Add a classification test for Inception v3

.github/workflows/main.yml

@@ -22,7 +22,7 @@ jobs:
 
   build:
     name: Build from OpenVINO source
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-18.04
     steps:
     - uses: actions/checkout@v2
       with:

crates/openvino/tests/classify-inception.rs

@@ -9,38 +9,25 @@ use std::fs;
 #[test]
 fn classify_inception() {
     let mut core = Core::new(None).unwrap();
-    let mut network = core
+    let network = core
         .read_network_from_file(
             &Fixture::graph().to_string_lossy(),
             &Fixture::weights().to_string_lossy(),
         )
         .unwrap();
 
     let input_name = &network.get_input_name(0).unwrap();
-    assert_eq!(input_name, "image_tensor");
+    assert_eq!(input_name, "input");
     let output_name = &network.get_output_name(0).unwrap();
-    assert_eq!(output_name, "DetectionOutput");
-
-    // Prepare inputs and outputs for resizing, since our input tensor is not the size the model expects.
-    network
-        .set_input_resize_algorithm(input_name, ResizeAlgorithm::RESIZE_BILINEAR)
-        .unwrap();
-    network.set_input_layout(input_name, Layout::NHWC).unwrap();
-    network
-        .set_input_precision(input_name, Precision::U8)
-        .unwrap();
-    network
-        .set_output_precision(output_name, Precision::FP32)
-        .unwrap();
+    assert_eq!(output_name, "InceptionV3/Predictions/Softmax");
 
     // Load the network.
     let mut executable_network = core.load_network(&network, "CPU").unwrap();
     let mut infer_request = executable_network.create_infer_request().unwrap();
-    // TODO eventually, this should not panic: infer_request.set_batch_size(1).unwrap();
 
     // Read the image.
     let tensor_data = fs::read(Fixture::tensor()).unwrap();
-    let tensor_desc = TensorDesc::new(Layout::NHWC, &[1, 3, 481, 640], Precision::U8);
+    let tensor_desc = TensorDesc::new(Layout::NHWC, &[1, 3, 299, 299], Precision::FP32);
     let blob = Blob::new(tensor_desc, &tensor_data).unwrap();
 
     // Execute inference.
@@ -60,30 +47,31 @@ fn classify_inception() {
     assert_eq!(
         &results[..5],
         &[
-            Result(15, 59.0),
-            Result(1, 1.0),
-            Result(8, 1.0),
-            Result(12, 1.0),
-            Result(16, 0.9939936),
+            Result(964, 0.9648312),
+            Result(763, 0.0015633557),
+            Result(412, 0.0007776478),
+            Result(814, 0.0006391522),
+            Result(924, 0.0006150733),
         ][..]
     )
 
-    // This above results match the output of running OpenVINO's `hello_classification` with the same inputs:
-    // $ bin/intel64/Debug/hello_classification /tmp/fixture/frozen_inference_graph.xml /tmp/fixture/000000062808.jpg CPU
+    // The results above almost match the output of OpenVINO's `hello_classification` with similar
+    // inputs:
+    // $ bin/intel64/Debug/hello_classification ../inception.xml ../pizza.jpg CPU
     // Top 10 results:
-    // Image /tmp/fixture/000000062808.jpg
+    // Image ../pizza.jpg
     // classid probability
     // ------- -----------
-    // 15      59.0000000
-    // 1       1.0000000
-    // 12      1.0000000
-    // 8       1.0000000
-    // 16      0.9939936
-    // 2       0.9750488
-    // 9       0.9535966
-    // 20      0.8796915
-    // 13      0.8178773
-    // 6       0.8092338
+    // 964     0.9656160
+    // 763     0.0015505
+    // 412     0.0007806
+    // 924     0.0006135
+    // 814     0.0006102
+    // 966     0.0005903
+    // 960     0.0004972
+    // 522     0.0003951
+    // 927     0.0003644
+    // 923     0.0002908
 }
 
 /// A structure for holding the `(category, probability)` pair extracted from the output tensor of

crates/openvino/tests/detect-inception.rs

@@ -0,0 +1,110 @@
+//! Demonstrates using `openvino-rs` to classify an image using an Inception SSD model and a prepared input tensor. See
+//! [README](fixtures/inception/README.md) for details on how this test fixture was prepared.
+mod fixtures;
+
+use fixtures::inception_ssd::Fixture;
+use openvino::{Blob, Core, Layout, Precision, ResizeAlgorithm, TensorDesc};
+use std::fs;
+
+#[test]
+fn detect_inception() {
+    let mut core = Core::new(None).unwrap();
+    let mut network = core
+        .read_network_from_file(
+            &Fixture::graph().to_string_lossy(),
+            &Fixture::weights().to_string_lossy(),
+        )
+        .unwrap();
+
+    let input_name = &network.get_input_name(0).unwrap();
+    assert_eq!(input_name, "image_tensor");
+    let output_name = &network.get_output_name(0).unwrap();
+    assert_eq!(output_name, "DetectionOutput");
+
+    // Prepare inputs and outputs for resizing, since our input tensor is not the size the model expects.
+    network
+        .set_input_resize_algorithm(input_name, ResizeAlgorithm::RESIZE_BILINEAR)
+        .unwrap();
+    network.set_input_layout(input_name, Layout::NHWC).unwrap();
+    network
+        .set_input_precision(input_name, Precision::U8)
+        .unwrap();
+    network
+        .set_output_precision(output_name, Precision::FP32)
+        .unwrap();
+
+    // Load the network.
+    let mut executable_network = core.load_network(&network, "CPU").unwrap();
+    let mut infer_request = executable_network.create_infer_request().unwrap();
+
+    // Read the image.
+    let tensor_data = fs::read(Fixture::tensor()).unwrap();
+    let tensor_desc = TensorDesc::new(Layout::NHWC, &[1, 3, 481, 640], Precision::U8);
+    let blob = Blob::new(tensor_desc, &tensor_data).unwrap();
+
+    // Execute inference.
+    infer_request.set_blob(input_name, blob).unwrap();
+    infer_request.infer().unwrap();
+    let mut results = infer_request.get_blob(output_name).unwrap();
+    let buffer = unsafe { results.buffer_mut_as_type::<f32>().unwrap().to_vec() };
+
+    // Sort results (TODO extract bounding boxes instead).
+    let mut results: Results = buffer
+        .iter()
+        .enumerate()
+        .map(|(c, p)| Result(c, *p))
+        .collect();
+    results.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap());
+
+    assert_eq!(
+        &results[..5],
+        &[
+            Result(15, 59.0),
+            Result(1, 1.0),
+            Result(8, 1.0),
+            Result(12, 1.0),
+            Result(16, 0.9939936),
+        ][..]
+    )
+
+    // This above results should match the output of running OpenVINO's
+    // `object_detection_sample_ssd` with the same inputs. This test incorrectly uses result
+    // sorting instead of extracting the bounding boxes like `object_detection_sample_ssd` does
+    // (FIXME):
+    // $ bin/intel64/Debug/object_detection_sample_ssd -m ../inception-ssd.xml -i ../pizza.jpg
+    // [ INFO ] InferenceEngine:
+    //     API version ............ 2.1
+    //     Build .................. custom_master_a1d858c5028c1a26d37286913d64028849454b75
+    //     Description ....... API
+    // Parsing input parameters
+    // [ INFO ] Files were added: 1
+    // [ INFO ]     ../pizza.jpg
+    // [ INFO ] Loading Inference Engine
+    // [ INFO ] Device info:
+    //     CPU
+    //     MKLDNNPlugin version ......... 2.1
+    //     Build ........... custom_master_a1d858c5028c1a26d37286913d64028849454b75
+    // [ INFO ] Loading network files:
+    //     ../inception-ssd.xml
+    // [ INFO ] Preparing input blobs
+    // [ INFO ] Batch size is 1
+    // [ INFO ] Preparing output blobs
+    // [ INFO ] Loading model to the device
+    // [ INFO ] Create infer request
+    // [ WARNING ] Image is resized from (640, 481) to (300, 300)
+    // [ INFO ] Batch size is 1
+    // [ INFO ] Start inference
+    // [ INFO ] Processing output blobs
+    // [0,1] element, prob = 0.975312    (1,19)-(270,389) batch id : 0 WILL BE PRINTED!
+    // [1,1] element, prob = 0.953244    (368,17)-(640,393) batch id : 0 WILL BE PRINTED!
+    // [2,59] element, prob = 0.993812    (143,280)-(502,423) batch id : 0 WILL BE PRINTED!
+    // [3,67] element, prob = 0.301402    (5,369)-(582,480) batch id : 0
+    // [ INFO ] Image out_0.bmp created!
+    // [ INFO ] Execution successful
+}
+
+/// A structure for holding the `(category, probability)` pair extracted from the output tensor of
+/// the OpenVINO classification.
+#[derive(Debug, PartialEq)]
+struct Result(usize, f32);
+type Results = Vec<Result>;

crates/openvino/tests/fixtures/inception/README.md

@@ -1,9 +1,11 @@
 # Inception SSD Test Fixture
 
-In order to test the use of `openvino-rs` in the real world, here we include the necessary files for performing the
-classification in [classify-inception.rs](../classify-inception.rs). The artifacts are included in-tree and can be
-rebuilt using the [build.sh] script (with the right system dependencies). The artifacts include:
- - the Inception SSD inference model, converted to OpenVINO IR format (`*.bin`, `*.mapping`, `*.xml`)
+In order to test the use of `openvino-rs` in the real world, here we include the necessary files for
+performing the classification in [classify-inception.rs](../classify-inception.rs). The artifacts
+are included in-tree and can be rebuilt using the [build.sh] script (with the right system
+dependencies). The artifacts include:
+ - the Inception v3 classification model, converted to OpenVINO IR format (`*.bin`, `*.mapping`,
+   `*.xml`)
  - an image from the COCO dataset transformed into the correct tensor format (`tensor-*.bgr`)
 
-The [mod.sh] `Fixture` provides the correct artifact paths in Rust.
+The [mod.sh] `Fixture` provides the correct artifact paths in Rust.

crates/openvino/tests/fixtures/inception/build.sh

@@ -10,20 +10,24 @@ OPENVINO_DIR=${OPENVINO_DIR:-$(realpath $FIXTURE_DIR/../../../../../../openvino)
 PYTHON=${PYTHON:-python3}
 pushd $TMP_DIR
 
-# Retrieve the MobileNet model from the TensorFlow 
-wget --no-clobber http://download.tensorflow.org/models/object_detection/ssd_inception_v2_coco_2018_01_28.tar.gz
-tar xzvf ssd_*.tar.gz
-ln -sf ssd_inception_v2_coco_2018_01_28 model
+# Retrieve the Inception model from the TensorFlow 
+wget --no-clobber https://storage.googleapis.com/download.tensorflow.org/models/inception_v3_2016_08_28_frozen.pb.tar.gz
+tar xzvf inception_*.tar.gz
 
-# Convert the model to OpenVINO IR using the model-optimizer.
+# Convert the model to OpenVINO IR using the model-optimizer; see
+# https://github.com/openvinotoolkit/open_model_zoo/blob/master/models/public/googlenet-v3/model.yml.
 pip install --user -r $OPENVINO_DIR/model-optimizer/requirements_tf.txt 
 $PYTHON $OPENVINO_DIR/model-optimizer/mo_tf.py \
-  --input_model model/frozen_inference_graph.pb \
-  --transformations_config $OPENVINO_DIR/model-optimizer/extensions/front/tf/ssd_v2_support.json \
-  --tensorflow_object_detection_api_pipeline_config model/pipeline.config 
-cp $TMP_DIR/frozen_inference_graph.bin $FIXTURE_DIR/inception-ssd.bin
-cp $TMP_DIR/frozen_inference_graph.mapping $FIXTURE_DIR/inception-ssd.mapping
-cp $TMP_DIR/frozen_inference_graph.xml $FIXTURE_DIR/inception-ssd.xml
+  --reverse_input_channels \
+  --input_shape=[1,299,299,3] \
+  --input=input \
+  --mean_values=input[127.5,127.5,127.5] \
+  --scale_values=input[127.5] \
+  --output=InceptionV3/Predictions/Softmax \
+  --input_model=inception_v3_2016_08_28_frozen.pb
+cp $TMP_DIR/inception_v3_2016_08_28_frozen.bin $FIXTURE_DIR/inception.bin
+cp $TMP_DIR/inception_v3_2016_08_28_frozen.mapping $FIXTURE_DIR/inception.mapping
+cp $TMP_DIR/inception_v3_2016_08_28_frozen.xml $FIXTURE_DIR/inception.xml
 
 # Retrieve the first 10 images of the COCO dataset.
 wget --no-clobber http://images.cocodataset.org/zips/val2017.zip
@@ -32,7 +36,7 @@ unzip -Z1 val2017.zip | head -n 10 | xargs unzip val2017.zip
 popd
 
 # Convert an image to raw tensor format
-cargo run -p openvino-tensor-converter -- $TMP_DIR/val2017/000000062808.jpg $FIXTURE_DIR/tensor-1x3x640x481-u8.bgr 481x640x3xu8
+cargo run -p openvino-tensor-converter -- $TMP_DIR/val2017/000000062808.jpg $FIXTURE_DIR/tensor-1x3x299x299-fp32.bgr 299x299x3xfp32
 
 # Clean up.
 rm -rf $TMP_DIR