Make MPS not the default (#19)

Let's make MPS not the default but just fallback to portable ops.

Author: mergennachin

mv3/apple/ExecuTorchDemo/README.md

@@ -48,27 +48,14 @@ The steps to add the necessary package dependency are available in the
 
 ### 4. Set Up ExecuTorch
 
-Clone ExecuTorch and set up the environment as explained in the [Building from Source tutorial](https://pytorch.org/executorch/main/using-executorch-building-from-source.html):
-
 ```bash
-git clone -b viable/strict https://github.com/pytorch/executorch.git && cd executorch
-
-python3 -m venv .venv && source .venv/bin/activate && pip install --upgrade pip
-
-./install_executorch.sh
+pip install executorch
 ```
 
-### 5. Backend Dependencies
-
-Install additional dependencies for [Core ML](https://pytorch.org/executorch/main/backends-coreml.html) and [MPS](https://pytorch.org/executorch/main/backends-mps.html) backends:
-
-```bash
-./backends/apple/coreml/scripts/install_requirements.sh
+Alternatively, clone ExecuTorch and set up the environment as explained in the [Building from Source tutorial](https://pytorch.org/executorch/main/using-executorch-building-from-source.html):
 
-./backends/apple/mps/install_requirements.sh
-```
 
-### 6. Clone the Demo App
+### 5. Clone the Demo App
 
 ```bash
 git clone --depth 1 https://github.com/pytorch-labs/executorch-examples.git
@@ -80,21 +67,19 @@ Now, let's move on to exporting and bundling the MobileNet v3 model.
 
 ### 1. Export Model
 
-Export the MobileNet v3 model using the command line and move the exported model to
-a specific location where the Demo App will pick them up:
+Export the MobileNet v3 model using the command line with Core ML, MPS and XNNPACK backends
 
 ```bash
-python3 -m examples.portable.scripts.export --model_name="mv3"
-
-APP_PATH="executorch-examples/mv3/apple/ExecuTorchDemo/ExecuTorchDemo"
-mkdir -p "$APP_PATH/Resources/Models/MobileNet/"
-mv mv3.pte "$APP_PATH/Resources/Models/MobileNet/"
+cd ~/executorch-examples
+python3 mv3/python/export.py
 ```
 
-Next, export the MobileNet v3 model with Core ML, MPS and XNNPACK backends by runnning this [`export.py`](https://github.com/pytorch-labs/executorch-examples/mv3/apple/ExecuTorchDemo/export.py) script.
+Move the exported model to a specific location where the Demo App will pick them up:
 
-```bash
-python3 export.py
+```
+cd ~/executorch-examples
+mkdir -p mv3/apple/ExecuTorchDemo/ExecuTorchDemo/Resources/Models/MobileNet/
+mv *.pte mv3/apple/ExecuTorchDemo/ExecuTorchDemo/Resources/Models/MobileNet/
 ```
 
 
@@ -104,7 +89,7 @@ Download the MobileNet model labels required for image classification:
 
 ```bash
 curl https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt \
-  -o "$APP_PATH/Resources/Models/MobileNet/imagenet_classes.txt"
+  -o mv3/apple/ExecuTorchDemo/ExecuTorchDemo/Resources/Models/MobileNet/imagenet_classes.txt
 ```
 
 ## Final Steps
@@ -118,7 +103,7 @@ Double-click on the project file under
 `executorch-examples/mv3/apple/ExecuTorchDemo/ExecuTorchDemo` to openit with Xcode, or run the command:
 
 ```bash
-open $APP_PATH.xcodeproj
+open mv3/apple/ExecuTorchDemo/ExecuTorchDemo.xcodeproj
 ```
 
 ### 2. Run Tests

mv3/python/export.py

@@ -8,39 +8,54 @@
 import torch
 import torchvision.models as models
 from executorch.backends.apple.coreml.partition import CoreMLPartitioner
-from executorch.backends.apple.mps.partition import MPSPartitioner
-from executorch.exir.backend.backend_details import CompileSpec 
+from executorch.exir import to_edge 
 from executorch.backends.xnnpack.partition.xnnpack_partitioner import XnnpackPartitioner
 from executorch.exir import to_edge_transform_and_lower
-import ssl
-import certifi
 
 
 def main() -> None:
     model = models.mobilenet_v3_small(weights="DEFAULT").eval()
     sample_inputs = (torch.randn(1, 3, 224, 224),)
 
+    et_program_portable = to_edge(
+        torch.export.export(model, sample_inputs),
+    ).to_executorch() 
+
     et_program_coreml = to_edge_transform_and_lower(
         torch.export.export(model, sample_inputs),
         partitioner=[CoreMLPartitioner()],
     ).to_executorch()
 
-    et_program_mps = to_edge_transform_and_lower(
+    
+    # MPS backend doesn't work yet with pip install today.
+    # Currently, it is just falling back to portable ops instead.
+    # 
+    # Please install ExecuTorch from source if you want to run to MPS backend:
+    # 
+    # https://pytorch.org/executorch/main/using-executorch-building-from-source.html#environment-setup
+    # ./install_executorch.sh --pybind mps coreml xnnpack
+    # ./backends/apple/mps/install_requirements.sh
+    # 
+    # et_program_mps = to_edge_transform_and_lower(
+    #   torch.export.export(model, sample_inputs),
+    #   partitioner=[MPSPartitioner([CompileSpec("use_fp16", bytes([True]))])], 
+    # ).to_executorch()
+    et_program_mps = to_edge(
         torch.export.export(model, sample_inputs),
-        partitioner=[MPSPartitioner([CompileSpec("use_fp16", bytes([True]))])], 
     ).to_executorch()
 
     et_program_xnnpack = to_edge_transform_and_lower(
         torch.export.export(model, sample_inputs),
         partitioner=[XnnpackPartitioner()],
     ).to_executorch()
 
-    save_path = "executorch-examples/mv3/apple/ExecuTorchDemo/ExecuTorchDemo/Resources/Models/MobileNet/"
-    with open(save_path+"mv3_coreml_all.pte", "wb") as file:
+    with open("mv3.pte", "wb") as file:
+        et_program_portable.write_to_file(file)
+    with open("mv3_coreml_all.pte", "wb") as file:
         et_program_coreml.write_to_file(file)
-    with open(save_path+"mv3_mps_float16.pte", "wb") as file:
+    with open("mv3_mps_float16.pte", "wb") as file:
         et_program_mps.write_to_file(file)
-    with open(save_path+"mv3_xnnpack_fp32.pte", "wb") as file:
+    with open("mv3_xnnpack_fp32.pte", "wb") as file:
         et_program_xnnpack.write_to_file(file)