Upload dnmetis testtool for NPU inference

dnmetis/backend/backend.py

@@ -0,0 +1,26 @@
+"""
+abstract backend class
+"""
+
+
+# pylint: disable=unused-argument,missing-docstring
+
+class Backend():
+    def __init__(self):
+        self.inputs = []
+        self.outputs = []
+
+    def version(self):
+        raise NotImplementedError("Backend:version")
+
+    def name(self):
+        raise NotImplementedError("Backend:name")
+
+    def load(self, args):
+        raise NotImplementedError("Backend:load")
+
+    def predict(self, feed):
+        raise NotImplementedError("Backend:predict")
+
+    def get_predict_time(self):
+        return 0

dnmetis/backend/backend_acl.py

@@ -0,0 +1,64 @@
+"""
+acl backend
+"""
+
+
+import dnmetis_backend as dnmetis_backend
+import backend.backend as backend
+import numpy as np
+import os
+import pdb
+
+class AclBackend(backend.Backend):
+    def __init__(self):
+        super(AclBackend, self).__init__()
+        self.ACL=5
+        self.outputs = ""
+        self.inputs = ""
+        self.model_path = ""
+        self.cfg_path = ""
+
+    def version(self):
+        return "1.0"
+
+    def name(self):
+        return "AclBackend"
+
+    def image_format(self):
+        # By default tensorflow uses NHWC (and the cpu implementation only does NHWC)
+        return "NHWC"
+
+    def load(self, args):
+        # there is no input/output meta data i the graph so it need to come from config.
+        if not args.inputs:
+            raise ValueError("AclBackend needs inputs")
+        if not args.outputs:
+            raise ValueError("AclBackend needs outputs")
+        self.outputs = args.outputs
+        self.inputs = args.inputs
+        self.model_path = args.model
+        self.cfg_path = args.cfg_path
+        #s.path.join(args.pwd, 'backend_cfg/built-in_config.txt')
+        dnmetis_backend.backend_setconfig(self.cfg_path)
+        dnmetis_backend.backend_load(self.ACL,self.model_path,"")
+        return self
+
+    def predict(self, feed):
+        #fed=feed[self.inputs[0]]
+        result_list=[]
+        result = dnmetis_backend.backend_predict(self.ACL,self.model_path,feed)
+
+        for _ in range(len(self.outputs)):
+            #resnet50 tf & caffe
+            if 'softmax_tensor' in self.outputs[_] or 'prob' in self.outputs[_]:
+                result_list.append(np.argmax(result[_],1))
+            # resnet50 tf
+            if 'ArgMax' in self.outputs[_]:
+                result_list.append(result[_])
+        if result_list == []:
+            # ssd-resnet34 tf
+            result_list = result
+        return result_list
+
+    def unload(self):
+        return dnmetis_backend.backend_unload(self.ACL,self.model_path,"")

dnmetis/backend_cfg/built-in_config.txt

@@ -0,0 +1,2 @@
+aclrtMemMallocPolicy=2
+backend_loglevel=3

dnmetis/datasets/imagenet_10/val_map.txt

@@ -0,0 +1,10 @@
+ILSVRC2012_val_00000001.JPEG 65
+ILSVRC2012_val_00000002.JPEG 970
+ILSVRC2012_val_00000003.JPEG 230
+ILSVRC2012_val_00000004.JPEG 809
+ILSVRC2012_val_00000005.JPEG 516
+ILSVRC2012_val_00000006.JPEG 57
+ILSVRC2012_val_00000007.JPEG 334
+ILSVRC2012_val_00000008.JPEG 415
+ILSVRC2012_val_00000009.JPEG 674
+ILSVRC2012_val_00000010.JPEG 332