pytorch2keras

PyTorch to Keras model converter.

Installation

pip install pytorch2keras

Important notice

To use the converter properly, please, make changes in your ~/.keras/keras.json:

...
"backend": "tensorflow",
"image_data_format": "channels_first",
...

PyTorch 0.4.1 and greater

There is the problem related to a new version:

To make it work, please, cast all your .view() parameters to int. For example:

class ResNet(torchvision.models.resnet.ResNet):
    def __init__(self, *args, **kwargs):
        super(ResNet, self).__init__(*args, **kwargs)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        x = self.avgpool(x)
        x = x.view(int(x.size(0)), -1)  #  << Here
        x = self.fc(x)
        return x

Tensorflow.js

For the proper conversion to a tensorflow.js format, please use the new flag names='short'.

Here is a short instruction how to get a tensorflow.js model:

First of all, you have to convert your model to Keras with this converter:

k_model = pytorch_to_keras(model, input_var, [(10, 32, 32,)], verbose=True, names='short')

Now you have Keras model. You can save it as h5 file and then convert it with tensorflowjs_converter but it doesn't work sometimes. As alternative, you may get Tensorflow Graph and save it as a frozen model:

# Function below copied from here:
# https://stackoverflow.com/questions/45466020/how-to-export-keras-h5-to-tensorflow-pb 
def freeze_session(session, keep_var_names=None, output_names=None, clear_devices=True):
    """
    Freezes the state of a session into a pruned computation graph.

    Creates a new computation graph where variable nodes are replaced by
    constants taking their current value in the session. The new graph will be
    pruned so subgraphs that are not necessary to compute the requested
    outputs are removed.
    @param session The TensorFlow session to be frozen.
    @param keep_var_names A list of variable names that should not be frozen,
                          or None to freeze all the variables in the graph.
    @param output_names Names of the relevant graph outputs.
    @param clear_devices Remove the device directives from the graph for better portability.
    @return The frozen graph definition.
    """
    from tensorflow.python.framework.graph_util import convert_variables_to_constants
    graph = session.graph
    with graph.as_default():
        freeze_var_names = \
            list(set(v.op.name for v in tf.global_variables()).difference(keep_var_names or []))
        output_names = output_names or []
        output_names += [v.op.name for v in tf.global_variables()]
        input_graph_def = graph.as_graph_def()
        if clear_devices:
            for node in input_graph_def.node:
                node.device = ""
        frozen_graph = convert_variables_to_constants(session, input_graph_def,
                                                      output_names, freeze_var_names)
        return frozen_graph


from keras import backend as K
import tensorflow as tf
frozen_graph = freeze_session(K.get_session(),
                              output_names=[out.op.name for out in k_model.outputs])

tf.train.write_graph(frozen_graph, ".", "my_model.pb", as_text=False)
print([i for i in k_model.outputs])

You will see the output layer name, so, now it's time to convert my_model.pb to tfjs model:

tensorflowjs_converter  \
    --input_format=tf_frozen_model \
    --output_node_names='TANHTObs/Tanh' \
    my_model.pb \
    model_tfjs

Thats all!

const MODEL_URL = `model_tfjs/tensorflowjs_model.pb`;
const WEIGHTS_URL = `model_tfjs/weights_manifest.json`;
cont model = await tf.loadFrozenModel(MODEL_URL, WEIGHTS_URL);

How to use

It's the converter of PyTorch graph to a Keras (Tensorflow backend) model.

Firstly, we need to load (or create) a valid PyTorch model:

class TestConv2d(nn.Module):
    """
    Module for Conv2d testing
    """

    def __init__(self, inp=10, out=16, kernel_size=3):
        super(TestConv2d, self).__init__()
        self.conv2d = nn.Conv2d(inp, out, stride=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 shape:

input_np = np.random.uniform(0, 1, (1, 10, 32, 32))
input_var = Variable(torch.FloatTensor(input_np))

We use the dummy-variable to trace the model (with jit.trace):

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 (e.g. fully convolutional netowrk):

from pytorch2keras.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 the modules have converted properly, the Keras model will be stored in the k_model variable.

API

Here is the only method pytorch_to_keras from pytorch2keras module.

def pytorch_to_keras(
    model, args, input_shapes,
    change_ordering=False, training=False, verbose=False, names=False,
)

Options:

model -- a PyTorch module to convert;
args -- list of dummy variables with proper shapes;
input_shapes -- list with shape tuples;
change_ordering -- boolean, if enabled, the converter will try to change BCHW to BHWC
training -- boolean, switch model to training mode (never use it)
verbose -- boolean, verbose output
names -- choice from [keep, short, random]. The selector set the target layer naming policy.

Supported layers

Activations:
- ReLU
- LeakyReLU
- SELU
- Sigmoid
- Softmax
- Tanh
- HardTanh
Constants
Convolutions:
- Conv1d
- Conv2d
- ConvTrsnpose2d
Element-wise:
- Add
- Mul
- Sub
- Div
Embedding
Linear
Normalizations:
- BatchNorm2d
- InstanceNorm2d
- Dropout
Poolings:
- MaxPool2d
- AvgPool2d
- Global MaxPool2d (adaptive pooling to shape [1, 1])
- Global AvgPool2d (adaptive pooling to shape [1, 1])
Not tested yet:
- Upsampling
- Padding
- Reshape

Models converted with pytorch2keras

ResNet*
VGG*
PreResNet*
SqueezeNet (with ceil_mode=False)
SqueezeNext
DenseNet*
AlexNet
Inception
SeNet
Mobilenet v2
DiracNet
DARTS
DRNC

Model	Top1	Top5	Params	FLOPs	Source weights	Remarks
ResNet-10	37.09	15.55	5,418,792	892.62M	osmr's repo	Success
ResNet-12	35.86	14.46	5,492,776	1,124.23M	osmr's repo	Success
ResNet-14	32.85	12.41	5,788,200	1,355.64M	osmr's repo	Success
ResNet-16	30.68	11.10	6,968,872	1,586.95M	osmr's repo	Success
ResNet-18 x0.25	49.16	24.45	831,096	136.64M	osmr's repo	Success
ResNet-18 x0.5	36.54	14.96	3,055,880	485.22M	osmr's repo	Success
ResNet-18 x0.75	33.25	12.54	6,675,352	1,045.75M	osmr's repo	Success
ResNet-18	29.13	9.94	11,689,512	1,818.21M	osmr's repo	Success
ResNet-34	25.34	7.92	21,797,672	3,669.16M	osmr's repo	Success
ResNet-50	23.50	6.87	25,557,032	3,868.96M	osmr's repo	Success
ResNet-50b	22.92	6.44	25,557,032	4,100.70M	osmr's repo	Success
ResNet-101	21.66	5.99	44,549,160	7,586.30M	osmr's repo	Success
ResNet-101b	21.18	5.60	44,549,160	7,818.04M	osmr's repo	Success
ResNet-152	21.01	5.61	60,192,808	11,304.85M	osmr's repo	Success
ResNet-152b	20.54	5.37	60,192,808	11,536.58M	osmr's repo	Success
PreResNet-18	28.72	9.88	11,687,848	1,818.41M	osmr's repo	Success
PreResNet-34	25.88	8.11	21,796,008	3,669.36M	osmr's repo	Success
PreResNet-50	23.39	6.68	25,549,480	3,869.16M	osmr's repo	Success
PreResNet-50b	23.16	6.64	25,549,480	4,100.90M	osmr's repo	Success
PreResNet-101	21.45	5.75	44,541,608	7,586.50M	osmr's repo	Success
PreResNet-101b	21.73	5.88	44,541,608	7,818.24M	osmr's repo	Success
PreResNet-152	20.70	5.32	60,185,256	11,305.05M	osmr's repo	Success
PreResNet-152b	21.00	5.75	60,185,256	11,536.78M	Gluon Model Zoo	Success
PreResNet-200b	21.10	5.64	64,666,280	15,040.27M	tornadomeet/ResNet	Success
DenseNet-121	25.11	7.80	7,978,856	2,852.39M	Gluon Model Zoo	Success
DenseNet-161	22.40	6.18	28,681,000	7,761.25M	Gluon Model Zoo	Success
DenseNet-169	23.89	6.89	14,149,480	3,381.48M	Gluon Model Zoo	Success
DenseNet-201	22.71	6.36	20,013,928	4,318.75M	Gluon Model Zoo	Success
DarkNet Tiny	40.31	17.46	1,042,104	496.34M	osmr's repo	Success
DarkNet Ref	38.00	16.68	7,319,416	365.55M	osmr's repo	Success
SqueezeNet v1.0	40.97	18.96	1,248,424	828.30M	osmr's repo	Success
SqueezeNet v1.1	39.09	17.39	1,235,496	354.88M	osmr's repo	Success
MobileNet x0.25	45.78	22.18	470,072	42.30M	osmr's repo	Success
MobileNet x0.5	36.12	14.81	1,331,592	152.04M	osmr's repo	Success
MobileNet x0.75	32.71	12.28	2,585,560	329.22M	Gluon Model Zoo	Success
MobileNet x1.0	29.25	10.03	4,231,976	573.83M	Gluon Model Zoo	Success
FD-MobileNet x0.25	56.19	31.38	383,160	12.44M	osmr's repo	Success
FD-MobileNet x0.5	42.62	19.69	993,928	40.93M	osmr's repo	Success
FD-MobileNet x1.0	35.95	14.72	2,901,288	146.08M	clavichord93/FD-MobileNet	Success
MobileNetV2 x0.25	48.89	25.24	1,516,392	32.22M	Gluon Model Zoo	Success
MobileNetV2 x0.5	35.51	14.64	1,964,736	95.62M	Gluon Model Zoo	Success
MobileNetV2 x0.75	30.82	11.26	2,627,592	191.61M	Gluon Model Zoo	Success
MobileNetV2 x1.0	28.51	9.90	3,504,960	320.19M	Gluon Model Zoo	Success
InceptionV3	21.22	5.59	23,834,568	5,746.72M	Gluon Model Zoo	Success
DiracNetV2-18	31.47	11.70	11,511,784	1,798.43M	szagoruyko/diracnets	Success
DiracNetV2-34	28.75	9.93	21,616,232	3,649.37M	szagoruyko/diracnets	Success
DARTS	26.70	8.74	4,718,752	537.64M	szagoruyko/diracnets	Success

Usage

Look at the tests directory.

License

This software is covered by MIT License.

andrewssobral/pytorch2keras