lower default initial learning rate to what Jonathan Ho had in his original repo

dalle2_pytorch/optimizer.py

@@ -7,7 +7,7 @@ def separate_weight_decayable_params(params):
 
 def get_optimizer(
     params,
-    lr = 3e-4,
+    lr = 2e-5,
     wd = 1e-2,
     betas = (0.9, 0.999),
     filter_by_requires_grad = False

dalle2_pytorch/train.py

@@ -221,7 +221,7 @@ class DecoderTrainer(nn.Module):
         self,
         decoder,
         use_ema = True,
-        lr = 3e-4,
+        lr = 2e-5,
         wd = 1e-2,
         max_grad_norm = None,
         amp = False,

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.2.19',
+  version = '0.2.20',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',