov  4 17:12:50 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: false
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 8
+    batch_size_per_gpu: 24
+    pin_memory: true
+    persistent_workers: true
+    prefetch_mode: cuda
+    latent_dtype: bf16
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+    latent_dtype: bf16
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 8
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 180
+  num_heads:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  mlp_ratio: 2
+  resi_connection: 1conv
+  head_num_feat: 128
+  primary_head: x4
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth
+  strict_load_g: true
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: true
+  mode: auto
+  dynamic: true
+  fullgraph: false
+  backend: inductor
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.0005
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.995
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  l1_latent_x2_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x2_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x2
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x2_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x2
+    kernel_size: 5
+    sigma: 1.0
+  l1_latent_x4_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x4_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x4
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x4_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x4
+    kernel_size: 5
+    sigma: 1.0
+val:
+  val_freq: 500
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 3
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: '39'
+name: '39'

04_11_2025/39_archived_20251104_171656/train_39_20251104_171250.log

@@ -0,0 +1,631 @@
+2025-11-04 17:12:50,164 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 17:12:50,164 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: False
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 8
+      batch_size_per_gpu: 24
+      pin_memory: True
+      persistent_workers: True
+      prefetch_mode: cuda
+      latent_dtype: bf16
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+      latent_dtype: bf16
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 8
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 180
+    num_heads: [6, 6, 6, 6, 6, 6]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    head_num_feat: 128
+    primary_head: x4
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth
+    strict_load_g: True
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/visualization
+  ]
+  compile:[
+    enabled: True
+    mode: auto
+    dynamic: True
+    fullgraph: False
+    backend: inductor
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.0005
+      weight_decay: 0
+      betas: [0.9, 0.995]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    l1_latent_x2_opt:[
+      type: L1Loss
+      loss_weight: 1.0
+      reduction: mean
+      space: latent
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: latent
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    aux_downsample_x2_opt:[
+      type: DownsampleConsistencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: pixel
+      target: x2
+      down_factor: 2
+      mode: bicubic
+    ]
+    hf_pixel_x2_opt:[
+      type: HighFrequencyL1Loss
+      loss_weight: 0.05
+      reduction: mean
+      space: pixel
+      target: x2
+      kernel_size: 5
+      sigma: 1.0
+    ]
+    l1_latent_x4_opt:[
+      type: L1Loss
+      loss_weight: 1.0
+      reduction: mean
+      space: latent
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: latent
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    aux_downsample_x4_opt:[
+      type: DownsampleConsistencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: pixel
+      target: x4
+      down_factor: 2
+      mode: bicubic
+    ]
+    hf_pixel_x4_opt:[
+      type: HighFrequencyL1Loss
+      loss_weight: 0.05
+      reduction: mean
+      space: pixel
+      target: x4
+      kernel_size: 5
+      sigma: 1.0
+    ]
+  ]
+  val:[
+    val_freq: 500
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 3
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 39
+  name: 39
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 17:12:51,776 INFO: Use wandb logger with id=b01rpt88; project=Swin2SR-Latent-SR.
+2025-11-04 17:13:04,536 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 17:13:04,538 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 24
+	World size (gpu number): 6
+	Steps per epoch: 33740
+	Configured training steps: 125000
+	Approximate epochs to cover: 4.
+2025-11-04 17:13:04,541 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 17:13:04,541 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 17:13:04,543 WARNING: Forced find_unused_parameters=True for multi-head training because DDP requires it.
+2025-11-04 17:13:04,672 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 17:13:06,749 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 13,743,240
+2025-11-04 17:13:06,750 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=180, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=180, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(180, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(128, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(180, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(128, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 17:13:06,803 INFO: Loading SwinIRMultiHead from ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth [key=params].
+2025-11-04 17:13:06,825 WARNING: torch.compile failed for net_g; running in eager mode. Unrecognized mode=auto, should be one of: default, reduce-overhead, max-autotune-no-cudagraphs, max-autotune
+2025-11-04 17:13:06,826 INFO: Use EMA with decay: 0.999
+2025-11-04 17:13:06,935 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 17:13:06,996 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 17:13:06,997 INFO: Loading SwinIRMultiHead from ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth [key=params].
+2025-11-04 17:13:07,017 INFO: Torch.compile disabled for EMA network; validation runs in eager mode.
+2025-11-04 17:13:07,019 INFO: Loss [L1Loss] is created.
+2025-11-04 17:13:07,019 INFO: Initialized l1_latent_x2_opt in latent space (w=1.0).
+2025-11-04 17:13:07,020 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 17:13:07,021 INFO: Initialized fft_frequency_x2_opt in latent space (w=0.1).
+2025-11-04 17:13:07,022 INFO: Loss [DownsampleConsistencyLoss] is created.
+2025-11-04 17:13:07,023 INFO: Initialized aux_downsample_x2_opt in pixel space (w=0.1).
+2025-11-04 17:13:07,024 INFO: Loss [HighFrequencyL1Loss] is created.
+2025-11-04 17:13:07,025 INFO: Initialized hf_pixel_x2_opt in pixel space (w=0.05).
+2025-11-04 17:13:07,025 INFO: Loss [L1Loss] is created.
+2025-11-04 17:13:07,025 INFO: Initialized l1_latent_x4_opt in latent space (w=1.0).
+2025-11-04 17:13:07,026 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 17:13:07,027 INFO: Initialized fft_frequency_x4_opt in latent space (w=0.1).
+2025-11-04 17:13:07,027 INFO: Loss [DownsampleConsistencyLoss] is created.
+2025-11-04 17:13:07,028 INFO: Initialized aux_downsample_x4_opt in pixel space (w=0.1).
+2025-11-04 17:13:07,029 INFO: Loss [HighFrequencyL1Loss] is created.
+2025-11-04 17:13:07,030 INFO: Initialized hf_pixel_x4_opt in pixel space (w=0.05).
+2025-11-04 17:13:07,031 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 17:13:07,032 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 17:13:07,033 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 17:15:41,590 INFO: Use cuda prefetch dataloader
+2025-11-04 17:15:41,591 INFO: Start training from epoch: 0, step: 0
+2025-11-04 17:15:44,793 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae

04_11_2025/39_archived_20251104_172026/basicsr_options.yaml

@@ -0,0 +1,260 @@
+# GENERATE TIME: Tue Nov  4 17:16:56 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: true
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 8
+    batch_size_per_gpu: 20
+    pin_memory: true
+    persistent_workers: true
+    prefetch_mode: cuda
+    latent_dtype: bf16
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+    latent_dtype: bf16
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 8
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 180
+  num_heads:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  mlp_ratio: 2
+  resi_connection: 1conv
+  head_num_feat: 128
+  primary_head: x4
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth
+  strict_load_g: true
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: true
+  mode: auto
+  dynamic: true
+  fullgraph: false
+  backend: inductor
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.0005
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.995
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  l1_latent_x2_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x2_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x2
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x2_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x2
+    kernel_size: 5
+    sigma: 1.0
+  l1_latent_x4_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x4_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x4
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x4_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x4
+    kernel_size: 5
+    sigma: 1.0
+val:
+  val_freq: 500
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 3
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: '39'
+name: '39'

04_11_2025/39_archived_20251104_172026/train_39_20251104_171656.log

@@ -0,0 +1,630 @@
+2025-11-04 17:16:56,392 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 17:16:56,392 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: True
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 8
+      batch_size_per_gpu: 20
+      pin_memory: True
+      persistent_workers: True
+      prefetch_mode: cuda
+      latent_dtype: bf16
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+      latent_dtype: bf16
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 8
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 180
+    num_heads: [6, 6, 6, 6, 6, 6]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    head_num_feat: 128
+    primary_head: x4
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth
+    strict_load_g: True
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/visualization
+  ]
+  compile:[
+    enabled: True
+    mode: auto
+    dynamic: True
+    fullgraph: False
+    backend: inductor
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.0005
+      weight_decay: 0
+      betas: [0.9, 0.995]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    l1_latent_x2_opt:[
+      type: L1Loss
+      loss_weight: 1.0
+      reduction: mean
+      space: latent
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: latent
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    aux_downsample_x2_opt:[
+      type: DownsampleConsistencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: pixel
+      target: x2
+      down_factor: 2
+      mode: bicubic
+    ]
+    hf_pixel_x2_opt:[
+      type: HighFrequencyL1Loss
+      loss_weight: 0.05
+      reduction: mean
+      space: pixel
+      target: x2
+      kernel_size: 5
+      sigma: 1.0
+    ]
+    l1_latent_x4_opt:[
+      type: L1Loss
+      loss_weight: 1.0
+      reduction: mean
+      space: latent
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: latent
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    aux_downsample_x4_opt:[
+      type: DownsampleConsistencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: pixel
+      target: x4
+      down_factor: 2
+      mode: bicubic
+    ]
+    hf_pixel_x4_opt:[
+      type: HighFrequencyL1Loss
+      loss_weight: 0.05
+      reduction: mean
+      space: pixel
+      target: x4
+      kernel_size: 5
+      sigma: 1.0
+    ]
+  ]
+  val:[
+    val_freq: 500
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 3
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 39
+  name: 39
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 17:16:58,092 INFO: Use wandb logger with id=9lx2vhyx; project=Swin2SR-Latent-SR.
+2025-11-04 17:17:10,541 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 17:17:10,542 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 20
+	World size (gpu number): 6
+	Steps per epoch: 40488
+	Configured training steps: 125000
+	Approximate epochs to cover: 4.
+2025-11-04 17:17:10,544 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 17:17:10,545 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 17:17:10,672 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 17:17:12,762 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 13,743,240
+2025-11-04 17:17:12,763 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=180, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=180, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(180, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(128, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(180, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(128, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 17:17:12,813 INFO: Loading SwinIRMultiHead from ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth [key=params].
+2025-11-04 17:17:12,835 WARNING: torch.compile failed for net_g; running in eager mode. Unrecognized mode=auto, should be one of: default, reduce-overhead, max-autotune-no-cudagraphs, max-autotune
+2025-11-04 17:17:12,837 INFO: Use EMA with decay: 0.999
+2025-11-04 17:17:12,945 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 17:17:13,005 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 17:17:13,006 INFO: Loading SwinIRMultiHead from ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth [key=params].
+2025-11-04 17:17:13,028 INFO: Torch.compile disabled for EMA network; validation runs in eager mode.
+2025-11-04 17:17:13,030 INFO: Loss [L1Loss] is created.
+2025-11-04 17:17:13,030 INFO: Initialized l1_latent_x2_opt in latent space (w=1.0).
+2025-11-04 17:17:13,031 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 17:17:13,032 INFO: Initialized fft_frequency_x2_opt in latent space (w=0.1).
+2025-11-04 17:17:13,033 INFO: Loss [DownsampleConsistencyLoss] is created.
+2025-11-04 17:17:13,034 INFO: Initialized aux_downsample_x2_opt in pixel space (w=0.1).
+2025-11-04 17:17:13,035 INFO: Loss [HighFrequencyL1Loss] is created.
+2025-11-04 17:17:13,036 INFO: Initialized hf_pixel_x2_opt in pixel space (w=0.05).
+2025-11-04 17:17:13,037 INFO: Loss [L1Loss] is created.
+2025-11-04 17:17:13,038 INFO: Initialized l1_latent_x4_opt in latent space (w=1.0).
+2025-11-04 17:17:13,038 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 17:17:13,038 INFO: Initialized fft_frequency_x4_opt in latent space (w=0.1).
+2025-11-04 17:17:13,038 INFO: Loss [DownsampleConsistencyLoss] is created.
+2025-11-04 17:17:13,038 INFO: Initialized aux_downsample_x4_opt in pixel space (w=0.1).
+2025-11-04 17:17:13,039 INFO: Loss [HighFrequencyL1Loss] is created.
+2025-11-04 17:17:13,040 INFO: Initialized hf_pixel_x4_opt in pixel space (w=0.05).
+2025-11-04 17:17:13,042 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 17:17:13,042 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 17:17:13,043 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 17:19:39,512 INFO: Use cuda prefetch dataloader
+2025-11-04 17:19:39,513 INFO: Start training from epoch: 0, step: 0
+2025-11-04 17:19:42,688 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae

04_11_2025/39_archived_20251104_172358/basicsr_options.yaml

@@ -0,0 +1,260 @@
+# GENERATE TIME: Tue Nov  4 17:20:26 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: true
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 8
+    batch_size_per_gpu: 16
+    pin_memory: true
+    persistent_workers: true
+    prefetch_mode: cuda
+    latent_dtype: bf16
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+    latent_dtype: bf16
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 8
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 180
+  num_heads:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  mlp_ratio: 2
+  resi_connection: 1conv
+  head_num_feat: 128
+  primary_head: x4
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth
+  strict_load_g: true
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: true
+  mode: auto
+  dynamic: true
+  fullgraph: false
+  backend: inductor
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.0005
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.995
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  l1_latent_x2_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x2_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x2
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x2_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x2
+    kernel_size: 5
+    sigma: 1.0
+  l1_latent_x4_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x4_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x4
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x4_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x4
+    kernel_size: 5
+    sigma: 1.0
+val:
+  val_freq: 500
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 3
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: '39'
+name: '39'

04_11_2025/39_archived_20251104_172358/train_39_20251104_172026.log

@@ -0,0 +1,630 @@
+2025-11-04 17:20:26,274 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 17:20:26,274 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: True
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 8
+      batch_size_per_gpu: 16
+      pin_memory: True
+      persistent_workers: True
+      prefetch_mode: cuda
+      latent_dtype: bf16
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+      latent_dtype: bf16
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 8
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 180
+    num_heads: [6, 6, 6, 6, 6, 6]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    head_num_feat: 128
+    primary_head: x4
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth
+    strict_load_g: True
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/visualization
+  ]
+  compile:[
+    enabled: True
+    mode: auto
+    dynamic: True
+    fullgraph: False
+    backend: inductor
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.0005
+      weight_decay: 0
+      betas: [0.9, 0.995]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    l1_latent_x2_opt:[
+      type: L1Loss
+      loss_weight: 1.0
+      reduction: mean
+      space: latent
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: latent
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    aux_downsample_x2_opt:[
+      type: DownsampleConsistencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: pixel
+      target: x2
+      down_factor: 2
+      mode: bicubic
+    ]
+    hf_pixel_x2_opt:[
+      type: HighFrequencyL1Loss
+      loss_weight: 0.05
+      reduction: mean
+      space: pixel
+      target: x2
+      kernel_size: 5
+      sigma: 1.0
+    ]
+    l1_latent_x4_opt:[
+      type: L1Loss
+      loss_weight: 1.0
+      reduction: mean
+      space: latent
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: latent
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    aux_downsample_x4_opt:[
+      type: DownsampleConsistencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: pixel
+      target: x4
+      down_factor: 2
+      mode: bicubic
+    ]
+    hf_pixel_x4_opt:[
+      type: HighFrequencyL1Loss
+      loss_weight: 0.05
+      reduction: mean
+      space: pixel
+      target: x4
+      kernel_size: 5
+      sigma: 1.0
+    ]
+  ]
+  val:[
+    val_freq: 500
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 3
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 39
+  name: 39
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 17:20:27,941 INFO: Use wandb logger with id=mg0c5tu9; project=Swin2SR-Latent-SR.
+2025-11-04 17:20:40,469 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 17:20:40,470 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 16
+	World size (gpu number): 6
+	Steps per epoch: 50610
+	Configured training steps: 125000
+	Approximate epochs to cover: 3.
+2025-11-04 17:20:40,474 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 17:20:40,475 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 17:20:40,602 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 17:20:42,650 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 13,743,240
+2025-11-04 17:20:42,651 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=180, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=180, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(180, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(128, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(180, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(128, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 17:20:42,724 INFO: Loading SwinIRMultiHead from ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth [key=params].
+2025-11-04 17:20:42,754 WARNING: torch.compile failed for net_g; running in eager mode. Unrecognized mode=auto, should be one of: default, reduce-overhead, max-autotune-no-cudagraphs, max-autotune
+2025-11-04 17:20:42,756 INFO: Use EMA with decay: 0.999
+2025-11-04 17:20:42,957 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 17:20:43,049 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 17:20:43,050 INFO: Loading SwinIRMultiHead from ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth [key=params].
+2025-11-04 17:20:43,079 INFO: Torch.compile disabled for EMA network; validation runs in eager mode.
+2025-11-04 17:20:43,082 INFO: Loss [L1Loss] is created.
+2025-11-04 17:20:43,083 INFO: Initialized l1_latent_x2_opt in latent space (w=1.0).
+2025-11-04 17:20:43,085 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 17:20:43,086 INFO: Initialized fft_frequency_x2_opt in latent space (w=0.1).
+2025-11-04 17:20:43,087 INFO: Loss [DownsampleConsistencyLoss] is created.
+2025-11-04 17:20:43,088 INFO: Initialized aux_downsample_x2_opt in pixel space (w=0.1).
+2025-11-04 17:20:43,089 INFO: Loss [HighFrequencyL1Loss] is created.
+2025-11-04 17:20:43,090 INFO: Initialized hf_pixel_x2_opt in pixel space (w=0.05).
+2025-11-04 17:20:43,091 INFO: Loss [L1Loss] is created.
+2025-11-04 17:20:43,092 INFO: Initialized l1_latent_x4_opt in latent space (w=1.0).
+2025-11-04 17:20:43,093 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 17:20:43,094 INFO: Initialized fft_frequency_x4_opt in latent space (w=0.1).
+2025-11-04 17:20:43,095 INFO: Loss [DownsampleConsistencyLoss] is created.
+2025-11-04 17:20:43,096 INFO: Initialized aux_downsample_x4_opt in pixel space (w=0.1).
+2025-11-04 17:20:43,097 INFO: Loss [HighFrequencyL1Loss] is created.
+2025-11-04 17:20:43,098 INFO: Initialized hf_pixel_x4_opt in pixel space (w=0.05).
+2025-11-04 17:20:43,100 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 17:20:43,100 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 17:20:43,101 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 17:23:12,237 INFO: Use cuda prefetch dataloader
+2025-11-04 17:23:12,238 INFO: Start training from epoch: 0, step: 0
+2025-11-04 17:23:15,243 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae

04_11_2025/39_archived_20251104_174404/basicsr_options.yaml

@@ -0,0 +1,260 @@
+# GENERATE TIME: Tue Nov  4 17:23:58 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: true
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 12
+    pin_memory: true
+    persistent_workers: true
+    prefetch_mode: cuda
+    latent_dtype: bf16
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+    latent_dtype: bf16
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 8
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 180
+  num_heads:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  mlp_ratio: 2
+  resi_connection: 1conv
+  head_num_feat: 128
+  primary_head: x4
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth
+  strict_load_g: true
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: true
+  mode: auto
+  dynamic: true
+  fullgraph: false
+  backend: inductor
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.0005
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.995
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  l1_latent_x2_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x2_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x2
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x2_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x2
+    kernel_size: 5
+    sigma: 1.0
+  l1_latent_x4_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x4_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x4
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x4_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x4
+    kernel_size: 5
+    sigma: 1.0
+val:
+  val_freq: 500
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 3
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: '39'
+name: '39'

04_11_2025/39_archived_20251104_174404/train_39_20251104_172358.log

@@ -0,0 +1,645 @@
+2025-11-04 17:23:58,544 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 17:23:58,544 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 6
+  manual_seed: 0
+  find_unused_parameters: True
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 12
+      pin_memory: True
+      persistent_workers: True
+      prefetch_mode: cuda
+      latent_dtype: bf16
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+      latent_dtype: bf16
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 8
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 180
+    num_heads: [6, 6, 6, 6, 6, 6]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    head_num_feat: 128
+    primary_head: x4
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth
+    strict_load_g: True
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/visualization
+  ]
+  compile:[
+    enabled: True
+    mode: auto
+    dynamic: True
+    fullgraph: False
+    backend: inductor
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.0005
+      weight_decay: 0
+      betas: [0.9, 0.995]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    l1_latent_x2_opt:[
+      type: L1Loss
+      loss_weight: 1.0
+      reduction: mean
+      space: latent
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: latent
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    aux_downsample_x2_opt:[
+      type: DownsampleConsistencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: pixel
+      target: x2
+      down_factor: 2
+      mode: bicubic
+    ]
+    hf_pixel_x2_opt:[
+      type: HighFrequencyL1Loss
+      loss_weight: 0.05
+      reduction: mean
+      space: pixel
+      target: x2
+      kernel_size: 5
+      sigma: 1.0
+    ]
+    l1_latent_x4_opt:[
+      type: L1Loss
+      loss_weight: 1.0
+      reduction: mean
+      space: latent
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: latent
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    aux_downsample_x4_opt:[
+      type: DownsampleConsistencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: pixel
+      target: x4
+      down_factor: 2
+      mode: bicubic
+    ]
+    hf_pixel_x4_opt:[
+      type: HighFrequencyL1Loss
+      loss_weight: 0.05
+      reduction: mean
+      space: pixel
+      target: x4
+      kernel_size: 5
+      sigma: 1.0
+    ]
+  ]
+  val:[
+    val_freq: 500
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 3
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 39
+  name: 39
+  dist: True
+  rank: 0
+  world_size: 6
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 17:24:00,577 INFO: Use wandb logger with id=809fuhnl; project=Swin2SR-Latent-SR.
+2025-11-04 17:24:13,622 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 17:24:13,623 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 12
+	World size (gpu number): 6
+	Steps per epoch: 67480
+	Configured training steps: 125000
+	Approximate epochs to cover: 2.
+2025-11-04 17:24:13,627 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 17:24:13,627 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 17:24:13,762 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 17:24:15,809 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 13,743,240
+2025-11-04 17:24:15,810 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=180, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=180, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(180, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(128, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(180, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(128, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 17:24:15,862 INFO: Loading SwinIRMultiHead from ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth [key=params].
+2025-11-04 17:24:15,884 WARNING: torch.compile failed for net_g; running in eager mode. Unrecognized mode=auto, should be one of: default, reduce-overhead, max-autotune-no-cudagraphs, max-autotune
+2025-11-04 17:24:15,885 INFO: Use EMA with decay: 0.999
+2025-11-04 17:24:15,993 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 17:24:16,054 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 17:24:16,055 INFO: Loading SwinIRMultiHead from ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth [key=params].
+2025-11-04 17:24:16,076 INFO: Torch.compile disabled for EMA network; validation runs in eager mode.
+2025-11-04 17:24:16,078 INFO: Loss [L1Loss] is created.
+2025-11-04 17:24:16,078 INFO: Initialized l1_latent_x2_opt in latent space (w=1.0).
+2025-11-04 17:24:16,079 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 17:24:16,080 INFO: Initialized fft_frequency_x2_opt in latent space (w=0.1).
+2025-11-04 17:24:16,082 INFO: Loss [DownsampleConsistencyLoss] is created.
+2025-11-04 17:24:16,083 INFO: Initialized aux_downsample_x2_opt in pixel space (w=0.1).
+2025-11-04 17:24:16,084 INFO: Loss [HighFrequencyL1Loss] is created.
+2025-11-04 17:24:16,085 INFO: Initialized hf_pixel_x2_opt in pixel space (w=0.05).
+2025-11-04 17:24:16,087 INFO: Loss [L1Loss] is created.
+2025-11-04 17:24:16,087 INFO: Initialized l1_latent_x4_opt in latent space (w=1.0).
+2025-11-04 17:24:16,088 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 17:24:16,089 INFO: Initialized fft_frequency_x4_opt in latent space (w=0.1).
+2025-11-04 17:24:16,090 INFO: Loss [DownsampleConsistencyLoss] is created.
+2025-11-04 17:24:16,091 INFO: Initialized aux_downsample_x4_opt in pixel space (w=0.1).
+2025-11-04 17:24:16,091 INFO: Loss [HighFrequencyL1Loss] is created.
+2025-11-04 17:24:16,092 INFO: Initialized hf_pixel_x4_opt in pixel space (w=0.05).
+2025-11-04 17:24:16,095 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 17:24:16,095 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 17:24:16,096 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 17:25:32,845 INFO: Use cuda prefetch dataloader
+2025-11-04 17:25:32,846 INFO: Start training from epoch: 0, step: 0
+2025-11-04 17:25:34,862 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae
+2025-11-04 17:27:59,195 INFO: [39..][epoch:  0, step:     100, lr:(5.000e-04,)] [eta: 1 day, 21:51:56, time (data): 1.463 (0.014)] l1_latent_x2_opt: 7.2477e-01 fft_frequency_x2_opt: 5.1790e-01 aux_downsample_x2_opt: 5.1585e-02 hf_pixel_x2_opt: 3.2817e-02 l1_latent_x4_opt: 8.1886e-01 fft_frequency_x4_opt: 6.0872e-01 aux_downsample_x4_opt: 6.5579e-02 hf_pixel_x4_opt: 3.4039e-02 
+2025-11-04 17:30:13,555 INFO: [39..][epoch:  0, step:     200, lr:(5.000e-04,)] [eta: 1 day, 22:12:05, time (data): 1.404 (0.007)] l1_latent_x2_opt: 7.2126e-01 fft_frequency_x2_opt: 5.1986e-01 aux_downsample_x2_opt: 5.6676e-02 hf_pixel_x2_opt: 3.7261e-02 l1_latent_x4_opt: 8.2266e-01 fft_frequency_x4_opt: 6.1523e-01 aux_downsample_x4_opt: 7.1628e-02 hf_pixel_x4_opt: 3.9245e-02 
+2025-11-04 17:32:28,228 INFO: [39..][epoch:  0, step:     300, lr:(5.000e-04,)] [eta: 1 day, 22:19:31, time (data): 1.347 (0.000)] l1_latent_x2_opt: 7.2014e-01 fft_frequency_x2_opt: 5.1628e-01 aux_downsample_x2_opt: 5.6589e-02 hf_pixel_x2_opt: 3.3866e-02 l1_latent_x4_opt: 8.2612e-01 fft_frequency_x4_opt: 6.0872e-01 aux_downsample_x4_opt: 7.1573e-02 hf_pixel_x4_opt: 3.5252e-02 
+2025-11-04 17:34:42,753 INFO: [39..][epoch:  0, step:     400, lr:(5.000e-04,)] [eta: 1 day, 22:21:21, time (data): 1.346 (0.000)] l1_latent_x2_opt: 7.1569e-01 fft_frequency_x2_opt: 5.2409e-01 aux_downsample_x2_opt: 5.2701e-02 hf_pixel_x2_opt: 3.3162e-02 l1_latent_x4_opt: 8.2035e-01 fft_frequency_x4_opt: 6.1784e-01 aux_downsample_x4_opt: 6.6638e-02 hf_pixel_x4_opt: 3.5358e-02 
+2025-11-04 17:36:57,066 INFO: [39..][epoch:  0, step:     500, lr:(5.000e-04,)] [eta: 1 day, 22:20:42, time (data): 1.343 (0.000)] l1_latent_x2_opt: 7.2576e-01 fft_frequency_x2_opt: 5.2441e-01 aux_downsample_x2_opt: 4.9497e-02 hf_pixel_x2_opt: 3.1129e-02 l1_latent_x4_opt: 8.2374e-01 fft_frequency_x4_opt: 6.1068e-01 aux_downsample_x4_opt: 6.0516e-02 hf_pixel_x4_opt: 3.1338e-02 
+2025-11-04 17:39:09,132 INFO: Validation val_x2
+	 # l1_latent: 0.7635	Best: 0.7635 @ 500 iter
+	 # pixel_psnr_pt: 28.7287	Best: 28.7287 @ 500 iter
+
+2025-11-04 17:41:18,080 INFO: Validation val_x4
+	 # l1_latent: 0.8521	Best: 0.8521 @ 500 iter
+	 # l2_latent: 1.1937	Best: 1.1937 @ 500 iter
+	 # pixel_psnr_pt: 26.2661	Best: 26.2661 @ 500 iter
+
+2025-11-04 17:43:32,744 INFO: [39..][epoch:  0, step:     600, lr:(5.000e-04,)] [eta: 2 days, 13:21:09, time (data): 1.345 (0.000)] l1_latent_x2_opt: 7.3411e-01 fft_frequency_x2_opt: 5.2734e-01 aux_downsample_x2_opt: 5.4268e-02 hf_pixel_x2_opt: 3.5028e-02 l1_latent_x4_opt: 8.4425e-01 fft_frequency_x4_opt: 6.2565e-01 aux_downsample_x4_opt: 7.0657e-02 hf_pixel_x4_opt: 3.7107e-02 

04_11_2025/39_archived_20251104_212958/basicsr_options.yaml

@@ -0,0 +1,260 @@
+# GENERATE TIME: Tue Nov  4 17:44:04 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: true
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 12
+    pin_memory: true
+    persistent_workers: true
+    prefetch_mode: cuda
+    latent_dtype: bf16
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+    latent_dtype: bf16
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 8
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 180
+  num_heads:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  mlp_ratio: 2
+  resi_connection: 1conv
+  head_num_feat: 128
+  primary_head: x4
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth
+  strict_load_g: true
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: true
+  mode: auto
+  dynamic: true
+  fullgraph: false
+  backend: inductor
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.0005
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.995
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  l1_latent_x2_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x2_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x2
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x2_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x2
+    kernel_size: 5
+    sigma: 1.0
+  l1_latent_x4_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x4_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x4
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x4_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x4
+    kernel_size: 5
+    sigma: 1.0
+val:
+  val_freq: 5000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 3
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: '39'
+name: '39'

04_11_2025/39_archived_20251104_212958/train_39_20251104_174404.log

@@ -0,0 +1,690 @@
+2025-11-04 17:44:04,636 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 17:44:04,636 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 8
+  manual_seed: 0
+  find_unused_parameters: True
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 12
+      pin_memory: True
+      persistent_workers: True
+      prefetch_mode: cuda
+      latent_dtype: bf16
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+      latent_dtype: bf16
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 8
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 180
+    num_heads: [6, 6, 6, 6, 6, 6]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    head_num_feat: 128
+    primary_head: x4
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth
+    strict_load_g: True
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/visualization
+  ]
+  compile:[
+    enabled: True
+    mode: auto
+    dynamic: True
+    fullgraph: False
+    backend: inductor
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.0005
+      weight_decay: 0
+      betas: [0.9, 0.995]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    l1_latent_x2_opt:[
+      type: L1Loss
+      loss_weight: 1.0
+      reduction: mean
+      space: latent
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: latent
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    aux_downsample_x2_opt:[
+      type: DownsampleConsistencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: pixel
+      target: x2
+      down_factor: 2
+      mode: bicubic
+    ]
+    hf_pixel_x2_opt:[
+      type: HighFrequencyL1Loss
+      loss_weight: 0.05
+      reduction: mean
+      space: pixel
+      target: x2
+      kernel_size: 5
+      sigma: 1.0
+    ]
+    l1_latent_x4_opt:[
+      type: L1Loss
+      loss_weight: 1.0
+      reduction: mean
+      space: latent
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: latent
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    aux_downsample_x4_opt:[
+      type: DownsampleConsistencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: pixel
+      target: x4
+      down_factor: 2
+      mode: bicubic
+    ]
+    hf_pixel_x4_opt:[
+      type: HighFrequencyL1Loss
+      loss_weight: 0.05
+      reduction: mean
+      space: pixel
+      target: x4
+      kernel_size: 5
+      sigma: 1.0
+    ]
+  ]
+  val:[
+    val_freq: 5000
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 3
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 39
+  name: 39
+  dist: True
+  rank: 0
+  world_size: 8
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 17:44:06,403 INFO: Use wandb logger with id=j67rxaon; project=Swin2SR-Latent-SR.
+2025-11-04 17:44:19,059 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 17:44:19,060 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 12
+	World size (gpu number): 8
+	Steps per epoch: 50610
+	Configured training steps: 125000
+	Approximate epochs to cover: 3.
+2025-11-04 17:44:19,064 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 17:44:19,065 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 17:44:19,195 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 17:44:21,462 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 13,743,240
+2025-11-04 17:44:21,463 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=180, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=180, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(180, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(128, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(180, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(128, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)
+2025-11-04 17:44:21,542 INFO: Loading SwinIRMultiHead from ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth [key=params].
+2025-11-04 17:44:21,574 WARNING: torch.compile failed for net_g; running in eager mode. Unrecognized mode=auto, should be one of: default, reduce-overhead, max-autotune-no-cudagraphs, max-autotune
+2025-11-04 17:44:21,576 INFO: Use EMA with decay: 0.999
+2025-11-04 17:44:21,758 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 17:44:21,851 INFO: Loading: params_ema does not exist, use params.
+2025-11-04 17:44:21,852 INFO: Loading SwinIRMultiHead from ./pretrained_weights/01_11_2025/31/models/net_g_110000.pth [key=params].
+2025-11-04 17:44:21,881 INFO: Torch.compile disabled for EMA network; validation runs in eager mode.
+2025-11-04 17:44:21,884 INFO: Loss [L1Loss] is created.
+2025-11-04 17:44:21,884 INFO: Initialized l1_latent_x2_opt in latent space (w=1.0).
+2025-11-04 17:44:21,885 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 17:44:21,886 INFO: Initialized fft_frequency_x2_opt in latent space (w=0.1).
+2025-11-04 17:44:21,887 INFO: Loss [DownsampleConsistencyLoss] is created.
+2025-11-04 17:44:21,888 INFO: Initialized aux_downsample_x2_opt in pixel space (w=0.1).
+2025-11-04 17:44:21,890 INFO: Loss [HighFrequencyL1Loss] is created.
+2025-11-04 17:44:21,891 INFO: Initialized hf_pixel_x2_opt in pixel space (w=0.05).
+2025-11-04 17:44:21,892 INFO: Loss [L1Loss] is created.
+2025-11-04 17:44:21,893 INFO: Initialized l1_latent_x4_opt in latent space (w=1.0).
+2025-11-04 17:44:21,894 INFO: Loss [FFTFrequencyLoss] is created.
+2025-11-04 17:44:21,895 INFO: Initialized fft_frequency_x4_opt in latent space (w=0.1).
+2025-11-04 17:44:21,895 INFO: Loss [DownsampleConsistencyLoss] is created.
+2025-11-04 17:44:21,895 INFO: Initialized aux_downsample_x4_opt in pixel space (w=0.1).
+2025-11-04 17:44:21,896 INFO: Loss [HighFrequencyL1Loss] is created.
+2025-11-04 17:44:21,897 INFO: Initialized hf_pixel_x4_opt in pixel space (w=0.05).
+2025-11-04 17:44:21,899 INFO: Precision configuration — train: bf16, eval: fp32
+2025-11-04 17:44:21,899 INFO: Gradient clipping enabled (G:norm(max_norm=0.4, norm_type=2.0)).
+2025-11-04 17:44:21,899 INFO: Model [SwinIRLatentModelMultiHead] is created.
+2025-11-04 17:45:39,919 INFO: Use cuda prefetch dataloader
+2025-11-04 17:45:39,920 INFO: Start training from epoch: 0, step: 0
+2025-11-04 17:45:42,256 INFO: Loading VAE(name=flux_vae, kind=kl) from wolfgangblack/flux_vae
+2025-11-04 17:48:06,672 INFO: [39..][epoch:  0, step:     100, lr:(5.000e-04,)] [eta: 1 day, 22:03:03, time (data): 1.468 (0.015)] l1_latent_x2_opt: 7.1538e-01 fft_frequency_x2_opt: 5.1611e-01 aux_downsample_x2_opt: 5.2473e-02 hf_pixel_x2_opt: 3.3547e-02 l1_latent_x4_opt: 8.1259e-01 fft_frequency_x4_opt: 5.9814e-01 aux_downsample_x4_opt: 6.4306e-02 hf_pixel_x4_opt: 3.4876e-02 
+2025-11-04 17:50:21,752 INFO: [39..][epoch:  0, step:     200, lr:(5.000e-04,)] [eta: 1 day, 22:25:06, time (data): 1.409 (0.007)] l1_latent_x2_opt: 7.2462e-01 fft_frequency_x2_opt: 5.1978e-01 aux_downsample_x2_opt: 5.8646e-02 hf_pixel_x2_opt: 3.8853e-02 l1_latent_x4_opt: 8.2175e-01 fft_frequency_x4_opt: 6.0938e-01 aux_downsample_x4_opt: 7.1363e-02 hf_pixel_x4_opt: 3.9566e-02 
+2025-11-04 17:52:36,893 INFO: [39..][epoch:  0, step:     300, lr:(5.000e-04,)] [eta: 1 day, 22:31:27, time (data): 1.351 (0.000)] l1_latent_x2_opt: 7.1494e-01 fft_frequency_x2_opt: 5.2539e-01 aux_downsample_x2_opt: 5.5444e-02 hf_pixel_x2_opt: 3.6638e-02 l1_latent_x4_opt: 8.1160e-01 fft_frequency_x4_opt: 6.2158e-01 aux_downsample_x4_opt: 7.0688e-02 hf_pixel_x4_opt: 3.8187e-02 
+2025-11-04 17:54:51,895 INFO: [39..][epoch:  0, step:     400, lr:(5.000e-04,)] [eta: 1 day, 22:32:46, time (data): 1.351 (0.000)] l1_latent_x2_opt: 7.2170e-01 fft_frequency_x2_opt: 5.2588e-01 aux_downsample_x2_opt: 5.4698e-02 hf_pixel_x2_opt: 3.4865e-02 l1_latent_x4_opt: 8.2728e-01 fft_frequency_x4_opt: 6.1133e-01 aux_downsample_x4_opt: 6.7423e-02 hf_pixel_x4_opt: 3.6045e-02 
+2025-11-04 17:57:07,534 INFO: [39..][epoch:  0, step:     500, lr:(5.000e-04,)] [eta: 1 day, 22:35:19, time (data): 1.357 (0.000)] l1_latent_x2_opt: 7.1057e-01 fft_frequency_x2_opt: 5.1440e-01 aux_downsample_x2_opt: 5.4125e-02 hf_pixel_x2_opt: 3.3586e-02 l1_latent_x4_opt: 8.1176e-01 fft_frequency_x4_opt: 6.1230e-01 aux_downsample_x4_opt: 7.0314e-02 hf_pixel_x4_opt: 3.7586e-02 
+2025-11-04 17:59:23,634 INFO: [39..][epoch:  0, step:     600, lr:(5.000e-04,)] [eta: 1 day, 22:37:50, time (data): 1.359 (0.000)] l1_latent_x2_opt: 7.1819e-01 fft_frequency_x2_opt: 5.1831e-01 aux_downsample_x2_opt: 5.4263e-02 hf_pixel_x2_opt: 3.5276e-02 l1_latent_x4_opt: 8.2229e-01 fft_frequency_x4_opt: 6.1182e-01 aux_downsample_x4_opt: 7.0281e-02 hf_pixel_x4_opt: 3.8503e-02 
+2025-11-04 18:01:38,566 INFO: [39..][epoch:  0, step:     700, lr:(5.000e-04,)] [eta: 1 day, 22:35:33, time (data): 1.349 (0.000)] l1_latent_x2_opt: 7.3096e-01 fft_frequency_x2_opt: 5.3174e-01 aux_downsample_x2_opt: 5.4279e-02 hf_pixel_x2_opt: 3.4074e-02 l1_latent_x4_opt: 8.2776e-01 fft_frequency_x4_opt: 6.2500e-01 aux_downsample_x4_opt: 6.8199e-02 hf_pixel_x4_opt: 3.5843e-02 
+2025-11-04 18:03:53,449 INFO: [39..][epoch:  0, step:     800, lr:(5.000e-04,)] [eta: 1 day, 22:33:09, time (data): 1.349 (0.000)] l1_latent_x2_opt: 7.3620e-01 fft_frequency_x2_opt: 5.2246e-01 aux_downsample_x2_opt: 5.5530e-02 hf_pixel_x2_opt: 3.6471e-02 l1_latent_x4_opt: 8.3013e-01 fft_frequency_x4_opt: 6.1523e-01 aux_downsample_x4_opt: 7.0553e-02 hf_pixel_x4_opt: 3.8146e-02 
+2025-11-04 18:06:08,534 INFO: [39..][epoch:  0, step:     900, lr:(5.000e-04,)] [eta: 1 day, 22:31:14, time (data): 1.351 (0.000)] l1_latent_x2_opt: 7.2444e-01 fft_frequency_x2_opt: 5.2930e-01 aux_downsample_x2_opt: 4.8316e-02 hf_pixel_x2_opt: 3.1387e-02 l1_latent_x4_opt: 8.1942e-01 fft_frequency_x4_opt: 6.1182e-01 aux_downsample_x4_opt: 6.1079e-02 hf_pixel_x4_opt: 3.2247e-02 
+2025-11-04 18:08:23,633 INFO: [39..][epoch:  0, step:   1,000, lr:(5.000e-04,)] [eta: 1 day, 22:29:18, time (data): 1.351 (0.000)] l1_latent_x2_opt: 7.3483e-01 fft_frequency_x2_opt: 5.1465e-01 aux_downsample_x2_opt: 5.8090e-02 hf_pixel_x2_opt: 3.7956e-02 l1_latent_x4_opt: 8.3406e-01 fft_frequency_x4_opt: 6.0596e-01 aux_downsample_x4_opt: 7.0483e-02 hf_pixel_x4_opt: 3.7886e-02 
+2025-11-04 18:10:38,554 INFO: [39..][epoch:  0, step:   1,100, lr:(5.000e-04,)] [eta: 1 day, 22:26:57, time (data): 1.349 (0.000)] l1_latent_x2_opt: 7.1678e-01 fft_frequency_x2_opt: 5.1831e-01 aux_downsample_x2_opt: 5.2998e-02 hf_pixel_x2_opt: 3.3298e-02 l1_latent_x4_opt: 8.0588e-01 fft_frequency_x4_opt: 6.0791e-01 aux_downsample_x4_opt: 6.3533e-02 hf_pixel_x4_opt: 3.4202e-02 
+2025-11-04 18:12:53,495 INFO: [39..][epoch:  0, step:   1,200, lr:(5.000e-04,)] [eta: 1 day, 22:24:40, time (data): 1.349 (0.000)] l1_latent_x2_opt: 7.1548e-01 fft_frequency_x2_opt: 5.1685e-01 aux_downsample_x2_opt: 4.7315e-02 hf_pixel_x2_opt: 2.9344e-02 l1_latent_x4_opt: 8.0816e-01 fft_frequency_x4_opt: 6.0254e-01 aux_downsample_x4_opt: 5.9042e-02 hf_pixel_x4_opt: 3.0520e-02 
+2025-11-04 18:15:08,326 INFO: [39..][epoch:  0, step:   1,300, lr:(5.000e-04,)] [eta: 1 day, 22:22:13, time (data): 1.348 (0.000)] l1_latent_x2_opt: 7.3182e-01 fft_frequency_x2_opt: 5.2222e-01 aux_downsample_x2_opt: 5.0035e-02 hf_pixel_x2_opt: 3.3350e-02 l1_latent_x4_opt: 8.2970e-01 fft_frequency_x4_opt: 6.1279e-01 aux_downsample_x4_opt: 6.3545e-02 hf_pixel_x4_opt: 3.4917e-02 
+2025-11-04 18:17:23,751 INFO: [39..][epoch:  0, step:   1,400, lr:(5.000e-04,)] [eta: 1 day, 22:20:40, time (data): 1.351 (0.000)] l1_latent_x2_opt: 7.2209e-01 fft_frequency_x2_opt: 5.1953e-01 aux_downsample_x2_opt: 4.8760e-02 hf_pixel_x2_opt: 3.2254e-02 l1_latent_x4_opt: 8.1701e-01 fft_frequency_x4_opt: 6.1035e-01 aux_downsample_x4_opt: 6.1903e-02 hf_pixel_x4_opt: 3.3500e-02 
+2025-11-04 18:19:39,029 INFO: [39..][epoch:  0, step:   1,500, lr:(5.000e-04,)] [eta: 1 day, 22:18:49, time (data): 1.351 (0.000)] l1_latent_x2_opt: 7.2367e-01 fft_frequency_x2_opt: 5.1733e-01 aux_downsample_x2_opt: 5.2854e-02 hf_pixel_x2_opt: 3.4301e-02 l1_latent_x4_opt: 8.1990e-01 fft_frequency_x4_opt: 6.0938e-01 aux_downsample_x4_opt: 6.7018e-02 hf_pixel_x4_opt: 3.5185e-02 
+2025-11-04 18:21:53,877 INFO: [39..][epoch:  0, step:   1,600, lr:(5.000e-04,)] [eta: 1 day, 22:16:22, time (data): 1.350 (0.000)] l1_latent_x2_opt: 7.2551e-01 fft_frequency_x2_opt: 5.2417e-01 aux_downsample_x2_opt: 5.1626e-02 hf_pixel_x2_opt: 3.4591e-02 l1_latent_x4_opt: 8.2464e-01 fft_frequency_x4_opt: 6.1279e-01 aux_downsample_x4_opt: 6.5784e-02 hf_pixel_x4_opt: 3.7126e-02 
+2025-11-04 18:24:11,059 INFO: [39..][epoch:  0, step:   1,700, lr:(5.000e-04,)] [eta: 1 day, 22:16:45, time (data): 1.374 (0.000)] l1_latent_x2_opt: 7.0685e-01 fft_frequency_x2_opt: 5.1196e-01 aux_downsample_x2_opt: 4.6546e-02 hf_pixel_x2_opt: 3.1026e-02 l1_latent_x4_opt: 7.9315e-01 fft_frequency_x4_opt: 5.9766e-01 aux_downsample_x4_opt: 6.0185e-02 hf_pixel_x4_opt: 3.2489e-02 
+2025-11-04 18:26:26,875 INFO: [39..][epoch:  0, step:   1,800, lr:(5.000e-04,)] [eta: 1 day, 22:15:17, time (data): 1.366 (0.000)] l1_latent_x2_opt: 7.2728e-01 fft_frequency_x2_opt: 5.2441e-01 aux_downsample_x2_opt: 5.4100e-02 hf_pixel_x2_opt: 3.5413e-02 l1_latent_x4_opt: 8.2565e-01 fft_frequency_x4_opt: 6.1328e-01 aux_downsample_x4_opt: 6.7567e-02 hf_pixel_x4_opt: 3.5887e-02 
+2025-11-04 18:28:41,834 INFO: [39..][epoch:  0, step:   1,900, lr:(5.000e-04,)] [eta: 1 day, 22:12:49, time (data): 1.349 (0.000)] l1_latent_x2_opt: 7.3979e-01 fft_frequency_x2_opt: 5.3149e-01 aux_downsample_x2_opt: 5.5061e-02 hf_pixel_x2_opt: 3.6172e-02 l1_latent_x4_opt: 8.3253e-01 fft_frequency_x4_opt: 6.1816e-01 aux_downsample_x4_opt: 6.8824e-02 hf_pixel_x4_opt: 3.7857e-02 
+2025-11-04 18:30:56,836 INFO: [39..][epoch:  0, step:   2,000, lr:(5.000e-04,)] [eta: 1 day, 22:10:25, time (data): 1.350 (0.000)] l1_latent_x2_opt: 7.1404e-01 fft_frequency_x2_opt: 5.1465e-01 aux_downsample_x2_opt: 4.7809e-02 hf_pixel_x2_opt: 3.0376e-02 l1_latent_x4_opt: 8.1420e-01 fft_frequency_x4_opt: 6.0547e-01 aux_downsample_x4_opt: 6.1369e-02 hf_pixel_x4_opt: 3.1697e-02 
+2025-11-04 18:33:12,865 INFO: [39..][epoch:  0, step:   2,100, lr:(5.000e-04,)] [eta: 1 day, 22:09:02, time (data): 1.362 (0.000)] l1_latent_x2_opt: 7.1883e-01 fft_frequency_x2_opt: 5.2393e-01 aux_downsample_x2_opt: 5.1495e-02 hf_pixel_x2_opt: 3.3885e-02 l1_latent_x4_opt: 8.1879e-01 fft_frequency_x4_opt: 6.1670e-01 aux_downsample_x4_opt: 6.6821e-02 hf_pixel_x4_opt: 3.6232e-02 
+2025-11-04 18:35:27,769 INFO: [39..][epoch:  0, step:   2,200, lr:(5.000e-04,)] [eta: 1 day, 22:06:31, time (data): 1.355 (0.000)] l1_latent_x2_opt: 7.2259e-01 fft_frequency_x2_opt: 5.2002e-01 aux_downsample_x2_opt: 5.5170e-02 hf_pixel_x2_opt: 3.5376e-02 l1_latent_x4_opt: 8.2379e-01 fft_frequency_x4_opt: 6.1475e-01 aux_downsample_x4_opt: 6.9685e-02 hf_pixel_x4_opt: 3.6974e-02 
+2025-11-04 18:37:42,780 INFO: [39..][epoch:  0, step:   2,300, lr:(5.000e-04,)] [eta: 1 day, 22:04:07, time (data): 1.350 (0.000)] l1_latent_x2_opt: 7.0540e-01 fft_frequency_x2_opt: 5.1172e-01 aux_downsample_x2_opt: 5.4046e-02 hf_pixel_x2_opt: 3.5699e-02 l1_latent_x4_opt: 8.0685e-01 fft_frequency_x4_opt: 6.0498e-01 aux_downsample_x4_opt: 6.9422e-02 hf_pixel_x4_opt: 3.7606e-02 
+2025-11-04 18:39:57,635 INFO: [39..][epoch:  0, step:   2,400, lr:(5.000e-04,)] [eta: 1 day, 22:01:36, time (data): 1.349 (0.000)] l1_latent_x2_opt: 7.2828e-01 fft_frequency_x2_opt: 5.2612e-01 aux_downsample_x2_opt: 5.1853e-02 hf_pixel_x2_opt: 3.4054e-02 l1_latent_x4_opt: 8.2290e-01 fft_frequency_x4_opt: 6.1963e-01 aux_downsample_x4_opt: 6.5827e-02 hf_pixel_x4_opt: 3.5909e-02 
+2025-11-04 18:42:12,603 INFO: [39..][epoch:  0, step:   2,500, lr:(5.000e-04,)] [eta: 1 day, 21:59:12, time (data): 1.350 (0.000)] l1_latent_x2_opt: 7.0629e-01 fft_frequency_x2_opt: 5.2148e-01 aux_downsample_x2_opt: 5.4957e-02 hf_pixel_x2_opt: 3.5025e-02 l1_latent_x4_opt: 8.1038e-01 fft_frequency_x4_opt: 6.1279e-01 aux_downsample_x4_opt: 6.9279e-02 hf_pixel_x4_opt: 3.6056e-02 
+2025-11-04 18:44:28,572 INFO: [39..][epoch:  0, step:   2,600, lr:(5.000e-04,)] [eta: 1 day, 21:57:35, time (data): 1.355 (0.000)] l1_latent_x2_opt: 7.1186e-01 fft_frequency_x2_opt: 5.0684e-01 aux_downsample_x2_opt: 5.3425e-02 hf_pixel_x2_opt: 3.4876e-02 l1_latent_x4_opt: 8.1505e-01 fft_frequency_x4_opt: 5.8838e-01 aux_downsample_x4_opt: 6.6993e-02 hf_pixel_x4_opt: 3.6462e-02 
+2025-11-04 18:46:43,332 INFO: [39..][epoch:  0, step:   2,700, lr:(5.000e-04,)] [eta: 1 day, 21:55:01, time (data): 1.348 (0.000)] l1_latent_x2_opt: 7.2911e-01 fft_frequency_x2_opt: 5.2881e-01 aux_downsample_x2_opt: 4.8840e-02 hf_pixel_x2_opt: 3.1548e-02 l1_latent_x4_opt: 8.2863e-01 fft_frequency_x4_opt: 6.1621e-01 aux_downsample_x4_opt: 6.1509e-02 hf_pixel_x4_opt: 3.2518e-02 
+2025-11-04 18:48:58,081 INFO: [39..][epoch:  0, step:   2,800, lr:(5.000e-04,)] [eta: 1 day, 21:52:28, time (data): 1.348 (0.000)] l1_latent_x2_opt: 7.2362e-01 fft_frequency_x2_opt: 5.2124e-01 aux_downsample_x2_opt: 5.2436e-02 hf_pixel_x2_opt: 3.5142e-02 l1_latent_x4_opt: 8.1948e-01 fft_frequency_x4_opt: 6.1133e-01 aux_downsample_x4_opt: 6.5940e-02 hf_pixel_x4_opt: 3.6466e-02 
+2025-11-04 18:51:14,591 INFO: [39..][epoch:  0, step:   2,900, lr:(5.000e-04,)] [eta: 1 day, 21:51:10, time (data): 1.368 (0.000)] l1_latent_x2_opt: 7.0759e-01 fft_frequency_x2_opt: 5.0098e-01 aux_downsample_x2_opt: 4.7543e-02 hf_pixel_x2_opt: 3.0715e-02 l1_latent_x4_opt: 8.0679e-01 fft_frequency_x4_opt: 5.9180e-01 aux_downsample_x4_opt: 6.3564e-02 hf_pixel_x4_opt: 3.2049e-02 
+2025-11-04 18:53:30,048 INFO: [39..][epoch:  0, step:   3,000, lr:(5.000e-04,)] [eta: 1 day, 21:49:06, time (data): 1.361 (0.000)] l1_latent_x2_opt: 7.0903e-01 fft_frequency_x2_opt: 5.1807e-01 aux_downsample_x2_opt: 5.4698e-02 hf_pixel_x2_opt: 3.7555e-02 l1_latent_x4_opt: 8.0087e-01 fft_frequency_x4_opt: 6.1133e-01 aux_downsample_x4_opt: 6.8103e-02 hf_pixel_x4_opt: 3.9139e-02 
+2025-11-04 18:55:46,300 INFO: [39..][epoch:  0, step:   3,100, lr:(5.000e-04,)] [eta: 1 day, 21:47:32, time (data): 1.362 (0.000)] l1_latent_x2_opt: 7.2234e-01 fft_frequency_x2_opt: 5.2539e-01 aux_downsample_x2_opt: 5.3446e-02 hf_pixel_x2_opt: 3.5487e-02 l1_latent_x4_opt: 8.2032e-01 fft_frequency_x4_opt: 6.1963e-01 aux_downsample_x4_opt: 6.7726e-02 hf_pixel_x4_opt: 3.7343e-02 
+2025-11-04 18:58:01,157 INFO: [39..][epoch:  0, step:   3,200, lr:(5.000e-04,)] [eta: 1 day, 21:45:02, time (data): 1.355 (0.000)] l1_latent_x2_opt: 7.2524e-01 fft_frequency_x2_opt: 5.2930e-01 aux_downsample_x2_opt: 5.6736e-02 hf_pixel_x2_opt: 3.9300e-02 l1_latent_x4_opt: 8.2499e-01 fft_frequency_x4_opt: 6.2646e-01 aux_downsample_x4_opt: 7.0797e-02 hf_pixel_x4_opt: 4.1326e-02 
+2025-11-04 19:00:15,895 INFO: [39..][epoch:  0, step:   3,300, lr:(5.000e-04,)] [eta: 1 day, 21:42:29, time (data): 1.347 (0.000)] l1_latent_x2_opt: 7.3721e-01 fft_frequency_x2_opt: 5.2734e-01 aux_downsample_x2_opt: 5.2932e-02 hf_pixel_x2_opt: 3.5381e-02 l1_latent_x4_opt: 8.3380e-01 fft_frequency_x4_opt: 6.1816e-01 aux_downsample_x4_opt: 6.6464e-02 hf_pixel_x4_opt: 3.6365e-02 
+2025-11-04 19:02:30,650 INFO: [39..][epoch:  0, step:   3,400, lr:(5.000e-04,)] [eta: 1 day, 21:39:57, time (data): 1.347 (0.000)] l1_latent_x2_opt: 7.2536e-01 fft_frequency_x2_opt: 5.2783e-01 aux_downsample_x2_opt: 5.3571e-02 hf_pixel_x2_opt: 3.5719e-02 l1_latent_x4_opt: 8.2371e-01 fft_frequency_x4_opt: 6.1865e-01 aux_downsample_x4_opt: 6.7078e-02 hf_pixel_x4_opt: 3.6596e-02 
+2025-11-04 19:04:46,275 INFO: [39..][epoch:  0, step:   3,500, lr:(5.000e-04,)] [eta: 1 day, 21:37:57, time (data): 1.357 (0.000)] l1_latent_x2_opt: 7.1946e-01 fft_frequency_x2_opt: 5.2759e-01 aux_downsample_x2_opt: 5.0332e-02 hf_pixel_x2_opt: 3.2460e-02 l1_latent_x4_opt: 8.2649e-01 fft_frequency_x4_opt: 6.2109e-01 aux_downsample_x4_opt: 6.5609e-02 hf_pixel_x4_opt: 3.4554e-02 
+2025-11-04 19:07:01,833 INFO: [39..][epoch:  0, step:   3,600, lr:(5.000e-04,)] [eta: 1 day, 21:35:54, time (data): 1.356 (0.000)] l1_latent_x2_opt: 7.1833e-01 fft_frequency_x2_opt: 5.1758e-01 aux_downsample_x2_opt: 5.4198e-02 hf_pixel_x2_opt: 3.5062e-02 l1_latent_x4_opt: 8.1729e-01 fft_frequency_x4_opt: 6.1035e-01 aux_downsample_x4_opt: 6.8125e-02 hf_pixel_x4_opt: 3.5862e-02 
+2025-11-04 19:09:18,544 INFO: [39..][epoch:  0, step:   3,700, lr:(5.000e-04,)] [eta: 1 day, 21:34:27, time (data): 1.371 (0.000)] l1_latent_x2_opt: 7.2597e-01 fft_frequency_x2_opt: 5.2295e-01 aux_downsample_x2_opt: 5.2231e-02 hf_pixel_x2_opt: 3.3028e-02 l1_latent_x4_opt: 8.2601e-01 fft_frequency_x4_opt: 6.1621e-01 aux_downsample_x4_opt: 6.7380e-02 hf_pixel_x4_opt: 3.5541e-02 
+2025-11-04 19:11:35,332 INFO: [39..][epoch:  0, step:   3,800, lr:(5.000e-04,)] [eta: 1 day, 21:33:01, time (data): 1.369 (0.000)] l1_latent_x2_opt: 7.1383e-01 fft_frequency_x2_opt: 5.2148e-01 aux_downsample_x2_opt: 4.9597e-02 hf_pixel_x2_opt: 3.2262e-02 l1_latent_x4_opt: 8.0896e-01 fft_frequency_x4_opt: 6.1523e-01 aux_downsample_x4_opt: 6.1547e-02 hf_pixel_x4_opt: 3.4062e-02 
+2025-11-04 19:13:50,277 INFO: [39..][epoch:  0, step:   3,900, lr:(5.000e-04,)] [eta: 1 day, 21:30:34, time (data): 1.349 (0.000)] l1_latent_x2_opt: 7.2276e-01 fft_frequency_x2_opt: 5.1953e-01 aux_downsample_x2_opt: 4.7234e-02 hf_pixel_x2_opt: 2.9734e-02 l1_latent_x4_opt: 8.2144e-01 fft_frequency_x4_opt: 6.1328e-01 aux_downsample_x4_opt: 6.2862e-02 hf_pixel_x4_opt: 3.2668e-02 
+2025-11-04 19:16:05,353 INFO: [39..][epoch:  0, step:   4,000, lr:(5.000e-04,)] [eta: 1 day, 21:28:13, time (data): 1.350 (0.000)] l1_latent_x2_opt: 7.3263e-01 fft_frequency_x2_opt: 5.3271e-01 aux_downsample_x2_opt: 5.2168e-02 hf_pixel_x2_opt: 3.4322e-02 l1_latent_x4_opt: 8.2936e-01 fft_frequency_x4_opt: 6.2354e-01 aux_downsample_x4_opt: 6.5301e-02 hf_pixel_x4_opt: 3.5689e-02 
+2025-11-04 19:18:20,526 INFO: [39..][epoch:  0, step:   4,100, lr:(5.000e-04,)] [eta: 1 day, 21:25:54, time (data): 1.352 (0.000)] l1_latent_x2_opt: 7.1286e-01 fft_frequency_x2_opt: 5.1074e-01 aux_downsample_x2_opt: 4.6884e-02 hf_pixel_x2_opt: 3.0640e-02 l1_latent_x4_opt: 8.1557e-01 fft_frequency_x4_opt: 5.9180e-01 aux_downsample_x4_opt: 5.8137e-02 hf_pixel_x4_opt: 3.0957e-02 
+2025-11-04 19:20:35,592 INFO: [39..][epoch:  0, step:   4,200, lr:(5.000e-04,)] [eta: 1 day, 21:23:33, time (data): 1.351 (0.000)] l1_latent_x2_opt: 7.2062e-01 fft_frequency_x2_opt: 5.2441e-01 aux_downsample_x2_opt: 5.2407e-02 hf_pixel_x2_opt: 3.4586e-02 l1_latent_x4_opt: 8.2534e-01 fft_frequency_x4_opt: 6.1279e-01 aux_downsample_x4_opt: 6.5348e-02 hf_pixel_x4_opt: 3.6032e-02 
+2025-11-04 19:22:50,603 INFO: [39..][epoch:  0, step:   4,300, lr:(5.000e-04,)] [eta: 1 day, 21:21:10, time (data): 1.350 (0.000)] l1_latent_x2_opt: 7.3037e-01 fft_frequency_x2_opt: 5.2295e-01 aux_downsample_x2_opt: 4.8673e-02 hf_pixel_x2_opt: 3.2038e-02 l1_latent_x4_opt: 8.3314e-01 fft_frequency_x4_opt: 6.0400e-01 aux_downsample_x4_opt: 6.1514e-02 hf_pixel_x4_opt: 3.2493e-02 
+2025-11-04 19:25:05,868 INFO: [39..][epoch:  0, step:   4,400, lr:(5.000e-04,)] [eta: 1 day, 21:18:54, time (data): 1.352 (0.000)] l1_latent_x2_opt: 7.3493e-01 fft_frequency_x2_opt: 5.3564e-01 aux_downsample_x2_opt: 5.3350e-02 hf_pixel_x2_opt: 3.4770e-02 l1_latent_x4_opt: 8.3158e-01 fft_frequency_x4_opt: 6.2598e-01 aux_downsample_x4_opt: 6.6108e-02 hf_pixel_x4_opt: 3.6103e-02 
+2025-11-04 19:27:22,034 INFO: [39..][epoch:  0, step:   4,500, lr:(5.000e-04,)] [eta: 1 day, 21:17:03, time (data): 1.365 (0.000)] l1_latent_x2_opt: 7.1594e-01 fft_frequency_x2_opt: 5.2026e-01 aux_downsample_x2_opt: 5.7206e-02 hf_pixel_x2_opt: 3.7309e-02 l1_latent_x4_opt: 8.1017e-01 fft_frequency_x4_opt: 6.1279e-01 aux_downsample_x4_opt: 7.1502e-02 hf_pixel_x4_opt: 3.8709e-02 
+2025-11-04 19:29:37,121 INFO: [39..][epoch:  0, step:   4,600, lr:(5.000e-04,)] [eta: 1 day, 21:14:43, time (data): 1.357 (0.000)] l1_latent_x2_opt: 7.1196e-01 fft_frequency_x2_opt: 5.2051e-01 aux_downsample_x2_opt: 4.9442e-02 hf_pixel_x2_opt: 3.2259e-02 l1_latent_x4_opt: 8.1157e-01 fft_frequency_x4_opt: 6.0547e-01 aux_downsample_x4_opt: 6.2155e-02 hf_pixel_x4_opt: 3.3576e-02 
+2025-11-04 19:31:52,148 INFO: [39..][epoch:  0, step:   4,700, lr:(5.000e-04,)] [eta: 1 day, 21:12:21, time (data): 1.350 (0.000)] l1_latent_x2_opt: 7.1185e-01 fft_frequency_x2_opt: 5.1611e-01 aux_downsample_x2_opt: 5.2363e-02 hf_pixel_x2_opt: 3.3759e-02 l1_latent_x4_opt: 8.1760e-01 fft_frequency_x4_opt: 6.1230e-01 aux_downsample_x4_opt: 6.7342e-02 hf_pixel_x4_opt: 3.7106e-02 
+2025-11-04 19:34:07,213 INFO: [39..][epoch:  0, step:   4,800, lr:(5.000e-04,)] [eta: 1 day, 21:10:00, time (data): 1.350 (0.000)] l1_latent_x2_opt: 7.1616e-01 fft_frequency_x2_opt: 5.2319e-01 aux_downsample_x2_opt: 5.3919e-02 hf_pixel_x2_opt: 3.5840e-02 l1_latent_x4_opt: 8.1068e-01 fft_frequency_x4_opt: 6.0986e-01 aux_downsample_x4_opt: 6.6938e-02 hf_pixel_x4_opt: 3.6482e-02 
+2025-11-04 19:36:22,310 INFO: [39..][epoch:  0, step:   4,900, lr:(5.000e-04,)] [eta: 1 day, 21:07:40, time (data): 1.351 (0.000)] l1_latent_x2_opt: 7.1679e-01 fft_frequency_x2_opt: 5.1318e-01 aux_downsample_x2_opt: 5.6177e-02 hf_pixel_x2_opt: 3.4949e-02 l1_latent_x4_opt: 8.1787e-01 fft_frequency_x4_opt: 6.0840e-01 aux_downsample_x4_opt: 7.0094e-02 hf_pixel_x4_opt: 3.6273e-02 
+2025-11-04 19:38:38,266 INFO: [39..][epoch:  0, step:   5,000, lr:(5.000e-04,)] [eta: 1 day, 21:05:41, time (data): 1.356 (0.000)] l1_latent_x2_opt: 7.1474e-01 fft_frequency_x2_opt: 5.1758e-01 aux_downsample_x2_opt: 5.4436e-02 hf_pixel_x2_opt: 3.5155e-02 l1_latent_x4_opt: 8.0913e-01 fft_frequency_x4_opt: 6.1084e-01 aux_downsample_x4_opt: 6.8508e-02 hf_pixel_x4_opt: 3.6914e-02 
+2025-11-04 19:38:38,267 INFO: Saving models and training states.
+2025-11-04 19:40:55,716 INFO: Validation val_x2
+	 # l1_latent: 0.7662	Best: 0.7662 @ 5000 iter
+	 # pixel_psnr_pt: 28.7766	Best: 28.7766 @ 5000 iter
+
+2025-11-04 19:43:28,886 INFO: Validation val_x4
+	 # l1_latent: 0.8514	Best: 0.8514 @ 5000 iter
+	 # l2_latent: 1.1908	Best: 1.1908 @ 5000 iter
+	 # pixel_psnr_pt: 26.3475	Best: 26.3475 @ 5000 iter
+

04_11_2025/39_archived_20251104_213142/basicsr_options.yaml

@@ -0,0 +1,260 @@
+# GENERATE TIME: Tue Nov  4 21:29:58 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: true
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 10
+    pin_memory: true
+    persistent_workers: true
+    prefetch_mode: cuda
+    latent_dtype: bf16
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+    latent_dtype: bf16
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 8
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 180
+  num_heads:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  mlp_ratio: 2
+  resi_connection: 1conv
+  head_num_feat: 128
+  primary_head: x4
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: ./runs/04_11_2025/39/models/net_g_5000.pth
+  strict_load_g: true
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025
+compile:
+  enabled: true
+  mode: auto
+  dynamic: true
+  fullgraph: false
+  backend: inductor
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.0005
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.995
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  l1_latent_x2_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x2_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x2
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x2_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x2
+    kernel_size: 5
+    sigma: 1.0
+  l1_latent_x4_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+    space: latent
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: latent
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  aux_downsample_x4_opt:
+    type: DownsampleConsistencyLoss
+    loss_weight: 0.1
+    reduction: mean
+    space: pixel
+    target: x4
+    down_factor: 2
+    mode: bicubic
+  hf_pixel_x4_opt:
+    type: HighFrequencyL1Loss
+    loss_weight: 0.05
+    reduction: mean
+    space: pixel
+    target: x4
+    kernel_size: 5
+    sigma: 1.0
+val:
+  val_freq: 5000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 3
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: '39'
+name: '39'

04_11_2025/39_archived_20251104_213142/train_39_20251104_212958.log

@@ -0,0 +1,601 @@
+2025-11-04 21:29:58,712 INFO: 
+                ____                _       _____  ____
+               / __ ) ____ _ _____ (_)_____/ ___/ / __ \
+              / __  |/ __ `// ___// // ___/\__ \ / /_/ /
+             / /_/ // /_/ /(__  )/ // /__ ___/ // _, _/
+            /_____/ \__,_//____//_/ \___//____//_/ |_|
+     ______                   __   __                 __      __
+    / ____/____   ____   ____/ /  / /   __  __ _____ / /__   / /
+   / / __ / __ \ / __ \ / __  /  / /   / / / // ___// //_/  / /
+  / /_/ // /_/ // /_/ // /_/ /  / /___/ /_/ // /__ / /<    /_/
+  \____/ \____/ \____/ \____/  /_____/\____/ \___//_/|_|  (_)
+    
+Version Information: 
+	BasicSR: 1.4.2
+	PyTorch: 2.9.0+cu129
+	TorchVision: 0.24.0+cpu
+2025-11-04 21:29:58,712 INFO: 
+  model_type: SwinIRLatentModelMultiHead
+  primary_head: x4
+  scale: 4
+  num_gpu: 8
+  manual_seed: 0
+  find_unused_parameters: True
+  precision:[
+    train: bf16
+    eval: fp32
+  ]
+  vae_sources:[
+    flux_vae:[
+      hf_repo: wolfgangblack/flux_vae
+      vae_kind: kl
+    ]
+  ]
+  datasets:[
+    train:[
+      name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+      type: MultiScaleLatentCacheDataset
+      scales: [128, 256, 512]
+      cache_dirs: ['/data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae', '/data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: train
+      filename_tmpl: {}
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      num_worker_per_gpu: 4
+      batch_size_per_gpu: 10
+      pin_memory: True
+      persistent_workers: True
+      prefetch_mode: cuda
+      latent_dtype: bf16
+    ]
+    val:[
+      name: sdxk_120_1024x1024
+      type: MultiScaleLatentCacheDataset
+      scales: [256, 512, 1024]
+      cache_dirs: ['/data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae']
+      vae_names: ['flux_vae']
+      phase: val
+      io_backend:[
+        type: disk
+      ]
+      scale: 4
+      mean: None
+      std: None
+      batch_size_per_gpu: 16
+      num_worker_per_gpu: 4
+      pin_memory: True
+      latent_dtype: bf16
+    ]
+  ]
+  network_g:[
+    type: SwinIRMultiHead
+    in_chans: 16
+    img_size: 32
+    window_size: 8
+    img_range: 1.0
+    depths: [6, 6, 6, 6, 6, 6]
+    embed_dim: 180
+    num_heads: [6, 6, 6, 6, 6, 6]
+    mlp_ratio: 2
+    resi_connection: 1conv
+    head_num_feat: 128
+    primary_head: x4
+    heads: [OrderedDict({'name': 'x2', 'scale': 2, 'out_chans': 16}), OrderedDict({'name': 'x4', 'scale': 4, 'out_chans': 16, 'primary': True})]
+  ]
+  path:[
+    pretrain_network_g: ./runs/04_11_2025/39/models/net_g_5000.pth
+    strict_load_g: True
+    experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39
+    models: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/models
+    training_states: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/training_states
+    log: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39
+    visualization: /data/kazanplova/latent_vae_upscale_train/runs/04_11_2025/39/visualization
+  ]
+  compile:[
+    enabled: True
+    mode: auto
+    dynamic: True
+    fullgraph: False
+    backend: inductor
+  ]
+  train:[
+    ema_decay: 0.999
+    head_inputs:[
+      x2:[
+        lq: 256
+        gt: 512
+      ]
+      x4:[
+        lq: 128
+        gt: 512
+      ]
+    ]
+    optim_g:[
+      type: Adam
+      lr: 0.0005
+      weight_decay: 0
+      betas: [0.9, 0.995]
+    ]
+    grad_clip:[
+      enabled: True
+      generator:[
+        type: norm
+        max_norm: 0.4
+        norm_type: 2.0
+      ]
+    ]
+    scheduler:[
+      type: MultiStepLR
+      milestones: [62500, 93750, 112500]
+      gamma: 0.5
+    ]
+    total_steps: 125000
+    warmup_iter: -1
+    l1_latent_x2_opt:[
+      type: L1Loss
+      loss_weight: 1.0
+      reduction: mean
+      space: latent
+      target: x2
+    ]
+    fft_frequency_x2_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: latent
+      target: x2
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    aux_downsample_x2_opt:[
+      type: DownsampleConsistencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: pixel
+      target: x2
+      down_factor: 2
+      mode: bicubic
+    ]
+    hf_pixel_x2_opt:[
+      type: HighFrequencyL1Loss
+      loss_weight: 0.05
+      reduction: mean
+      space: pixel
+      target: x2
+      kernel_size: 5
+      sigma: 1.0
+    ]
+    l1_latent_x4_opt:[
+      type: L1Loss
+      loss_weight: 1.0
+      reduction: mean
+      space: latent
+      target: x4
+    ]
+    fft_frequency_x4_opt:[
+      type: FFTFrequencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: latent
+      target: x4
+      norm: ortho
+      use_log_amplitude: False
+      alpha: 0.0
+      normalize_weight: True
+      eps: 1e-8
+    ]
+    aux_downsample_x4_opt:[
+      type: DownsampleConsistencyLoss
+      loss_weight: 0.1
+      reduction: mean
+      space: pixel
+      target: x4
+      down_factor: 2
+      mode: bicubic
+    ]
+    hf_pixel_x4_opt:[
+      type: HighFrequencyL1Loss
+      loss_weight: 0.05
+      reduction: mean
+      space: pixel
+      target: x4
+      kernel_size: 5
+      sigma: 1.0
+    ]
+  ]
+  val:[
+    val_freq: 5000
+    save_img: True
+    head_evals:[
+      x2:[
+        save_img: True
+        label: val_x2
+        val_sizes:[
+          lq: 512
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+      x4:[
+        save_img: True
+        label: val_x4
+        val_sizes:[
+          lq: 256
+          gt: 1024
+        ]
+        metrics:[
+          l1_latent:[
+            type: L1Loss
+            space: latent
+          ]
+          l2_latent:[
+            type: MSELoss
+            space: latent
+          ]
+          pixel_psnr_pt:[
+            type: calculate_psnr_pt
+            space: pixel
+            crop_border: 2
+            test_y_channel: False
+          ]
+        ]
+      ]
+    ]
+  ]
+  logger:[
+    print_freq: 100
+    save_checkpoint_freq: 5000
+    use_tb_logger: True
+    wandb:[
+      project: Swin2SR-Latent-SR
+      entity: kazanplova-it-more
+      resume_id: None
+      max_val_images: 3
+    ]
+  ]
+  dist_params:[
+    backend: nccl
+    port: 29500
+    dist: True
+  ]
+  load_networks_only: False
+  exp_name: 39
+  name: 39
+  dist: True
+  rank: 0
+  world_size: 8
+  auto_resume: False
+  is_train: True
+  root_path: /data/kazanplova/latent_vae_upscale_train
+
+2025-11-04 21:30:00,381 INFO: Use wandb logger with id=u9urcy9z; project=Swin2SR-Latent-SR.
+2025-11-04 21:30:13,034 INFO: Dataset [MultiScaleLatentCacheDataset] - gpt4_nanobana_midjourney_recraft_CropsFluxVAE is built.
+2025-11-04 21:30:13,035 INFO: Training statistics:
+	Number of train images: 4858507
+	Dataset enlarge ratio: 1
+	Batch size per gpu: 10
+	World size (gpu number): 8
+	Steps per epoch: 60732
+	Configured training steps: 125000
+	Approximate epochs to cover: 3.
+2025-11-04 21:30:13,039 INFO: Dataset [MultiScaleLatentCacheDataset] - sdxk_120_1024x1024 is built.
+2025-11-04 21:30:13,039 INFO: Number of val images/folders in sdxk_120_1024x1024: 153
+2025-11-04 21:30:13,169 INFO: Network [SwinIRMultiHead] is created.
+2025-11-04 21:30:15,363 INFO: Network: DistributedDataParallel - SwinIRMultiHead, with parameters: 13,743,240
+2025-11-04 21:30:15,364 INFO: SwinIRMultiHead(
+  (conv_first): Conv2d(16, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (patch_embed): PatchEmbed(
+    (norm): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+  )
+  (patch_unembed): PatchUnEmbed()
+  (pos_drop): Dropout(p=0.0, inplace=False)
+  (layers): ModuleList(
+    (0): RSTB(
+      (residual_group): BasicLayer(
+        dim=180, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): Identity()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+    (1-5): 5 x RSTB(
+      (residual_group): BasicLayer(
+        dim=180, input_resolution=(32, 32), depth=6
+        (blocks): ModuleList(
+          (0): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (1): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (2): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (3): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (4): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=0, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+          (5): SwinTransformerBlock(
+            dim=180, input_resolution=(32, 32), num_heads=6, window_size=8, shift_size=4, mlp_ratio=2.0
+            (norm1): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (attn): WindowAttention(
+              dim=180, window_size=(8, 8), num_heads=6
+              (qkv): Linear(in_features=180, out_features=540, bias=True)
+              (attn_drop): Dropout(p=0.0, inplace=False)
+              (proj): Linear(in_features=180, out_features=180, bias=True)
+              (proj_drop): Dropout(p=0.0, inplace=False)
+              (softmax): Softmax(dim=-1)
+            )
+            (drop_path): DropPath()
+            (norm2): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+            (mlp): Mlp(
+              (fc1): Linear(in_features=180, out_features=360, bias=True)
+              (act): GELU(approximate='none')
+              (fc2): Linear(in_features=360, out_features=180, bias=True)
+              (drop): Dropout(p=0.0, inplace=False)
+            )
+          )
+        )
+      )
+      (conv): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+      (patch_embed): PatchEmbed()
+      (patch_unembed): PatchUnEmbed()
+    )
+  )
+  (norm): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
+  (conv_after_body): Conv2d(180, 180, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+  (heads): ModuleDict(
+    (x2): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(180, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(128, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+    (x4): _SwinIRPixelShuffleHead(
+      (conv_before): Sequential(
+        (0): Conv2d(180, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): LeakyReLU(negative_slope=0.01, inplace=True)
+      )
+      (upsample): Upsample(
+        (0): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (1): PixelShuffle(upscale_factor=2)
+        (2): Conv2d(128, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+        (3): PixelShuffle(upscale_factor=2)
+      )
+      (conv_last): Conv2d(128, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+    )
+  )
+)

05_11_2025/40/basicsr_options.yaml

@@ -0,0 +1,243 @@
+# GENERATE TIME: Wed Nov  5 16:51:50 2025
+# CMD:
+# train_vae.py -opt /data/kazanplova/latent_vae_upscale_train/runs/05_11_2025/basicsr_options.yaml --launcher pytorch --local_rank 0
+
+model_type: SwinIRLatentModelMultiHead
+primary_head: x4
+scale: 4
+num_gpu: auto
+manual_seed: 0
+find_unused_parameters: true
+precision:
+  train: bf16
+  eval: fp32
+vae_sources:
+  flux_vae:
+    hf_repo: wolfgangblack/flux_vae
+    vae_kind: kl
+datasets:
+  train:
+    name: gpt4_nanobana_midjourney_recraft_CropsFluxVAE
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 128
+    - 256
+    - 512
+    cache_dirs:
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/nano_banana_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/midjourney_full_dataset_crops_new/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/gpt4_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/recraft_data_crops/embeddings/flux_vae
+    - /data/kazanplova/datasets/full_latent_upscale_dataset_train/mjv5_crops_fix/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: train
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 12
+    pin_memory: true
+    persistent_workers: true
+    latent_dtype: bf16
+  val:
+    name: sdxk_120_1024x1024
+    type: MultiScaleLatentCacheDataset
+    scales:
+    - 256
+    - 512
+    - 1024
+    cache_dirs:
+    - /data/kazanplova/datasets/latent_upscale_validation_120_samples/embeddings/flux_vae
+    vae_names:
+    - flux_vae
+    phase: val
+    io_backend:
+      type: disk
+    scale: 4
+    mean: null
+    std: null
+    batch_size_per_gpu: 16
+    num_worker_per_gpu: 4
+    pin_memory: true
+    latent_dtype: bf16
+network_g:
+  type: SwinIRMultiHead
+  in_chans: 16
+  img_size: 32
+  window_size: 8
+  img_range: 1.0
+  depths:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  embed_dim: 180
+  num_heads:
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  - 6
+  mlp_ratio: 2
+  resi_connection: 1conv
+  primary_head: x4
+  head_num_feat: 128
+  heads:
+  - name: x2
+    scale: 2
+    out_chans: 16
+  - name: x4
+    scale: 4
+    out_chans: 16
+    primary: true
+path:
+  pretrain_network_g: runs/04_11_2025/39/models/net_g_45000.pth
+  strict_load_g: true
+  experiments_root: /data/kazanplova/latent_vae_upscale_train/runs/05_11_2025
+compile:
+  enabled: true
+  mode: auto
+  dynamic: true
+  fullgraph: false
+  backend: inductor
+train:
+  ema_decay: 0.999
+  head_inputs:
+    x2:
+      lq: 256
+      gt: 512
+    x4:
+      lq: 128
+      gt: 512
+  optim_g:
+    type: Adam
+    lr: 0.0002
+    weight_decay: 0
+    betas:
+    - 0.9
+    - 0.99
+  grad_clip:
+    enabled: true
+    generator:
+      type: norm
+      max_norm: 0.4
+      norm_type: 2.0
+  scheduler:
+    type: MultiStepLR
+    milestones:
+    - 62500
+    - 93750
+    - 112500
+    gamma: 0.5
+  total_steps: 125000
+  warmup_iter: -1
+  eagle_pixel_x2_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x2
+  l1_pixel_x2_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x2
+  fft_frequency_x2_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x2
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+  eagle_pixel_x4_opt:
+    type: Eagle_Loss
+    loss_weight: 5.0e-05
+    reduction: mean
+    space: pixel
+    patch_size: 3
+    cutoff: 0.5
+    target: x4
+  l1_pixel_x4_opt:
+    type: L1Loss
+    loss_weight: 10.0
+    reduction: mean
+    space: pixel
+    target: x4
+  fft_frequency_x4_opt:
+    type: FFTFrequencyLoss
+    loss_weight: 1.0
+    reduction: mean
+    space: pixel
+    target: x4
+    norm: ortho
+    use_log_amplitude: false
+    alpha: 0.0
+    normalize_weight: true
+    eps: 1e-8
+val:
+  val_freq: 5000
+  save_img: true
+  head_evals:
+    x2:
+      save_img: true
+      label: val_x2
+      val_sizes:
+        lq: 512
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+    x4:
+      save_img: true
+      label: val_x4
+      val_sizes:
+        lq: 256
+        gt: 1024
+      metrics:
+        l1_latent:
+          type: L1Loss
+          space: latent
+        l2_latent:
+          type: MSELoss
+          space: latent
+        pixel_psnr_pt:
+          type: calculate_psnr_pt
+          space: pixel
+          crop_border: 2
+          test_y_channel: false
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  use_tb_logger: true
+  wandb:
+    project: Swin2SR-Latent-SR
+    entity: kazanplova-it-more
+    resume_id: null
+    max_val_images: 10
+dist_params:
+  backend: nccl
+  port: 29500
+  dist: true
+load_networks_only: false
+exp_name: '40'
+name: '40'