GRU of PyDirectML throw error use the parameters of the ONNX example are used.

I wrote a simple example of using GRU operator:

import pydirectml as dml
import numpy as np

steps = 2
batchSize = 3
inputSize = 3
hiddenSize = 5
numDirections = 1

input_bindings = []

def append_input_tensor(builder: dml.GraphBuilder, input_bindings: list, input_tensor: dml.TensorDesc):
    tensor = dml.input_tensor(builder, len(input_bindings), input_tensor)
    input_bindings.append(dml.Binding(tensor, np.zeros(tensor.get_output_desc().sizes)))
    return tensor

# Create a GPU device, and build a model graph.
device = dml.Device(True, True)
builder = dml.GraphBuilder(device)
data_type = dml.TensorDataType.FLOAT32
flags = dml.TensorFlags.OWNED_BY_DML

input_data = [1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15, 16, 17, 18]
input_array = np.array(input_data, np.float32)
input = dml.input_tensor(builder, 0, dml.TensorDesc(data_type, [1, steps, batchSize, inputSize]))
input_bindings.append(dml.Binding(input, input_array))
weight = append_input_tensor(builder, input_bindings, dml.TensorDesc(data_type, flags, [1, numDirections, 3 * hiddenSize, inputSize]))
recurrentWeight = append_input_tensor(builder, input_bindings, dml.TensorDesc(data_type, flags, [1, numDirections, 3 * hiddenSize, hiddenSize]))

gru = dml.gru(
  input = input,
  weight = weight,
  recurrence = recurrentWeight,
  activation_descs = [dml.FusedActivation(dml.OperatorType.ACTIVATION_SIGMOID), dml.FusedActivation(dml.OperatorType.ACTIVATION_SIGMOID)],
  direction = dml.RecurrentNetworkDirection.FORWARD,
  output_options = dml.OutputOptions.Single
)

# Compile the expression graph into a compiled operator
op = builder.build(dml.ExecutionFlags.NONE, [gru])

# Compute the result
output_data = device.compute(op, input_bindings, [gru])
output_tensor = np.array(output_data[0], np.float32)

print(output_tensor)

But it will throw the following error:

Traceback (most recent call last):
  File ".\Python\samples\gru.py", line 36, in <module>
    output_options = dml.OutputOptions.Single
RuntimeError: m_device->CreateOperator(&opDesc, IID_PPV_ARGS(&op))

Is there a problem with the example I wrote? BTW, are there any sample code for GRU usage?

[miaobin]

@miaobin Did the debug layer enlighten?

Yes, with DirectML debug layer we found two issues. We found a bug about creating the inputs of Gru. I submitted a PR to fix this and it has been merged. In addition, we note that DML_TENSOR_FLAGS of all structure DML_GRU_OPERATOR_DESC members must be set to DML_TENSOR_FLAG_NONE. I thank this issue can be closed. Thanks a lot!

[huningxin]

@miaobin, did you happen to try DirectML debug layer and find more details?