pixellot fixes

onnx2kerastl/elementwise_layers.py

@@ -381,8 +381,11 @@ def convert_where(node, params, layers, lambda_func, node_name, keras_name):
                                      layers[node.input[1]],
                                      tf_name=f"{params['cleaned_name']}_where_1")
     else:
-        layers[node_name] = tf_where(casted, layers[node.input[1]], layers[node.input[2]],
+        try:
+            layers[node_name] = tf_where(casted, layers[node.input[1]], layers[node.input[2]],
                                          tf_name=f"{params['cleaned_name']}_where_2")
+        except Exception as e:
+            print(1)
 
 
 def convert_scatter_nd(node, params, layers, lambda_func, node_name, keras_name):

onnx2kerastl/layers.py

@@ -21,7 +21,7 @@
     convert_tile, convert_gather_elements
 from .constant_layers import convert_constant, convert_constant_of_shape, convert_one_hot
 from .normalization_layers import convert_batchnorm, convert_instancenorm, convert_dropout, convert_lrn, convert_layernorm
-from .pooling_layers import convert_avgpool, convert_maxpool, convert_global_avg_pool, convert_topk, convert_roi_align
+from .pooling_layers import convert_avgpool, convert_global_max_pool, convert_maxpool, convert_global_avg_pool, convert_topk, convert_roi_align
 from .padding_layers import convert_padding
 from .upsampling_layers import convert_upsample
 from .caffe2_layers import convert_alias_with_name, convert_resize_nearest
@@ -81,6 +81,7 @@
     'Dropout': convert_dropout,
     'LRN': convert_lrn,
     'MaxPool': convert_maxpool,
+    'GlobalMaxPool': convert_global_max_pool,
     'AveragePool': convert_avgpool,
     'GlobalAveragePool': convert_global_avg_pool,
     'Shape': convert_shape,

onnx2kerastl/pooling_layers.py

@@ -95,6 +95,39 @@ def convert_maxpool(node, params, layers, lambda_func, node_name, keras_name):
             input_0 = layers[node_name + "_pre_" + rand_string]
     layers[node_name] = pooling(input_0)
 
+def convert_global_max_pool(node, params, layers, lambda_func, node_name, keras_name):
+    """
+    Convert GlobalMaxPool 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
+    """
+    input_0 = ensure_tf_type(layers[node.input[0]], name="%s_const" % keras_name)
+    tensor_dim = len(input_0.shape)
+    if tensor_dim == 3:
+        global_pool = keras.layers.GlobalMaxPooling1D(data_format='channels_first',
+                                                      name=f"{params['cleaned_name']}_global_max_pool_3")
+    elif tensor_dim == 4:
+        global_pool = keras.layers.GlobalMaxPooling2D(data_format='channels_first',
+                                                      name=f"{params['cleaned_name']}_global_max_pool_4")
+    elif tensor_dim == 5:
+        global_pool = keras.layers.GlobalMaxPooling3D(data_format='channels_first',
+                                                      name=f"{params['cleaned_name']}_global_max_pool_5")
+    else:
+        raise NotImplementedError("Global max pooling of dims < 3 or dims > 5 is not supported")
+    input_0 = global_pool(input_0)
+    new_shape = input_0.shape.as_list()
+    new_shape = new_shape[1:]
+    new_shape.extend([1] * (tensor_dim - 2))
+    reshape_layer = keras.layers.Reshape(new_shape, name=f"{params['cleaned_name']}_global_max_pool_reshape")
+    input_0 = reshape_layer(input_0)
+
+    layers[node_name] = input_0
+
 
 def convert_avgpool(node, params, layers, lambda_func, node_name, keras_name):
     """

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "onnx2kerastl"
-version = "0.0.157"
+version = "0.0.158"
 description = ""
 authors = ["dorhar <doron.harnoy@tensorleap.ai>"]
 license = "MIT"