Add hard sigmoid

onnx2kerastl/activation_layers.py

@@ -1,3 +1,5 @@
+import torch.nn
+import tensorflow as tf
 from tensorflow import keras
 import logging
 from .utils import ensure_tf_type, ensure_numpy_type
@@ -85,6 +87,27 @@ def convert_sigmoid(node, params, layers, lambda_func, node_name, keras_name):
     layers[node_name] = sigmoid(input_0)
 
 
+def convert_hard_sigmoid(node, params, layers, lambda_func, node_name, keras_name, alpha=0.167, beta=0.5):
+    """
+    Convert Hard Sigmoid activation layer
+    :param node: current operation node
+    :param params: operation attributes
+    :param layers: available keras layers
+    :param lambda_func: function for keras Lambda layer
+    :param node_name: internal converter name
+    :param keras_name: resulting layer name
+    :return: None
+    """
+    if len(node.input) != 1:
+        assert AttributeError('More than 1 input for an activation layer.')
+
+    input_0 = ensure_tf_type(layers[node.input[0]], name="%s_const" % keras_name)
+
+    input_0 = tf.multiply(input_0, 5/6)     # TODO: conversion from torch HardSigmoid to tf version
+    hard_sigmoid = keras.layers.Activation('hard_sigmoid', name=keras_name)
+    layers[node_name] = hard_sigmoid(input_0)
+
+
 def convert_tanh(node, params, layers, lambda_func, node_name, keras_name):
     """
     Convert Tanh activation layer

onnx2kerastl/layers.py

@@ -1,6 +1,6 @@
 from .convolution_layers import convert_conv, convert_convtranspose
 from .activation_layers import convert_relu, convert_elu, convert_lrelu, convert_selu, \
-    convert_sigmoid, convert_tanh, convert_softmax, convert_prelu
+    convert_sigmoid, convert_hard_sigmoid, convert_tanh, convert_softmax, convert_prelu
 from .operation_layers import convert_clip, convert_exp, convert_reduce_sum, convert_reduce_mean, \
     convert_log, convert_pow, convert_sqrt, convert_split, convert_cast, convert_floor, convert_identity, \
     convert_argmax, convert_reduce_l2, convert_reduce_max
@@ -22,6 +22,7 @@
     'Elu': convert_elu,
     'LeakyRelu': convert_lrelu,
     'Sigmoid': convert_sigmoid,
+    'HardSigmoid': convert_hard_sigmoid,
     'Tanh': convert_tanh,
     'Selu': convert_selu,
     'Clip': convert_clip,

onnx2kerastl/utils.py

@@ -117,3 +117,6 @@ def check_torch_keras_error(model, k_model, input_np, epsilon=1e-5, change_order
             max_error = error
 
     return max_error
+
+
+

test/layers/activations/test_hard_sigmoid.py

@@ -0,0 +1,46 @@
+import torch.nn as nn
+import numpy as np
+import pytest
+
+from test.utils import convert_and_test
+
+
+class LayerHardSigmoid(nn.Module):
+    """
+    Test for nn.layers based types
+    """
+    def __init__(self):
+        super(LayerHardSigmoid, self).__init__()
+        self.hard_sig = nn.Hardsigmoid()
+
+    def forward(self, x):
+        x = self.hard_sig(x)
+        return x
+
+
+class FHardSigmoid(nn.Module):
+    """
+    Test for nn.functional types
+    """
+    def __init__(self):
+        super(FHardSigmoid, self).__init__()
+
+    def forward(self, x):
+        from torch.nn import functional as F
+        return F.hardsigmoid(x)
+
+
+@pytest.mark.parametrize('change_ordering', [True, False])
+def test_layer_sigmoid(change_ordering):
+    model = LayerHardSigmoid()
+    model.eval()
+    input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
+    error = convert_and_test(model, input_np, verbose=False, change_ordering=change_ordering)
+
+
+@pytest.mark.parametrize('change_ordering', [True, False])
+def test_f_hard_sigmoid(change_ordering):
+    model = FHardSigmoid()
+    model.eval()
+    input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
+    error = convert_and_test(model, input_np, verbose=False, change_ordering=change_ordering)