resnet/dssm/roformer修改onnx节点

目录

resnet101 修改内部节点

dssm_part 修改内部节点

修改前：

修改后：

其他解决方案：

resnet101 切图

roformer 输入修改

切图

---

resnet101 修改内部节点

文件链接：https://1drv.ms/u/s!ArOkNLeDa1G_iivsT-9z0uHyzVtH?e=bbb4gD

运行测试文件出现 Segmentation fault (core dumped) 错误，查看 relay 发现是 dyn.nn.pad 导致动态 shape 所致

手动修改 onnx 模型，将 pad 和 下面的 max_pool2d 合并，问题解决

model = onnx.load("resnet101_keras.onnx") graph = model.graph node = graph.node """ op_type (string): The name of the operator to construct inputs (list of string): list of input names outputs (list of string): list of output names name (string, default None): optional unique identifier for NodeProto doc_string (string, default None): optional documentation string for NodeProto domain (string, default None): optional domain for NodeProto.                 If it's None, we will just use default domain (which is empty) **kwargs (dict): the attributes of the node.  The acceptable values                 are documented in :func:`make_attribute`. """ new_maxPool_node = onnx.helper.make_node(     op_type="MaxPool",     inputs=["relu0/Relu:0"],     outputs=["zero_padding2d_2/Pad:0_pooling0"],     strides=[2,2],     kernel_shape=[3,3],     pads=[1,1,1,1],     ) graph.node.remove(old_pad_node) graph.node.remove(old_maxPool_node) graph.node.insert(4, new_maxPool_node) onnx.checker.check_model(model) onnx.save(model, "resnet101_keras_modified.onnx")

dssm_part 修改内部节点

文件链接：https://1drv.ms/u/s!ArOkNLeDa1G_iixncRNKGJJuMc67?e=W05D6D

报错 Check failed: (false) is false: relay.concatenate requires all tensors have the same dtype，查看 relay 发现 dyn.tile 产生了动态 shape

修改 onnx 模型，将 Initializer 类型常量改为 Constant 类型图节点，问题解决

修改前：

修改后：

import onnx model = onnx.load("./test_dssm_part.onnx") graph = model.graph node = graph.node # initializer = graph.initializer # print("============ initializer[1] ============") # print(initializer[1]) # print("============ initializer[2]============") # print(initializer[2])    # old_const_fold_opt__23 = initializer[1] # old_const_fold_opt__22 = initializer[2] # """ # make_tensor(name, data_type, dims, vals, raw=False) #     Make a TensorProto with specified arguments.  If raw is False, this #     function will choose the corresponding proto field to store the #     values based on data_type. If raw is True, use "raw_data" proto #     field to store the values, and values should be of type bytes in #     this case. # """ # new_const_fold_opt__23 = onnx.helper.make_tensor( #     name="const_fold_opt__23", #     data_type=onnx.TensorProto.INT64, #     dims=[4], #     vals=b"\001\000\000\000\000\000\000\000\001\000\000\000\000\000\000\000\031\000\000\000\000\000\000\000\001\000\000\000\000\000\000\000" #     raw=True, #     )    # graph.node.remove(old_const_fold_opt__23) # graph.node.insert(1, new_const_fold_opt__23) print("============ node[12] ============") print(node[12]) print("============ node[34] ============") print(node[34]) old_tile_1 = node[12] old_tile = node[34] """ op_type (string): The name of the operator to construct inputs (list of string): list of input names outputs (list of string): list of output names name (string, default None): optional unique identifier for NodeProto doc_string (string, default None): optional documentation string for NodeProto domain (string, default None): optional domain for NodeProto.         If it's None, we will just use default domain (which is empty) **kwargs (dict): the attributes of the node.  The acceptable values         are documented in :func:`make_attribute`. """ new_tile_const = onnx.helper.make_node(     name = "tile_const",     op_type="Constant",     inputs=[],     outputs=["tile_const:0"],     value=onnx.helper.make_tensor(         'value',         onnx.TensorProto.INT64,         [4],         b"\001\000\000\000\000\000\000\000\001\000\000\000\000\000\000\000\031\000\000\000\000\000\000\000\001\000\000\000\000\000\000\000",         raw=True)     ) new_tile = onnx.helper.make_node(     name="Tile",     op_type="Tile",     inputs=["Reshape_1:0", "tile_const:0"],     outputs=["Tile:0"], ) graph.node.remove(old_tile) graph.node.insert(34, new_tile) graph.node.insert(34, new_tile_const) new_tile_1_const = onnx.helper.make_node(     name = "tile_1_const",     op_type="Constant",     inputs=[],     outputs=["tile_1_const:0"],     value=onnx.helper.make_tensor(         'value',         onnx.TensorProto.INT64,         [3],         b"\001\000\000\000\000\000\000\000\001\000\000\000\000\000\000\000\001\000\000\000\000\000\000\000",         raw=True)     ) new_tile_1 = onnx.helper.make_node(     name="Tile_1",     op_type="Tile",     inputs=["ExpandDims:0", "tile_1_const:0"],     outputs=["Tile_1:0"], ) graph.node.remove(old_tile_1) graph.node.insert(12, new_tile_1) graph.node.insert(12, new_tile_1_const) onnx.checker.check_model(model) onnx.save(model, "./test_dssm_part_modified.onnx")

其他解决方案：

from_onnx 中添加参数 freeze_params=True

mod, params = relay.frontend.from_onnx(graph_def, shape=input_shape, freeze_params=True)

resnet101 切图

文件链接：https://1drv.ms/u/s!ArOkNLeDa1G_ii310CmCT6y6rXW0?e=q9E7O6

resnet101 编译成 stcobj 后，给定同样的输入，输出的内容有差异。为定位该精度问题，对 onnx 模型进行切图操作，通过指定新的 output 节点，对比输出内容来判断出错节点。切图输出示例：

正在上传…重新上传取消

def cut_onnx_model():     model = onnx.load(model_path_modified_onnx)     new_output = onnx.helper.make_tensor_value_info("output", 1, ['N', 3, 224, 224])     LAST_NODE_IDX = 1     model.graph.output[0].CopyFrom(new_output)     model.graph.node[LAST_NODE_IDX].CopyFrom(         onnx.helper.make_node(             name="xxx",             op_type="BatchNormalization",             inputs=["data:01_transposed","scale101","bias101","mean101","var101"],             outputs=["output"],             epsilon=1.999999987845058e-08,             momentum=0.9900000095367432,             domain=""         )     )     for i in range(len(model.graph.node)-1, LAST_NODE_IDX, -1):         n = model.graph.node[i]         model.graph.node.remove(n)           onnx.checker.check_model(model)     onnx.save(model, model_path_cut_onnx)

roformer 输入修改

输入 input_token 为 float16，转 int 出现精度问题，手动修改模型输入接受 int32 类型的 input_token

def modify_onnx_model(model_path):     print("====================== modify onnx model ======================")     model_path_modified = "model_zoo/models/roformer/open_roformer-tf-fp32_modified.onnx"     model = onnx.load(model_path)     graph = model.graph     node = graph.node     model.graph.input[1].type.tensor_type.elem_type = onnx.TensorProto.INT32     old_cast = node[9]     old_gather = node[13]     old_equal = node[0]     new_gather = onnx.helper.make_node(         op_type="Gather",         inputs=["Func/StatefulPartitionedCall/input/_2:0", "input_token:0"],         outputs=["StatefulPartitionedCall/model_1/Embedding-Token/embedding_lookup:0"],         name="StatefulPartitionedCall/model_1/Embedding-Token/embedding_lookup",         axis=0,     )     old_gather.CopyFrom(new_gather)     new_cast = onnx.helper.make_node(         op_type="Cast",         inputs=["input_token:0"],         outputs=["StatefulPartitionedCall/model_1/Embedding-Token/Cast:0"],         name="StatefulPartitionedCall/model_1/Embedding-Token/Cast",         to=onnx.TensorProto.FLOAT,     )     old_cast.CopyFrom(new_cast)     new_equal = onnx.helper.make_node(         op_type="Equal",         inputs=["StatefulPartitionedCall/model_1/Embedding-Token/Cast:0", "zero_reduce__65"],         outputs=["StatefulPartitionedCall/model_1/Embedding-Token/NotEqual:0"],         name="StatefulPartitionedCall/model_1/Embedding-Token/NotEqual",     )     old_equal.CopyFrom(new_equal)     new_cast = node[9]     node.remove(new_cast)     node.insert(0, new_cast)     onnx.checker.check_model(model)     onnx.save(model, model_path_modified)     return model_path_modified

切图

onnx.utils.extract_model(         model_path,         model_path2,         ["input_segment:0", "input_token:0"],         ["StatefulPartitionedCall/model_1/Embedding-Norm/sub:0"]     )