Merge pull request #748 from mlcommons/variant_bsz_fix

override imagenet_resnet silu and gelu bsz in configs
mlcommons · Apr 2, 2024 · d329e42 · d329e42
2 parents e9bdbd8 + e203a42
commit d329e42
Show file tree

Hide file tree

Showing 17 changed files with 66 additions and 1 deletion.
diff --git a/prize_qualification_baselines/external_tuning/jax_nadamw_full_budget.py b/prize_qualification_baselines/external_tuning/jax_nadamw_full_budget.py
@@ -307,6 +307,10 @@ def get_batch_size(workload_name):
     return 32
   elif workload_name == 'imagenet_resnet':
     return 1024
+  elif workload_name == 'imagenet_resnet_silu':
+    return 512
+  elif workload_name == 'imagenet_resnet_gelu':
+    return 512
   elif workload_name == 'imagenet_vit':
     return 1024
   elif workload_name == 'librispeech_conformer':

diff --git a/prize_qualification_baselines/external_tuning/jax_nadamw_target_setting.py b/prize_qualification_baselines/external_tuning/jax_nadamw_target_setting.py
@@ -307,6 +307,10 @@ def get_batch_size(workload_name):
     return 32
   elif workload_name == 'imagenet_resnet':
     return 1024
+  elif workload_name == 'imagenet_resnet_silu':
+    return 512
+  elif workload_name == 'imagenet_resnet_gelu':
+    return 512
   elif workload_name == 'imagenet_vit':
     return 1024
   elif workload_name == 'librispeech_conformer':

diff --git a/prize_qualification_baselines/external_tuning/pytorch_nadamw_full_budget.py b/prize_qualification_baselines/external_tuning/pytorch_nadamw_full_budget.py
@@ -309,6 +309,10 @@ def get_batch_size(workload_name):
     return 32
   elif workload_name == 'imagenet_resnet':
     return 1024
+  elif workload_name == 'imagenet_resnet_silu':
+    return 512
+  elif workload_name == 'imagenet_resnet_gelu':
+    return 512
   elif workload_name == 'imagenet_vit':
     return 1024
   elif workload_name == 'librispeech_conformer':

diff --git a/prize_qualification_baselines/external_tuning/pytorch_nadamw_target_setting.py b/prize_qualification_baselines/external_tuning/pytorch_nadamw_target_setting.py
@@ -309,6 +309,10 @@ def get_batch_size(workload_name):
     return 32
   elif workload_name == 'imagenet_resnet':
     return 1024
+  elif workload_name == 'imagenet_resnet_silu':
+    return 512
+  elif workload_name == 'imagenet_resnet_gelu':
+    return 512
   elif workload_name == 'imagenet_vit':
     return 1024
   elif workload_name == 'librispeech_conformer':

diff --git a/prize_qualification_baselines/self_tuning/jax_nadamw_full_budget.py b/prize_qualification_baselines/self_tuning/jax_nadamw_full_budget.py
@@ -322,6 +322,10 @@ def get_batch_size(workload_name):
     return 32
   elif workload_name == 'imagenet_resnet':
     return 1024
+  elif workload_name == 'imagenet_resnet_silu':
+    return 512
+  elif workload_name == 'imagenet_resnet_gelu':
+    return 512
   elif workload_name == 'imagenet_vit':
     return 1024
   elif workload_name == 'librispeech_conformer':

diff --git a/prize_qualification_baselines/self_tuning/jax_nadamw_target_setting.py b/prize_qualification_baselines/self_tuning/jax_nadamw_target_setting.py
@@ -322,6 +322,10 @@ def get_batch_size(workload_name):
     return 32
   elif workload_name == 'imagenet_resnet':
     return 1024
+  elif workload_name == 'imagenet_resnet_silu':
+    return 512
+  elif workload_name == 'imagenet_resnet_gelu':
+    return 512
   elif workload_name == 'imagenet_vit':
     return 1024
   elif workload_name == 'librispeech_conformer':

diff --git a/prize_qualification_baselines/self_tuning/pytorch_nadamw_full_budget.py b/prize_qualification_baselines/self_tuning/pytorch_nadamw_full_budget.py
@@ -324,6 +324,10 @@ def get_batch_size(workload_name):
     return 32
   elif workload_name == 'imagenet_resnet':
     return 1024
+  elif workload_name == 'imagenet_resnet_silu':
+    return 512
+  elif workload_name == 'imagenet_resnet_gelu':
+    return 512
   elif workload_name == 'imagenet_vit':
     return 1024
   elif workload_name == 'librispeech_conformer':

diff --git a/prize_qualification_baselines/self_tuning/pytorch_nadamw_target_setting.py b/prize_qualification_baselines/self_tuning/pytorch_nadamw_target_setting.py
@@ -324,6 +324,10 @@ def get_batch_size(workload_name):
     return 32
   elif workload_name == 'imagenet_resnet':
     return 1024
+  elif workload_name == 'imagenet_resnet_silu':
+    return 512
+  elif workload_name == 'imagenet_resnet_gelu':
+    return 512
   elif workload_name == 'imagenet_vit':
     return 1024
   elif workload_name == 'librispeech_conformer':

diff --git a/reference_algorithms/paper_baselines/adamw/jax/submission.py b/reference_algorithms/paper_baselines/adamw/jax/submission.py
@@ -165,6 +165,10 @@ def get_batch_size(workload_name):
     return 32
   elif workload_name == 'imagenet_resnet':
     return 1024
+  elif workload_name == 'imagenet_resnet_silu':
+    return 512
+  elif workload_name == 'imagenet_resnet_gelu':
+    return 512
   elif workload_name == 'imagenet_vit':
     return 1024
   elif workload_name == 'librispeech_conformer':

diff --git a/reference_algorithms/paper_baselines/adamw/pytorch/submission.py b/reference_algorithms/paper_baselines/adamw/pytorch/submission.py
@@ -133,6 +133,10 @@ def get_batch_size(workload_name):
     return 32
   elif workload_name == 'imagenet_resnet':
     return 1024
+  elif workload_name == 'imagenet_resnet_silu':
+    return 512
+  elif workload_name == 'imagenet_resnet_gelu':
+    return 512
   elif workload_name == 'imagenet_vit':
     return 1024
   elif workload_name == 'librispeech_conformer':

diff --git a/reference_algorithms/paper_baselines/momentum/jax/submission.py b/reference_algorithms/paper_baselines/momentum/jax/submission.py
@@ -199,6 +199,10 @@ def get_batch_size(workload_name):
     return 32
   elif workload_name == 'imagenet_resnet':
     return 1024
+  elif workload_name == 'imagenet_resnet_silu':
+    return 512
+  elif workload_name == 'imagenet_resnet_gelu':
+    return 512
   elif workload_name == 'imagenet_vit':
     return 1024
   elif workload_name == 'librispeech_conformer':

diff --git a/reference_algorithms/paper_baselines/momentum/pytorch/submission.py b/reference_algorithms/paper_baselines/momentum/pytorch/submission.py
@@ -152,6 +152,10 @@ def get_batch_size(workload_name):
     return 32
   elif workload_name == 'imagenet_resnet':
     return 1024
+  elif workload_name == 'imagenet_resnet_silu':
+    return 512
+  elif workload_name == 'imagenet_resnet_gelu':
+    return 512
   elif workload_name == 'imagenet_vit':
     return 1024
   elif workload_name == 'librispeech_conformer':

diff --git a/reference_algorithms/paper_baselines/nadamw/jax/submission.py b/reference_algorithms/paper_baselines/nadamw/jax/submission.py
@@ -307,6 +307,10 @@ def get_batch_size(workload_name):
     return 32
   elif workload_name == 'imagenet_resnet':
     return 1024
+  elif workload_name == 'imagenet_resnet_silu':
+    return 512
+  elif workload_name == 'imagenet_resnet_gelu':
+    return 512
   elif workload_name == 'imagenet_vit':
     return 1024
   elif workload_name == 'librispeech_conformer':

diff --git a/reference_algorithms/paper_baselines/nadamw/pytorch/submission.py b/reference_algorithms/paper_baselines/nadamw/pytorch/submission.py
@@ -309,6 +309,10 @@ def get_batch_size(workload_name):
     return 32
   elif workload_name == 'imagenet_resnet':
     return 1024
+  elif workload_name == 'imagenet_resnet_silu':
+    return 512
+  elif workload_name == 'imagenet_resnet_gelu':
+    return 512
   elif workload_name == 'imagenet_vit':
     return 1024
   elif workload_name == 'librispeech_conformer':

diff --git a/reference_algorithms/paper_baselines/nesterov/jax/submission.py b/reference_algorithms/paper_baselines/nesterov/jax/submission.py
@@ -199,6 +199,10 @@ def get_batch_size(workload_name):
     return 32
   elif workload_name == 'imagenet_resnet':
     return 1024
+  elif workload_name == 'imagenet_resnet_silu':
+    return 512
+  elif workload_name == 'imagenet_resnet_gelu':
+    return 512
   elif workload_name == 'imagenet_vit':
     return 1024
   elif workload_name == 'librispeech_conformer':

diff --git a/reference_algorithms/paper_baselines/nesterov/pytorch/submission.py b/reference_algorithms/paper_baselines/nesterov/pytorch/submission.py
@@ -152,6 +152,10 @@ def get_batch_size(workload_name):
     return 32
   elif workload_name == 'imagenet_resnet':
     return 1024
+  elif workload_name == 'imagenet_resnet_silu':
+    return 512
+  elif workload_name == 'imagenet_resnet_gelu':
+    return 512
   elif workload_name == 'imagenet_vit':
     return 1024
   elif workload_name == 'librispeech_conformer':

diff --git a/submissions/template/submission.py b/submissions/template/submission.py
@@ -49,7 +49,8 @@ def get_batch_size(workload_name):
     Args: 
       workload_name (str): Valid workload_name values are: "wmt", "ogbg", 
         "criteo1tb", "fastmri", "imagenet_resnet", "imagenet_vit", 
-        "librispeech_deepspeech", "librispeech_conformer".
+        "librispeech_deepspeech", "librispeech_conformer" or any of the
+        variants.
     Returns:
       int: batch_size 
     Raises: