From 08ab1b15b7f8da4de2ad04264a03d6e02fd56916 Mon Sep 17 00:00:00 2001 From: YannAhlgrim Date: Fri, 17 Jul 2026 09:57:42 +0200 Subject: [PATCH] micro batches --- configs/supervised_vitb16_448.yaml | 4 ++-- configs/supervised_vitg16_224_in22k.yaml | 4 ++-- configs/supervised_vith14_224.yaml | 4 ++-- configs/supervised_vith14_224_in1k.yaml | 4 ++-- configs/supervised_vith14_224_in22k.yaml | 4 ++-- configs/supervised_vith16_448.yaml | 4 ++-- configs/supervised_vith16_448_in1k.yaml | 4 ++-- 7 files changed, 14 insertions(+), 14 deletions(-) diff --git a/configs/supervised_vitb16_448.yaml b/configs/supervised_vitb16_448.yaml index 80f0e5a..e0f7cbc 100644 --- a/configs/supervised_vitb16_448.yaml +++ b/configs/supervised_vitb16_448.yaml @@ -10,7 +10,7 @@ meta: head_type: linear data: - batch_size: 128 + batch_size: 16 root_path: ./wilds_data num_workers: 10 pin_mem: true @@ -31,7 +31,7 @@ optimization: lr: 0.01 weight_decay: 5.0e-4 lr_schedule: cosine - gradient_accumulation_steps: 1 + gradient_accumulation_steps: 8 final_lr: 0.0 early_stopping: enabled: true diff --git a/configs/supervised_vitg16_224_in22k.yaml b/configs/supervised_vitg16_224_in22k.yaml index d7a5ef4..fc58d0f 100644 --- a/configs/supervised_vitg16_224_in22k.yaml +++ b/configs/supervised_vitg16_224_in22k.yaml @@ -12,7 +12,7 @@ meta: use_gradient_checkpointing: true data: - batch_size: 256 + batch_size: 16 root_path: ./wilds_data num_workers: 10 pin_mem: true @@ -33,7 +33,7 @@ optimization: lr: 0.001 weight_decay: 5.0e-4 lr_schedule: cosine - gradient_accumulation_steps: 1 + gradient_accumulation_steps: 16 final_lr: 0.0 early_stopping: enabled: true diff --git a/configs/supervised_vith14_224.yaml b/configs/supervised_vith14_224.yaml index d3b9f07..b0d312f 100644 --- a/configs/supervised_vith14_224.yaml +++ b/configs/supervised_vith14_224.yaml @@ -10,7 +10,7 @@ meta: head_type: bn_linear data: - batch_size: 512 + batch_size: 16 root_path: ./wilds_data num_workers: 10 pin_mem: true @@ -31,7 +31,7 @@ optimization: lr: 0.001 weight_decay: 5.0e-4 lr_schedule: cosine - gradient_accumulation_steps: 1 + gradient_accumulation_steps: 32 final_lr: 0.0 early_stopping: enabled: true diff --git a/configs/supervised_vith14_224_in1k.yaml b/configs/supervised_vith14_224_in1k.yaml index 6dcc577..b1de2af 100644 --- a/configs/supervised_vith14_224_in1k.yaml +++ b/configs/supervised_vith14_224_in1k.yaml @@ -11,7 +11,7 @@ meta: head_type: linear data: - batch_size: 1024 + batch_size: 16 root_path: ./wilds_data num_workers: 10 pin_mem: true @@ -32,7 +32,7 @@ optimization: lr: 0.01 weight_decay: 0 lr_schedule: cosine - gradient_accumulation_steps: 1 + gradient_accumulation_steps: 64 final_lr: 0.0 early_stopping: enabled: true diff --git a/configs/supervised_vith14_224_in22k.yaml b/configs/supervised_vith14_224_in22k.yaml index 89756e2..d9a59e7 100644 --- a/configs/supervised_vith14_224_in22k.yaml +++ b/configs/supervised_vith14_224_in22k.yaml @@ -11,7 +11,7 @@ meta: head_type: bn_linear data: - batch_size: 512 + batch_size: 16 root_path: ./wilds_data num_workers: 10 pin_mem: true @@ -32,7 +32,7 @@ optimization: lr: 0.001 weight_decay: 5.0e-4 lr_schedule: cosine - gradient_accumulation_steps: 1 + gradient_accumulation_steps: 32 final_lr: 0.0 early_stopping: enabled: true diff --git a/configs/supervised_vith16_448.yaml b/configs/supervised_vith16_448.yaml index e7ce62f..3e4b95f 100644 --- a/configs/supervised_vith16_448.yaml +++ b/configs/supervised_vith16_448.yaml @@ -10,7 +10,7 @@ meta: head_type: bn_linear data: - batch_size: 1024 + batch_size: 16 root_path: ./wilds_data num_workers: 10 pin_mem: true @@ -31,7 +31,7 @@ optimization: lr: 0.01 weight_decay: 5.0e-4 lr_schedule: cosine - gradient_accumulation_steps: 1 + gradient_accumulation_steps: 64 final_lr: 0.0 early_stopping: enabled: true diff --git a/configs/supervised_vith16_448_in1k.yaml b/configs/supervised_vith16_448_in1k.yaml index fc8f82b..5d0bd85 100644 --- a/configs/supervised_vith16_448_in1k.yaml +++ b/configs/supervised_vith16_448_in1k.yaml @@ -12,7 +12,7 @@ meta: use_gradient_checkpointing: true data: - batch_size: 512 + batch_size: 16 root_path: ./wilds_data num_workers: 10 pin_mem: true @@ -33,7 +33,7 @@ optimization: lr: 0.001 weight_decay: 0 lr_schedule: cosine - gradient_accumulation_steps: 1 + gradient_accumulation_steps: 32 final_lr: 0.0 early_stopping: enabled: true