Add weights and configs

hiera/configs/hiera_base_1e-4_hyper_224_448.yaml

@@ -0,0 +1,63 @@
+base_configs:
+  - config/base.yaml
+  - config/cluster_paths_inat.yaml
+name: hiera_base_1e-4_hyper_224_448
+data:
+  dataset: inaturalist
+  crop_size: 448
+  val_crop_size: 448
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: false
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+  - Reptilia
+model:
+  name: EncoderDecoder
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: get_hiera_model
+  backbone:
+    img_size: 224
+    input_dim: 3
+  cls_head: xl
+  context:
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 2
+optimizer:
+  name: adamw
+  base_lr: 1e-4
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 16
+  val_batch_size: 1
+  freeze_epochs: 0
+  freeze_bn: false
+  test_every: 1
+  test_reset: true
+  clip_grad: 5.0
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params:
+        field: label
+      weight: 1.0
+      display: on

hiera/configs/hiera_base_1e-4_nonxl_224.yaml

@@ -0,0 +1,63 @@
+base_configs:
+  - config/base.yaml
+  - config/cluster_paths_inat.yaml
+name: hiera_base_1e-4_nonxl_224
+data:
+  dataset: inaturalist
+  crop_size: 224
+  val_crop_size: 224
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: false
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+  - Reptilia
+model:
+  name: EncoderDecoder
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: get_hiera_model
+  backbone:
+    img_size: 224
+    input_dim: 3
+  cls_head: naive
+  context:
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 2
+optimizer:
+  name: adamw
+  base_lr: 1e-4
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 16
+  val_batch_size: 1
+  freeze_epochs: 0
+  freeze_bn: false
+  test_every: 1
+  test_reset: true
+  clip_grad: 5.0
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params:
+        field: label
+      weight: 1.0
+      display: on

hiera/configs/hiera_baseplus_1e-4_nonxl_224.yaml

@@ -0,0 +1,63 @@
+base_configs:
+  - config/base.yaml
+  - config/cluster_paths_inat.yaml
+name: hiera_baseplus_1e-4_nonxl_224
+data:
+  dataset: inaturalist
+  crop_size: 224
+  val_crop_size: 224
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: false
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+  - Reptilia
+model:
+  name: EncoderDecoder
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: get_hiera_model_base_plus
+  backbone:
+    img_size: 224
+    input_dim: 3
+  cls_head: naive
+  context:
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 2
+optimizer:
+  name: adamw
+  base_lr: 1e-4
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 16
+  val_batch_size: 1
+  freeze_epochs: 0
+  freeze_bn: false
+  test_every: 1
+  test_reset: true
+  clip_grad: 5.0
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params:
+        field: label
+      weight: 1.0
+      display: on

hiera/weights/hiera-b-hyper-224-448-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf7c07c912366183353161ecf2bb1ebe5bb4b003add32940aa3624c21965c453
+size 281939643

hiera/weights/hiera-b-nonxl-224-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff2e83379a7ecb4d068999831f9dbefcaad6082145ffb66ec05a4f4c36d107bc
+size 244166884

hiera/weights/hiera-bplus-nonxl-224-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:012cc98fd445a02d2a8902fec1b8486263f5fcaf5807953aacc8e88b6e683201
+size 322400822

swin-b/configs/swin_v2_base_1e-4_nonxl_256.yaml

@@ -0,0 +1,69 @@
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_base_1e-4_nonxl_256
+data:
+  dataset: inaturalist
+  crop_size: 256 
+  val_crop_size: 256
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_base_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_base_patch4_window16_256.pth
+    upsample: False
+  cls_head: naive
+  xl_context:
+    skip_connection: off
+    enabled: off
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 2
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-4
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 60
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-b/configs/swin_v2_base_1e-4_xl_512_256_context_2chips.yaml

@@ -0,0 +1,75 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/cluster_paths_inat.yaml
+name: swin_v2_base_1e-4_xl_512_256_context_2chips
+fsdp: off
+fp16: off
+data:
+  dataset: inaturalist
+  crop_size: 512 
+  val_crop_size: 512
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_base_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_base_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    hidden_size: 768
+    classification_mode: off
+    in_context_patches: 128
+    tiling: naive_two_stream
+    n_layer: 2
+    mem_chip: 2
+  resume: ''
+optimizer:
+  name: adamw
+  lr: 1e-4
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 14
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-b/configs/swin_v2_base_1e-5_hyper-1layer_512_256.yaml

@@ -0,0 +1,71 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_base_1e-5_hyper-1layer_512_256
+fsdp: off
+data:
+  dataset: inaturalist
+  crop_size: 512 
+  val_crop_size: 512
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_base_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_base_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 1
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-5
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 14
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-b/configs/swin_v2_base_1e-5_hyper_256_256.yaml

@@ -0,0 +1,71 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_base_1e-5_hyper_256_256
+fsdp: off
+data:
+  dataset: inaturalist
+  crop_size: 256
+  val_crop_size: 256
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_base_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_base_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 2
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-5
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 14
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-b/configs/swin_v2_base_1e-5_hyper_512_256.yaml

@@ -0,0 +1,71 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_base_1e-5_hyper_512_256
+fsdp: off
+data:
+  dataset: inaturalist
+  crop_size: 512 
+  val_crop_size: 512
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_base_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_base_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 2
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-5
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 14
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-b/configs/swin_v2_base_1e-5_mamba_512_256.yaml

@@ -0,0 +1,75 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_base_1e-5_mamba_512_256
+fsdp: off
+fp16: off
+data:
+  dataset: inaturalist
+  crop_size: 512 
+  val_crop_size: 512
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_base_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_base_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    hidden_size: 768
+    classification_mode: off
+    attention_method: mamba
+    tiling: naive_two_stream
+    n_layer: 4
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-5
+  lr: 1e-5
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 16
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-b/weights/swin-b-hyper-1layer-512-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f78ad762b9e6862c262402f128d9e3638279ac3fb2b09ff3a862123291857cdc
+size 389571612

swin-b/weights/swin-b-hyper-256-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:51d688d4f7eebba35371b5b55f77b52e9bac671a8e6a458422beaffea0f679b8
+size 427339589

swin-b/weights/swin-b-hyper-512-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1b655efd85c4ffd6605bbb5e40edd0c27810e7daf52fd71d213858417f09d59
+size 427339589

swin-b/weights/swin-b-mamba-512-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f5dde760dd280ca06440ce561969c8f9af795fae5b11a0ff8ed955b997c40d18
+size 412174185

swin-b/weights/swin-b-nonxl-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5acf64cb87dea60dd49618bfc58987a217ba45fa8c3369b1667a8ba4f1ad6d1d
+size 369244131

swin-l/configs/swin_v2_large_1e-5_hyper-1layer_512_256.yaml

@@ -0,0 +1,71 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_large_1e-5_hyper-1layer_512_256
+fsdp: off
+data:
+  dataset: inaturalist
+  crop_size: 512 
+  val_crop_size: 512
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_large_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_base_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 1
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-5
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 8
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-l/configs/swin_v2_large_1e-5_hyper_256_256.yaml

@@ -0,0 +1,71 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_large_1e-5_hyper_256_256
+fsdp: off
+data:
+  dataset: inaturalist
+  crop_size: 256
+  val_crop_size: 256
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_large_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_base_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 2
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-5
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 32
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-l/configs/swin_v2_large_1e-5_hyper_512_256.yaml

@@ -0,0 +1,71 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_large_1e-5_hyper_512_256
+fsdp: off
+data:
+  dataset: inaturalist
+  crop_size: 512 
+  val_crop_size: 512
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_large_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_base_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 2
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-5
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 8
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-l/configs/swin_v2_large_1e-5_mamba_512_256.yaml

@@ -0,0 +1,75 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_large_1e-5_mamba_512_256
+fsdp: off
+fp16: off
+data:
+  dataset: inaturalist
+  crop_size: 512 
+  val_crop_size: 512
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_large_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_large_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    hidden_size: 768
+    classification_mode: off
+    attention_method: mamba
+    tiling: naive_two_stream
+    n_layer: 4
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-5
+  lr: 1e-5
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 8
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-l/configs/swin_v2_large_1e-5_nonxl_256.yaml

@@ -0,0 +1,70 @@
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_large_1e-5_nonxl_256
+fsdp: off
+data:
+  dataset: inaturalist
+  crop_size: 256 
+  val_crop_size: 256
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_large_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_base_patch4_window16_256.pth
+    upsample: False
+  cls_head: naive
+  xl_context:
+    skip_connection: off
+    enabled: off
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 2
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-5
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 2
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-l/configs/swin_v2_large_1e-5_xl_512_256_context_2chips.yaml

@@ -0,0 +1,75 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/cluster_paths_inat.yaml
+name: swin_v2_large_1e-5_xl_512_256_context_2chips
+fsdp: off
+fp16: off
+data:
+  dataset: inaturalist
+  crop_size: 512 
+  val_crop_size: 512
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_large_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_base_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    hidden_size: 768
+    classification_mode: off
+    in_context_patches: 128
+    tiling: naive_two_stream
+    n_layer: 2
+    mem_chip: 2
+  resume: ''
+optimizer:
+  name: adamw
+  lr: 1e-5
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 8
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-l/weights/swin-l-hyper-1layer-512-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f6c66c2cb5b50e5624317f5ed9b38c95d2242b05ef69e7b90d97125bf760a12
+size 824377372

swin-l/weights/swin-l-hyper-256-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:138028b3ebe1136e54b0cf9e2465a723a4cc80c852bad1e96d82ba9e225e78b2
+size 862144325

swin-l/weights/swin-l-hyper-512-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:741ade1df8e6af59fff9ab7b10a87243f88365fdf16c612d2698b5712d8b7477
+size 862144325

swin-l/weights/swin-l-mamba-512-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aecf734bc26d77012e3b7148f9396bc9ca5e7c5d225e3136589170453f19a9c1
+size 846978921

swin-l/weights/swin-l-nonxl-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53b5e9cfdb075bc422128002318f86a494235754e8e34ab4589f5fd40c2fb9c5
+size 825798627

swin-s/configs/swin_v2_small_1e-4_hyper-1layer_512_256.yaml

@@ -0,0 +1,71 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_small_1e-4_hyper-1layer_512_256
+fsdp: off
+data:
+  dataset: inaturalist
+  crop_size: 512
+  val_crop_size: 256
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_small_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_small_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 1
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-4
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 20
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-s/configs/swin_v2_small_1e-4_hyper_256_256.yaml

@@ -0,0 +1,71 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_small_1e-4_hyper_256_256
+fsdp: off
+data:
+  dataset: inaturalist
+  crop_size: 256 
+  val_crop_size: 256
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_small_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_small_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 2
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-4
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 80
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-s/configs/swin_v2_small_1e-4_hyper_512_256.yaml

@@ -0,0 +1,71 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_small_1e-4_hyper_512_256
+fsdp: off
+data:
+  dataset: inaturalist
+  crop_size: 512
+  val_crop_size: 256
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_small_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_small_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 2
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-4
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 20
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-s/configs/swin_v2_small_1e-4_nonxl_256.yaml

@@ -0,0 +1,69 @@
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_small_1e-4_nonxl_256
+data:
+  dataset: inaturalist
+  crop_size: 256 
+  val_crop_size: 256
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_small_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_small_patch4_window16_256.pth
+    upsample: False
+  cls_head: naive
+  xl_context:
+    skip_connection: off
+    enabled: off
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 2
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-4
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 80
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-s/configs/swin_v2_small_1e-4_xl_512_256_context_2chips.yaml

@@ -0,0 +1,75 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/cluster_paths_inat.yaml
+name: swin_v2_small_1e-4_xl_512_256_context_2chips
+fsdp: off
+fp16: off
+data:
+  dataset: inaturalist
+  crop_size: 512 
+  val_crop_size: 512
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_small_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_base_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    hidden_size: 768
+    classification_mode: off
+    in_context_patches: 128
+    tiling: naive_two_stream
+    n_layer: 2
+    mem_chip: 2
+  resume: ''
+optimizer:
+  name: adamw
+  lr: 1e-4
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 20
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-s/weights/swin-s-hyper-1layer-512-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:281efaeaa28fcb2209f89dd1aa073acb65e7caf319091b7e9ea00876b819992a
+size 237049820

swin-s/weights/swin-s-hyper-256-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a91cd7dee38c2ff336b0303250a568d9c61ed0101ef4c2b788457d1beb5781a6
+size 274817797

swin-s/weights/swin-s-hyper-512-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b78ca95892d7e1e3e1cc2719d476bd06db961ba4935ae01cf4556192d4f5b82
+size 274817797

swin-s/weights/swin-s-nonxl-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b045b5f3e67f4345a693d7f3f5654fc2d3a89d1d2943b1169f280b363537851a
+size 208993187

swin-s/weights/swin-s-xl-512-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fcee8718f03ec05e2f85627509303581309b197a9f5b5096eea62eb2d61f67e
+size 274816453

swin-t/configs/swin_v2_tiny_1e-4_hyper-1layer_512_256.yaml

@@ -0,0 +1,71 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_tiny_1e-4_hyper-1layer_512_256
+fsdp: off
+data:
+  dataset: inaturalist
+  crop_size: 512
+  val_crop_size: 256
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_tiny_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_tiny_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 1
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-4
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 32
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-t/configs/swin_v2_tiny_1e-4_hyper_256_256.yaml

@@ -0,0 +1,71 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_tiny_1e-4_hyper_256_256
+fsdp: off
+data:
+  dataset: inaturalist
+  crop_size: 256 
+  val_crop_size: 256
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_tiny_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_tiny_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 2
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-4
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 120
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-t/configs/swin_v2_tiny_1e-4_hyper_512_256.yaml

@@ -0,0 +1,71 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_tiny_1e-4_hyper_512_256
+fsdp: off
+data:
+  dataset: inaturalist
+  crop_size: 512
+  val_crop_size: 256
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_tiny_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_tiny_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 2
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-4
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 32
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-t/configs/swin_v2_tiny_1e-4_mamba_512_256.yaml

@@ -0,0 +1,75 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_tiny_1e-4_mamba_512_256
+fsdp: off
+fp16: off
+data:
+  dataset: inaturalist
+  crop_size: 512 
+  val_crop_size: 512
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_tiny_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_tiny_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    hidden_size: 768
+    classification_mode: off
+    attention_method: mamba
+    tiling: naive_two_stream
+    n_layer: 4
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-4
+  lr: 1e-4
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 32
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-t/configs/swin_v2_tiny_1e-4_nonxl_256.yaml

@@ -0,0 +1,69 @@
+base_configs: 
+  - config/base.yaml
+  - config/paths.yaml
+name: swin_v2_tiny_1e-4_nonxl_256
+data:
+  dataset: inaturalist
+  crop_size: 256 
+  val_crop_size: 256
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_tiny_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_tiny_patch4_window16_256.pth
+    upsample: False
+  cls_head: naive
+  xl_context:
+    skip_connection: off
+    enabled: off
+    classification_mode: on
+    tiling: naive_two_stream
+    n_layer: 2
+  resume: ''
+optimizer:
+  name: adamw
+  base_lr: 1e-4
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 120
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-t/configs/swin_v2_tiny_1e-4_xl_512_256_context_2chips.yaml

@@ -0,0 +1,75 @@
+
+base_configs: 
+  - config/base.yaml
+  - config/cluster_paths_inat.yaml
+name: swin_v2_tiny_1e-4_xl_512_256_context_2chips
+fsdp: off
+fp16: off
+data:
+  dataset: inaturalist
+  crop_size: 512 
+  val_crop_size: 512
+  batch_size: ${train.batch_size}
+  val_batch_size: ${train.val_batch_size}
+  num_workers: 1
+  num_classes: 284
+  interpolation: bilinear
+  test_crop: False
+  aug:
+    auto_augment: rand-m9-mstd0.5-inc1
+    color_jitter: 0.4
+    reprob: 0.0
+    remode: pixel
+    recount: 1
+    mixup: 0.0
+    cutmix: 0.0
+    label_smoothing: 0.3
+    random_resized_crop: False
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
+  supercategories:
+    - Reptilia
+model: 
+  name: EncoderDecoderV2
+  num_classes: ${data.num_classes}
+  mlp_ratio: 4
+  backbone_class: swinv2_tiny_window16_256_timm
+  backbone:
+    img_size: 256
+    input_dim: 3
+    use_vanilla_backward: False
+    pretrained: ${oc.env:PRETRAINED_CKPT_PATH, "./ckpts"}/swinv2_base_patch4_window16_256.pth
+    upsample: False
+  cls_head: xl
+  xl_context:
+    skip_connection: off
+    enabled: off
+    hidden_size: 768
+    classification_mode: off
+    in_context_patches: 128
+    tiling: naive_two_stream
+    n_layer: 2
+    mem_chip: 2
+  resume: ''
+optimizer:
+  name: adamw
+  lr: 1e-4
+  classifier_ratio: 1.0
+  warmup_epochs: 0
+train:
+  epochs: 100
+  batch_size: 30
+  val_batch_size: 1
+  freeze_epochs: 0
+  test_every: 1
+  test_reset: True
+  clip_grad: 5.0
+val: False
+losses:
+  losses:
+    - name: cls
+      type: CrossEntropy
+      params: 
+        field: label 
+      weight: 1.0
+      display: on

swin-t/weights/swin-t-hyper-1layer-512-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ba925dd7f4d0ce068de1d500d3628ae96a9f38311c5c0dd5302221bd345e8bb
+size 151428048

swin-t/weights/swin-t-hyper-256-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95db1742267db95e0d0b24337f3ddb14787675b8b4ec5456b98a85837ebe576a
+size 189197453

swin-t/weights/swin-t-hyper-512-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c71ac4871b2989d159418e8de49430e9a162e163727fb196e73f3b4739e65811
+size 189197453

swin-t/weights/swin-t-mamba-512-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ad07cf9a5ddc2865bdc6a88aaebc205b4ec14dac308dda4c59ac82bc2143104
+size 174032049

swin-t/weights/swin-t-nonxl-256-top1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e12d61f071a805dd57b299f4dd7df9ada90c1b2d243f8f7f5780345320f4c6c
+size 123373595