diff --git a/codes/models/archs/ChainedEmbeddingGen.py b/codes/models/archs/ChainedEmbeddingGen.py
index e1fce993..f283e1b3 100644
--- a/codes/models/archs/ChainedEmbeddingGen.py
+++ b/codes/models/archs/ChainedEmbeddingGen.py
@@ -128,7 +128,10 @@ class MultifacetedChainedEmbeddingGen(nn.Module):
 
         self.initial_conv = ConvGnLelu(3, 64, kernel_size=7, bias=True, norm=False, activation=False)
 
-        self.teco_recurrent_process = ConvGnLelu(3, 64, kernel_size=3, stride=2, norm=False, bias=True, activation=False)
+        if scale == 2:
+            self.teco_recurrent_process = ConvGnLelu(3, 64, kernel_size=3, stride=2, norm=False, bias=True, activation=False)
+        else:
+            self.teco_recurrent_process = ConvGnLelu(3, 64, kernel_size=7, stride=4, norm=False, bias=True, activation=False)
         self.teco_recurrent_join = ReferenceJoinBlock(64, residual_weight_init_factor=.01, final_norm=False, kernel_size=1, depth=3, join=False)
 
         self.prog_recurrent_process = ConvGnLelu(64, 64, kernel_size=3, stride=1, norm=False, bias=True, activation=False)
diff --git a/codes/models/base_model.py b/codes/models/base_model.py
index ea08aecc..47664e8a 100644
--- a/codes/models/base_model.py
+++ b/codes/models/base_model.py
@@ -108,7 +108,9 @@ class BaseModel():
 
         load_net_clean = OrderedDict()  # remove unnecessary 'module.'
         for k, v in load_net.items():
-            if k.startswith('module.'):
+            if 'teco_recurrent_process' in k:
+                continue
+            elif k.startswith('module.'):
                 load_net_clean[k[7:]] = v
             else:
                 load_net_clean[k] = v