add mobilenet example

Author: dskkato

Cargo.toml

@@ -7,5 +7,6 @@ edition = "2018"
 
 [dependencies]
 image = "0.23.14"
+ndarray = "0.15.3"
 random = "0.12.2"
-tensorflow = "0.17.0"
+tensorflow = {version="0.17.0", features = ["ndarray"]}

examples/zenn.rs

@@ -0,0 +1,64 @@
+use std::error::Error;
+use std::fs::File;
+use std::io::Read;
+use std::path::Path;
+use std::result::Result;
+use tensorflow::Code;
+use tensorflow::Graph;
+use tensorflow::ImportGraphDefOptions;
+use tensorflow::Session;
+use tensorflow::SessionOptions;
+use tensorflow::SessionRunArgs;
+use tensorflow::Status;
+use tensorflow::Tensor;
+
+use ndarray;
+
+use image::io::Reader as ImageReader;
+use image::GenericImageView;
+
+fn main() -> Result<(), Box<dyn Error>> {
+    let filename = "examples/zenn/mobilenetv3large.pb";
+    if !Path::new(filename).exists() {
+        return Err(Box::new(
+            Status::new_set(
+                Code::NotFound,
+                &format!(
+                    "Run 'python examples/zenn/zenn.py' to generate {} \
+                     and try again.",
+                    filename
+                ),
+            )
+            .unwrap(),
+        ));
+    }
+
+    // Create input variables for our addition
+    let mut x = Tensor::new(&[1, 224, 224, 3]);
+    let img = ImageReader::open("examples/zenn/sample.png")?.decode()?;
+    for (i, (_, _, pixel)) in img.pixels().enumerate() {
+        x[3 * i] = pixel.0[0] as f32;
+        x[3 * i + 1] = pixel.0[1] as f32;
+        x[3 * i + 2] = pixel.0[2] as f32;
+    }
+
+    // Load the computation graph defined by addition.py.
+    let mut graph = Graph::new();
+    let mut proto = Vec::new();
+    File::open(filename)?.read_to_end(&mut proto)?;
+    graph.import_graph_def(&proto, &ImportGraphDefOptions::new())?;
+    let session = Session::new(&SessionOptions::new(), &graph)?;
+
+    // Run the graph.
+    let mut args = SessionRunArgs::new();
+    args.add_feed(&graph.operation_by_name_required("x")?, 0, &x);
+    let output = args.request_fetch(&graph.operation_by_name_required("Identity")?, 0);
+    session.run(&mut args)?;
+
+    // Check our results.
+    let output: Tensor<f32> = args.fetch(output)?;
+    let res: ndarray::Array<f32, _> = output.into();
+    println!("{:?}", res);
+
+    Ok(())
+}

examples/zenn/imagenet_class_index.json

@@ -0,0 +1,28 @@
+import tensorflow as tf
+from tensorflow.python.framework.convert_to_constants import (
+    convert_variables_to_constants_v2,
+)
+from numpy.testing import assert_almost_equal
+
+# default input shape 224x224x3
+model = tf.keras.applications.MobileNetV3Large()
+
+x = tf.TensorSpec(model.input_shape, tf.float32)
+concrete_function = tf.function(lambda x: model(x)).get_concrete_function(x)
+# now all variables are converted to constants.
+# if this step is omitted, dumped graph does not include trained weights
+frozen_model = convert_variables_to_constants_v2(concrete_function)
+directory = "examples/zenn"
+tf.io.write_graph(frozen_model.graph, directory, "mobilenetv3large.pb", as_text=False)
+
+# original prediction
+buf = tf.io.read_file("examples/zenn/sample.png")
+img = tf.image.decode_png(buf)
+sample = img[tf.newaxis, :, :, :]
+original_pred = model(sample)
+
+frozen_model_pred = frozen_model(tf.cast(sample, tf.float32))[0]
+assert_almost_equal(original_pred.numpy(), frozen_model_pred.numpy())
+decoded = tf.keras.applications.mobilenet_v3.decode_predictions(original_pred.numpy())[0]
+print(decoded)
+