Pytorch to Keras model convertor. Still beta for now.
pip install pytorch2keras
In that moment the only PyTorch 0.2 (deprecated) and PyTorch 0.4.0 (latest stable) are supported.
To use the converter properly, please, make changes in your ~/.keras/keras.json:
...
"backend": "tensorflow",
"image_data_format": "channels_first",
...
The latest version of PyTorch (0.4.1) isn't supported yet.
There are some problem related to a new version:
Q. PyTorch 0.4 hadn't released wheel package for Python 3.7
A. You can build it from source:
git clone https://github.com/pytorch/pytorch
cd pytorch
git checkout v0.4.0
NO_CUDA=1 python setup.py install
Q. Tensorflow isn't available for Python 3.7
A. Yes, we're waiting for it.
For the proper convertion to the tensorflow.js format, please use a new flag short_names=True.
Please, follow this guide to compile the latest version.
Additional information for Arch Linux users:
- the latest gcc8 is incompatible with actual nvcc version
- the legacy gcc54 can't compile C/C++ modules because of compiler flags
It's a convertor of pytorch graph to a Keras (Tensorflow backend) graph.
Firstly, we need to load (or create) pytorch model:
class TestConv2d(nn.Module):
"""Module for Conv2d convertion testing
"""
def __init__(self, inp=10, out=16, kernel_size=3):
super(TestConv2d, self).__init__()
self.conv2d = nn.Conv2d(inp, out, stride=(inp % 3 + 1), kernel_size=kernel_size, bias=True)
def forward(self, x):
x = self.conv2d(x)
return x
model = TestConv2d()
# load weights here
# model.load_state_dict(torch.load(path_to_weights.pth))
The next step - create a dummy variable with correct shapes:
input_np = np.random.uniform(0, 1, (1, 10, 32, 32))
input_var = Variable(torch.FloatTensor(input_np))
We're using dummy-variable in order to trace the model.
from converter import pytorch_to_keras
# we should specify shape of the input tensor
k_model = pytorch_to_keras(model, input_var, [(10, 32, 32,)], verbose=True)
You can also set H and W dimensions to None to make your model shape-agnostic:
from converter import pytorch_to_keras
# we should specify shape of the input tensor
k_model = pytorch_to_keras(model, input_var, [(10, None, None,)], verbose=True)
That's all! If all is ok, the Keras model is stores into the k_model variable.
Layers:
- Linear
- Conv2d (also with groups)
- DepthwiseConv2d (with limited parameters)
- Conv3d
- ConvTranspose2d
- MaxPool2d
- MaxPool3d
- AvgPool2d
- Global average pooling (as special case of AdaptiveAvgPool2d)
- Embedding
- UpsamplingNearest2d
- BatchNorm2d
- InstanceNorm2d
Reshape:
- View
- Reshape
- Transpose
Activations:
- ReLU
- LeakyReLU
- PReLU (only with 0.2)
- SELU (only with 0.2)
- Tanh
- HardTanh (clamp)
- Softmax
- Softplus (only with 0.2)
- Softsign (only with 0.2)
- Sigmoid
Element-wise:
- Addition
- Multiplication
- Subtraction
Misc:
- reduce sum ( .sum() method)
- Pooling: count_include_pad, dilation, ceil_mode
- ResNet*
- PreResNet*
- SqueezeNet (with ceil_mode=False)
- SqueezeNext
- DenseNet*
- AlexNet
- Inception
- SeNet
- Mobilenet v2
Look at the tests directory.
This software is covered by MIT License.