Merge branch 'develop' of github.com:PaddlePaddle/GraphNet into device_rewrite_sample_pass

Author: lixinqi

graph_net/test/graph_variable_rename_test.sh

@@ -2,6 +2,7 @@
 
 GRAPH_NET_ROOT=$(python3 -c "import graph_net; import os; print(
 os.path.dirname(graph_net.__file__))")
+WORKSPACE=/tmp/graph_variable_rename_workspace
 
 # input model path
 MODEL_NAME=resnet18
@@ -16,7 +17,7 @@ config_json_str=$(cat <<EOF
         "data_input_predicator_class_name": "NaiveDataInputPredicator",
         "model_runnable_predicator_filepath": "$GRAPH_NET_ROOT/torch/constraint_util.py",
         "model_runnable_predicator_class_name": "ModelRunnablePredicator",
-        "output_dir": "/tmp/graph_variable_rename_workspace"
+        "output_dir": "$WORKSPACE"
     }
 }
 EOF
@@ -25,3 +26,23 @@ CONFIG=$(echo $config_json_str | base64 -w 0)
 
 python3 -m graph_net.model_path_handler --model-path samples/$MODEL_PATH_IN_SAMPLES --handler-config=$CONFIG
 # python3 -m graph_net.model_path_handler --model-path-list $GRAPH_NET_ROOT/config/decomposition_error_tmp_torch_samples_list.txt --handler-config=$CONFIG
+
+test_compiler_config_json_str=$(cat <<EOF
+{
+    "model_path_prefix": "$GRAPH_NET_ROOT",
+    "renamed_root": "$WORKSPACE"
+}
+EOF
+)
+TEST_COMPILER_CONFIG=$(echo $test_compiler_config_json_str | base64 -w 0)
+
+python3 -m graph_net.torch.test_compiler \
+    --model-path $GRAPH_NET_ROOT/../samples/$MODEL_PATH_IN_SAMPLES \
+    --compiler graph_variable_renamer_validator \
+    --device cuda \
+    --config $TEST_COMPILER_CONFIG \
+    > "$WORKSPACE/validation.log" 2>&1
+
+python3 -m graph_net.plot_ESt \
+    --benchmark-path "$WORKSPACE/validation.log" \
+    --output-dir "$WORKSPACE"

graph_net/torch/backend/graph_variable_renamer_validator_backend.py

@@ -0,0 +1,90 @@
+import torch
+from pathlib import Path
+from typing import Dict
+from graph_net.tensor_meta import TensorMeta
+import os
+import importlib.util
+
+
+class RenamedModelAdapter(torch.nn.Module):
+    def __init__(self, renamed_model: torch.nn.Module, mapping: Dict[str, str]):
+        super().__init__()
+        self.model = renamed_model
+        self.mapping = mapping
+        if hasattr(renamed_model, "__graph_net_file_path__"):
+            self.__graph_net_file_path__ = renamed_model.__graph_net_file_path__
+
+    def forward(self, **kwargs):
+        new_kwargs = self._convert_by_name_mapping(kwargs)
+        return self.model(**new_kwargs)
+
+    def _convert_by_name_mapping(self, kwargs):
+        new_kwargs = {}
+        for old_name, value in kwargs.items():
+            if old_name in self.mapping:
+                new_name = self.mapping[old_name]
+                new_kwargs[new_name] = value
+        return new_kwargs
+
+
+class GraphVariableRenamerValidatorBackend:
+    def _get_rename_mapping(self, model_dir: Path):
+        mapping = {}
+        for meta_file in ["input_meta.py", "weight_meta.py"]:
+            meta_path = model_dir / meta_file
+            if not meta_path.exists():
+                continue
+            metas = TensorMeta.unserialize_from_py_file(str(meta_path))
+            for m in metas:
+                if m.original_name:
+                    mapping[m.original_name] = m.name
+        return mapping
+
+    def _load_model_instance(self, path: str, device: str) -> torch.nn.Module:
+        class_name = "GraphModule"
+        model_file = os.path.join(path, "model.py")
+
+        spec = importlib.util.spec_from_file_location(class_name, model_file)
+        module = importlib.util.module_from_spec(spec)
+        spec.loader.exec_module(module)
+
+        ModelClass = getattr(module, class_name)
+        instance = ModelClass().to(device)
+        return instance
+
+    def _make_config(
+        self,
+        model_path_prefix: str,
+        renamed_root: str,
+        renamed_dentry: str = "_renamed",
+    ):
+        return {
+            "model_path_prefix": model_path_prefix,
+            "renamed_root": renamed_root,
+            "renamed_dentry": renamed_dentry,
+        }
+
+    def __call__(self, model: torch.nn.Module) -> torch.nn.Module:
+        config = self._make_config(**self.config)
+        model_path = os.path.dirname(model.__class__.__graph_net_file_path__)
+        model_name = os.path.basename(model_path)
+        renamed_dir_name = f"{model_name}_renamed"
+        renamed_model_dir = os.path.join(config["renamed_root"], renamed_dir_name)
+
+        print(f"[GraphVariableRenamerValidatorBackend] Processing: {model_name}")
+        print(
+            f"[GraphVariableRenamerValidatorBackend] Loading from: {renamed_model_dir}"
+        )
+
+        device = model.__class__.__graph_net_device__
+        renamed_model = self._load_model_instance(renamed_model_dir, device)
+        mapping = self._get_rename_mapping(Path(renamed_model_dir))
+        assert (
+            mapping
+        ), f"Mapping is empty for {renamed_dir_name} at {renamed_model_dir}"
+        adapter = RenamedModelAdapter(renamed_model, mapping)
+        return adapter.eval()
+
+    def synchronize(self):
+        if torch.cuda.is_available():
+            torch.cuda.synchronize()

graph_net/torch/graph_variable_renamer.py

@@ -79,8 +79,10 @@ def __call__(self, rel_model_path):
         module, inputs = get_torch_module_and_inputs(src_model_path)
         gm = parse_sole_graph_module(module, inputs)
         gm = self.rename_graph_variables(gm, inputs, src_model_path)
+        model_name = os.path.basename(rel_model_path.rstrip(os.sep))
+        new_rel_path = f"{model_name}_renamed"
         dst_model_path = os.path.realpath(
-            os.path.join(self.config["output_dir"], rel_model_path)
+            os.path.join(self.config["output_dir"], new_rel_path)
         )
         Path(dst_model_path).parent.mkdir(parents=True, exist_ok=True)
         shutil.copytree(src_model_path, dst_model_path, dirs_exist_ok=True)
@@ -158,45 +160,47 @@ def _get_input_names_from_signature(self, module):
     def rename_graph_variables(
         self, gm: torch.fx.GraphModule, sample_inputs, model_path
     ):
-        in_cnt = 0
-        w_cnt = 0
-        tmp_cnt = 0
-
-        arg_iter = iter(sample_inputs)
+        counters = {"in": 0, "w": 0, "tmp": 0}
+        # graph may not have input, only contain weights
+        arg_iter = iter(sample_inputs) if sample_inputs else iter([])
         for node in gm.graph.nodes:
-            if "original_name" not in node.meta:
-                node.meta["original_name"] = node.name
-
-            if node.op == "placeholder":
-                real_arg = next(arg_iter)
-                is_weight = not self.data_input_predicator(model_path, node.name)
-                if node.type is not None:
-                    if isinstance(node.type, type) and issubclass(
-                        node.type, torch.nn.parameter.Parameter
-                    ):
-                        is_weight = True
-                elif real_arg is not None:
-                    if isinstance(real_arg, torch.nn.Parameter):
-                        is_weight = True
-
-                if is_weight:
-                    new_name = f"w_{w_cnt}"
-                    w_cnt += 1
-                else:
-                    new_name = f"in_{in_cnt}"
-                    in_cnt += 1
-
-                node.name = new_name
-                node.target = new_name
-
-            elif node.op == "get_attr":
-                node.name = f"w_{w_cnt}"
-                w_cnt += 1
-
-            elif node.op != "output":
-                node.name = f"tmp_{tmp_cnt}"
-                tmp_cnt += 1
-
+            self._process_single_node(node, arg_iter, counters, model_path)
         gm.graph.lint()
         gm.recompile()
         return gm
+
+    def _process_single_node(self, node, arg_iter, counters, model_path):
+        if "original_name" not in node.meta:
+            node.meta["original_name"] = node.name
+        if node.op == "placeholder":
+            self._handle_placeholder(node, arg_iter, counters, model_path)
+        elif node.op == "get_attr":
+            self._apply_rename(node, "w", counters)
+        elif node.op != "output":
+            self._apply_rename(node, "tmp", counters)
+        else:
+            # Do nothing
+            pass
+
+    def _handle_placeholder(self, node, arg_iter, counters, model_path):
+        real_arg = next(arg_iter, None)
+        is_weight = self._is_weight_node(node, real_arg, model_path)
+        prefix = "w" if is_weight else "in"
+        self._apply_rename(node, prefix, counters, update_target=True)
+
+    def _apply_rename(self, node, prefix, counters, update_target=False):
+        new_name = f"{prefix}_{counters[prefix]}"
+        counters[prefix] += 1
+        node.name = new_name
+        if update_target:
+            node.target = new_name
+
+    def _is_weight_node(self, node, real_arg, model_path):
+        is_not_data_input = not self.data_input_predicator(model_path, node.name)
+        is_parameter_type = (
+            node.type is not None
+            and isinstance(node.type, type)
+            and issubclass(node.type, torch.nn.parameter.Parameter)
+        )
+        is_parameter_value = isinstance(real_arg, torch.nn.Parameter)
+        return is_not_data_input or is_parameter_type or is_parameter_value

graph_net/torch/test_compiler.py

@@ -25,6 +25,9 @@
 from graph_net.torch.backend.range_decomposer_validator_backend import (
     RangeDecomposerValidatorBackend,
 )
+from graph_net.torch.backend.graph_variable_renamer_validator_backend import (
+    GraphVariableRenamerValidatorBackend,
+)
 from graph_net import test_compiler_util
 from graph_net import path_utils
 
@@ -38,6 +41,7 @@
     "nope": NopeBackend(),
     "unstable_to_stable": UnstableToStableBackend(),
     "range_decomposer_validator": RangeDecomposerValidatorBackend(),
+    "graph_variable_renamer_validator": GraphVariableRenamerValidatorBackend(),
 }