I try to register a api to implement the gradient of relay.concatenate, Below
is the code I implemented. but i think there is missing some data.
The following code is the c++ code I implemented for concatenate grad
namespace tvm {
namespace relay {
//concatenate split
bool ConcatenateGradRel(const Array<Type>& types,
int num_inputs,
const Attrs& attrs,
const TypeReporter& reporter) {
// types = [ograd, tensor_tuple, result]
CHECK_EQ(types.size(), 3);
**DLOG(INFO) << "input type: " << types <<std::endl;**
const auto* ograd = types[0].as<TensorTypeNode>();
const auto* tensor_tuple = types[1].as<TupleTypeNode>();
DLOG(INFO) << "tensor_tuple: " << tensor_tuple->type_key() <<std::endl;
DLOG(INFO) << "tensor_tuple fields type: " <<
tensor_tuple->fields->type_key() <<std::endl;
DLOG(INFO) << "tensor_tuple size: " << tensor_tuple->fields.size()
<<std::endl;
DLOG(INFO) << "tensor_tuple 1: " << tensor_tuple->fields[0]->type_key()
<<std::endl;
DLOG(INFO) << "tensor_tuple 2: " << tensor_tuple->fields[1]->type_key()
<<std::endl;
const auto& first = Downcast<TensorType>(tensor_tuple->fields[0]);
const auto param = attrs.as<ConcatenateAttrs>();
CHECK(param != nullptr);
auto axis = param->axis;
CHECK(-ndim <= axis && axis < ndim)
<< "concatenate only accepts `axis` in [-ndim, ndim)"
<< ", but got axis = " << axis
<< ", and ndim = " << ndim;
axis = axis < 0 ? ndim + axis : axis;
// Calculate shape
reporter->Assign(types[2], types[1]);
return true;
}
Array<Tensor> ConcatenateGradCompute(const Attrs& attrs,
const Array<Tensor>& inputs,
const Type& out_type,
const Target& target) {
DLOG(INFO) << "Grad Compute" <<std::endl;
const auto* param = attrs.as<ConcatenateAttrs>();
//const auto* param = attrs.as<SplitAttrs>();
CHECK(param != nullptr);
auto axis = param->axis;
const auto test = out_type.as<TupleTypeNode>();
int64_t num_sections = 2;
return Array<Tensor>{
topi::split_sections(inputs[0], num_sections, param->axis) };
}
Expr MakeConcatenateGrad(Expr ograd,
Expr tensor_tuple,
int axis
)
{
//auto attrs = make_node<SplitAttrs>();
auto attrs = make_node<ConcatenateAttrs>();
attrs->axis = axis;
DLOG(INFO) << "tensor_tuple type: " << tensor_tuple->type_key()
<<std::endl;
const auto tensors = tensor_tuple.as<TupleTypeNode>();
static const Op& op = Op::Get("concatenate_grad");
return CallNode::make(op, {ograd, tensor_tuple}, Attrs(attrs), {});
}
TVM_REGISTER_API("relay.op._make.concatenate_grad")
.set_body_typed(MakeConcatenateGrad);
RELAY_REGISTER_OP("concatenate_grad")
.describe(R"code(concatenate grad.
)code" TVM_ADD_FILELINE)
.set_attrs_type_key("relay.attrs.ConcatenateAttrs")
.set_num_inputs(2)
.add_argument("ograd", "Tensor", "The gradient of output.")
.add_argument("tensor_tuple", "Tensor", "The input list of tensors.")
.set_support_level(1)
.add_type_rel("ConcatenateGradRel", ConcatenateGradRel)
.set_attr<FTVMCompute>("FTVMCompute", ConcatenateGradCompute)
.set_attr<TOpPattern>("TOpPattern", kInjective);
} // namespace relay
} // namespace tvm
I use relay.concatenate connected 2 matrices. one shape is 2x3 and the other is
3x3, the concatenate axis is 0, so the result is a 5x3 matrices. At the bottom
there is my test case code.
In line 13 of the code, I print out the type information, I hope i could get
[TensorType([5, 3], float32),
TupleTypeNode([TensorType([2, 3], float32), RefTypeNode(TensorType([2, 3],
float32))],
[TensorType([3, 3], float32), RefTypeNode(TensorType([3, 3],
float32))]), IncompleteTypeNode(0, 0x3b4f2f0)]
But actually what I got is
`[IncompleteTypeNode(0, 0x50d7f50),
TupleTypeNode([TensorType([2, 3], float32), RefTypeNode(TensorType([2, 3],
float32))]), IncompleteTypeNode(0, 0x516ada0)]`
i find that the type[1] value is just the first data of the concatenated
matrices, there is missing a value of the 3x3 matrices. I want to know how to
get the full value of the Tuple type.
Below is a partial error log
TVMError:
Error(s) have occurred. We have annotated the program with them:
In main:
v0.0.1
fn (%x: Tensor[(2, 3), float32], %y: Tensor[(3, 3), float32]) {
%0 = fn () -> () {
()
}
let %x1 = ref(%0)
%1 = zeros_like(%x)
%2 = ref(%1)
let %x2 = (%x, %2)
%3 = zeros_like(%y)
%4 = ref(%3)
let %x3 = (%y, %4)
%12 = fn (%x5: (Tensor[(2, 3), float32], ref(Tensor[(2, 3), float32])),
%y1: (Tensor[(3, 3), float32], ref(Tensor[(3, 3), float32]))) {
let %x6 = (%x5, %y1)
%5 = %x6.0
let %x7 = concatenate(%5)an internal invariant was violated while
typechecking your program [18:00:48]
/home/rui.huang/tvm-git/3rdparty/HalideIR/src/tvm/node/node.h:264: Check
failed: ref->template is_type<typename SubRef::ContainerType>() ||
ref->template derived_from<typename SubRef::ContainerType>(): Downcast from
relay.RefType to relay.TensorType failed.
; an internal invariant was violated while typechecking your program
[18:00:48] /home/rui.huang/tvm-git/3rdparty/HalideIR/src/tvm/node/node.h:264:
Check failed: ref->template is_type<typename SubRef::ContainerType>() ||
ref->template derived_from<typename SubRef::ContainerType>(): Downcast from
relay.RefType to relay.TensorType failed.
; an internal invariant was violated while typechecking your program
[18:00:48] /home/rui.huang/tvm-git/3rdparty/HalideIR/src/tvm/node/node.h:264:
Check failed: ref->template is_type<typename SubRef::ContainerType>() ||
ref->template derived_from<typename SubRef::ContainerType>(): Downcast from
relay.RefType to relay.TensorType failed.
; an internal invariant was violated while typechecking your program
[18:00:48] /home/rui.huang/tvm-git/3rdparty/HalideIR/src/tvm/node/node.h:264:
Check failed: ref->template is_type<typename SubRef::ContainerType>() ||
ref->template derived_from<typename SubRef::ContainerType>(): Downcast from
relay.RefType to relay.TensorType failed.
;
%6 = zeros_like(%x7)
let %x8 = ref(%6)
let %x9 = %x1^
%11 = fn () -> () {
let %x11 = %x8^
%7 = %x6.1an internal invariant was violated while typechecking your
program [18:00:48] /home/rui.huang/tvm-git/src/relay/pass/type_solver.cc:119:
Check failed: resolved.defined(): Unable to unify parent types:
TupleTypeNode([TensorType([3, 3], float32), RefTypeNode(TensorType([3, 3],
float32))]) and RefTypeNode(IncompleteTypeNode(0, 0x3a00b10))
let %x12 = %7^
%8 = %x6.1an internal invariant was violated while typechecking your
program [18:00:48] /home/rui.huang/tvm-git/src/relay/pass/type_solver.cc:119:
Check failed: resolved.defined(): Unable to unify parent types:
TupleTypeNode([TensorType([3, 3], float32), RefTypeNode(TensorType([3, 3],
float32))]) and RefTypeNode(IncompleteTypeNode(0, 0x3bc3680))
;
%9 = concatenate_grad(%x11, %5)...
In other operators, all operation parameters are single values, but in
concatenate operations, relay.concatenate uses Tuple(list([x, y, z...])) as a
parameter to pass forward operating. Concatenate's forward operation code is
located _tvm/relay/op/tensor.py( line:719 )_
The following code is the concatenate grad interface I registered.
@register_gradient("concatenate")
def concatenate_grad(orig, grad):
axis = orig.attrs.axis
data = orig.args[0]
return [_make.concatenate_grad(grad, data, axis)]
The following code is my test case
def relay_concatenate_grad(func, x, y, axis=0):
dshape = (2, 3)
oshape = (3, 3)
dtype = 'float32'
x = np.random.rand(*dshape).astype(dtype)
y = np.random.rand(*oshape).astype(dtype)
input_parameters = {'x': x, 'y': y}
intrp = relay.create_executor(ctx=tvm.context('llvm', 0), target='llvm')
x_var = relay.var('x', relay.TensorType(x.shape, 'float32'))
y_var = relay.var('y', relay.TensorType(y.shape, 'float32'))
fwd_func = relay.Function([x_var, y_var], relay.concatenate([x_var,
y_var], axis))
op_res = intrp.evaluate(fwd_func)(**input_parameters)
print("grad fun:\n", gradient(fwd_func))
bwd_func = infer_type(gradient(fwd_func))
_, grads = intrp.evaluate(bwd_func)(**input_parameters)
print("relay_ grad:\n", grads)
return grads
---
[Visit
Topic](https://discuss.tvm.ai/t/relay-concatenate-missing-data-when-development-the-gradient-of-relay-concatenate/3595/1)
to respond.
You are receiving this because you enabled mailing list mode.
To unsubscribe from these emails, [click
here](https://discuss.tvm.ai/email/unsubscribe/c16ae972300dba127f342101a08156da3ba673fbc4d280f539292bfc4eefe19d).
Tianqi Chen, UW, Seattle, WA, 98105, United States
http://tracking.discuss.tvm.ai/tracking/unsubscribe?msgid=iwr5OpZi_MfkRYrBpsRv8w2
