rm conf

leffa/conf/TARGETS

@@ -1,17 +0,0 @@
-load("//gen_ai/genie/components:macros.bzl", "genie_hydra_config_bundle")
-
-oncall("genads_infra")
-
-# All configs including dataloading, torchtnt, and profiling.
-# Note you need all dependencies for hydra instantiation here.
-genie_hydra_config_bundle(
-    name = "idm_vton_hydra_configs",
-    srcs = glob(["**/*.yaml"]),
-    deps = [
-        "//caffe2:torch",
-        "//genads/common/data:transforms",
-        "//genads/idm_vton:idm_vton_lib",
-        "//media_dataloader/api:api",
-        "//torchmultimodal/fb/genai/transforms:transforms",
-    ],
-)

leffa/conf/constants/base.yaml

@@ -1,31 +0,0 @@
-# for virtual try-on
-# height: 512
-# width: 384
-# batch_size: 8
-height: 1024
-width: 768
-batch_size: 2
-
-# for pose transfer
-# height: 256
-# width: 176
-# batch_size: 8
-# height: 512
-# width: 352
-# batch_size: 4
-# height: 1024
-# width: 704
-# batch_size: 1
-
-precision: bf16
-
-max_steps: null
-max_epochs: 200
-max_train_steps_per_epoch: null
-
-evaluate_every_n_train_steps: null
-evaluate_every_n_train_epochs: null
-max_eval_steps_per_eval_epoch: null
-
-use_torchsnapshot: false
-checkpoint_every_n_steps: 500

leffa/conf/datasets/deepfashion_test.yaml

@@ -1,47 +0,0 @@
-deepfashion_test:
-  dataset:
-    _target_: media_dataloader.api.EnrichingDataset
-    datasource:
-      _target_: media_dataloader.api.LazyHiveDataSource
-      namespace: mgenai
-      table: deepfashion_pose_transfer
-      partition_filter_predicate_list: ["ds = '2024-08-15' AND set_name = 'val'"]
-    enrichments:
-      - _target_: media_dataloader.api.media_lookups.ManifoldLookups
-        lookup_handle_to_media_columns:
-          to_img_manifold_path: "image"
-          from_img_manifold_path: "cloth"
-          to_img_iuv_manifold_path: "image_densepose"
-          from_img_iuv_manifold_path: "cloth_densepose"
-    collate_fn:
-      - _target_: media_dataloader.api.Collate
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image
-        blob_field: image
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: cloth
-        blob_field: cloth
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image_densepose
-        blob_field: image_densepose
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: cloth_densepose
-        blob_field: cloth_densepose
-      - _target_: leffa.datasets.transform.VtonTransform
-        height: ${constants.height}
-        width: ${constants.width}
-        is_train: false
-        dataset: deepfashion
-        aug_garment_ratio: 0.0
-        get_garment_from_person_ratio: 0.0
-        aug_mask_ratio: 0.0
-
-  dataloader:
-    _target_: media_dataloader.api.StatefulDataLoader
-    dataset: ${datasets.deepfashion_test.dataset}
-    batch_size: ${constants.batch_size}
-    num_workers: 0
-    prefetch_factor: null
-    pin_memory: true
-    persistent_workers: false
-    multiprocessing_context: null

leffa/conf/datasets/deepfashion_train.yaml

@@ -1,47 +0,0 @@
-deepfashion_train:
-  dataset:
-    _target_: media_dataloader.api.EnrichingDataset
-    datasource:
-      _target_: media_dataloader.api.LazyHiveDataSource
-      namespace: mgenai
-      table: deepfashion_pose_transfer
-      partition_filter_predicate_list: ["ds = '2024-08-15' AND set_name = 'train'"]
-    enrichments:
-      - _target_: media_dataloader.api.media_lookups.ManifoldLookups
-        lookup_handle_to_media_columns:
-          to_img_manifold_path: "image"
-          from_img_manifold_path: "cloth"
-          to_img_iuv_manifold_path: "image_densepose"
-          from_img_iuv_manifold_path: "cloth_densepose"
-    collate_fn:
-      - _target_: media_dataloader.api.Collate
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image
-        blob_field: image
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: cloth
-        blob_field: cloth
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image_densepose
-        blob_field: image_densepose
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: cloth_densepose
-        blob_field: cloth_densepose
-      - _target_: leffa.datasets.transform.VtonTransform
-        height: ${constants.height}
-        width: ${constants.width}
-        is_train: true
-        dataset: deepfashion
-        aug_garment_ratio: 0.0
-        get_garment_from_person_ratio: 0.0
-        aug_mask_ratio: 0.0
-
-  dataloader:
-    _target_: media_dataloader.api.StatefulDataLoader
-    dataset: ${datasets.deepfashion_train.dataset}
-    batch_size: ${constants.batch_size}
-    num_workers: 4
-    prefetch_factor: 2
-    pin_memory: true
-    persistent_workers: true
-    multiprocessing_context: forkserver

leffa/conf/datasets/dress_code_test.yaml

@@ -1,56 +0,0 @@
-dress_code_test:
-  dataset:
-    _target_: media_dataloader.api.EnrichingDataset
-    datasource:
-      _target_: media_dataloader.api.LazyHiveDataSource
-      namespace: ad_metrics
-      table: vton_public_dataset_dress_code_test_paired_v2
-      # table: vton_public_dataset_dress_code_test_unpaired_v2
-      # table: vton_public_dataset_dress_code_test_upper_body_paired_v2
-      # table: vton_public_dataset_dress_code_test_upper_body_unpaired_v2
-      partition_filter_predicate_list: ["ds = '2024-09-14'"]
-      # table: vton_public_dataset_dress_code_test_lower_body_paired_v2
-      # table: vton_public_dataset_dress_code_test_lower_body_unpaired_v2
-      # table: vton_public_dataset_dress_code_test_dresses_paired_v2
-      # table: vton_public_dataset_dress_code_test_dresses_unpaired_v2
-      # partition_filter_predicate_list: ["ds = '2024-09-16'"]
-    enrichments:
-      - _target_: media_dataloader.api.media_lookups.ManifoldLookups
-        lookup_handle_to_media_columns:
-          image_manifold_path: "image"
-          cloth_manifold_path: "cloth"
-          agnostic_mask_manifold_path: "agnostic_mask"
-          dense_manifold_path: "image_densepose"
-          label_map_manifold_path: "image_parse"
-    collate_fn:
-      - _target_: media_dataloader.api.Collate
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image
-        blob_field: image
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: cloth
-        blob_field: cloth
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: agnostic_mask
-        blob_field: agnostic_mask
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image_densepose
-        blob_field: image_densepose
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image_parse
-        blob_field: image_parse
-      - _target_: leffa.datasets.transform.VtonTransform
-        height: ${constants.height}
-        width: ${constants.width}
-        is_train: false
-        dataset: dress_code
-
-  dataloader:
-    _target_: media_dataloader.api.StatefulDataLoader
-    dataset: ${datasets.dress_code_test.dataset}
-    batch_size: ${constants.batch_size}
-    num_workers: 0
-    prefetch_factor: null
-    pin_memory: true
-    persistent_workers: false
-    multiprocessing_context: null

leffa/conf/datasets/dress_code_train.yaml

@@ -1,55 +0,0 @@
-dress_code_train:
-  dataset:
-    _target_: media_dataloader.api.EnrichingDataset
-    datasource:
-      _target_: media_dataloader.api.LazyHiveDataSource
-      namespace: ad_metrics
-      table: vton_public_dataset_dress_code_train_v2
-      # table: vton_public_dataset_dress_code_train_upper_body_v2
-      partition_filter_predicate_list: ["ds = '2024-09-14'"]
-      # table: vton_public_dataset_dress_code_train_lower_body_v2
-      # table: vton_public_dataset_dress_code_train_dresses_v2
-      # partition_filter_predicate_list: ["ds = '2024-09-15'"]
-    enrichments:
-      - _target_: media_dataloader.api.media_lookups.ManifoldLookups
-        lookup_handle_to_media_columns:
-          image_manifold_path: "image"
-          cloth_manifold_path: "cloth"
-          agnostic_mask_manifold_path: "agnostic_mask"
-          dense_manifold_path: "image_densepose"
-          label_map_manifold_path: "image_parse"
-    collate_fn:
-      - _target_: media_dataloader.api.Collate
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image
-        blob_field: image
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: cloth
-        blob_field: cloth
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: agnostic_mask
-        blob_field: agnostic_mask
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image_densepose
-        blob_field: image_densepose
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image_parse
-        blob_field: image_parse
-      - _target_: leffa.datasets.transform.VtonTransform
-        height: ${constants.height}
-        width: ${constants.width}
-        is_train: true
-        dataset: dress_code
-        aug_garment_ratio: 0.0
-        get_garment_from_person_ratio: 0.0
-        aug_mask_ratio: 0.0
-
-  dataloader:
-    _target_: media_dataloader.api.StatefulDataLoader
-    dataset: ${datasets.dress_code_train.dataset}
-    batch_size: ${constants.batch_size}
-    num_workers: 4
-    prefetch_factor: 2
-    pin_memory: true
-    persistent_workers: true
-    multiprocessing_context: forkserver

leffa/conf/datasets/viton_hd_test.yaml

@@ -1,54 +0,0 @@
-viton_hd_test:
-  dataset:
-    _target_: media_dataloader.api.EnrichingDataset
-    datasource:
-      _target_: media_dataloader.api.LazyHiveDataSource
-      namespace: ad_metrics
-      table: vton_public_dataset_viton_hd_test_paired_v2
-      partition_filter_predicate_list: ["ds = '2024-10-30'"]
-      # table: vton_public_dataset_viton_hd_test_unpaired_v1
-      # partition_filter_predicate_list: ["ds = '2024-09-12'"]
-    enrichments:
-      - _target_: media_dataloader.api.media_lookups.ManifoldLookups
-        lookup_handle_to_media_columns:
-          image_manifold_path: "image"
-          cloth_manifold_path: "cloth"
-          agnostic_mask_manifold_path: "agnostic_mask"
-          image_densepose_manifold_path: "image_densepose"
-          cloth_mask_manifold_path: "cloth_mask"
-          image_parse_v3_manifold_path: "image_parse"
-    collate_fn:
-      - _target_: media_dataloader.api.Collate
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image
-        blob_field: image
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: cloth
-        blob_field: cloth
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: agnostic_mask
-        blob_field: agnostic_mask
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image_densepose
-        blob_field: image_densepose
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: cloth_mask
-        blob_field: cloth_mask
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image_parse
-        blob_field: image_parse
-      - _target_: leffa.datasets.transform.VtonTransform
-        height: ${constants.height}
-        width: ${constants.width}
-        is_train: false
-        dataset: viton_hd
-
-  dataloader:
-    _target_: media_dataloader.api.StatefulDataLoader
-    dataset: ${datasets.viton_hd_test.dataset}
-    batch_size: ${constants.batch_size}
-    num_workers: 0
-    prefetch_factor: null
-    pin_memory: true
-    persistent_workers: false
-    multiprocessing_context: null

leffa/conf/datasets/viton_hd_test_local.yaml

@@ -1,20 +0,0 @@
-viton_hd_test_local:
-  dataset:
-    _target_: leffa.datasets.viton_hd.VitonHDLocalDataset
-    # _target_: leffa.datasets.random_dataset.RandomDataset
-    dataroot_path: /home/zijianzhou/data/viton_hd
-    phase: test
-    order: unpaired
-    size:
-      - ${constants.height}
-      - ${constants.width}
-
-  dataloader:
-    _target_: torch.utils.data.DataLoader
-    dataset: ${datasets.viton_hd_test_local.dataset}
-    batch_size: ${constants.batch_size}
-    num_workers: 0
-    prefetch_factor: null
-    pin_memory: true
-    persistent_workers: false
-    multiprocessing_context: null

leffa/conf/datasets/viton_hd_train.yaml

@@ -1,55 +0,0 @@
-viton_hd_train:
-  dataset:
-    _target_: media_dataloader.api.EnrichingDataset
-    datasource:
-      _target_: media_dataloader.api.LazyHiveDataSource
-      namespace: ad_metrics
-      table: vton_public_dataset_viton_hd_train_v2
-      partition_filter_predicate_list: ["ds = '2024-10-30'"]
-    enrichments:
-      - _target_: media_dataloader.api.media_lookups.ManifoldLookups
-        lookup_handle_to_media_columns:
-          image_manifold_path: "image"
-          cloth_manifold_path: "cloth"
-          agnostic_mask_manifold_path: "agnostic_mask"
-          image_densepose_manifold_path: "image_densepose"
-          cloth_mask_manifold_path: "cloth_mask"
-          image_parse_v3_manifold_path: "image_parse"
-    collate_fn:
-      - _target_: media_dataloader.api.Collate
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image
-        blob_field: image
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: cloth
-        blob_field: cloth
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: agnostic_mask
-        blob_field: agnostic_mask
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image_densepose
-        blob_field: image_densepose
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: cloth_mask
-        blob_field: cloth_mask
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image_parse
-        blob_field: image_parse
-      - _target_: leffa.datasets.transform.VtonTransform
-        height: ${constants.height}
-        width: ${constants.width}
-        is_train: true
-        dataset: viton_hd
-        aug_garment_ratio: 0.0
-        get_garment_from_person_ratio: 0.0
-        aug_mask_ratio: 0.0
-
-  dataloader:
-    _target_: media_dataloader.api.StatefulDataLoader
-    dataset: ${datasets.viton_hd_train.dataset}
-    batch_size: ${constants.batch_size}
-    num_workers: 4
-    prefetch_factor: 2
-    pin_memory: true
-    persistent_workers: true
-    multiprocessing_context: forkserver

leffa/conf/eval/base.yaml

@@ -1,4 +0,0 @@
-warmup_iters: 0
-evaluate_every_n_train_steps: ${constants.evaluate_every_n_train_steps}
-evaluate_every_n_train_epochs: ${constants.evaluate_every_n_train_epochs}
-max_eval_steps_per_eval_epoch: ${constants.max_eval_steps_per_eval_epoch}

leffa/conf/model/cat_vton_flux.yaml

@@ -1,19 +0,0 @@
-_target_: leffa.models.cat_vton_model.CatVtonModel
-# FLUX.1-dev
-pretrained_model_name_or_path: manifold://genads_models/tree/zijianzhou/model/FLUX.1-dev
-new_in_channels: 33  # 16+1+16
-height: ${constants.height}
-width: ${constants.width}
-garment_dropout_ratio: 0.1
-use_dream: false
-dream_detail_preservation: 10.0
-use_garment_mask: false
-only_optimize_unet_attn1: true
-use_learning_flow_in_attention: false
-learning_flow_in_attention_lambda: 0.001
-learning_flow_in_attention_stop_timestep: 500
-use_attention_flow_loss: false
-attention_flow_loss_lambda: 0.001
-use_pixel_space_supervision: false
-pixel_space_supervision_lambda: 10.0
-use_densepose: false

leffa/conf/model/cat_vton_sd15.yaml

@@ -1,20 +0,0 @@
-_target_: leffa.models.cat_vton_model.CatVtonModel
-# SD1.5
-pretrained_model_name_or_path: manifold://genads_models/tree/zijianzhou/model/stable-diffusion-inpainting
-pretrained_vae_name_or_path: manifold://genads_models/tree/zijianzhou/model/sd-vae-ft-mse
-new_in_channels: 9
-height: ${constants.height}
-width: ${constants.width}
-garment_dropout_ratio: 0.1
-use_dream: true
-dream_detail_preservation: 10.0
-use_garment_mask: false
-only_optimize_unet_attn1: true
-use_learning_flow_in_attention: false
-learning_flow_in_attention_lambda: 0.001
-learning_flow_in_attention_stop_timestep: 500
-use_attention_flow_loss: false
-attention_flow_loss_lambda: 0.001
-use_pixel_space_supervision: true
-pixel_space_supervision_lambda: 10.0
-use_densepose: false

leffa/conf/model/cat_vton_sdxl.yaml

@@ -1,11 +0,0 @@
-_target_: leffa.models.cat_vton_model.CatVtonModel
-# SDXL
-pretrained_model_name_or_path: manifold://genads_models/tree/zijianzhou/model/stable-diffusion-xl-1.0-inpainting-0.1
-new_in_channels: 9
-height: ${constants.height}
-width: ${constants.width}
-garment_dropout_ratio: 0.1
-use_dream: true
-dream_detail_preservation: 10.0
-use_garment_mask: false
-only_optimize_unet_attn1: false

leffa/conf/model/idm_vton_sd15.yaml

@@ -1,16 +0,0 @@
-_target_: leffa.models.idm_vton_model.IdmVtonModel
-pretrained_model_name_or_path: manifold://genads_models/tree/zijianzhou/model/stable-diffusion-inpainting
-pretrained_vae_name_or_path: manifold://genads_models/tree/zijianzhou/model/sd-vae-ft-mse
-pretrained_garmentnet_path: manifold://genads_models/tree/zijianzhou/model/stable-diffusion-inpainting
-new_in_channels: 13
-height: ${constants.height}
-width: ${constants.width}
-garment_dropout_ratio: 0.1
-use_dream: false
-dream_detail_preservation: 10.0
-skip_cross_attention: true
-skip_cross_attention_garmentnet: true
-copy_unet_to_unet_encoder: false
-only_optimize_unet_attn1: false
-optimize_unet: true
-optimize_unet_encoder: true

leffa/conf/model/idm_vton_sdxl.yaml

@@ -1,17 +0,0 @@
-_target_: leffa.models.idm_vton_model.IdmVtonModel
-pretrained_model_name_or_path: manifold://genads_models/tree/zijianzhou/model/stable-diffusion-xl-1.0-inpainting-0.1
-pretrained_garmentnet_path: manifold://genads_models/tree/zijianzhou/model/stable-diffusion-xl-base-1.0
-pretrained_image_encoder_path: manifold://genads_models/tree/zijianzhou/model/IP-Adapter/models/image_encoder
-pretrained_ip_adapter_path: manifold://genads_models/tree/zijianzhou/model/IP-Adapter/sdxl_models/ip-adapter-plus_sdxl_vit-h.bin
-new_in_channels: 13
-height: ${constants.height}
-width: ${constants.width}
-garment_dropout_ratio: 0.1
-use_dream: false
-dream_detail_preservation: 10.0
-skip_cross_attention: false
-skip_cross_attention_garmentnet: false
-copy_unet_to_unet_encoder: false
-only_optimize_unet_attn1: false
-optimize_unet: true
-optimize_unet_encoder: false

leffa/conf/model/simple_vton_sd15.yaml

@@ -1,23 +0,0 @@
-_target_: leffa.models.simple_vton_model.SimpleVtonModel
-pretrained_model_name_or_path: manifold://genads_models/tree/zijianzhou/model/stable-diffusion-inpainting
-pretrained_vae_name_or_path: manifold://genads_models/tree/zijianzhou/model/sd-vae-ft-mse
-pretrained_garmentnet_path: manifold://genads_models/tree/zijianzhou/model/stable-diffusion-inpainting
-new_in_channels: 12
-height: ${constants.height}
-width: ${constants.width}
-garment_dropout_ratio: 0.1
-use_dream: false
-dream_detail_preservation: 10.0
-skip_cross_attention: true
-skip_cross_attention_garmentnet: true
-copy_unet_to_unet_encoder: false
-only_optimize_unet_attn1: false
-optimize_unet: true
-optimize_unet_encoder: true
-use_learning_flow_in_attention: false
-learning_flow_in_attention_lambda: 0.001
-learning_flow_in_attention_stop_timestep: 500
-use_attention_flow_loss: false
-attention_flow_loss_lambda: 0.001
-use_pixel_space_supervision: false
-pixel_space_supervision_lambda: 10.0

leffa/conf/model/simple_vton_sdxl.yaml

@@ -1,22 +0,0 @@
-_target_: leffa.models.simple_vton_model.SimpleVtonModel
-pretrained_model_name_or_path: manifold://genads_models/tree/zijianzhou/model/stable-diffusion-xl-1.0-inpainting-0.1
-pretrained_garmentnet_path: manifold://genads_models/tree/zijianzhou/model/stable-diffusion-xl-base-1.0
-new_in_channels: 12
-height: ${constants.height}
-width: ${constants.width}
-garment_dropout_ratio: 0.1
-use_dream: false
-dream_detail_preservation: 10.0
-skip_cross_attention: true
-skip_cross_attention_garmentnet: true
-copy_unet_to_unet_encoder: false
-only_optimize_unet_attn1: false
-optimize_unet: true
-optimize_unet_encoder: true
-use_learning_flow_in_attention: false
-learning_flow_in_attention_lambda: 0.001
-learning_flow_in_attention_stop_timestep: 500
-use_attention_flow_loss: false
-attention_flow_loss_lambda: 0.001
-use_pixel_space_supervision: false
-pixel_space_supervision_lambda: 10.0

leffa/conf/predict.yaml

@@ -1,66 +0,0 @@
-# File used for running pure prediction with torchtnt, including a custom score computation.
-
-seed: 42
-
-defaults:
-  - constants: base
-  # - model: idm_vton_sd15
-  # - model: idm_vton_sdxl
-  # - model: cat_vton_sd15
-  # - model: cat_vton_sdxl
-  - model: simple_vton_sd15
-  # - model: simple_vton_sdxl
-  # - model: simple_vton_sd15_fluxvae
-  # - model: simple_vton_sdxl_fluxvae
-  - datasets:
-    - viton_hd_test
-    - dress_code_test
-    - deepfashion_test
-  - _self_
-
-constants:
-  # for virtual try-on
-  # height: 512
-  # width: 384
-  height: 1024
-  width: 768
-  # for pose transfer
-  # height: 256
-  # width: 176
-  # height: 512
-  # width: 352
-  # height: 1024
-  # width: 704
-  batch_size: 1
-
-# null to go through whole dataloader.
-max_steps_per_epoch: null
-dataloader: ${datasets.viton_hd_test.dataloader}
-# dataloader: ${datasets.dress_code_test.dataloader}
-# dataloader: ${datasets.deepfashion_test.dataloader}
-manifold_log_dir: manifold://genads_models/tree/zijianzhou/output/simple_vton/tmp
-
-inference:
-  # _target_: leffa.inference.IdmVtonInference
-  # _target_: leffa.inference.CatVtonInference
-  _target_: leffa.inference.SimpleVtonInference
-  _partial_: True
-  model: ${model}
-  model_entity_id: null
-  checkpoint_version: null
-
-callbacks:
-  - _target_: leffa.callbacks.save_image_callback.SaveImageCallback
-    manifold_path: ${manifold_log_dir}/generated_paired
-    # manifold_path: ${manifold_log_dir}/generated_unpaired
-    task_type: vton
-    # manifold_path: ${manifold_log_dir}
-    # task_type: pose_transfer
-
-unit:
-  _target_: leffa.vton_pred_unit.VtonPredUnit
-  inference_fn: ${inference}
-
-hydra:
-  run:
-    dir: /tmp/

leffa/conf/test.yaml

@@ -1,89 +0,0 @@
-### Main entry for the training config in hydra.
-### Only top level configurations can run, we decompose the full
-### config to multiple subfolders for better reusability.
-
-seed: 42
-
-defaults:
-  - constants: base
-  - model: simple_vton_sd15
-  - trainer: base
-  - eval: base
-  - datasets:
-    - viton_hd_train
-    - viton_hd_test
-    - viton_hd_test_local
-    - dress_code_train
-    - dress_code_test
-    - deepfashion_train
-    - deepfashion_test
-  - _self_
-
-constants:
-  height: 1024
-  width: 768
-  batch_size: 1
-
-train_dataset: ${datasets.viton_hd_train}
-# train_dataset: ${datasets.dress_code_train}
-# train_dataset: ${datasets.deepfashion_train}
-eval_dataset: ${datasets.viton_hd_test}
-
-unit:
-  _target_: leffa.vton_unit.VtonUnit
-  _partial_: True
-  model: ${model}
-  # strategy: ddp
-  strategy:
-    _target_: leffa.utils.create_fsdp_strategy
-    sharding_strategy: SHARD_GRAD_OP
-    state_dict_type: SHARDED_STATE_DICT
-    mixed_precision:
-      param_dtype: ${constants.precision}
-      reduce_dtype: ${constants.precision}
-      cast_forward_inputs: True
-    class_paths:
-      # For VAE (first stage)
-      - diffusers.models.unets.unet_2d_blocks.DownEncoderBlock2D
-      - diffusers.models.unets.unet_2d_blocks.UNetMidBlock2D
-      - diffusers.models.unets.unet_2d_blocks.UpDecoderBlock2D
-      # For UNet (unet stage) IdmVton
-      - leffa.models.diffusion_model.attentionhacked_tryon.BasicTransformerBlock
-      - leffa.models.diffusion_model.attentionhacked_garment.BasicTransformerBlock
-      # For UNet (unet stage) CatVton
-      - diffusers.models.attention.BasicTransformerBlock
-      # For CLIP (condition stage)
-      - transformers.CLIPTextModel
-      - transformers.CLIPTextModelWithProjection
-      - transformers.CLIPVisionModelWithProjection
-  optim_fn:
-    _target_: torch.optim.AdamW
-    _partial_: True
-    lr: 1.0e-5
-    betas: [0.9, 0.999]
-    eps: 1.0e-8
-    weight_decay: 1.0e-2
-    amsgrad: false
-  lr_scheduler_fn:
-    _target_: torch.optim.lr_scheduler.ConstantLR
-    _partial_: True
-    factor: 1.0
-  swa_params:
-    _target_: torchtnt.framework.auto_unit.SWAParams
-    warmup_steps_or_epochs: 0
-    step_or_epoch_update_freq: 1
-    averaging_method: ema
-    ema_decay: 0.9999
-    use_lit: True
-  precision: ${constants.precision}
-  clip_grad_norm: 1.0
-
-umm_metadata:
-  model_type_name: ads_genads_ldm
-  model_series_name: ads_genads_ldm
-  oncall: ai_genads
-
-checkpoint:
-  checkpoint_dir: null
-  checkpoint_path: null
-  checkpoint_every_n_steps: ${constants.checkpoint_every_n_steps}

leffa/conf/train.yaml

@@ -1,84 +0,0 @@
-### Main entry for the training config in hydra.
-### Only top level configurations can run, we decompose the full
-### config to multiple subfolders for better reusability.
-
-seed: 42
-
-defaults:
-  - constants: base
-  - model: simple_vton_sd15
-  - trainer: base
-  - eval: base
-  - datasets:
-    - viton_hd_train
-    - viton_hd_test
-    - viton_hd_test_local
-    - dress_code_train
-    - dress_code_test
-    - deepfashion_train
-    - deepfashion_test
-  - _self_
-
-train_dataset: ${datasets.viton_hd_train}
-# train_dataset: ${datasets.dress_code_train}
-# train_dataset: ${datasets.deepfashion_train}
-eval_dataset: null
-
-unit:
-  _target_: leffa.vton_unit.VtonUnit
-  _partial_: True
-  model: ${model}
-  # strategy: ddp
-  strategy:
-    _target_: leffa.utils.create_fsdp_strategy
-    sharding_strategy: SHARD_GRAD_OP
-    state_dict_type: SHARDED_STATE_DICT
-    mixed_precision:
-      param_dtype: ${constants.precision}
-      reduce_dtype: ${constants.precision}
-      cast_forward_inputs: True
-    class_paths:
-      # For VAE (first stage)
-      - diffusers.models.unets.unet_2d_blocks.DownEncoderBlock2D
-      - diffusers.models.unets.unet_2d_blocks.UNetMidBlock2D
-      - diffusers.models.unets.unet_2d_blocks.UpDecoderBlock2D
-      # For UNet (unet stage) IdmVton
-      - leffa.models.diffusion_model.attentionhacked_tryon.BasicTransformerBlock
-      - leffa.models.diffusion_model.attentionhacked_garment.BasicTransformerBlock
-      # For UNet (unet stage) CatVton
-      - diffusers.models.attention.BasicTransformerBlock
-      # For CLIP (condition stage)
-      - transformers.CLIPTextModel
-      - transformers.CLIPTextModelWithProjection
-      - transformers.CLIPVisionModelWithProjection
-  optim_fn:
-    _target_: torch.optim.AdamW
-    _partial_: True
-    lr: 1.0e-5
-    betas: [0.9, 0.999]
-    eps: 1.0e-8
-    weight_decay: 1.0e-2
-    amsgrad: false
-  lr_scheduler_fn:
-    _target_: torch.optim.lr_scheduler.ConstantLR
-    _partial_: True
-    factor: 1.0
-  swa_params:
-    _target_: torchtnt.framework.auto_unit.SWAParams
-    warmup_steps_or_epochs: 0
-    step_or_epoch_update_freq: 1
-    averaging_method: ema
-    ema_decay: 0.9999
-    use_lit: True
-  precision: ${constants.precision}
-  clip_grad_norm: 1.0
-
-umm_metadata:
-  model_type_name: ads_genads_ldm
-  model_series_name: ads_genads_ldm
-  oncall: ai_genads
-
-checkpoint:
-  checkpoint_dir: null
-  checkpoint_path: null
-  checkpoint_every_n_steps: ${constants.checkpoint_every_n_steps}

leffa/conf/train_local.yaml

@@ -1,9 +0,0 @@
-## Starter configuration to be ran locally for testing purpose.
-
-defaults:
-  - train
-  - _self_
-
-trainer:
-  resume_from_last_ckpt: false
-  model_entity_id: null

leffa/conf/train_mae.yaml

@@ -1,159 +0,0 @@
-constants:
-  img_size: 512
-  batch_size: 16
-  precision: fp32
-  max_epochs: 1000
-  max_steps: null
-  max_train_steps_per_epoch: null
-  evaluate_every_n_train_steps: null
-  evaluate_every_n_train_epochs: 10
-  max_eval_steps_per_eval_epoch: null
-  use_torchsnapshot: false
-  checkpoint_every_n_steps: 500
-model:
-  _target_: leffa.models.idm_vton_model.Mae4BgGen
-  img_size: ${constants.img_size}
-  patch_size: 16
-  embed_dim: 1024
-  depth: 24
-  num_heads: 16
-  # pretrained_path: manifold://genads_models/tree/zijianzhou/model/mae/mae_pretrain_vit_large.pth
-  pretrained_path: null
-  bg_masking_type: min
-trainer:
-  max_epochs: ${constants.max_epochs}
-  max_steps: ${constants.max_steps}
-  max_train_steps_per_epoch: ${constants.max_train_steps_per_epoch}
-  checkpoint_every_n_steps: ${constants.checkpoint_every_n_steps}
-  model_entity_id: null
-  resume_from_last_ckpt: true
-  model_store_checkpoint_version: null
-  garbage_collector_interval: 5001
-  pretrained_weights: null
-  log_dir: manifold://fblearner_flow_run_metrics/tree/torchmultimodal/idm_vton/logs/
-  use_pt2: false
-  memory_snapshot: false
-eval:
-  warmup_iters: 0
-  evaluate_every_n_train_steps: ${constants.evaluate_every_n_train_steps}
-  evaluate_every_n_train_epochs: ${constants.evaluate_every_n_train_epochs}
-  max_eval_steps_per_eval_epoch: ${constants.max_eval_steps_per_eval_epoch}
-datasets:
-  mae_train:
-    dataset:
-      _target_: media_dataloader.api.EnrichingDataset
-      datasource:
-        _target_: media_dataloader.api.LazyHiveDataSource
-        namespace: ad_metrics
-        table: hybrid_3_0_1st_shein_data
-        partition_filter_predicate_list:
-        - ds = '2024-07-20'
-      enrichments:
-      - _target_: media_dataloader.api.media_lookups.EverstoreLookups
-        lookup_handle_to_media_columns:
-          everstore_handle: "image"
-      - _target_: media_dataloader.api.media_lookups.ManifoldLookups
-        lookup_handle_to_media_columns:
-          binary_mask_manifold_path: bg_mask
-      collate_fn:
-      - _target_: media_dataloader.api.Collate
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image
-        blob_field: image
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: bg_mask
-        blob_field: bg_mask
-      - _target_: leffa.datasets.transform.MaeTransform
-        input_size: ${constants.img_size}
-        is_train: true
-    dataloader:
-      _target_: media_dataloader.api.StatefulDataLoader
-      dataset: ${datasets.mae_train.dataset}
-      batch_size: ${constants.batch_size}
-      num_workers: 8
-      prefetch_factor: 2
-      pin_memory: true
-      persistent_workers: true
-      multiprocessing_context: forkserver
-  mae_test:
-    dataset:
-      _target_: media_dataloader.api.EnrichingDataset
-      datasource:
-        _target_: media_dataloader.api.LazyHiveDataSource
-        namespace: ad_metrics
-        table: hybrid_3_0_1st_shein_data
-        partition_filter_predicate_list:
-        - ds = '2024-07-20'
-      enrichments:
-      - _target_: media_dataloader.api.media_lookups.EverstoreLookups
-        lookup_handle_to_media_columns:
-          everstore_handle: "image"
-      - _target_: media_dataloader.api.media_lookups.ManifoldLookups
-        lookup_handle_to_media_columns:
-          binary_mask_manifold_path: bg_mask
-      collate_fn:
-      - _target_: media_dataloader.api.Collate
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: image
-        blob_field: image
-      - _target_: torchmultimodal.fb.genai.transforms.hive_transforms.EverstoreImageToPILTransform
-        image_field: bg_mask
-        blob_field: bg_mask
-      - _target_: leffa.datasets.transform.MaeTransform
-        input_size: ${constants.img_size}
-        is_train: false
-    dataloader:
-      _target_: media_dataloader.api.StatefulDataLoader
-      dataset: ${datasets.mae_train.dataset}
-      batch_size: ${constants.batch_size}
-      num_workers: 0
-      prefetch_factor: null
-      pin_memory: true
-      persistent_workers: false
-      multiprocessing_context: null
-seed: 42
-train_dataset: ${datasets.mae_train}
-eval_dataset: null
-# eval_dataset: ${datasets.mae_test}
-unit:
-  _target_: leffa.vton_unit.VtonUnit
-  _partial_: true
-  model: ${model}
-  strategy: ddp
-  # strategy:
-  #   _target_: leffa.utils.create_fsdp_strategy
-  #   sharding_strategy: FULL_SHARD
-  #   state_dict_type: SHARDED_STATE_DICT
-  #   class_paths:
-  #   - leffa.models.idm_vton_model.MaskedAutoencoderViT
-  optim_fn:
-    _target_: torch.optim.AdamW
-    _partial_: true
-    lr: 1.0e-05
-    betas:
-    - 0.9
-    - 0.999
-    eps: 1.0e-08
-    weight_decay: 0.01
-    amsgrad: false
-  lr_scheduler_fn:
-    _target_: torch.optim.lr_scheduler.ConstantLR
-    _partial_: true
-    factor: 1.0
-  swa_params:
-    _target_: torchtnt.framework.auto_unit.SWAParams
-    warmup_steps_or_epochs: 0
-    step_or_epoch_update_freq: 1
-    averaging_method: ema
-    ema_decay: 0.9999
-    use_lit: true
-  precision: ${constants.precision}
-  clip_grad_norm: 1.0
-umm_metadata:
-  model_type_name: ads_genads_ldm
-  model_series_name: ads_genads_ldm
-  oncall: ai_genads
-checkpoint:
-  checkpoint_dir: null
-  checkpoint_path: null
-  checkpoint_every_n_steps: ${constants.checkpoint_every_n_steps}

leffa/conf/trainer/base.yaml

@@ -1,14 +0,0 @@
-max_epochs: ${constants.max_epochs}
-max_steps: ${constants.max_steps}
-max_train_steps_per_epoch: ${constants.max_train_steps_per_epoch}
-
-checkpoint_every_n_steps: ${constants.checkpoint_every_n_steps}
-
-resume_from_last_ckpt: True
-model_entity_id: null
-model_store_checkpoint_version: null
-garbage_collector_interval: 5001
-pretrained_weights: null
-log_dir: manifold://fblearner_flow_run_metrics/tree/torchmultimodal/idm_vton/logs
-use_pt2: False
-memory_snapshot: False