inisis / brocolli Goto Github PK

如题。。。

Traceback (most recent call last):
File "/media/3255035b-77ce-41e8-8665-808657363219/deep-learning/predict_floder.py", line 147, in
main()
File "/media/3255035b-77ce-41e8-8665-808657363219/deep-learning/predict_floder.py", line 104, in main
pytorch_parser.convert()
File "/home/miniconda3/envs/py39/lib/python3.9/site-packages/brocolli/converter/pytorch_caffe_parser.py", line 51, in convert
self.text_net, self.binary_weights = self.gen_ir()
File "/home/miniconda3/envs/py39/lib/python3.9/site-packages/brocolli/converter/pytorch_caffe_parser.py", line 363, in gen_ir
raise NotImplementedError(
NotImplementedError: method sqrt is not implemented

models.py中的f"conv_{module_i}“ 出现该错误，将f去掉即可

Traceback (most recent call last):
File "pytorch_model_converter.py", line 51, in
pytorch_parser = PytorchParser(model_file, [3, 300, 300])
File "/brocolli/tests/../converter/pytorch/pytorch_parser.py", line 104, in init
self.pytorch_graph.build(self.input_shape)
File "/brocolli/tests/../converter/pytorch/pytorch_graph.py", line 115, in build
trace, output = torch.jit.get_trace_graph(self.model, (dummy_input, ))
File "/usr/lib/python3.5/contextlib.py", line 77, in exit
self.gen.throw(type, value, traceback)
File "/brocolli/tests/../converter/pytorch/pytorch_graph.py", line 100, in set_training
yield
File "/brocolli/tests/../converter/pytorch/pytorch_graph.py", line 115, in build
trace, output = torch.jit.get_trace_graph(self.model, (dummy_input, ))
File "/usr/local/lib/python3.5/dist-packages/torch/jit/init.py", line 255, in get_trace_graph
return LegacyTracedModule(f, nderivs=nderivs)(*args, **kwargs)
File "/usr/local/lib/python3.5/dist-packages/torch/nn/modules/module.py", line 491, in call
result = self.forward(*input, **kwargs)
File "/usr/local/lib/python3.5/dist-packages/torch/jit/init.py", line 288, in forward
out = self.inner(*trace_inputs)
File "/usr/local/lib/python3.5/dist-packages/torch/nn/modules/module.py", line 489, in call
result = self._slow_forward(*input, **kwargs)
File "/usr/local/lib/python3.5/dist-packages/torch/nn/modules/module.py", line 479, in _slow_forward
result = self.forward(*input, **kwargs)
File "/brocolli/tests/ssd.py", line 163, in forward
self.priors.type(x.data.type()) # default boxes
RuntimeError: Attempted to trace Detect, but tracing of legacy functions is not supported

While I was trying to convert SSD300_vgg16 model to Caffe with phase = 'test' instead of 'export' I got the following error :
File "/root/brocolli/converter/pytorch/pytorch_caffe_parser.py", line 138, in gen_IR
node_type = layer_map[onnx_node_type]
KeyError: 'prim::PythonOp'

recently I research scnn model ,I have trained it with the pytorch ,now I am facing a question that I need to convert the pytorch model to caffe model ,I have tried the tools you recommend to me ,the scnn pytorch net link is :"https://github.com/harryhan618/SCNN_Pytorch "and the convert tool link is :"https://github.com/hahnyuan/nn_tools"
I use it to transform it ,the output error is like this:.....140674339123632:batch_norm_blob15 getting
relu15 was added to layers
140674339123704:relu_blob15 was added to blobs
Add blob relu_blob15 : torch.Size([1, 128, 36, 100])
140674339123704:relu_blob15 getting
:::x.shape::: torch.Size([1, 128, 36, 100])
conv16 was added to layers
140674339126368:conv_blob16 was added to blobs
Add blob conv_blob16 : torch.Size([1, 128, 1, 100])
Traceback (most recent call last):
File "example/scnn_pytorch_to_caffe.py", line 26, in
pytorch_to_caffe.trans_net(net,input,name)
File "./pytorch_to_caffe.py", line 716, in trans_net
out = net.forward(input_var)
File "./model.py", line 36, in forward
x = self.message_passing_forward(x)
File "./model.py", line 62, in message_passing_forward
x = self.message_passing_once(x, ms_conv, v, r)
File "./model.py", line 89, in message_passing_once
out.append(slices[i] + F.relu(conv(out[i - 1])))
The general meaning is when dealing with the slice operation error occors,can you help me to have a look at the question ,thank you

While trying to convert MMDetection SSD300 model the following error occurred:
TypeError: forward() missing 1 required positional argument: 'img_metas'

RuntimeError: Exporting the operator max_unpool2d to ONNX opset version 13 is not supported. Please feel free to request support or submit a pull request on PyTorch

I want to convert siamrpn which has two inputs to caffe , how to do it? thank you

大佬，后续会支持pytorch 1.0版本么

Urlopen error when downloading the mini imagenet dataset

@Inis 你好，我在转换模型时候报output 2 of traced region did not have observable data dependence with trace inputs; this probably indicates your program cannot be understood by the tracer.请问你有遇见过吗？

Dear sir :
I want to convert the scnn pytorch model to caffemodel,does your tool work?or can you give me some advice ,thank you !

老哥，有个问题，其实在pytorch_model_converter里的pytorch_parser.run(model_file)之后，ssd的caffemodel和prototxt就已经生成了。
后面为什么要再跑ssd_layer.py example.json pytorch_model/best.pth.prototxt new.prototxt ，其中example.json文件是起什么作用呢。

多谢。

RuntimeError: Exporting the operator linalg_norm to ONNX opset version 13 is not supported. Please feel free to request support or submit a pull request on PyTorch GitHub.

你好，请问ssd.py文件的下载链接哪里有呀

up主您好，如果我想要把您仓库的代码进行扩充，令其可以实现 LSTM 神经网络的量化，并且想把量化后的 LSTM 层放到 GPU 上进行推理加速。
在技术上这是否是可行的呢？如果可行的话，up 您可以大致分享以下实现路线吗？我也想为您的项目添砖加瓦

1.win10/11情况下,需要删除dataloader的 num_workers=8
否则会出以下错
E AttributeError: Can't pickle local object 'test_mnist..MNISTCHINA'

2.#121 pytorch_quantizer.prepare()改成 pytorch_quantizer.prepare_calibration()

Traceback (most recent call last):
File "/media/3255035b-77ce-41e8-8665-808657363219/deep-learning/predict_floder.py", line 147, in
main()
File "/media/3255035b-77ce-41e8-8665-808657363219/deep-learning/predict_floder.py", line 104, in main
pytorch_parser.convert()
File "/home/miniconda3/envs/py39/lib/python3.9/site-packages/brocolli/converter/pytorch_caffe_parser.py", line 51, in convert
self.text_net, self.binary_weights = self.gen_ir()
File "/home/miniconda3/envs/py39/lib/python3.9/site-packages/brocolli/converter/pytorch_caffe_parser.py", line 396, in gen_ir
raise NotImplementedError(
NotImplementedError: method pow is not implemented

Hi, when i run the demo

from brocolli.converter.pytorch_onnx_parser import PytorchOnnxParser
model = models.resnet18(pretrained=False)
x = torch.rand(1 ,3 ,224, 224)
runner = PytorchOnnxParser(model, x)
runner.convert()
runner.save("resnet18.onnx")
runner.check_result()

it tell me:
---> 36 self.input_shape = map_replace(input_shape, get_torch_size)

/opt/conda/lib/python3.7/site-packages/brocolli/converter/utils.py in map_replace(args, fn)
44 return list(map_replace(elem, fn) for elem in args)
45 else:
---> 46 return fn(args)

/opt/conda/lib/python3.7/site-packages/brocolli/converter/utils.py in get_torch_size(obj)
25
26 def get_torch_size(obj):
---> 27 return torch.Size(obj)

TypeError: torch.Size() takes an iterable of 'int' (item 0 is 'Tensor')

i wonder what cause the problem

hi,
i want to test your code and then convert my model.but i could not find your "test_pytorch.py" file. and may some errors:No module named model.classifier.could you tell me how to solve this problem? thanks

Dear sir :
I want to convert the scnn pytorch model to caffemodel,does your tool work?or can you give me some advice ,thank you !

This is very good work, I would like to ask if the efficientnet will be supported？

大佬，请问ssd.py在哪个链接下载，能给个链接吗，没找到

python环境为3.7.5
建立venv虚拟环境后，在虚拟环境中安装
pip install brocolli
pip install brocolli-caffe
报错：

Something wrong about "align_corners" in function "upsample".

Detail:
TypeError: upsample_bilinear2d() got an unexpected keyword argument 'align_corners' (occurred when translating upsample_bilinear2d)

http://120.224.26.73:15030 地址无法访问，麻烦大佬看看

Traceback (most recent call last):
File "/self/pytorch_quan/brocolli-master/brocolli/converter/pytorch_onnx_parser.py", line 13, in
from .onnx_layers import *
ImportError: attempted relative import with no known parent package

While trying to convert the YOLOv5 model, I'm getting KeyError for 'onnx:ScatterND', 'onnx:Range', 'onnx:Expand', 'onnx:Where' and 'onnx:Equal'. Could you please help me on how to get rid this error? I have got the same error when tried with the docker image also.

我编译安装了https://github.com/inisis/caffe.git过后，在brocolli/tests下运行 pytorch_model_converter.py出现No module named 'ssd'的错误，import caffe是成功的，请问应该怎么解决呀？

$ python3 pytorch_model_converter.py
Traceback (most recent call last):
File "pytorch_model_converter.py", line 19, in
from ssd import *
ModuleNotFoundError: No module named 'ssd'
free(): invalid pointer
Aborted (core dumped)

调试的时候，运行到pytorch_parser.run(model_file)报错：
pytorch_parser.run(model_file)
text_net,binary_weight = self.genIR()
layer-data=fun(current_node)
weight = self.state_dict[weight_name]
keyerror: '.weight'

您知道问题出在哪里吗？
我的环境是：torch0.4
numpy'1.14.3'(这个和要求的不一样)

need support onnxruntime 1.15.1

当前的master分支的代码在运行ptq_example.py的时候会出现如下问题

Traceback (most recent call last):
  File "/home/p4ai/model_test/test/brocolli/ptq_example_402.py", line 126, in <module>
    pytorch_quantizer.convert()
  File "/home/p4ai/model_test/test/brocolli/brocolli/quantization/quantizer.py", line 359, in convert
    quantized = Linear.from_float(module)
  File "/home/p4ai/model_test/test/brocolli/brocolli/quantization/quantization_layers/linear.py", line 55, in from_float
    qlinear.act_scale = torch.Tensor(act_scale).to(qweight.device)
IndexError: slice() cannot be applied to a 0-dim tensor.

环境为python 3.9;使用cpu;pytorch版本为1.9.0+cu102;

Thank you for sharing such an excellent open source component！

I encountered a problem:
File "/pycode/brocolli/brocolli/converter/pytorch_caffe_parser.py", line 51, in convert
self.text_net, self.binary_weights = self.gen_ir()
File "/pycode/brocolli/brocolli/converter/pytorch_caffe_parser.py", line 230, in gen_ir
raise NotImplementedError("module %s is not implemented" % (module))
NotImplementedError: module InstanceNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=False) is not implemented

May I ask how to handle it?

Hi, I'm newer to use caffe, so when I run this project some problem comes out. There is no ssd directory or package in project, so whether it's code by author or need to install ? confusions

hello，使用的是你提供的docker环境，
需要转换的模型是语义分割任务，上采样时使用的是：self.upsample = nn.Upsample(scale_factor=scale_factor, mode="bilinear", align_corners=False)。
在转caffe时使用下面的代码调用的你接口：
runner = Runner("parsing_net", model, [1, 3, 480, 640], 13, fuse=True)
runner.pyotrch_inference()
runner.convert()
runner.caffe_inference()
runner.check_result()

会报下面的错误，是不是不支持双线性插值上采样呢？

I was trying to convertYOLOx-Darknet53 and got the KeyError: 'onnx::Range'.
I tried with the latest docker image. Doesn't it support conversion of YOLOX- Darknet53?

hello ，有遇到过这个问题不？？

pytorch_graph.py->extract()->line 130

因为不在同一台电脑，以上是出问题的定位，请见谅

output = self.detect(
loc.view(loc.size(0), -1, 4), # loc preds
self.softmax(conf.view(conf.size(0), -1,
self.num_classes)), # conf preds
self.priors.type(type(x.data.type())) # default boxes
)
output 2 of traced region did not have observable data dependence with trace inputs; this probably indicates your program cannot be understood by the tracer.

but when I Modify the code
output = (
loc.view(loc.size(0), -1, 4), # loc preds
self.softmax(conf.view(-1, self.num_classes)), # conf preds
)
Traceback (most recent call last):
File "/home/zt/brocolli-master/tests/pytorch_model_converter.py", line 55, in
pytorch_parser.run(model_file)
File "/home/zt/brocolli-master/converter/core/parser.py", line 18, in run
text_net, binary_weights = self.gen_IR()
File "/home/zt/brocolli-master/converter/pytorch/pytorch_parser.py", line 126, in gen_IR
layer_data = func(current_node)
File "/home/zt/brocolli-master/converter/pytorch/pytorch_parser.py", line 228, in rename_Conv
weight = self.state_dict[weights_name]
KeyError: 'base.weight'

Please help me.

您好，inisis，您的工程非常不错，我的问题如下：1、您工程里支持的版本是0.4.0，我的pytorch项目中包含upsample layer ，torch.onnx.export()貌似不支持upsample，但是我现在用的最新版1.3可以导出，您的建议是什么呢？2、我的工程结构类似于retinanet，不知道您的项目能否完美转换。很期待您的解惑，谢谢！

Hi ，Thank you for your wonderful and impressive project！ seen the test/quantization/test_quanzation_nets.py use the cpu to load nets and datasets, I'd like follow your idea to trace my model and quantizate with GPU ways,but when i set model.cuda() and use the input_data.cuda() then the error raised:RuntimeError: Input type (torch.FloatTensor) and weight type (torch.cuda.FloatTensor) should be the same
then go to the erro raised location: dummy_input = self.gen_input_tensor(self.input_shape)
def gen_input_tensor(self, shapes):
input_tensor = []
for shape in shapes:
if isinstance(shape, (tuple, list)):
if all(isinstance(element, int) for element in shape):
input_tensor.append(torch.rand(shape).to(torch.float32))
else:
input_tensor.append(self.gen_input_tensor(shape))
else:
input_tensor.append(torch.rand(shape).to(torch.float32))

    return input_tensor

is this func determine the inputs's type from my origin "input_data.cuda()" to other kind types that finally caused the error?
and would pls tell some details about how to use your demo with our own datasets and nets in a GPU way？
thank you very much！