corrected imagenet learning rate

thomasverelst · Jun 18, 2020 · b586cd5 · b586cd5
1 parent a29b33f
commit b586cd5
Show file tree

Hide file tree

Showing 2 changed files with 8 additions and 7 deletions.
diff --git a/classification/README.md b/classification/README.md
@@ -68,30 +68,31 @@ smaller models are Resnet 26, 20, 14 and 8, e.g.
 
 ## ResNet-101 on ImageNet-1K
 
+Note: Specify the path of your imagenet dataset using the `--dataset-root` argument, or change the path in the `main_imagenet.y` file. 
 
 ### Evaluate a pretrained sparse DynConv network
 
-    python main_imagenet.py --batchsize 64 -r exp/imagenet/resnet101/sparse03/checkpoint_best.pth --budget 0.3 -e
+    python main_imagenet.py -r exp/imagenet/resnet101/sparse03/checkpoint_best.pth --budget 0.3 -e
 
 should result in 
 
 >\* Epoch 97 - Prec@1 75.710
 >\* FLOPS (multiply-accumulates, MACs) per image:  2997.180928 MMac
 
 use the `--plot_ponder` flag to visualize the ponder cost maps (computation heatmaps):
-    python main_imagenet.py --batchsize 64 -r exp/imagenet/resnet101/sparse03/checkpoint_best.pth --budget 0.3 -e --plot_ponder
+    python main_imagenet.py -r exp/imagenet/resnet101/sparse03/checkpoint_best.pth --budget 0.3 -e --plot_ponder
 
 
 Likewise:
 
-    python main_imagenet.py --batchsize 64 -r exp/imagenet/resnet101/sparse05/checkpoint_best.pth --budget 0.5 -e    
+    python main_imagenet.py -r exp/imagenet/resnet101/sparse05/checkpoint_best.pth --budget 0.5 -e    
 
-    python main_imagenet.py --batchsize 64 -r exp/imagenet/resnet101/sparse07/checkpoint_best.pth --budget 0.7 -e
+    python main_imagenet.py -r exp/imagenet/resnet101/sparse07/checkpoint_best.pth --budget 0.7 -e
 
-    python main_imagenet.py --batchsize 64 -r exp/imagenet/resnet101/sparse08/checkpoint_best.pth --budget 0.8 -e
+    python main_imagenet.py -r exp/imagenet/resnet101/sparse08/checkpoint_best.pth --budget 0.8 -e
 
 
 
 ### Evaluate a pretrained baseline
 
-    python main_imagenet.py --batchsize 64 -r exp/imagenet/resnet101/base/checkpoint_best.pth -e
+    python main_imagenet.py -r exp/imagenet/resnet101/base/checkpoint_best.pth -e
diff --git a/classification/main_imagenet.py b/classification/main_imagenet.py
@@ -24,7 +24,7 @@
 
 def main():
     parser = argparse.ArgumentParser(description='PyTorch ImageNet Training with sparse masks')
-    parser.add_argument('--lr', default=0.1, type=float, help='learning rate')
+    parser.add_argument('--lr', default=0.025, type=float, help='learning rate')
     parser.add_argument('--lr_decay', default=[30,60,90], nargs='+', type=int, help='learning rate decay epochs')
     parser.add_argument('--momentum', default=0.9, type=float, help='momentum')
     parser.add_argument('--weight_decay', default=1e-4, type=float, help='weight decay')