Restructure ingress/torch-mlir

Signed-off-by: dchigarev <dmitry.chigarev@intel.com>

Author: dchigarev

.gitignore

@@ -0,0 +1,2 @@
+__pycache__
+ingress/Torch-MLIR/examples/**/dumps/*.mlir

ingress/Torch-MLIR/README.md

@@ -0,0 +1,49 @@
+Using scripts in this directory one can convert a Torch Model to a MLIR module.
+
+The conversion script is written in python and is basically a wrapper around [`torch-mlir` library](https://github.com/llvm/torch-mlir). One need to setup a python virtual environment with torch and torch-mlir libraries
+(`./scripts/install-virtualenv.sh`) to use the script.
+
+In order to convert a model the script has to recieve:
+1. An instance of `torch.nn.Model` with proper state (weights).
+2. Sample input arguments to the model (e.g. empty tensor with proper shape and dtype).
+
+There are two options of how this info can be provided to the converter:
+
+### 1. Instantiate a model in your own script and use a function from the `py_src/export_lib` (recomended)
+
+In this scenario a user is responsible for instantiating a model with proper state in their
+own python script. Then they should import a `generate_mlir` function from `py_src.export_lib`
+and call it in order to get a MLIR module:
+
+```python
+model : nn.Model = get_model()
+sample_args = (get_sample_tensor(),)
+
+# PYTHONPATH=$(pwd)/py_src/
+from export_lib import generate_mlir
+
+mlir_module = generate_mlir(model, sample_args, dialect="linalg")
+print(mlir_module)
+```
+
+### 2. Use `py_src/main.py` or `scripts/generate-mlir.sh` and pass Torch Model parameters via CLI
+
+In this scenario the `py_src/main.py` script is fully responsible for instantiating a torch model
+and converting it to MLIR. User has to pass a proper python entrypoint for model's factory,
+its parameters if needed (`--model-args & --model-kwargs`), and sample model arguments (either
+as `--sample-shapes` or as an entrypoint to a function returning args and kwargs `--sample-fn`).
+
+```
+# note that 'my_module' has to be in $PYTHONPATH
+python py_src/main.py --model-entrypoint my_module:my_factory   \
+                      --module-state-path path/to/state.pth     \
+                      --sample-shapes '1,2,324,float32'         \
+                      --out-mlir res.mlir
+
+# note that 'my_module' has to be in $PYTHONPATH
+./scripts/generate-mlir.sh --model-entrypoint torchvision.models:resnet18      \ 
+                           --sample-fn my_module:generate_resnet18_sample_args \
+                           --out-mlir res.mlir
+```
+
+Look into `examples/` folder for more info.

ingress/Torch-MLIR/examples/dummy_mlp_cli/dummy_mlp.pth

@@ -0,0 +1,23 @@
+import torch
+import torch.nn as nn
+
+import os
+
+class DummyMLP(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Linear(10, 32),
+            nn.ReLU(),
+            nn.Linear(32, 2)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+def make_dummy_mlp():
+    return DummyMLP()
+
+if __name__ == "__main__":
+    script_dir = os.path.dirname(os.path.abspath(__file__))
+    torch.save(make_dummy_mlp().state_dict(), os.path.join(script_dir, "dummy_mlp.pth"))

ingress/Torch-MLIR/examples/dummy_mlp_cli/dummy_mlp_factory.py

@@ -0,0 +1,10 @@
+#!/usr/bin/env bash
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+ROOT_DIR=$SCRIPT_DIR/../../scripts/
+
+PYTHONPATH=$PYTHONPATH:$SCRIPT_DIR $ROOT_DIR/generate-mlir.sh --model-entrypoint dummy_mlp_factory:make_dummy_mlp \
+  --model-state-path $SCRIPT_DIR/dummy_mlp.pth \
+  --sample-shapes "1,10,float32" \
+  --dialect linalg \
+  --out-mlir $SCRIPT_DIR/dummy_mlp_sh.mlir

ingress/Torch-MLIR/examples/dummy_mlp_cli/export_bash.sh

@@ -0,0 +1,10 @@
+#!/usr/bin/env bash
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+ROOT_DIR=$SCRIPT_DIR/../../py_src/
+
+PYTHONPATH=$PYTHONPATH:$ROOT_DIR:$SCRIPT_DIR python $ROOT_DIR/main.py --model-entrypoint dummy_mlp_factory:make_dummy_mlp \
+  --model-state-path $SCRIPT_DIR/dummy_mlp.pth \
+  --sample-shapes "1,10,float32" \
+  --dialect linalg \
+  --out-mlir $SCRIPT_DIR/dummy_mlp.mlir

ingress/Torch-MLIR/examples/dummy_mlp_cli/export_py.sh

@@ -0,0 +1,25 @@
+import torch
+import torch.nn as nn
+
+from export_lib.export import generate_mlir
+
+class DummyMLP(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Linear(10, 32),
+            nn.ReLU(),
+            nn.Linear(32, 2)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+def main():
+    model = DummyMLP()
+    dummy_input = torch.randn(1, 10)
+    mlir_mod = generate_mlir(model, (dummy_input,), {})
+    print(mlir_mod)
+
+if __name__ == "__main__":
+    main()

ingress/Torch-MLIR/examples/dummy_mlp_python/export.py

@@ -0,0 +1,4 @@
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+ROOT_DIR=$SCRIPT_DIR/../../py_src/
+
+PYTHONPATH=$ROOT_DIR python $SCRIPT_DIR/export.py

ingress/Torch-MLIR/examples/dummy_mlp_python/run.sh

@@ -0,0 +1,9 @@
+#!/usr/bin/env bash
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+ROOT_DIR=$SCRIPT_DIR/../../scripts/
+
+$ROOT_DIR/generate-mlir.sh --model-entrypoint torchvision.models:resnet18 \
+  --sample-shapes "1,3,224,224,float32" \
+  --dialect linalg \
+  --out-mlir $SCRIPT_DIR/resnet_18_sh.mlir

ingress/Torch-MLIR/examples/resnet_18_cli/export_bash.sh

@@ -0,0 +1,9 @@
+#!/usr/bin/env bash
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+ROOT_DIR=$SCRIPT_DIR/../../py_src/
+
+python $ROOT_DIR/main.py --model-entrypoint torchvision.models:resnet18 \
+  --sample-shapes "1,3,224,224,float32" \
+  --dialect linalg \
+  --out-mlir $SCRIPT_DIR/resnet_18.mlir