Merge branch 'keras_interface' of github.com:ModECI/MDF into keras_in…

…terface

examples/TensorFlow/Keras/keras_to_MDF.py

@@ -0,0 +1,53 @@
+import numpy as np
+import tensorflow as tf
+
+from modeci_mdf.interfaces.keras import keras_to_mdf
+from modelspec.utils import _val_info
+from modeci_mdf.execution_engine import EvaluableGraph
+
+# load the keras model
+model = tf.keras.models.load_model("kr_N_model.h5")
+
+# get one of the test images from the mnnist test dataset
+_, (x_test, y_test) = tf.keras.datasets.mnist.load_data()
+x_test = tf.keras.utils.normalize(x_test, axis=1)
+one_x_test = x_test[0, :, :]
+
+# reshape the dimension of the test image to 2-D image
+one_x_test = one_x_test.reshape(1, 28, 28)
+
+# get the output of predicting with the keras model
+output = model.predict(one_x_test)
+
+# Convert the Keras model to MDF
+mdf_model, params_dict = keras_to_mdf(model=model, args=one_x_test)
+
+# Get mdf graph
+mdf_graph = mdf_model.graphs[0]
+
+# visualize mdf graph-image
+mdf_model.to_graph_image(
+    engine="dot",
+    output_format="png",
+    view_on_render=False,
+    level=1,
+    filename_root="keras_to_MDF",
+    is_horizontal=True,
+    solid_color=True,
+)
+from IPython.display import Image
+
+Image(filename="Keras_to_MDF.png")
+
+# Evaluate the model via the MDF scheduler
+eg = EvaluableGraph(graph=mdf_graph, verbose=False)
+eg.evaluate()
+output_mdf = eg.output_enodes[0].get_output()
+print("Evaluated the graph in Keras, output: %s" % (_val_info(output_mdf)))
+
+# Assert that the results are the same for Keras and MDF
+assert np.allclose(
+    output,
+    output_mdf,
+)
+print("Passed all comparison tests!")

src/modeci_mdf/interfaces/keras/__init__.py

@@ -0,0 +1,3 @@
+"""Import and export code for `Keras <https://keras.io/>`_ models"""
+
+from .importer import keras_to_mdf

src/modeci_mdf/interfaces/keras/importer.py

@@ -0,0 +1,199 @@
+from inspect import Parameter
+from pyclbr import Function
+import random
+
+from typing import Union, Dict, Any, Tuple, List, Callable
+from xml.dom import Node
+
+import tensorflow as tf
+from tensorflow import keras
+from tensorflow.keras.layers import Dense, Flatten, Input
+
+from modeci_mdf.execution_engine import EvaluableGraph
+from modeci_mdf.mdf import *
+from modeci_mdf.utils import simple_connect
+
+import numpy as np
+
+
+def get_weights_and_activation(layers, model):
+
+    params = {}
+    activations = []
+    for layer in layers:
+        n = {}
+        lyr = model.get_layer(layer)
+        if lyr.weights != []:
+            wgts, bias = lyr.weights
+            n["weights"], n["bias"] = wgts.numpy(), bias.numpy()
+            params[layer] = n
+            activations.append(str(lyr.activation).split()[1])
+    return params, activations
+
+
+def init_model_with_graph(model_id, graph_id):
+    mod = Model(id=model_id)
+    mod_graph = Graph(id=graph_id)
+    mod.graphs.append(mod_graph)
+    return mod, mod_graph
+
+
+def create_input_node(node_id, value):  # , reshape=False):
+    # if reshape == True:
+    #     value = np.array(value).flatten()
+    # else:
+    #     value = np.array(value)
+    input_node = Node(id=node_id)
+    input_node.parameters.append(
+        Parameter(id=f"{node_id}_in", value=np.array(value).tolist())
+    )
+    input_node.output_ports.append(
+        OutputPort(id=f"{node_id}_out", value=f"{node_id}_in")
+    )
+    return input_node
+
+
+def create_flatten_node(node_id):
+    flatten_node = Node(id=node_id)
+    flatten_node.input_ports.append(InputPort(id=f"{node_id}_in"))
+
+    # args for onnx::reshape function
+    args = {"data": f"{node_id}_in", "shape": np.array([-1], dtype=np.int64)}
+
+    # application of the onnx::reshape function to the input
+    flatten_node.functions.append(
+        Function(id="onnx_Reshape", function="onnx::Reshape", args=args)
+    )
+    flatten_node.output_ports.append(
+        OutputPort(id=f"{node_id}_out", value="onnx_Reshape")
+    )
+    return flatten_node
+
+
+def create_dense_node(node_id, weights, bias):
+    node = Node(id=node_id)
+    node.input_ports.append(InputPort(id=f"{node_id}_in"))
+    # Weights
+    node.parameters.append(Parameter(id="wgts", value=weights))
+    # bias
+    node.parameters.append(Parameter(id="bias", value=bias))
+    # Value Weights + bias
+    node.parameters.append(
+        Parameter(id="Output", value=f"({node_id}_in @ wgts) + bias")
+    )
+
+    node.output_ports.append(Parameter(id=f"{node_id}_out", value="Output"))
+    return node
+
+
+def create_activation_node(node_id, activation_name):
+    activation = Node(id=node_id)
+    activation.input_ports.append(InputPort(id=f"{node_id}_in"))
+
+    # Functionality of relu
+    if activation_name == "relu":
+        # Value of relu function
+        relu_ = f"({node_id}_in * ({node_id}_in > 0 ))"
+        activation.parameters.append(Parameter(id="Output", value=relu_))
+
+    # Functionality of sigmoid
+    elif activation_name == "sigmoid":
+        # args for exponential function
+        args = {"variable0": "pos_in", "scale": 1, "rate": 1, "bias": 0, "offset": 0}
+
+        # this will make x => x
+        activation.parameters.append(Parameter(id="pos_in", value=f"{node_id}_in"))
+        # value of e^x
+        activation.functions.append(
+            Function(id="exp", function="exponential", args=args)
+        )
+        # value of sigmoid
+        activation.functions.append(Function(id="Output", value="1 / (1 + exp)"))
+
+    elif activation_name == "softmax":
+        # args for exponential function
+        args = {
+            "variable0": f"{node_id}_in",
+            "scale": 1,
+            "rate": 1,
+            "bias": 0,
+            "offset": 0,
+        }
+
+        # exponential of each value
+        activation.functions.append(
+            Function(id="exp", function="exponential", args=args)
+        )
+        # sum of all exponentials
+        activation.functions.append(Function(id="exp_sum", value="sum(exp)"))
+        # normalizing results
+        activation.functions.append(Function(id="Output", value="exp / exp_sum"))
+
+    activation.output_ports.append(OutputPort(id=f"{node_id}_out", value="Output"))
+    return activation
+
+
+def keras_to_mdf(
+    model: Union[Callable, tf.keras.Model, tf.Module],
+    args: Union[None, np.ndarray, tf.Tensor] = None,
+    # trace: bool = False,
+) -> Union[Model, Graph]:
+    r"""
+    Convert a Keras model to an MDF model.
+    Args:
+        model: The model to translate into MDF.
+        args: The input arguments for this model.
+
+    Returns:
+        The translated MDF model
+    """
+
+    # create mdf model and graph
+    mdf_model, mdf_model_graph = init_model_with_graph(
+        f"{model.name}".capitalize(), f"{model.name}_Graph".capitalize()
+    )
+
+    # create the input node
+    input_node = create_input_node("Input_0", args)
+    mdf_model_graph.nodes.append(input_node)
+
+    # create other nodes needed for mdf graph using the type of layers from the keras model
+    # get layers in the keras model
+    layers = []
+    for layer in model.layers:
+        layers.append(layer.name)
+
+    # get the parameters and activation in each dense layer
+    params, activations = get_weights_and_activation(layers, model)
+
+    node_count = 1
+    for layer in layers:
+        if layer == "flatten":
+            # input_node = create_input_node(f"{layer.capitalize()}_{node_count}", args, reshape=True)
+            flatten_node = create_flatten_node(f"{layer.capitalize()}_{node_count}")
+            mdf_model_graph.nodes.append(flatten_node)
+            node_count += 1
+
+        elif "dense" in layer:
+            weights = params[f"{layer}"]["weights"]
+            bias = params[f"{layer}"]["bias"]
+            dense_node = create_dense_node(
+                f"{layer[:5].capitalize()}_{node_count}", weights, bias
+            )
+            mdf_model_graph.nodes.append(dense_node)
+            node_count += 1
+
+            activation = str(model.get_layer(layer).activation).split()[1]
+            if activation != "linear":
+                activation_node = create_activation_node(
+                    f"{activation.capitalize()}_{node_count}", activation
+                )
+                mdf_model_graph.nodes.append(activation_node)
+                node_count += 1
+
+    for i in range(len(mdf_model_graph.nodes) - 1):
+        e1 = simple_connect(
+            mdf_model_graph.nodes[i], mdf_model_graph.nodes[i + 1], mdf_model_graph
+        )
+
+    return mdf_model, params