[Convert to onnx]

[trungna1374]

Have anyone tried to convert .pth to .onnx?
I tried, but outputs were unchanged for difference inputs.

[SealQ]

Hello, has your problem been solved?

[visiont3lab]

Hi any news?

[arshv27]

Hey, did anyone succeed at resolving this issue?

[DietDietDiet]

Hey, did anyone succeed at resolving this issue?
@trungna1374 @SealQ @arshv27 @visiont3lab

[xuhuaren]

has anyone solved it :)

[xuhuaren]

I successfully convert pytorch model to onnx one.

    self.workers = 1
    self.batch_size = 192
    self.onnx = os.path.join(main_path, "squeezenet1_1.onnx")
    self.batch_max_length = 25
    self.imgH = 32
    self.imgW = 100
    self.rgb = False
    self.character = '0123456789abcdefghijklmnopqrstuvwxyz'
    self.sensitive = False
    self.PAD = False
    self.Transformation = 'TPS'
    self.FeatureExtraction = 'ResNet'
    self.SequenceModeling = 'BiLSTM'
    self.Prediction = 'Attn'
    self.num_fiducial = 20
    self.input_channel = 1
    self.output_channel = 512
    self.hidden_size = 256
    self.num_gpu = 0

No better method, just step by step debug.

[aisensiy]

@xuhuaren

    self.workers = 1
    self.batch_size = 192
    self.onnx = os.path.join(main_path, "squeezenet1_1.onnx")
    self.batch_max_length = 25
    self.imgH = 32
    self.imgW = 100
    self.rgb = False
    self.character = '0123456789abcdefghijklmnopqrstuvwxyz'
    self.sensitive = False
    self.PAD = False
    self.Transformation = 'TPS'
    self.FeatureExtraction = 'ResNet'
    self.SequenceModeling = 'BiLSTM'
    self.Prediction = 'Attn'
    self.num_fiducial = 20
    self.input_channel = 1
    self.output_channel = 512
    self.hidden_size = 256
    self.num_gpu = 0

Could you please share some insights about this? That will be quite helpful.

[xuhuaren]

self.Transformation = 'None'

…

----- 原始邮件 ----- 发件人: "aisensiy" <[email protected]> 收件人: "clovaai/deep-text-recognition-benchmark" <[email protected]> 抄送: "Xuhua Ren" <[email protected]>, "Mention" <[email protected]> 发送时间: 星期四, 2020年 5 月 14日 下午 4:22:48 主题: Re: [clovaai/deep-text-recognition-benchmark] [Convert to onnx] (#76) @xuhuaren ``` self.workers = 1 self.batch_size = 192 self.onnx = os.path.join(main_path, "squeezenet1_1.onnx") self.batch_max_length = 25 self.imgH = 32 self.imgW = 100 self.rgb = False self.character = '0123456789abcdefghijklmnopqrstuvwxyz' self.sensitive = False self.PAD = False self.Transformation = 'TPS' self.FeatureExtraction = 'ResNet' self.SequenceModeling = 'BiLSTM' self.Prediction = 'Attn' self.num_fiducial = 20 self.input_channel = 1 self.output_channel = 512 self.hidden_size = 256 self.num_gpu = 0 ``` Could you please share some insights about this? That will be quite helpful.

-- You are receiving this because you were mentioned. Reply to this email directly or view it on GitHub: #76 (comment)

[aisensiy]

@xuhuaren

Wow, thanks for your super fast reply!

You mean not self.Transformation = 'TPS' but self.Transformation = 'None' ?

[cowyyy]

@xuhuaren @aisensiy
Hi, I'm also confronted with the problem of onnx export.
1 TPS has f.grid_sample which has no op in onnx
2 'Attn' has a loop
could you please share some insights about this?

[xuhuaren]

sub_hidden = torch.unsqueeze(hidden[0], 1) output_hiddens = torch.cat((output_hiddens[:, :i, :], sub_hidden), dim=1) disable the TPS module.

…

----- 原始邮件 ----- 发件人: "cowyyy" <[email protected]> 收件人: "clovaai/deep-text-recognition-benchmark" <[email protected]> 抄送: "Xuhua Ren" <[email protected]>, "Mention" <[email protected]> 发送时间: 星期二, 2020年 5 月 19日 下午 5:36:12 主题: Re: [clovaai/deep-text-recognition-benchmark] [Convert to onnx] (#76) @xuhuaren @aisensiy Hi, I'm also confronted with the problem of onnx export. 1 TPS has f.grid_sample which has no op in onnx 2 'Attn' has a loop could you please share some insights about this?

-- You are receiving this because you were mentioned. Reply to this email directly or view it on GitHub: #76 (comment)

[ark-42]

@xuhuaren So this cannot be ported with the TPS module?

[DietDietDiet]

@xuhuaren could you share your code of export onnx

[jusim]

Code would be very helpful, i have difficulties with this as well

[Greg-Tarr]

Yes, please share the code.

[SunAnLan]

@xuhuaren could you please tell me the version of your pytorch and cuda? As I just tried with the config like yours, there comes an error RuntimeError: Exporting the operator _thnn_fused_lstm_cell to ONNX opset version 9 is not supported. Please open a bug to request ONNX export support for the missing operator.

[kishcs]

I am trying to convert model to onnx or TRT.
I am facing issue while converting to onnx or TRT.

Anyone tried and succeeded ?