merge develop

Author: lixinqi

graph_net/test/dev_model_list/get_fusible_subgraph_sample_list.txt

@@ -0,0 +1,100 @@
+samples/timm/crossvit_small_240.in1k
+samples/timm/poolformerv2_s12.sail_in1k
+samples/timm/regnety_080.pycls_in1k
+samples/timm/dla46x_c.in1k
+samples/timm/mobilenetv1_100.ra4_e3600_r224_in1k
+samples/timm/efficientnetv2_rw_s.ra2_in1k
+samples/timm/vit_base_patch16_rope_ape_224.naver_in1k
+samples/timm/fastvit_t8.apple_dist_in1k
+samples/timm/test_byobnet.r160_in1k
+samples/timm/mambaout_base.in1k
+samples/timm/davit_small
+samples/timm/resnet61q.ra2_in1k
+samples/timm/coat_tiny
+samples/timm/regnetx_004.pycls_in1k
+samples/timm/convnextv2_large.fcmae
+samples/timm/regnety_640.seer
+samples/timm/repvit_m1_1.dist_300e_in1k
+samples/timm/tinynet_d.in1k
+samples/timm/resnetrs270.tf_in1k
+samples/timm/cait_m48_448
+samples/timm/legacy_seresnet50.in1k
+samples/timm/tinynet_a.in1k
+samples/timm/convnext_small.fb_in1k
+samples/timm/vit_huge_patch14_clip_quickgelu_224.dfn5b
+samples/timm/dpn131.mx_in1k
+samples/timm/convnextv2_large.fcmae_ft_in1k
+samples/timm/convnextv2_small
+samples/timm/repvit_m1.dist_in1k
+samples/timm/cs3darknet_s
+samples/timm/resnet50d.a1_in1k
+samples/timm/dm_nfnet_f6
+samples/timm/coatnet_1_rw_224
+samples/timm/lcnet_050.ra2_in1k
+samples/timm/efficientnet_em.ra2_in1k
+samples/timm/dpn48b
+samples/timm/semnasnet_075.rmsp_in1k
+samples/timm/skresnet34.ra_in1k
+samples/timm/crossvit_15_dagger_240.in1k
+samples/timm/mnasnet_100.rmsp_in1k
+samples/timm/mobilenetv3_rw.rmsp_in1k
+samples/timm/xception65p.ra3_in1k
+samples/timm/coatnet_0_rw_224
+samples/timm/eca_nfnet_l3
+samples/timm/deit3_base_patch16_224.fb_in1k
+samples/timm/mambaout_base_short_rw.sw_e500_in1k
+samples/timm/mobilenetv4_conv_small.e1200_r224_in1k
+samples/timm/xception71.tf_in1k
+samples/timm/dla60.in1k
+samples/timm/repghostnet_130.in1k
+samples/timm/mambaout_base_plus_rw.sw_e150_in12k
+samples/timm/poolformerv2_s36.sail_in1k
+samples/timm/deit3_huge_patch14_224.fb_in1k
+samples/timm/vit_base_patch32_clip_224.datacompxl
+samples/timm/poolformer_m48.sail_in1k
+samples/timm/regnety_006.pycls_in1k
+samples/timm/starnet_s4.in1k
+samples/timm/poolformer_m36.sail_in1k
+samples/timm/vit_huge_patch14_gap_224.in1k_ijepa
+samples/timm/efficientnet_b3.ra2_in1k
+samples/timm/mobilenetv3_large_150d.ra4_e3600_r256_in1k
+samples/timm/hgnetv2_b0.ssld_stage1_in22k_in1k
+samples/timm/convnextv2_huge.fcmae
+samples/timm/davit_huge
+samples/timm/regnetx_004_tv.tv2_in1k
+samples/timm/dla34.in1k
+samples/timm/convnext_xlarge.fb_in22k
+samples/timm/resmlp_12_224.fb_dino
+samples/timm/fasternet_t1.in1k
+samples/timm/resnetblur50.bt_in1k
+samples/timm/res2net50d.in1k
+samples/timm/vit_base_patch32_224.augreg_in1k
+samples/timm/mambaout_base_wide_rw.sw_e500_in1k
+samples/timm/vgg19_bn.tv_in1k
+samples/timm/vit_small_patch16_rope_ape_224.naver_in1k
+samples/timm/hardcorenas_b.miil_green_in1k
+samples/timm/vgg16.tv_in1k
+samples/timm/xception41p.ra3_in1k
+samples/timm/efficientnet_lite0.ra_in1k
+samples/timm/regnetv_064.ra3_in1k
+samples/timm/regnety_320.pycls_in1k
+samples/timm/convnext_pico.d1_in1k
+samples/timm/repvit_m1_0.dist_300e_in1k
+samples/timm/resnet50c.gluon_in1k
+samples/timm/mobileone_s4.apple_in1k
+samples/timm/ghostnet_100.in1k
+samples/timm/deit_base_distilled_patch16_384
+samples/timm/dpn68b.mx_in1k
+samples/timm/dla60_res2next
+samples/timm/resnet101d.gluon_in1k
+samples/timm/eva02_large_patch14_clip_224.merged2b
+samples/timm/fasternet_m.in1k
+samples/timm/mobilenetv2_110d.ra_in1k
+samples/timm/regnetx_064.pycls_in1k
+samples/timm/cspresnet50.ra_in1k
+samples/timm/resmlp_24_224.fb_dino
+samples/timm/mobileone_s3.apple_in1k
+samples/timm/mobileone_s2.apple_in1k
+samples/timm/res2net101d
+samples/timm/hardcorenas_f.miil_green_in1k
+samples/timm/hrnet_w18_ssld.paddle_in1k

graph_net/test/dev_model_list/small_sample_list_for_get_fusible_subgraph.txt

@@ -0,0 +1,10 @@
+#samples/timm/crossvit_small_240.in1k
+#samples/timm/poolformerv2_s12.sail_in1k
+#samples/timm/regnety_080.pycls_in1k
+#samples/timm/dla46x_c.in1k
+#samples/timm/mobilenetv1_100.ra4_e3600_r224_in1k
+samples/timm/efficientnetv2_rw_s.ra2_in1k
+#samples/timm/vit_base_patch16_rope_ape_224.naver_in1k
+#samples/timm/fastvit_t8.apple_dist_in1k
+#samples/timm/test_byobnet.r160_in1k
+#samples/timm/mambaout_base.in1k

graph_net/test/fully_fusible_subgraph_extractor_test.sh

@@ -6,8 +6,8 @@ os.path.dirname(graph_net.__file__))")
 # input model path
 MODEL_NAME=resnet18
 MODEL_PATH_IN_SAMPLES=/timm/$MODEL_NAME
-# INPUT_MODEL_LIST=$GRAPH_NET_ROOT/config/get_fusible_subgraph_sample_list.txt
-INPUT_MODEL_LIST=$GRAPH_NET_ROOT/config/small_sample_list_for_get_fusible_subgraph.txt
+# INPUT_MODEL_LIST=$GRAPH_NET_ROOT/test/dev_model_list/get_fusible_subgraph_sample_list.txt
+INPUT_MODEL_LIST=$GRAPH_NET_ROOT/test/dev_model_list/small_sample_list_for_get_fusible_subgraph.txt
 
 OUTPUT_DIR="/tmp/find_fully_fusible_output"
 config_json_str=$(cat <<EOF

graph_net/torch/fully_fusible_subgraph_extractor.py

@@ -114,9 +114,7 @@ def __call__(self, rel_model_path):
             model_path
         )
         for start_pos, end_pos in self._get_sub_ranges():
-            logger.warning("fully_fusible_subgraph_predicator-begin")
             success = fully_fusible_subgraph_predicator(start_pos, end_pos)
-            logger.warning("fully_fusible_subgraph_predicator-end")
             if not success:
                 continue
             with tempfile.TemporaryDirectory(
@@ -126,9 +124,7 @@ def __call__(self, rel_model_path):
                     temp_dir, start_pos, end_pos
                 )
                 naive_graph_decomposer = NaiveDecomposerExtractor(decomposer_config)
-                logger.warning("naive_graph_decomposer-begin")
                 naive_graph_decomposer(rel_model_path)
-                logger.warning("naive_graph_decomposer-end")
                 fully_fusible_destination_path = self._copy_from_tmp_dir_to_output_dir(
                     temp_dir, rel_model_path
                 )

graph_net/torch/graph_decomposer.py

@@ -48,13 +48,8 @@ def make_config(
         output_dir="./tmp/naive_decomposer_dir",
         filter_path=None,
         filter_config=None,
-        post_extract_process_path=None,
-        post_extract_process_class_name=None,
-        post_extract_process_config=None,
         **kwargs,
     ):
-        if post_extract_process_config is None:
-            post_extract_process_config = {}
         for pos in split_positions:
             assert isinstance(
                 pos, int
@@ -66,9 +61,6 @@ def make_config(
             "output_dir": output_dir,
             "filter_path": filter_path,
             "filter_config": filter_config if filter_config is not None else {},
-            "post_extract_process_path": post_extract_process_path,
-            "post_extract_process_class_name": post_extract_process_class_name,
-            "post_extract_process_config": post_extract_process_config,
         }
 
     def __call__(self, gm: torch.fx.GraphModule, sample_inputs):
@@ -111,14 +103,9 @@ def _make_config(
         chain_style=False,
         filter_path=None,
         filter_config=None,
-        post_extract_process_path=None,
-        post_extract_process_class_name=None,
-        post_extract_process_config=None,
         model_path_prefix="",
         **kwargs,
     ):
-        if post_extract_process_config is None:
-            post_extract_process_config = {}
         for pos in split_positions:
             assert isinstance(
                 pos, int
@@ -130,9 +117,6 @@ def _make_config(
             "output_dir": output_dir,
             "filter_path": filter_path,
             "filter_config": filter_config if filter_config is not None else {},
-            "post_extract_process_path": post_extract_process_path,
-            "post_extract_process_class_name": post_extract_process_class_name,
-            "post_extract_process_config": post_extract_process_config,
             "model_path_prefix": model_path_prefix,
         }
 
@@ -146,15 +130,16 @@ def __call__(self, rel_model_path):
         module, inputs = get_torch_module_and_inputs(model_path)
         gm = parse_immutable_model_path_into_sole_graph_module(model_path)
         try:
-            logger.warning("convert_to_submodules_graph-call-begin")
+            # logger.warning("convert_to_submodules_graph-call-begin")
             rewrited_gm: torch.fx.GraphModule = convert_to_submodules_graph(
                 gm,
                 submodule_hook=self.get_naive_decomposer_extractor(model_path),
                 **config,
             )
             rewrited_gm(*inputs)
         finally:
-            logger.warning("convert_to_submodules_graph-call-end")
+            pass
+            # logger.warning("convert_to_submodules_graph-call-end")
 
     def get_naive_decomposer_extractor(self, model_path):
         def fn(submodule, seq_no):
@@ -186,9 +171,6 @@ def _make_config(
         output_dir="./tmp/naive_decomposer_dir",
         filter_path=None,
         filter_config=None,
-        post_extract_process_path=None,
-        post_extract_process_class_name=None,
-        post_extract_process_config=None,
         model_path_prefix="",
         **kwargs,
     ):
@@ -198,18 +180,13 @@ def _make_config(
             raise ValueError(
                 f"split_results_path should be a valid JSON file path, but got {split_results_path=}"
             )
-        if post_extract_process_config is None:
-            post_extract_process_config = {}
         return {
             "split_results_path": split_results_path,
             "group_head_and_tail": group_head_and_tail,
             "chain_style": chain_style,
             "output_dir": output_dir,
             "filter_path": filter_path,
             "filter_config": filter_config if filter_config is not None else {},
-            "post_extract_process_path": post_extract_process_path,
-            "post_extract_process_class_name": post_extract_process_class_name,
-            "post_extract_process_config": post_extract_process_config,
             "model_path_prefix": model_path_prefix,
         }
 
@@ -274,7 +251,13 @@ def __init__(
             ),
         )
         self.filter = self.make_filter(self.config)
-        self.post_extract_process = self.make_post_extract_process(self.config)
+
+    def _get_model_path(self):
+        return os.path.join(
+            self.config["output_dir"],
+            f"{self.parent_graph_name}_decomposed",
+            self.model_name,
+        )
 
     def _get_model_path(self):
         return os.path.join(
@@ -290,27 +273,15 @@ def forward(self, *args):
                 self.builtin_extractor(self.submodule, args)
                 self._post_extract_process()
             self.extracted = True
-        logger.warning("naive decomposer end")
         return self.submodule(*args)
 
     def need_extract(self, gm, sample_inputs):
         if self.filter is None:
             return True
         return self.filter(gm, sample_inputs)
 
-    def _post_extract_process(self):
-        model_path = self._get_model_path()
-        return self.post_extract_process(model_path)
-
     def make_filter(self, config):
         if config["filter_path"] is None:
             return None
         module = imp_util.load_module(config["filter_path"])
         return module.GraphFilter(config["filter_config"])
-
-    def make_post_extract_process(self, config):
-        if config.get("post_extract_process_path") is None:
-            return lambda *args, **kwargs: None
-        module = imp_util.load_module(config["post_extract_process_path"])
-        cls = getattr(module, config["post_extract_process_class_name"])
-        return cls(config["post_extract_process_config"])