Adding example for SDv2

examples/kohya_train_db_fixed_with-reg_SDv2 512 base.ps1

@@ -0,0 +1,64 @@
+# This powershell script will create a model using the fine tuning dreambooth method. It will require landscape,
+# portrait and square images.
+#
+# Adjust the script to your own needs
+
+# variable values
+$pretrained_model_name_or_path = "D:\models\512-base-ema.ckpt"
+$data_dir = "D:\models\dariusz_zawadzki\kohya_reg\data"
+$reg_data_dir = "D:\models\dariusz_zawadzki\kohya_reg\reg"
+$logging_dir = "D:\models\dariusz_zawadzki\logs"
+$output_dir = "D:\models\dariusz_zawadzki\train_db_fixed_model_reg_v2"
+$resolution = "512,512"
+$lr_scheduler="polynomial"
+$cache_latents = 1 # 1 = true, 0 = false
+
+$image_num = Get-ChildItem $data_dir -Recurse -File -Include *.png, *.jpg, *.webp | Measure-Object | %{$_.Count}
+
+Write-Output "image_num: $image_num"
+
+$dataset_repeats = 200
+$learning_rate = 2e-6
+$train_batch_size = 4
+$epoch = 1
+$save_every_n_epochs=1
+$mixed_precision="bf16"
+$num_cpu_threads_per_process=6
+
+# You should not have to change values past this point
+if ($cache_latents -eq 1) {
+    $cache_latents_value="--cache_latents"
+}
+else {
+    $cache_latents_value=""
+}
+
+$repeats = $image_num * $dataset_repeats
+$mts = [Math]::Ceiling($repeats / $train_batch_size * $epoch)
+
+Write-Output "Repeats: $repeats"
+
+cd D:\kohya_ss
+.\venv\Scripts\activate
+
+accelerate launch --num_cpu_threads_per_process $num_cpu_threads_per_process train_db_fixed.py `
+    --v2 `
+    --pretrained_model_name_or_path=$pretrained_model_name_or_path `
+    --train_data_dir=$data_dir `
+    --output_dir=$output_dir `
+    --resolution=$resolution `
+    --train_batch_size=$train_batch_size `
+    --learning_rate=$learning_rate `
+    --max_train_steps=$mts `
+    --use_8bit_adam `
+    --xformers `
+    --mixed_precision=$mixed_precision `
+    $cache_latents_value `
+    --save_every_n_epochs=$save_every_n_epochs `
+    --logging_dir=$logging_dir `
+    --save_precision="fp16" `
+    --reg_data_dir=$reg_data_dir `
+    --seed=494481440 `
+    --lr_scheduler=$lr_scheduler
+
+# Add the inference yaml file along with the model for proper loading. Need to have the same name as model... Most likelly "last.yaml" in our case.

v2_inference/v2-inpainting-inference.yaml

@@ -0,0 +1,158 @@
+model:
+  base_learning_rate: 5.0e-05
+  target: ldm.models.diffusion.ddpm.LatentInpaintDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: "jpg"
+    cond_stage_key: "txt"
+    image_size: 64
+    channels: 4
+    cond_stage_trainable: false
+    conditioning_key: hybrid
+    scale_factor: 0.18215
+    monitor: val/loss_simple_ema
+    finetune_keys: null
+    use_ema: False
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        use_checkpoint: True
+        image_size: 32 # unused
+        in_channels: 9
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [ 4, 2, 1 ]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 4, 4 ]
+        num_head_channels: 64 # need to fix for flash-attn
+        use_spatial_transformer: True
+        use_linear_in_transformer: True
+        transformer_depth: 1
+        context_dim: 1024
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          #attn_type: "vanilla-xformers"
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+            - 1
+            - 2
+            - 4
+            - 4
+          num_res_blocks: 2
+          attn_resolutions: [ ]
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder
+      params:
+        freeze: True
+        layer: "penultimate"
+
+
+data:
+  target: ldm.data.laion.WebDataModuleFromConfig
+  params:
+    tar_base: null  # for concat as in LAION-A
+    p_unsafe_threshold: 0.1
+    filter_word_list: "data/filters.yaml"
+    max_pwatermark: 0.45
+    batch_size: 8
+    num_workers: 6
+    multinode: True
+    min_size: 512
+    train:
+      shards:
+        - "pipe:aws s3 cp s3://stability-aws/laion-a-native/part-0/{00000..18699}.tar -"
+        - "pipe:aws s3 cp s3://stability-aws/laion-a-native/part-1/{00000..18699}.tar -"
+        - "pipe:aws s3 cp s3://stability-aws/laion-a-native/part-2/{00000..18699}.tar -"
+        - "pipe:aws s3 cp s3://stability-aws/laion-a-native/part-3/{00000..18699}.tar -"
+        - "pipe:aws s3 cp s3://stability-aws/laion-a-native/part-4/{00000..18699}.tar -"  #{00000-94333}.tar"
+      shuffle: 10000
+      image_key: jpg
+      image_transforms:
+      - target: torchvision.transforms.Resize
+        params:
+          size: 512
+          interpolation: 3
+      - target: torchvision.transforms.RandomCrop
+        params:
+          size: 512
+      postprocess:
+        target: ldm.data.laion.AddMask
+        params:
+          mode: "512train-large"
+          p_drop: 0.25
+    # NOTE use enough shards to avoid empty validation loops in workers
+    validation:
+      shards:
+        - "pipe:aws s3 cp s3://deep-floyd-s3/datasets/laion_cleaned-part5/{93001..94333}.tar - "
+      shuffle: 0
+      image_key: jpg
+      image_transforms:
+      - target: torchvision.transforms.Resize
+        params:
+          size: 512
+          interpolation: 3
+      - target: torchvision.transforms.CenterCrop
+        params:
+          size: 512
+      postprocess:
+        target: ldm.data.laion.AddMask
+        params:
+          mode: "512train-large"
+          p_drop: 0.25
+
+lightning:
+  find_unused_parameters: True
+  modelcheckpoint:
+    params:
+      every_n_train_steps: 5000
+
+  callbacks:
+    metrics_over_trainsteps_checkpoint:
+      params:
+        every_n_train_steps: 10000
+
+    image_logger:
+      target: main.ImageLogger
+      params:
+        enable_autocast: False
+        disabled: False
+        batch_frequency: 1000
+        max_images: 4
+        increase_log_steps: False
+        log_first_step: False
+        log_images_kwargs:
+          use_ema_scope: False
+          inpaint: False
+          plot_progressive_rows: False
+          plot_diffusion_rows: False
+          N: 4
+          unconditional_guidance_scale: 5.0
+          unconditional_guidance_label: [""]
+          ddim_steps: 50  # todo check these out for depth2img,
+          ddim_eta: 0.0   # todo check these out for depth2img,
+
+  trainer:
+    benchmark: True
+    val_check_interval: 5000000
+    num_sanity_val_steps: 0
+    accumulate_grad_batches: 1

v2_inference/v2-midas-inference.yaml

@@ -0,0 +1,74 @@
+model:
+  base_learning_rate: 5.0e-07
+  target: ldm.models.diffusion.ddpm.LatentDepth2ImageDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: "jpg"
+    cond_stage_key: "txt"
+    image_size: 64
+    channels: 4
+    cond_stage_trainable: false
+    conditioning_key: hybrid
+    scale_factor: 0.18215
+    monitor: val/loss_simple_ema
+    finetune_keys: null
+    use_ema: False
+
+    depth_stage_config:
+      target: ldm.modules.midas.api.MiDaSInference
+      params:
+        model_type: "dpt_hybrid"
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        use_checkpoint: True
+        image_size: 32 # unused
+        in_channels: 5
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [ 4, 2, 1 ]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 4, 4 ]
+        num_head_channels: 64 # need to fix for flash-attn
+        use_spatial_transformer: True
+        use_linear_in_transformer: True
+        transformer_depth: 1
+        context_dim: 1024
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          #attn_type: "vanilla-xformers"
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+            - 1
+            - 2
+            - 4
+            - 4
+          num_res_blocks: 2
+          attn_resolutions: [ ]
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder
+      params:
+        freeze: True
+        layer: "penultimate"
+
+

v2_inference/x4-upscaling.yaml

@@ -0,0 +1,76 @@
+model:
+  base_learning_rate: 1.0e-04
+  target: ldm.models.diffusion.ddpm.LatentUpscaleDiffusion
+  params:
+    parameterization: "v"
+    low_scale_key: "lr"
+    linear_start: 0.0001
+    linear_end: 0.02
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: "jpg"
+    cond_stage_key: "txt"
+    image_size: 128
+    channels: 4
+    cond_stage_trainable: false
+    conditioning_key: "hybrid-adm"
+    monitor: val/loss_simple_ema
+    scale_factor: 0.08333
+    use_ema: False
+
+    low_scale_config:
+      target: ldm.modules.diffusionmodules.upscaling.ImageConcatWithNoiseAugmentation
+      params:
+        noise_schedule_config: # image space
+          linear_start: 0.0001
+          linear_end: 0.02
+        max_noise_level: 350
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        use_checkpoint: True
+        num_classes: 1000  # timesteps for noise conditioning (here constant, just need one)
+        image_size: 128
+        in_channels: 7
+        out_channels: 4
+        model_channels: 256
+        attention_resolutions: [ 2,4,8]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 2, 4]
+        disable_self_attentions: [True, True, True, False]
+        disable_middle_self_attn: False
+        num_heads: 8
+        use_spatial_transformer: True
+        transformer_depth: 1
+        context_dim: 1024
+        legacy: False
+        use_linear_in_transformer: True
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        ddconfig:
+          # attn_type: "vanilla-xformers" this model needs efficient attention to be feasible on HR data, also the decoder seems to break in half precision (UNet is fine though)
+          double_z: True
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult: [ 1,2,4 ]  # num_down = len(ch_mult)-1
+          num_res_blocks: 2
+          attn_resolutions: [ ]
+          dropout: 0.0
+
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder
+      params:
+        freeze: True
+        layer: "penultimate"
+