NicolaLancellotti commented on a change in pull request #9508:
URL: https://github.com/apache/tvm/pull/9508#discussion_r750174449
##########
File path: tests/python/contrib/test_ethosu/test_legalize.py
##########
@@ -221,135 +221,135 @@ def get_shape_expr(in_expr, out_expr):
return shape
+def compute_ofm_shape(ifm_shape, padding, kernel_shape, strides, dilation):
+ if padding.lower() == "valid":
+ h = math.ceil((ifm_shape[1] - (kernel_shape[0] - 1) * dilation[0]) /
strides[0])
+ w = math.ceil((ifm_shape[2] - (kernel_shape[1] - 1) * dilation[1]) /
strides[1])
+ if padding.lower() == "same":
+ h = math.ceil(ifm_shape[1] / strides[0])
+ w = math.ceil(ifm_shape[2] / strides[1])
+ ofm_shape = [ifm_shape[0], h, w, kernel_shape[3]]
+ return ofm_shape
+
+
INVERSE_LAYOUT_TRANSFORM_OHWI_MAP = {
"HWIO": [1, 2, 3, 0],
"HWOI": [1, 2, 0, 3],
"OWHI": [0, 1, 2, 3],
}
-def test_ethosu_conv2d_legalize():
- def create_graph_single(input_tensor_name, input_tensor_shape,
input_tensor_dtype):
- c1_params = relay_ir_builder.QnnConv2DParams(input_tensor_dtype)
- c1_params.ifm.shape = input_tensor_shape
- c1_params.kernel.shape = (3, 3, c1_params.ifm.shape[3], 32)
- c1_params.strides = (1, 1)
- c1_params.pad = "VALID"
- c1_params.activation = "CLIP"
- c1_params.clip_min = 23
- c1_params.clip_max = 180
- input0 = relay.var(input_tensor_name, shape=c1_params.ifm.shape,
dtype=c1_params.ifm.dtype)
- c1, new_params = relay_ir_builder.create_qnn_conv2d(c1_params, input0)
- c1_params.ofm.shape = get_shape_expr(input0, c1)
-
- f = relay.Function([input0], c1)
- mod = tvm.IRModule()
- mod["main"] = f
- return mod, [c1_params]
-
- def create_graph_double(input_tensor_name, input_tensor_shape,
input_tensor_dtype):
- c1_params = relay_ir_builder.QnnConv2DParams(input_tensor_dtype)
- c1_params.ifm.shape = input_tensor_shape
- c1_params.kernel.shape = (7, 7, c1_params.ifm.shape[3], 8)
- c1_params.strides = (2, 2)
- c1_params.pad = "VALID"
- c1_params.activation = "CLIP"
- c1_params.clip_min = 10
- c1_params.clip_max = 240
- input0 = relay.var(input_tensor_name, shape=c1_params.ifm.shape,
dtype=c1_params.ifm.dtype)
- c1, new_params = relay_ir_builder.create_qnn_conv2d(c1_params, input0)
- c1_params.ofm.shape = get_shape_expr(input0, c1)
-
- c2_params = relay_ir_builder.QnnConv2DParams(input_tensor_dtype)
- c2_params.ifm.shape = c1_params.ofm.shape
- c2_params.kernel.shape = (5, 5, c2_params.ifm.shape[3], 16)
- c2_params.strides = (1, 1)
- c2_params.pad = "SAME"
- c2, new_params = relay_ir_builder.create_qnn_conv2d(c2_params, c1)
- c2_params.ofm.shape = get_shape_expr(input0, c2)
-
- f = relay.Function([input0], c2)
- mod = tvm.IRModule()
- mod["main"] = f
- return mod, [c2_params, c1_params]
[email protected]("ifm_shape", [(1, 299, 299, 3), (1, 55, 55, 3)])
[email protected]("kernel_shape", [(3, 2, 3, 3), (1, 3, 3, 3)])
[email protected]("padding", ["SAME", "VALID"])
[email protected]("strides, dilation", [((1, 1), (2, 1)), ((3, 2), (1,
1))])
[email protected]("activation", [None, None])
Review comment:
I think Elen meant "RELU", that is a string. By the way, I have just
noted that we model the no activation in different ways, the pooling and binary
elementwise tests use the string "NONE" and the depthwise conv2d, and this
patch use the None value. For me, it is the same using the former or the
latter, but I think we should uniform the tests.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
