Merge pull request ARM-software#30 from ARM-software/add_ssd_mobilene…

…t_v1_int8

Added SSD MobileNet v1 INT8

README.md

@@ -227,6 +227,15 @@
         <th width="100">Mali GPU</th>
         <th width="100">Ethos U</th>
     </tr>
+    <tr>
+        <td><a href = "models/object_detection/ssd_mobilenet_v1/tflite_int8">SSD MobileNet v1 INT8 *</a></td>
+        <td align="center">INT8</td>
+        <td align="center">TensorFlow Lite</td>
+        <td align="center">:heavy_check_mark:</td>
+        <td align="center">:heavy_multiplication_x:</td>
+        <td align="center">:heavy_check_mark:</td>
+        <td align="center">:heavy_multiplication_x:</td>
+    </tr>
     <tr>
         <td><a href="models/object_detection/ssd_mobilenet_v1/tflite_fp32">SSD MobileNet v1 FP32 *</a></td>
         <td align="center">FP32</td>

models/object_detection/ssd_mobilenet_v1/tflite_fp32/get_class_labels.sh

@@ -19,7 +19,7 @@
 git clone --depth 1 https://github.com/tensorflow/models.git ./tf_models
 cp tf_models/research/object_detection/data/mscoco_label_map.pbtxt .
 
-python scripts/export_labels.py --path mscoco_label_map.pbtxt
+python scripts/export_labels.py --path mscoco_label_map.pbtxt --num_classes 90
 tr -d \" < temp.txt > labelmapping.txt
 rm -rf temp.txt mscoco_label_map.pbtxt
 rm -rf ./tf_models

models/object_detection/ssd_mobilenet_v1/tflite_fp32/scripts/export_labels.py

@@ -16,7 +16,6 @@
 
 import argparse
 import collections
-import sys
 
 def read_label_map(label_map_path):
     item_id = None
@@ -42,24 +41,30 @@ def read_label_map(label_map_path):
 
     return items
 
-def convert_dictionary_to_list(d):
+def convert_dictionary_to_list(d, num_classes):
     output_list = []
     # order dictionary by keys
     d = collections.OrderedDict(sorted(d.items()))
-    for k, v in d.items():
-        output_list.append(v)
+
+    for c in range(num_classes):
+        if c + 1 in d:
+            output_list.append(d[c + 1])
+        else:
+            output_list.append('')
 
     return output_list
 
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(description="Process ImageNet labels.")
     parser.add_argument("--path", type=str, required=True)
+    parser.add_argument("--num_classes", type=int, required=True)
 
     args = parser.parse_args()
 
     items = read_label_map(args.path)
-    items = convert_dictionary_to_list(items)
+    items = convert_dictionary_to_list(items, args.num_classes)
+
     with open("temp.txt", "w") as f:
         for item in items:
             f.write("%s\n" % item)

models/object_detection/ssd_mobilenet_v1/tflite_int8/README.md

@@ -0,0 +1,91 @@
+# SSD MobileNet v1 INT8
+
+## Description
+SSD MobileNet v1 is a object detection network, that localizes and identifies objects in an input image. This is a TF Lite quantized version that takes a 300x300 input image and outputs detections for this image. This model is converted from FP32 to INT8 using post-training quantization.
+
+## License
+[Apache-2.0](https://spdx.org/licenses/Apache-2.0.html)
+
+## Related Materials
+### Class Labels
+The class labels associated with this model can be downloaded by running the script `get_class_labels.sh`.
+
+### Model Recreation Code
+Code to recreate this model can be found [here](recreate_model/).
+
+
+## Network Information
+| Network Information |  Value         |
+|---------------------|----------------|
+|  Framework          | TensorFlow Lite |
+|  SHA-1 Hash         | fef68428bd439b70eb983b57d6a342871fa0deaa |
+|  Size (Bytes)       | 7311392 |
+|  Provenance         | https://arxiv.org/abs/1512.02325 |
+|  Paper              | https://arxiv.org/abs/1512.02325 |
+
+## Performance
+| Platform | Optimized |
+|----------|:---------:|
+| Cortex-A |:heavy_check_mark:         |
+| Cortex-M |:heavy_multiplication_x:         |
+| Mali GPU |:heavy_check_mark:         |
+| Ethos U  |:heavy_multiplication_x:         |
+
+### Key
+* :heavy_check_mark: - Will run on this platform.
+* :heavy_multiplication_x: - Will not run on this platform.
+
+## Accuracy
+Dataset: COCO 2017 Validation
+
+| Metric | Value |
+|--------|-------|
+| mAP | 0.234 |
+
+## Optimizations
+| Optimization |  Value  |
+|--------------|---------|
+| Quantization | INT8 |
+
+## Network Inputs
+<table>
+    <tr>
+        <th width="200">Input Node Name</th>
+        <th width="100">Shape</th>
+        <th width="300">Description</th>
+    </tr>
+    <tr>
+        <td>tfl.quantize</td>
+        <td>(1, 300, 300, 3)</td>
+        <td>A resized and normalized input image.</td> 
+    </tr>
+</table>
+
+## Network Outputs
+<table>
+    <tr>
+        <th width="200">Output Node Name</th>
+        <th width="100">Shape</th>
+        <th width="300">Description</th>
+    </tr>
+    <tr>
+        <td>TFLite_Detection_PostProcess:01</td>
+        <td>()</td>
+        <td>The y1, x1, y2, x2 coordinates of the bounding boxes for each detection</td> 
+    </tr>
+    <tr>
+        <td>TFLite_Detection_PostProcess:02</td>
+        <td>()</td>
+        <td>The class of each detection</td> 
+    </tr>
+    <tr>
+        <td>TFLite_Detection_PostProcess:03</td>
+        <td>()</td>
+        <td>The probability score for each classification</td> 
+    </tr>
+    <tr>
+        <td>TFLite_Detection_PostProcess:04</td>
+        <td>()</td>
+        <td>A vector containing a number corresponding to the number of detections</td> 
+    </tr>
+</table>

models/object_detection/ssd_mobilenet_v1/tflite_int8/definition.yaml

@@ -0,0 +1,57 @@
+benchmark:
+  COCO 2017 Validation:
+    mAP: '0.234'
+description: SSD MobileNet v1 is a object detection network, that localizes and identifies
+  objects in an input image. This is a TF Lite quantized version that takes a 300x300
+  input image and outputs detections for this image. This model is converted from
+  FP32 to INT8 using post-training quantization.
+license:
+- Apache-2.0
+network:
+  file_size_bytes: 7311392
+  filename: ssd_mobilenet_v1.tflite
+  framework: TensorFlow Lite
+  hash:
+    algorithm: sha1
+    value: fef68428bd439b70eb983b57d6a342871fa0deaa
+  provenance: https://arxiv.org/abs/1512.02325
+network_parameters:
+  input_nodes:
+  - description: A resized and normalized input image.
+    example_input:
+      path: models/object_detection/ssd_mobilenet_v1/tflite_int8/testing_input/tfl.quantize
+    name: tfl.quantize
+    shape:
+    - 1
+    - 300
+    - 300
+    - 3
+  output_nodes:
+  - description: The y1, x1, y2, x2 coordinates of the bounding boxes for each detection
+    name: TFLite_Detection_PostProcess:01
+    shape: []
+    test_output_path: models/object_detection/ssd_mobilenet_v1/tflite_int8/testing_output/TFLite_Detection_PostProcess:01
+  - description: The class of each detection
+    name: TFLite_Detection_PostProcess:02
+    shape: []
+    test_output_path: models/object_detection/ssd_mobilenet_v1/tflite_int8/testing_output/TFLite_Detection_PostProcess:02
+  - description: The probability score for each classification
+    name: TFLite_Detection_PostProcess:03
+    shape: []
+    test_output_path: models/object_detection/ssd_mobilenet_v1/tflite_int8/testing_output/TFLite_Detection_PostProcess:03
+  - description: A vector containing a number corresponding to the number of detections
+    name: TFLite_Detection_PostProcess:04
+    shape: []
+    test_output_path: models/object_detection/ssd_mobilenet_v1/tflite_int8/testing_output/TFLite_Detection_PostProcess:04
+operators:
+  TensorFlow Lite:
+  - CONCATENATION
+  - CONV_2D
+  - CUSTOM
+  - DEPTHWISE_CONV_2D
+  - DEQUANTIZE
+  - LOGISTIC
+  - QUANTIZE
+  - RELU6
+  - RESHAPE
+paper: https://arxiv.org/abs/1512.02325

models/object_detection/ssd_mobilenet_v1/tflite_int8/get_class_labels.sh

@@ -0,0 +1,25 @@
+# Copyright (C) 2021 Arm Limited or its affiliates. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the License); you may
+# not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an AS IS BASIS, WITHOUT
+# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+#!/usr/bin/env bash
+
+git clone --depth 1 https://github.com/tensorflow/models.git ./tf_models
+cp tf_models/research/object_detection/data/mscoco_label_map.pbtxt .
+
+python scripts/export_labels.py --path mscoco_label_map.pbtxt --num_classes 90
+tr -d \" < temp.txt > labelmapping.txt
+rm -rf temp.txt mscoco_label_map.pbtxt
+rm -rf ./tf_models

models/object_detection/ssd_mobilenet_v1/tflite_int8/recreate_model/README.md

@@ -0,0 +1,14 @@
+# SSD MobileNet v1 INT8 Re-Creation
+This folder contains scripts that allow you to re-create the model and benchmark it's performance.
+
+## Requirements
+The scripts in this folder requires that the following must be installed:
+- Python 3.7
+- protoc
+
+## Running The Script
+### Recreate The Model
+Run the following command in a terminal: `./quantize_ssd_mobilenet_v1.sh`
+
+### Benchmarking The Model
+Run the following command in a terminal: `./benchmark_ssd_mobilenet_v1.sh`

models/object_detection/ssd_mobilenet_v1/tflite_int8/recreate_model/benchmark_model.py

@@ -0,0 +1,114 @@
+# Copyright (C) 2021 Arm Limited or its affiliates. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the License); you may
+# not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an AS IS BASIS, WITHOUT
+# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+
+import numpy as np
+import tensorflow_datasets as tfds
+import tensorflow as tf
+
+from pycocotools.coco import COCO
+from pycocotools.cocoeval import COCOeval
+
+detections = []
+
+# Yields the image pre-processed with it's class
+def coco_generator(coco_dataset, input_size=(300, 300)):
+    for item in coco_dataset:
+        image = item['image']
+        image = tf.image.resize(image, input_size)
+        image = tf.expand_dims(image, 0)
+
+        # MobileNet pre-processing
+        image = (image / 255. - 0.5) * 2
+
+        yield image, item['image/id'], item['image'].shape
+
+
+def __convert_to_coco_bbox(b, input_size):
+    # For COCO it is [x, y, width, height]
+    # The bounding box b is of type: [ymin, xmin, ymax, xmax]
+    x = b[1] * input_size[1]
+    y = b[0] * input_size[0]
+    width = (b[3] - b[1]) * input_size[1]
+    height = (b[2] - b[0]) * input_size[0]
+
+    return [x, y, width, height]
+
+
+def process_output(output, image_id, image_size):
+    detection_boxes, detection_classes, detection_scores, num_detections = output
+
+    detections_in_image = []
+    for i in range(int(num_detections[0])):
+        detections_in_image.append(
+            {
+                'image_id': image_id.numpy(),
+                'category_id': int(detection_classes[0, i]) + 1,
+                'bbox': __convert_to_coco_bbox(detection_boxes[0, i], image_size), 
+                'score': detection_scores[0, i]
+            }
+        )
+
+    return detections_in_image
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Benchmark SSD MobileNet v1.')
+    parser.add_argument('--path', type=str, help='Path to the model.')
+
+    args = parser.parse_args()
+
+    # Get the COCO 2017 validation set
+    coco_dataset = tfds.load('coco/2017', split='validation')
+
+    # Setup the TensorFlow Lite interpreter
+    interpreter = tf.lite.Interpreter(model_path=args.path)
+    interpreter.allocate_tensors()
+
+    input_node = interpreter.get_input_details()[0]
+    input_t = input_node['index']
+    output_t = [details['index'] for details in interpreter.get_output_details()]
+
+    for data, data_id, image_shape in coco_generator(coco_dataset):
+        # Quantize the input data
+        scale = input_node["quantization_parameters"]["scales"]
+        zero_point = input_node["quantization_parameters"]["zero_points"]
+
+        data = data / scale
+        data += zero_point
+
+        numpy_data = tf.cast(data, tf.int8).numpy()
+        interpreter.set_tensor(input_t, numpy_data)
+        interpreter.invoke()
+
+        output = [ interpreter.get_tensor(o) for o in output_t ]
+
+        detection_outputs = process_output(output, data_id, (image_shape[0], image_shape[1]))
+        detections += detection_outputs
+
+    # Use the COCO API to measure the accuracy on the annotations
+    coco_ground_truth = COCO('./annotations/instances_val2017.json')
+    coco_results = coco_ground_truth.loadRes(detections)
+
+    coco_eval = COCOeval(coco_ground_truth, coco_results, 'bbox')
+
+    image_ids = [d['image_id'] for d in detections]
+    coco_eval.params.imgIds = image_ids
+
+    coco_eval.evaluate()
+    coco_eval.accumulate()
+    coco_eval.summarize()