diff --git a/src/ml/neural_net/mps_layer_instance_norm_data_loader.mm b/src/ml/neural_net/mps_layer_instance_norm_data_loader.mm
index 60e49c4439..613596baf8 100644
--- a/src/ml/neural_net/mps_layer_instance_norm_data_loader.mm
+++ b/src/ml/neural_net/mps_layer_instance_norm_data_loader.mm
@@ -177,11 +177,11 @@ - (void) loadBeta:(float *)beta {
 
 - (float *) beta {
   NSUInteger previousStyle = _styleIndex;
-  _betaPlaceHolder = [NSMutableData data];
+  _betaPlaceHolder.length = 0;
   for (NSUInteger index = 0; index < _styles; index++) {
     _styleIndex = index;
     [self checkpointWithCommandQueue:_cq];
-    float* betaWeights = (float *) [[[_style_props objectAtIndex: _styleIndex] betaBuffer] contents];
+    float* betaWeights = (float *) [_style_props[_styleIndex].betaBuffer contents];
     [_betaPlaceHolder appendBytes:betaWeights length:sizeof(float)*_numberOfFeatureChannels];
   }
   _styleIndex = previousStyle;
@@ -196,12 +196,12 @@ - (void) loadGamma:(float *)gamma {
 
 // TODO: refactor for multiple indicies
 - (float *) gamma {
-  _gammaPlaceHolder = [NSMutableData data];
   NSUInteger previousStyle = _styleIndex;
+  _gammaPlaceHolder.length = 0;
   for (NSUInteger index = 0; index < _styles; index++) { 
     _styleIndex = index; 
     [self checkpointWithCommandQueue:_cq];
-    float* gammaWeights = (float *) [[[_style_props objectAtIndex: _styleIndex] gammaBuffer] contents];
+    float* gammaWeights = (float *) [_style_props[_styleIndex].gammaBuffer contents];
     [_gammaPlaceHolder appendBytes:gammaWeights length:sizeof(float)*_numberOfFeatureChannels];
   }
   _styleIndex = previousStyle;
diff --git a/src/ml/neural_net/style_transfer/mps_style_transfer.m b/src/ml/neural_net/style_transfer/mps_style_transfer.m
index 0517e62e4f..e74dff351e 100644
--- a/src/ml/neural_net/style_transfer/mps_style_transfer.m
+++ b/src/ml/neural_net/style_transfer/mps_style_transfer.m
@@ -438,6 +438,10 @@ - (void) setLearningRate:(float)lr {
   return [lossDict copy];
 }
 
+/**
+* HACK: this somehow checkpoints the model for weight exports and updating the
+*       data loaders. Following up internally for a proper fix to this issue.
+**/
 - (void) checkpoint {
   _inferenceGraph = [MPSNNGraph graphWithDevice:_dev
                                     resultImage:_model.forwardPass
diff --git a/src/python/turicreate/toolkits/style_transfer/style_transfer.py b/src/python/turicreate/toolkits/style_transfer/style_transfer.py
index 843287c34c..61665da47b 100644
--- a/src/python/turicreate/toolkits/style_transfer/style_transfer.py
+++ b/src/python/turicreate/toolkits/style_transfer/style_transfer.py
@@ -409,6 +409,18 @@ def create(style_dataset, content_dataset, style_feature=None,
 
         mps_mxnet_key_map = _MpsStyleGraphAPI.mps_mxnet_weight_dict()
 
+        # LOGIC
+        #
+        # The two layers we are concerned about are very different. The instance
+        # norm layer weights should have a dimensionality two. Wheras the 
+        # Convolutional Weights have a dimensionality of four. The Convolutional
+        # weights are in a different format in MxNet then they are in MPS. Since
+        # the arrays come back flattened in the MPS a reshape has to occur. But
+        # this reshape happens before the transpose therefore the shape itself
+        # has to be transposed. For the InstanceNorm Layer, however, the weights
+        # are passed back to MxNet in the correct format so just a simple
+        # reshape suffices.
+        #
         for key in mps_weights:
             if "transformer" in key:
                 weight = transformer.collect_params()[mps_mxnet_key_map[key]].data()