add examples cnn and onnx configs simple dnn, correct plotting layout

examples/cnn/create_small_cnn.py

@@ -0,0 +1,101 @@
+# coding: utf-8
+
+"""
+Test script to create a simple DNN model.
+Best to be executed in a CMSSW environment with TensorFlow and cmsml installed.
+Will install tf2onnx for the user if not present in environment.
+
+Signature: f32(64) -> f32(8)
+"""
+
+import os
+import subprocess
+import importlib.util
+
+import cmsml
+
+
+def create_model(
+    model_dir: str,
+    postfix: str = r"l{n_layers}k{kernel_size}f{n_filters}",
+    n_in_1: int = 28,
+    n_in_2: int = 28,
+    n_out: int = 8,
+    n_layers: int = 1,
+    kernel_size: int = 3,
+    n_filters: int = 4,
+    batch_norm: bool = True,
+    pooling: bool = True,
+) -> None:
+    # get tensorflow
+    tf, _, tf_version = cmsml.tensorflow.import_tf()
+    print("creating simple cnn model")
+    print(f"location  : {model_dir}")
+    print(f"TF version: {'.'.join(map(str, tf_version))}")
+
+    # set random seeds to get deterministic results for testing
+    tf.keras.utils.set_random_seed(1)
+
+    # define input layer
+    x = tf.keras.Input(shape=(n_in_1, n_in_2, 1), dtype=tf.float32, name="input")
+
+    # model layers
+    a = tf.keras.layers.BatchNormalization(axis=1, renorm=True)(x) if batch_norm else x
+    for _ in range(n_layers):
+        a = tf.keras.layers.Conv2D(n_filters, kernel_size, padding="same", activation="elu")(a)
+        if pooling:
+            a = tf.keras.layers.MaxPooling2D(pool_size=(2, 2), strides=(1, 1), padding="same")(a)
+        if batch_norm:
+            a = tf.keras.layers.BatchNormalization(axis=1, renorm=True)(a)
+
+    # output layer
+    b = tf.keras.layers.Flatten()(a)
+    y = tf.keras.layers.Dense(n_out, activation="softmax", name="output", dtype=tf.float32)(b)
+
+    # define the model
+    model = tf.keras.Model(inputs=[x], outputs=[y])
+
+    # test evaluation
+    print(model([tf.constant([[[[i] for i in range(n_in_2)] for _ in range(n_in_1)]], dtype=tf.float32)]))
+
+    # save it as a frozen graph
+    _postfix = postfix.format(n_in_1=n_in_1, n_in_2=n_in_2, n_out=n_out, n_layers=n_layers, kernel_size=kernel_size,
+                              n_filters=n_filters, batch_norm=batch_norm)
+    cmsml.tensorflow.save_graph(
+        os.path.join(model_dir, f"frozen_graph_{_postfix}.pb"),
+        model,
+        variables_to_constants=True,
+    )
+
+    # create a SavedModel
+    tf.saved_model.save(
+        model,
+        os.path.join(model_dir, f"saved_model_{_postfix}"),
+    )
+
+    # convert SavedModel to onnx
+    # install tf2onnx if necessary
+    if importlib.util.find_spec("tf2onnx") is None:
+        subprocess.run("pip3 install --user tf2onnx", shell=True)
+
+    # convert
+    subprocess.run(
+        f"""
+            python3 -m tf2onnx.convert \
+            --saved-model saved_model_{_postfix} \
+            --output onnx_graph_{_postfix}.onnx \
+        """,
+        shell=True,
+    )
+
+
+def main():
+    this_dir = os.path.dirname(os.path.abspath(__file__))
+    create_model(this_dir, n_layers=1, kernel_size=1)
+    create_model(this_dir, n_layers=1, kernel_size=3)
+    create_model(this_dir, n_layers=5, kernel_size=1)
+    create_model(this_dir, n_layers=5, kernel_size=3)
+
+
+if __name__ == "__main__":
+    main()

examples/cnn/model_onnx_l1k1f4.yaml

@@ -0,0 +1,11 @@
+model:
+  name: simple_cnn_onnx
+  label: ONNX 1x4x1
+  version: "1.0.0"
+  inference_engine: onnx
+  file: ./onnx_graph_l1k1f4.onnx
+  inputs:
+    - name: input
+      shape: [28, 28, 1]
+  outputs:
+    - name: output

examples/cnn/model_onnx_l1k3f4.yaml

@@ -0,0 +1,11 @@
+model:
+  name: simple_cnn_onnx
+  label: ONNX 1x4x3
+  version: "1.0.0"
+  inference_engine: onnx
+  file: ./onnx_graph_l1k3f4.onnx
+  inputs:
+    - name: input
+      shape: [28,28,1]
+  outputs:
+    - name: output

examples/cnn/model_onnx_l5k1f4.yaml

@@ -0,0 +1,11 @@
+model:
+  name: simple_cnn_onnx
+  label: ONNX 5x4x1
+  version: "1.0.0"
+  inference_engine: onnx
+  file: ./onnx_graph_l5k1f4.onnx
+  inputs:
+    - name: input
+      shape: [28,28,1]
+  outputs:
+    - name: output

examples/cnn/model_onnx_l5k3f4.yaml

@@ -0,0 +1,11 @@
+model:
+  name: simple_cnn_onnx
+  label: ONNX 5x4x3
+  version: "1.0.0"
+  inference_engine: onnx
+  file: ./onnx_graph_l5k3f4.onnx
+  inputs:
+    - name: input
+      shape: [28,28,1]
+  outputs:
+    - name: output

examples/cnn/model_tf_l1k1f4.yaml

@@ -0,0 +1,11 @@
+model:
+  name: simple_cnn_tf
+  label: TF 1x4x1
+  version: "1.0.0"
+  inference_engine: tf
+  file: ./frozen_graph_l1k1f4.pb
+  inputs:
+    - name: input
+      shape: [28,28,1]
+  outputs:
+    - name: Identity

examples/cnn/model_tf_l1k3f4.yaml

@@ -0,0 +1,11 @@
+model:
+  name: simple_cnn_tf
+  label: TF 1x4x3
+  version: "1.0.0"
+  inference_engine: tf
+  file: ./frozen_graph_l1k3f4.pb
+  inputs:
+    - name: input
+      shape: [28,28,1]
+  outputs:
+    - name: Identity

examples/cnn/model_tf_l5k1f4.yaml

@@ -0,0 +1,11 @@
+model:
+  name: simple_cnn_tf
+  label: TF 5x4x1
+  version: "1.0.0"
+  inference_engine: tf
+  file: ./frozen_graph_l5k1f4.pb
+  inputs:
+    - name: input
+      shape: [28,28,1]
+  outputs:
+    - name: Identity

examples/cnn/model_tf_l5k3f4.yaml

@@ -0,0 +1,11 @@
+model:
+  name: simple_cnn_tf
+  label: TF 5x4x3
+  version: "1.0.0"
+  inference_engine: tf
+  file: ./frozen_graph_l5k3f4.pb
+  inputs:
+    - name: input
+      shape: [28,28,1]
+  outputs:
+    - name: Identity

examples/cnn/model_tfaot_l1k1f4.yaml

@@ -0,0 +1,12 @@
+model:
+  name: simple_cnn_tfaot
+  label: AOT 1x4x1
+  version: "1.0.0"
+  inference_engine: tfaot
+  saved_model: ./saved_model_l1k1f4
+  serving_key: serving_default
+
+compilation:
+  batch_sizes: [1]
+  tf_xla_flags: []
+  xla_flags: []

examples/cnn/model_tfaot_l1k3f4.yaml

@@ -0,0 +1,12 @@
+model:
+  name: simple_cnn_tfaot
+  label: AOT 1x4x3
+  version: "1.0.0"
+  inference_engine: tfaot
+  saved_model: ./saved_model_l1k3f4
+  serving_key: serving_default
+
+compilation:
+  batch_sizes: [1]
+  tf_xla_flags: []
+  xla_flags: []

examples/cnn/model_tfaot_l5k1f4.yaml

@@ -0,0 +1,12 @@
+model:
+  name: simple_cnn_tfaot
+  label: AOT 5x4x1
+  version: "1.0.0"
+  inference_engine: tfaot
+  saved_model: ./saved_model_l5k1f4
+  serving_key: serving_default
+
+compilation:
+  batch_sizes: [1]
+  tf_xla_flags: []
+  xla_flags: []

examples/cnn/model_tfaot_l5k3f4.yaml

@@ -0,0 +1,12 @@
+model:
+  name: simple_cnn_tfaot
+  label: AOT 5x4x3
+  version: "1.0.0"
+  inference_engine: tfaot
+  saved_model: ./saved_model_l5k3f4
+  serving_key: serving_default
+
+compilation:
+  batch_sizes: [1]
+  tf_xla_flags: []
+  xla_flags: []

examples/cnn/model_tfaot_l5k3f4_bu.yaml

@@ -0,0 +1,12 @@
+model:
+  name: simple_cnn_tfaot
+  label: AOT 5x4x3 BS
+  version: "1.0.0"
+  inference_engine: tfaot
+  saved_model: ./saved_model_l5k3f4
+  serving_key: serving_default
+
+compilation:
+  batch_sizes: [1, 2, 4, 8, 16, 32, 64, 128, 256]
+  tf_xla_flags: []
+  xla_flags: []