diff --git a/docker/cpu/Dockerfile b/docker/cpu/Dockerfile
index f94c5ae..8f1e1f6 100755
--- a/docker/cpu/Dockerfile
+++ b/docker/cpu/Dockerfile
@@ -1,4 +1,4 @@
-# nvidia-docker build -t cresi_v3_image .
+# docker build -t cresi_v3_image_cpu .
 # nvidia-docker run -it --ipc=host -v /local_data:/local_data/ -p 9111:9111 --name cresi_v3 cresi_v3_image
 # To use jupyter notebook, inside container run:
 # jupyter notebook --ip 0.0.0.0 --no-browser --allow-root --port=9111
@@ -128,10 +128,13 @@ RUN apt install -y libgl1-mesa-glx
 # add a jupyter kernel for the conda environment in case it's wanted
 RUN source activate cresi && python -m ipykernel.kernelspec
 
+# # clone cresi
+RUN git clone https://github.com/avanetten/cresi.git /opt/cresi/
+
 # TensorBoard
 EXPOSE 6006
 # IPython
 EXPOSE 9111
 
-WORKDIR "/"
+WORKDIR "/opt/cresi"
 RUN ["/bin/bash"]
\ No newline at end of file
diff --git a/docker/gpu/Dockerfile b/docker/gpu/Dockerfile
index 327ed32..96a63b0 100755
--- a/docker/gpu/Dockerfile
+++ b/docker/gpu/Dockerfile
@@ -139,10 +139,13 @@ RUN apt install -y libgl1-mesa-glx
 # add a jupyter kernel for the conda environment in case it's wanted
 RUN source activate cresi && python -m ipykernel.kernelspec
 
+# # clone cresi
+RUN git clone https://github.com/avanetten/cresi.git /opt/cresi/
+
 # TensorBoard
 EXPOSE 6006
 # IPython
 EXPOSE 9111
 
-WORKDIR "/"
-RUN ["/bin/bash"]
+WORKDIR "/opt/cresi"
+RUN ["/bin/bash"]
\ No newline at end of file
diff --git a/notebooks/dar_tutorial_cpu/cresi_cpu_part2.ipynb b/notebooks/dar_tutorial_cpu/cresi_cpu_part2.ipynb
index ac0cdda..eb0ab0c 100644
--- a/notebooks/dar_tutorial_cpu/cresi_cpu_part2.ipynb
+++ b/notebooks/dar_tutorial_cpu/cresi_cpu_part2.ipynb
@@ -324,7 +324,7 @@
     "-----\n",
     "## 5. Run Inference\n",
     "\n",
-    "The simplest option is to sipmly run the test.sh script (e.g. ```./test.sh configs/dar_tutorial_cpu.json```) which will execute all of the various inference scipts in series.  For this exercise, we will instead run the individual commands to view what's going on at each step.\n",
+    "The simplest option is to sipmly run the test.sh script (e.g. ```./test.sh configs/dar_tutorial_cpu.json```) which will execute all of the various inference scripts in series.  For this exercise, we will instead run the individual commands to view what's going on at each step.\n",
     "\n",
     "### 5.1. 02_eval.py \n",
     "\n",
@@ -1515,7 +1515,7 @@
    "metadata": {},
    "source": [
     "----\n",
-    "Clearly, the optial route is very different when weighting by time rather than distance, underscoring the need to incorporate speed estimates into road graphs.  We encourage the interested reader to explore optimal routes between random nodes in the cell below (running multiple times will yield different results each time)."
+    "Clearly, the optimal route is very different when weighting by time rather than distance, underscoring the need to incorporate speed estimates into road graphs.  We encourage the interested reader to explore optimal routes between random nodes in the cell below (running multiple times will yield different results each time)."
    ]
   },
   {
@@ -1634,7 +1634,7 @@
     "----\n",
     "### Conclusion\n",
     "\n",
-    "In this tutorial we have shown how to extract a road network graph with speed / travel time estimates direclty from satellite imagery using only open source data and code.  Inference is relatively rapid, running at 0.7 square kilometers per minute even on a CPU.  GPU infrence times would be a minimum of 20X faster.  Even for a city unseen during training (Dar Es Salaam), the CRESI algorithm manages to return a road network that is sufficiently connected and complete to be routable.  We showed how to compute optimal routes, and demonstrated the importance of routing using speed versus geometric distance.  We encourage interested parties to explore new testing regions, dive deeper into the intricacies of road network extraction and routing, and to contribute to open source geospatial software if time and interest permits.  "
+    "In this tutorial we have shown how to extract a road network graph with speed / travel time estimates directly from satellite imagery using only open source data and code.  Inference is relatively rapid, running at 0.7 square kilometers per minute even on a CPU.  GPU inference times would be a minimum of 20X faster.  Even for a city unseen during training (Dar Es Salaam), the CRESI algorithm manages to return a road network that is sufficiently connected and complete to be routable.  We showed how to compute optimal routes, and demonstrated the importance of routing using speed versus geometric distance.  We encourage interested parties to explore new testing regions, dive deeper into the intricacies of road network extraction and routing, and to contribute to open source geospatial software if time and interest permits.  "
    ]
   },
   {
diff --git a/results/weights/ave_speed_mc_focal/fold0_best.pth.zip b/results/weights/aws_weights/fold0_best.pth.zip
similarity index 84%
rename from results/weights/ave_speed_mc_focal/fold0_best.pth.zip
rename to results/weights/aws_weights/fold0_best.pth.zip
index cc5db1f..fd1c6ff 100644
Binary files a/results/weights/ave_speed_mc_focal/fold0_best.pth.zip and b/results/weights/aws_weights/fold0_best.pth.zip differ