{ "id": 402545, "title": "Nobuco, - pytorch2tf conversion made easy.", "url": "/competitions/asl-signs/discussion/402545", "author_name": "", "post_date": "2023-04-18T19:59:22.157581700Z", "votes": 7, "comment_count": 3, "views": 0, "content": "

I stumbled upon a library which was released just couple of days ago, it's a library that converts pytorch graph op-by-op resulting in the equivalent keras graph, take a look nobuco

\n

I was trying to convert my model to TF via torch.onnx.export and onnx_tf, however I have some custom ops that were converted in not a very efficient way resulting in ~50% overall slowdown, and with nobuco I was able to successfully convert my model to keras without losing performance.

\n
dummy_input = torch.rand(size=(L, point_dim))\npytorch_module = MyModule().eval()\n\nkeras_model = nobuco.pytorch_to_keras(\n    pytorch_module,\n    args=[dummy_input],\n    inputs_channel_order=ChannelOrder.PYTORCH,\n    outputs_channel_order=ChannelOrder.TENSORFLOW\n)\n
", "messages": [ { "id": "2226286", "postDate": "04/18/2023 19:59:22", "content": "

I stumbled upon a library which was released just couple of days ago, it's a library that converts pytorch graph op-by-op resulting in the equivalent keras graph, take a look nobuco

\n

I was trying to convert my model to TF via torch.onnx.export and onnx_tf, however I have some custom ops that were converted in not a very efficient way resulting in ~50% overall slowdown, and with nobuco I was able to successfully convert my model to keras without losing performance.

\n
dummy_input = torch.rand(size=(L, point_dim))\npytorch_module = MyModule().eval()\n\nkeras_model = nobuco.pytorch_to_keras(\n    pytorch_module,\n    args=[dummy_input],\n    inputs_channel_order=ChannelOrder.PYTORCH,\n    outputs_channel_order=ChannelOrder.TENSORFLOW\n)\n
", "rawMarkdown": "I stumbled upon a library which was released just couple of days ago, it's a library that converts pytorch graph op-by-op resulting in the equivalent keras graph, take a look [nobuco](https://github.com/AlexanderLutsenko/nobuco)\n\nI was trying to convert my model to TF via torch.onnx.export and onnx_tf, however I have some custom ops that were converted in not a very efficient way resulting in ~50% overall slowdown, and with nobuco I was able to successfully convert my model to keras without losing performance.\n\n```\ndummy_input = torch.rand(size=(L, point_dim))\npytorch_module = MyModule().eval()\n\nkeras_model = nobuco.pytorch_to_keras(\n pytorch_module,\n args=[dummy_input],\n inputs_channel_order=ChannelOrder.PYTORCH,\n outputs_channel_order=ChannelOrder.TENSORFLOW\n)\n```", "votes": null }, { "id": "2237809", "postDate": "04/28/2023 01:20:02", "content": "

What is point_dim？Is that embedding size?

", "rawMarkdown": "What is point_dim？Is that embedding size?", "votes": null }, { "id": "2237900", "postDate": "04/28/2023 03:52:20", "content": "

It depends on what nn.Module you want to convert, - dummy_input is just a random tensor that has certain dimensions that correspond to the input shape of pytorch_module

", "rawMarkdown": "It depends on what nn.Module you want to convert, - dummy_input is just a random tensor that has certain dimensions that correspond to the input shape of pytorch_module", "votes": null }, { "id": "2246998", "postDate": "05/05/2023 15:54:40", "content": "

@martynoveduard thank you for sharing this! :) Nobuco is a great way to convert PyTorch models to TensorFlow. But sometimes a large number of errors appear during the conversion.

", "rawMarkdown": "martynoveduard thank you for sharing this! :) Nobuco is a great way to convert PyTorch models to TensorFlow. But sometimes a large number of errors appear during the conversion.", "votes": null } ], "comments": [ { "id": 2237809, "author_name": "jacksonyou", "author_url": "", "post_date": "04/28/2023 01:20:02", "content": "

What is point_dim？Is that embedding size?

", "votes": null, "replies": [ { "id": 2237900, "author_name": "martynoveduard", "author_url": "", "post_date": "04/28/2023 03:52:20", "content": "

It depends on what nn.Module you want to convert, - dummy_input is just a random tensor that has certain dimensions that correspond to the input shape of pytorch_module

", "votes": null, "replies": [] } ] }, { "id": 2246998, "author_name": "ivanisaev", "author_url": "", "post_date": "05/05/2023 15:54:40", "content": "

@martynoveduard thank you for sharing this! :) Nobuco is a great way to convert PyTorch models to TensorFlow. But sometimes a large number of errors appear during the conversion.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2226286": "I stumbled upon a library which was released just couple of days ago, it's a library that converts pytorch graph op-by-op resulting in the equivalent keras graph, take a look [nobuco](https://github.com/AlexanderLutsenko/nobuco)\n\nI was trying to convert my model to TF via torch.onnx.export and onnx_tf, however I have some custom ops that were converted in not a very efficient way resulting in ~50% overall slowdown, and with nobuco I was able to successfully convert my model to keras without losing performance.\n\n```\ndummy_input = torch.rand(size=(L, point_dim))\npytorch_module = MyModule().eval()\n\nkeras_model = nobuco.pytorch_to_keras(\n pytorch_module,\n args=[dummy_input],\n inputs_channel_order=ChannelOrder.PYTORCH,\n outputs_channel_order=ChannelOrder.TENSORFLOW\n)\n```", "2237809": "What is point_dim？Is that embedding size?", "2237900": "It depends on what nn.Module you want to convert, - dummy_input is just a random tensor that has certain dimensions that correspond to the input shape of pytorch_module", "2246998": "martynoveduard thank you for sharing this! :) Nobuco is a great way to convert PyTorch models to TensorFlow. But sometimes a large number of errors appear during the conversion." }, "source": "meta" }