• There is an error before this fix: https://colab.research.google.com/gist/AlexeyAB/5d398fd45903b16f4db3aafeec1dbab2/effnetlite_torchscript.ipynb

• No error after fix: https://colab.research.google.com/gist/AlexeyAB/efefb642ae4ae54301690fc19c67f999/effnetlite_torchscript_fixed.ipynb

• math.ceil(a / b) and math.ceil(torch.true_divide(a, b)) give the same result for int values: https://colab.research.google.com/gist/AlexeyAB/7edb6c5a324ff073628af2a80165e9d2/untitled7.ipynb

The simplest code to reproduce the error is:

import torch

model = torch.hub.load(
        "rwightman/gen-efficientnet-pytorch",
        "tf_efficientnet_lite3",
        pretrained=False,
    )

rand_example = torch.rand(1, 3, 256, 256)
model = torch.jit.trace(model, rand_example)

RuntimeError: Integer division of tensors using div or / is no longer supported, and in a future release div will perform true division as in Python 3. Use true_divide or floor_divide (// in Python) instead.

TorchScript converts the model into a graph-mode model - so it speeds up execution,allows you to use the model on smartphones, avoids some errors when using SWA, and much more ...

@rwightman Hi,

What is the simplest way to convert trained Pytorch-weights of EfficientNet-Lite model back to Tensorflow? If I want to train EfficientNet-Lite model in Pytorch and then use it in TensorFlow (pb/tflite)?

I have done it by using https://github.com/onnx/onnx-tensorflow but it adds extra transpose layers and use multiple conv2d-layers instead of one GroupedConv2D, so Pytorch->ONNX->TensorFlow-model is 10x times slower than the native TensorFlow model: https://github.com/onnx/onnx-tensorflow/issues/782

When I use this code

model = geffnet.create_model( 'efficientnet_b5', num_classes=1000, in_chans=3, pretrained=True, exportable=True)

But this error occured.

efficientnet b0~b3 => okay

efficientnet b4~b7 => URL is empty.

How to solved this problem?

Hi, thanks for the great repo! What's the reason that the x.mul_(inner) line is commented out? Does the commented line return different results? Thanks!

https://github.com/rwightman/gen-efficientnet-pytorch/blob/d2eadab6eb4427014c3b80eeaff1ff82d2d1f3f4/geffnet/activations/activations.py#L20-L25

Hi, thanks for the great work. Following my question here , I tried to convert to ONNX using this repo. But I got several errors.

By inputting this command just like your example, I got segmentation fault error :

sudo python3 onnx_export.py --model mobilenetv3_100 ./mobilenetv3_100.onnx
==> Creating PyTorch mobilenetv3_100 model
==> Exporting model to ONNX format at './mobilenetv3_100.onnx'
==> Loading and checking exported model from './mobilenetv3_100.onnx'
Segmentation fault

When I tried with efficientnet_b0 using checkpoint and not using checkpoint sudo python3 onnx_export.py --model efficientnet_b0 ./efficientnet.onnx or sudo python3 onnx_export.py --model efficientnet_b0 --checkpoint ../train/model_best.pth.tar --num-classes 30 ./efficientnet.onnx

I got Couldn't export Python operator SwishAutoFn error.

==> Creating PyTorch efficientnet_b0 model
=> Loading checkpoint '../train/20191015-Deepeye36k-efficientnet_b0-224/model_best.pth.tar'
=> Loaded checkpoint '../train/20191015-Deepeye36k-efficientnet_b0-224/model_best.pth.tar'
==> Exporting model to ONNX format at './efficientnet.onnx'
Traceback (most recent call last):
  File "onnx_export.py", line 75, in <module>
    main()
  File "onnx_export.py", line 59, in main
    input_names=input_names, output_names=output_names)
  File "/home/ivan/.local/lib/python3.6/site-packages/torch/onnx/__init__.py", line 26, in _export
    result = utils._export(*args, **kwargs)
  File "/home/ivan/.local/lib/python3.6/site-packages/torch/onnx/utils.py", line 394, in _export
    operator_export_type, strip_doc_string, val_keep_init_as_ip)
RuntimeError: ONNX export failed: Couldn't export Python operator SwishAutoFn

Any help would be appreciated, thanks

@rwightman Hi,

Some fixes for static-quantization of EfficientNet-Lite:

1. Use nn.Conv2d instead of F.conv2d for Conv2dSameExport

2. Use nn.quantized.FloatFunctional instead of += for Residual connections

3. There is still only 1 issue with nn.ZeroPad2d: https://colab.research.google.com/gist/AlexeyAB/820648ce1e6a8ae25ea5bdb49a0c7a81/pytorch_quant_pad_same.ipynb https://github.com/rwightman/gen-efficientnet-pytorch/blob/e84f5546ceb5be79a13d86209ecc93077f264057/geffnet/conv2d_layers.py#L110

   • Do you know simple solution for this issue?
   • Can I just comment this line x = self.pad(x) in my local copy if I use network resolution that is multiple of 32 = (2^5) for example 224x224 for models with 5 subsampling layers?

The model creation fails with current PyPI release of this library (0.9.8).

When installed from the sources

pip install git+https://github.com/rwightman/gen-efficientnet-pytorch.git

There isn't any error and model is being instantiated properly. Seems like some fixes are not yet published to PyPI

Hi @rwightman!

Thank you for your great implementation for efficientnet and its relatives.

I found the global param is _SCRIPTABLE and _EXPORTABLE in the config file. However, when I use torch hub to access the model, it is hard for me to acces these two variables and successful export the model. In this case I'm wondering is that possible to set _SCRIPTABLE and _EXPORTABLE as a parameter of EfficientNetBuilder?

Best,

I export efficientnet_b0 onnx and set the 640x640 size , it will happen the error ''' import onnx import onnx_tensorrt.backend as backend import numpy as np model = onnx.load("efficientnet_b0.onnx") engine = backend.prepare(model, device='CUDA:0') input_data = np.random.random(size=(1, 3, 640, 640)).astype(np.float32) output_data = engine.run(input_data)[0] print(output_data) print(output_data.shape) '''

[Error] [TensorRT] ERROR: Parameter check failed at: ../builder/Network.cpp::addPoolingNd::500, condition: allDimsGtEq(windowSize, 1) && volume(windowSize) < MAX_KERNEL_DIMS_PRODUCT Traceback (most recent call last): File "test_onnx.py", line 7, in engine = backend.prepare(model, device='CUDA:0') File "/opt/conda/lib/python3.6/site-packages/onnx_tensorrt-0.1.0-py3.6-linux-x86_64.egg/onnx_tensorrt/backend.py", line 218, in prepare return TensorRTBackendRep(model, device, **kwargs) File "/opt/conda/lib/python3.6/site-packages/onnx_tensorrt-0.1.0-py3.6-linux-x86_64.egg/onnx_tensorrt/backend.py", line 94, in init raise RuntimeError(msg) RuntimeError: While parsing node number 8: builtin_op_importers.cpp:1175 In function importGlobalAveragePool: [8] Assertion failed: layer_ptr

I downloaded models via the PyTorch Hub API and exported ONNX file well on a few weeks ago. But now, I cannot do that.

The geffnet 3edcd4d, which was commited a week ago, uses SiLU activation with PyTorch 1.7.0, but recent ONNX opset 12 hasn't supported SiLU yet. Therefore torch.onnx.export() of PyTorch 1.7.0 results in following error.

import torch

model = torch.hub.load(
    'rwightman/gen-efficientnet-pytorch', 'tf_efficientnet_b0',
    pretrained=True, exportable=True, verbose=False)
model.eval()
dummy_input = torch.randn(1,3,224,224)
model(dummy_input)
traced_model = torch.jit.trace(model, dummy_input)
torch_out = torch.onnx.export(
    model, dummy_input, 'tf_efficientnet_b0.onnx',
    export_params=True, verbose=False,opset_version=12)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/k-tanabe/.local/lib/python3.6/site-packages/torch/onnx/__init__.py", line 230, in export
    custom_opsets, enable_onnx_checker, use_external_data_format)
  File "/home/k-tanabe/.local/lib/python3.6/site-packages/torch/onnx/utils.py", line 91, in export
    use_external_data_format=use_external_data_format)
  File "/home/k-tanabe/.local/lib/python3.6/site-packages/torch/onnx/utils.py", line 639, in _export
    dynamic_axes=dynamic_axes)
  File "/home/k-tanabe/.local/lib/python3.6/site-packages/torch/onnx/utils.py", line 421, in _model_to_graph
    dynamic_axes=dynamic_axes, input_names=input_names)
  File "/home/k-tanabe/.local/lib/python3.6/site-packages/torch/onnx/utils.py", line 203, in _optimize_graph
    graph = torch._C._jit_pass_onnx(graph, operator_export_type)
  File "/home/k-tanabe/.local/lib/python3.6/site-packages/torch/onnx/__init__.py", line 263, in _run_symbolic_function
    return utils._run_symbolic_function(*args, **kwargs)
  File "/home/k-tanabe/.local/lib/python3.6/site-packages/torch/onnx/utils.py", line 930, in _run_symbolic_function
    symbolic_fn = _find_symbolic_in_registry(domain, op_name, opset_version, operator_export_type)
  File "/home/k-tanabe/.local/lib/python3.6/site-packages/torch/onnx/utils.py", line 888, in _find_symbolic_in_registry
    return sym_registry.get_registered_op(op_name, domain, opset_version)
  File "/home/k-tanabe/.local/lib/python3.6/site-packages/torch/onnx/symbolic_registry.py", line 111, in get_registered_op
    raise RuntimeError(msg)
RuntimeError: Exporting the operator silu to ONNX opset version 12 is not supported. Please open a bug to request ONNX export support for the missing operator

You have updated also geffnet/version.py from 1.0.0 to 1.0.1 in 3edcd4d. Could you tag v1.0.0 on the previous revision e84f554, which supports ONNX file export with PyTorch 1.7.0. Then I can down load v1.0.0 with a following command.

model = torch.hub.load(
    'rwightman/gen-efficientnet-pytorch:v1.0.0', 'tf_efficientnet_b0',
    pretrained=True, exportable=True, verbose=False)

Hi Ross,

First of all, thank you for the awesome repo, it has been wonderful to use this and pytorch-image-models so far. I just have a couple of questions about ONNX export.

1. I (mistakenly) trained my EfficientNet model on the models found in pytorch-image-models and then realised that to export it to ONNX I needed to have a) used gen-efficientnet-pytorch and b) set the geffnet.config.set_exportable(True) beforehand as you have stated in your export script.

Is there any way I could perhaps still transfer the weights from a pytorch-image-models version of EfficientNet to an equivalent gen-efficientnet-pytorch version AND at the same time make it compatible with ONNX?

2. Is there any chance you could also add ONNX export somehow to the pytorch-image-models versions of EfficientNet? My code base immensely prefers to work with the attributes and methods available for models in that repository compared with the models in geffnet

3. Seconding the suggestion in issue #32 and would love to have the option for that for pytorch-image-models as well.

Thank you

how to fuse conv and bn when using this repo to perform quantization aware training? show in the code snippet,like fuse_model func: import geffnet model = geffnet.efficientnet_lite0(pretrained=True, drop_rate=0.25, drop_connect_rate=0.2) mode.fuse_model()