[Models] add retinanet run inference demo

models/docker/Dockerfile

@@ -112,11 +112,15 @@ RUN mv /tmp/models/text/machine_comprehension/bert-squad/model/bertsquad-10.onnx
 
 RUN cd /tmp/models && \
     git checkout -b yolov4 625eaec && \
-    git-lfs pull -X="" -I="yolov4.onnx"
+    git-lfs pull -X="" -I="yolov4.onnx" && \
+    git-lfs pull -X="" -I="retinanet-9.onnx"
 
 WORKDIR /models/vision/detection/yolo
 RUN mv /tmp/models/vision/object_detection_segmentation/yolov4/model/yolov4.onnx .
 
+WORKDIR /models/vision/detection/retinanet
+RUN mv /tmp/models/vision/object_detection_segmentation/retinanet/model/retinanet-9.onnx .
+
 WORKDIR /tmp
 RUN git clone --depth=1 -b rel-1.8.0 https://github.com/onnx/onnx.git
 RUN mv onnx/onnx/backend/test/data/node /unittests
@@ -153,7 +157,8 @@ WORKDIR /tmp/models/vision/detection/yolo
 RUN curl -o test.jpg http://images.cocodataset.org/test2017/000000133861.jpg
 RUN curl -O https://raw.githubusercontent.com/AlexeyAB/darknet/master/data/person.jpg
 RUN curl -O https://raw.githubusercontent.com/ultralytics/yolov5/master/data/images/zidane.jpg
-
+WORKDIR /tmp/models/vision/detection/retinanet
+RUN curl -O https://raw.githubusercontent.com/onnx/models/625eaece0dc31ed1f02067307555321b8a8065cd/vision/object_detection_segmentation/retinanet/dependencies/demo.jpg
 
 # Collect all Models & Data
 FROM alpine

models/docker/build_image.sh

@@ -1,6 +1,6 @@
 #!/bin/bash 
 
-VER="v0.1"
+VER="v0.2"
 NAMESPACE="registry-intl.us-west-1.aliyuncs.com/computation"
 TAG=$NAMESPACE/halo:$VER-model-zoo
 

models/vision/detection/retinanet/detection.py

@@ -0,0 +1,102 @@
+#!/usr/bin/env python3
+import os
+import os.path
+from functools import reduce
+import ctypes
+import numpy as np
+from PIL import Image
+
+
+base_path = os.path.dirname(os.path.abspath(__file__))
+output_path = os.getenv("RETINANET_OUTPUT_PATH")
+model_path = os.getenv("RETINANET_MODEL_PATH")
+
+image_path = os.path.join(model_path, "demo.jpg")
+ref_output_path = os.path.join(base_path, "ref_out")
+
+so_exe = ctypes.CDLL(os.path.join(output_path, "retinanet-9.so"))
+
+
+def flatten(inputs):
+    return [[flatten(i) for i in inputs] if isinstance(inputs, (list, tuple)) else inputs]
+
+
+def update_flatten_list(inputs, res_list):
+    for i in inputs:
+        res_list.append(i) if not isinstance(i, (list, tuple)) else update_flatten_list(i, res_list)
+    return res_list
+
+
+def to_numpy(x):
+    if type(x) is not np.ndarray:
+        x = x.detach().cpu().numpy() if x.requires_grad else x.cpu().numpy()
+    return x
+
+
+def run_inference(image_arr, is_save=False):
+    image_arr = image_arr.flatten().astype(ctypes.c_float)
+
+    if is_save:
+        np.savetxt(os.path.join(output_path, "input_0.txt"), 
+                image_arr.flatten().astype(np.float32), 
+                fmt="%.17f", 
+                delimiter=',', 
+                encoding='utf-8')
+
+    def struct_out(out):
+        return (ctypes.c_float * reduce(lambda x, y: x * y, out))()
+
+    outputs_shape = [(1, 720, 60, 80), (1, 720, 30, 40), (1, 720, 15, 20), (1, 720, 8, 10), (1, 720, 4, 5), 
+                    (1, 36, 60, 80), (1, 36, 30, 40), (1, 36, 15, 20), (1, 36, 8, 10), (1, 36, 4, 5)]
+    outputs = [struct_out(o) for o in outputs_shape]
+
+    # run inference 
+    so_exe.model(image_arr.ctypes.data_as(ctypes.c_void_p), outputs[0], outputs[1], outputs[2], 
+        outputs[3], outputs[4], outputs[5], outputs[6], outputs[7], outputs[8], outputs[9])
+
+    for out in outputs:
+        _data = np.ctypeslib.as_array(out).astype(np.float32)
+        _index = outputs.index(out)
+        _shape = outputs_shape[_index]
+        outputs[_index] = np.reshape(_data, _shape)
+
+        output_filename = f"output_{_shape[1]}_{_shape[-1]}.txt"
+        if is_save:
+            np.savetxt(os.path.join(output_path, output_filename), 
+                _data.flatten(), 
+                fmt="%.17f", 
+                delimiter=',', 
+                encoding='utf-8')
+
+        # Checking model output
+        ref_output_file = os.path.join(ref_output_path, f"{output_filename}")
+        diff_result = np.allclose(_data.flatten(), np.loadtxt(ref_output_file, dtype=np.float32), rtol=1e-4, atol=1e-4)
+        print(f"{output_filename} {diff_result}")
+
+    return outputs
+
+
+def preprocess(is_preprocessed=True):
+    _preprocess_data = None
+    if is_preprocessed:
+        # read input data preprocessed from txt file
+        _preprocess_data = np.loadtxt(os.path.join(ref_output_path, "input_0.txt"), dtype=np.float32)
+    else:
+        # preprocess image
+        from torchvision import transforms
+
+        input_image = Image.open(image_path)
+        prepro = transforms.Compose([
+            transforms.ToTensor(),
+            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
+        ])
+        input_tensor = prepro(input_image)
+        _preprocess_data = input_tensor.unsqueeze(0).numpy()
+
+    return _preprocess_data
+
+
+if __name__ == "__main__":
+
+    input_data = preprocess()
+    run_inference(input_data, is_save=True)