readme fix and make cifar10 work off grenoble (#6187)

determined-ai · Mar 9, 2023 · 9563274 · 9563274
1 parent 46e9f74
commit 9563274
Show file tree

Hide file tree

Showing 3 changed files with 13 additions and 7 deletions.
diff --git a/ds_autotuning_prototype/README.md b/ds_autotuning_prototype/README.md
@@ -16,7 +16,7 @@ The basic steps:
 In the same directory as this `README`, run the following, for instance:
 
 ```bash
-python3 -m dsat.autotune examples/minimal_example/autotune_config.yaml examples/minimal_example
+python3 -m dsat.autotune examples/ffn_example/autotune_config.yaml examples/ffn_example
 ```
 
 (the config may need to be altered for your cluster.)
@@ -43,4 +43,4 @@ TODOs:
 - Not all native DS AT code paths are currently supported, e.g. providing explicit batch sizes to use,
   fast mode.
 - Benchmark against native DS AT. The 0.8.1 update which fixed the DS AT units issues might have also
-broken the `--autotuning run` flag?
+  broken the `--autotuning run` flag?
diff --git a/ds_autotuning_prototype/examples/cifar10_example/script.py b/ds_autotuning_prototype/examples/cifar10_example/script.py
@@ -38,10 +38,16 @@ def main(
     transform = transforms.Compose(
         [transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]
     )
-
-    trainset = torchvision.datasets.CIFAR10(
-        root=hparams.data.root, train=True, download=False, transform=transform
-    )
+    # Cluster specific code for Grenoble.
+    try:
+        # This works on GG's Grenoble setup, but not otherwise.
+        trainset = torchvision.datasets.CIFAR10(
+            root=hparams.data.root, train=True, download=False, transform=transform
+        )
+    except:
+        trainset = torchvision.datasets.CIFAR10(
+            root=".", train=True, download=True, transform=transform
+        )
 
     trainloader = torch.utils.data.DataLoader(trainset, batch_size=16, shuffle=True, num_workers=2)
 

diff --git a/ds_autotuning_prototype/examples/ffn_example/autotune_config.yaml b/ds_autotuning_prototype/examples/ffn_example/autotune_config.yaml
@@ -13,7 +13,7 @@ searcher:
   smaller_is_better: False
 hyperparameters:
   dim: 1024
-  layers: 16 # Total params will be layers * 1024 ** 2
+  layers: 4 # Total params will be layers * 1024 ** 2
   # NOTE: dsat code expects usual DS config dict to appear as in the below.
   ds_config:
     train_micro_batch_size_per_gpu: 128