Runtime Onnx Model Error

Our latency predictor only requries structures and attributes of onnx models. The weights are not required. For saving space, the uploaded onnx models do not contain weights. You can try to fill onnx models with random weights and infer them by onnxruntime, as follows,
import onnx
import numpy as np
import onnxruntime as ort

def fill_onnx_weight(G):
    for init in G.graph.initializer:
        fill_a_tensor(init)
    for node in G.graph.node:
        # conv weight may be constant node
        if node.op_type == "Constant":
           for attr in node.attribute:
               if attr.type == onnx.AttributeProto.AttributeType.TENSOR:
                   fill_a_tensor(attr.t)

def fill_a_tensor(T, min_len=10):

    def stat_size(T):
        L = 0
        L = max(L, len(T.float_data))
        L = max(L, len(T.double_data))
        L = max(L, len(T.int32_data))
        L = max(L, len(T.int64_data))
        L = max(L, len(T.uint64_data))
        L = max(L, len(T.raw_data) // 4)
        return L

    # in case of overlapping the shape constant & initializer
    length = np.prod(T.dims)
    # real size == 0 need to fill too
    if stat_size(T) > 0:
        return

    if T.data_type == onnx.TensorProto.FLOAT:
        T.float_data[:] = np.random.rand(length) - 0.5

    elif T.data_type == onnx.TensorProto.DOUBLE:
        T.double_data[:] = np.random.rand(length) - 0.5

    elif T.data_type == onnx.TensorProto.COMPLEX64:
        T.float_data[:] = np.random.rand(length * 2) - 0.5

    elif T.data_type == onnx.TensorProto.COMPLEX128:
        T.double_data[:] = np.random.rand(length * 2) - 0.5

    elif T.data_type == onnx.TensorProto.INT64:
        T.int64_data[:] = (np.random.rand(length) * 127).astype(np.int64) - 64

    elif T.data_type == onnx.TensorProto.UINT32 or \
        T.data_type == onnx.TensorProto.UINT64:
        T.uint64_data[:] = (np.random.rand(length) * 127).astype(np.uint64)

    elif T.data_type == onnx.TensorProto.UINT8 or \
        T.data_type == onnx.TensorProto.INT8 or \
        T.data_type == onnx.TensorProto.UINT16 or \
        T.data_type == onnx.TensorProto.INT16 or \
        T.data_type == onnx.TensorProto.INT32 or \
        T.data_type == onnx.TensorProto.BOOL or \
        T.data_type == onnx.TensorProto.FLOAT16 or \
        T.data_type == onnx.TensorProto.BFLOAT16:
        T.int32_data[:] = (np.random.rand(length) * 127).astype(np.int32) - 64

    else:
        #onnx.TensorProto.UNDEFINED,
        #onnx.TensorProto.STRING,
        pass

def pass_remove_ceil_mode_for_MaxPool(G):
    for node in G.graph.node:
        if node.op_type.endswith("Pool"):
            for attr in node.attribute:
                if attr.name == "ceil_mode":
                    print("remove attirbute ceil_mode for {}: {}".format(node.op_type, node.name))
                    node.attribute.remove(attr)

if __name__ == "__main__":
    onnx_path = "dataset/multi_platform/onnx/nnmeter_alexnet/nnmeter_alexnet_transform_0122.onnx"
    G = onnx.load(onnx_path)
    fill_onnx_weight(G)
    pass_remove_ceil_mode_for_MaxPool(G)
    # onnx.save(G, onnx_path.replace(".onnx", "_new.onnx"))

    sess = ort.InferenceSession(G.SerializeToString())
    sess.set_providers(['CPUExecutionProvider'])
    input_name = sess.get_inputs()[0].name
    input_shape = sess.get_inputs()[0].shape
    output_name = sess.get_outputs()[0].name
    data = (np.random.rand(*input_shape) - 0.5).astype(np.float32)

    print("run with input {}, shape={}".format(input_name, data.shape))
    outputs = sess.run([output_name], {input_name : data})
    for idx, o in enumerate(outputs):
        print("output {} shape: {}".format(idx, o.shape))
ModelTC / NNLQP

Runtime Onnx Model Error #5
