onnx importer backend pad reshape transpose

tools/nntool/importer/onnx/handlers/backend/pad.py

@@ -41,7 +41,7 @@ def _common(cls, node, mode='constant', pads=None, constant_value=0, **kwargs):
         x = inputs[0]
         x_shape = x[2].shape
 
-        apads = np.array(pads).reshape((-1, 2))
+        apads = np.array(pads).reshape((2,-1)).T
         if cls.is_constant(x):
             logger.info("reducing %s to a constant", valid_name)
             val = cls.get_constant(x)

tools/nntool/importer/onnx/handlers/backend/slice.py

@@ -60,7 +60,8 @@ def _common(cls, node, starts, ends, axes, steps, **kwargs):
                     if step < 0:
                         p_shape.append((begin - end)//-step)
                     else:
-                        p_shape.append((end - begin)//step)
+                        # p_shape.append((end - begin)//step)
+                        p_shape.append(int(np.ceil((end - begin)/step))) #test on yolox wrong index number
 
             except ValueError:
                 p_slices.append((0, dim, 1))

tools/nntool/importer/tflite2/handlers/backend/padv2.py

@@ -23,17 +23,15 @@
 
 @tflite_op("PADV2")
 class PadV2(PadMixin, BackendHandler):
-
+# xinglong debug use pad.py code here. can compile uint8 yolov3 tflite
     @classmethod
     def _common(cls, node: TFLiteNode, **kwargs):
 
         all_nodes = kwargs['all_nodes']
         inputs = [all_nodes[t] for t in node.input]
         pads = list(cls._verify_constant(inputs[1]))
         node.input[1].used = True
-        pad_vals = list((v, v) for v in cls._verify_constant(inputs[2]))
-        node.input[2].used = True
-        return super(PadV2, cls)._common(node, pads=pads, pad_vals=pad_vals, **kwargs)
+        return super(PadV2, cls)._common(node, pads=pads, **kwargs)
 
     @classmethod
     def version_1(cls, node: TFLiteNode, **kwargs):

tools/nntool/quantization/float/kernels/tensor_functions.py

@@ -46,7 +46,8 @@ def execute(cls, params,
                 in_shape = tuple(dim for dim in in_tensor.shape if dim > 1)
                 expected_shape = tuple(dim for dim in params.dims.shape if dim > 1)
                 if in_shape != expected_shape:
-                    raise ValueError(f'{params.name} received input of shape {in_tensor.shape} but expecting {params.dims.shape}')
+                    in_tensor = np.transpose(in_tensor, (2,0,1)) #input rgb (h,w,c)=>(c,h,w)
+                    # raise ValueError(f'{params.name} received input of shape {in_tensor.shape} but expecting {params.dims.shape}')
             in_tensor = in_tensor.reshape(params.dims.shape)
         else:
             in_tensor = resize(in_tensor, params.dims.shape)

tools/nntool/quantization/symmetric/kernels/pool.py

@@ -109,7 +109,7 @@ def execute(cls, params,
             in_tensor = np.pad(in_tensor,
                                params.padding.numpy_pad_shape(in_dims),
                                mode='constant',
-                               constant_values=qrec.in_qs[0].pad_zero_point)
+                               constant_values=qrec.in_qs[0].zero_point) #no pad_zero_point
             pad_w = params.padding.w
             pad_h = params.padding.h
         else:

tools/nntool/quantization/symmetric/kernels/tensor_functions.py

@@ -38,8 +38,9 @@ def execute(cls, params,
                 expected_shape = tuple(
                     dim for dim in params.dims.shape if dim > 1)
                 if in_shape != expected_shape:
-                    raise ValueError(
-                        f'{params.name} received input of shape {in_tensor.shape} but expecting {params.dims.shape}')
+                    in_tensor = np.transpose(in_tensor, (2,0,1)) #input rgb (h,w,c)=>(c,h,w)
+                    # raise ValueError(
+                    #     f'{params.name} received input of shape {in_tensor.shape} but expecting {params.dims.shape}')
             in_tensor = in_tensor.reshape(params.dims.shape)
         else:
             in_tensor = resize(in_tensor, params.dims.shape)