add logs for detecting oom layers

onnx2kerastl/converter.py

@@ -13,6 +13,8 @@
 import tensorflow as tf
 from keras.models import Model
 
+from onnx2kerastl.logger_helpers import log_oom_layers
+
 from .customonnxlayer import onnx_custom_objects_map
 from .exceptions import UnsupportedLayer, OnnxUnsupported
 from .layers import AVAILABLE_CONVERTERS
@@ -94,7 +96,7 @@ def onnx_to_keras(onnx_model, input_names, name_policy=None, verbose=True, chang
     logger = logging.getLogger('onnx2keras')
 
     logger.info('Converter is called.')
-
+    log_oom_layers(onnx_model, logger)
     onnx_weights = onnx_model.graph.initializer
     onnx_inputs = onnx_model.graph.input
     onnx_outputs = [i.name for i in onnx_model.graph.output]

onnx2kerastl/elementwise_layers.py

@@ -132,23 +132,7 @@ def convert_elementwise_add(node, params, layers, lambda_func, node_name, keras_
                     name=keras_name
                 )(variable_input)
     else:
-        # Both inputs are constants; perform addition at runtime without embedding the result
-        constant_value_0 = tf.constant(input_0)
-        constant_value_1 = tf.constant(input_1)
-
-        # Create a dummy input to satisfy Keras's requirements
-        dummy_input = keras.Input(shape=(1,), name='dummy_input')
-
-        # Define a function outside the Lambda to ensure serialization
-        def add_constants(x):
-            return constant_value_0 + constant_value_1 + x * 0  # x * 0 ensures connection to input
-
-        layers[node_name] = keras.layers.Lambda(
-            add_constants,
-            name=keras_name
-        )(dummy_input)
-
-        layers['dummy_input'] = dummy_input
+        layers[node_name] = input_0 + input_1
 
 
 

onnx2kerastl/logger_helpers.py

@@ -0,0 +1,46 @@
+def log_oom_layers(model, logger):
+    # Load the ONNX model
+    graph = model.graph
+
+    # Log general model information
+    logger.debug("IR_VERSION: {}".format(model.ir_version))
+    logger.debug("PRODUCER: {}".format(model.producer_name))
+    logger.debug("PRODUCER_VERSION: {}".format(model.producer_version))
+    logger.debug("DOMAIN: {}".format(model.domain))
+    logger.debug("MODEL_VERSION: {}".format(model.model_version))
+    logger.debug("OPSET_IMPORT: {}".format(str({opset.domain: opset.version for opset in model.opset_import})))
+
+    # Log input information
+    for input_tensor in graph.input:
+        logger.debug("Input: {}; Shape: {}; Type: {}".format(
+            input_tensor.name, 
+            str(input_tensor.type.tensor_type.shape.dim).replace('\n', ''),
+            input_tensor.type.tensor_type.elem_type
+        ))
+
+    # Log output information
+    for output_tensor in graph.output:
+        logger.debug("Output: {}; Shape: {}; Type: {}".format(
+            output_tensor.name,
+            str(output_tensor.type.tensor_type.shape.dim).replace('\n', ''), 
+            output_tensor.type.tensor_type.elem_type
+        ))
+
+    # Log initializers
+    for initializer in graph.initializer:
+        raw_data_log = initializer.raw_data if len(initializer.raw_data)<50 else ""
+        logger.debug("Initializer: {}; Shape: {}; Type: {}; Raw Data: {}".format(
+            initializer.name, 
+            initializer.dims,
+            initializer.data_type,
+            raw_data_log
+        ))
+
+    # Log each node's information
+    for node in graph.node:
+        logger.debug("Node: {}; Name: {}".format(node.op_type, node.name))
+        logger.debug("  Inputs: {}".format(node.input))
+        logger.debug("  Outputs: {}".format(node.output))
+        logger.debug("  Attributes:")
+        for attr in node.attribute:
+            logger.debug("    - {}: {}".format(attr.name, attr))

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "onnx2kerastl"
-version = "0.0.152"
+version = "0.0.153"
 description = ""
 authors = ["dorhar <[email protected]>"]
 license = "MIT"