Evaluate YOLO v4 on Inferentia

Note: this tutorial runs on tensorflow-neuron 1.x only

Introduction

This tutorial walks through compiling and evaluating YOLO v4 model on Inferentia using the AWS Neuron SDK 09/2020 release. We recommend running this tutorial on an EC2 inf1.2xlarge instance which contains one Inferentia and 8 vCPU cores, as well as 16 GB of memory.Verify that this Jupyter notebook is running the Python kernel environment that was set up according to the Tensorflow Installation Guide You can select the Kernel from the “Kernel -> Change Kernel” option on the top of this Jupyter notebook page.

Prerequisites

This demo requires the following pip packages:

neuron-cc tensorflow-neuron<2 requests pillow matplotlib pycocotools torch

and debian/rpm package aws-neuron-runtime.

On DLAMI, aws-neuron-runtime is already pre-installed.

!pip install tensorflow_neuron==1.15.5.2.8.9.0 neuron_cc==1.13.5.0 requests pillow matplotlib pycocotools==2.0.1 numpy==1.18.2 torch~=1.5.0 --force \
    --extra-index-url=https://pip.repos.neuron.amazonaws.com

Part 1: Download Dataset and Generate Pretrained SavedModel

Download COCO 2017 validation dataset

We start by downloading the COCO validation dataset, which we will use to validate our model. The COCO 2017 dataset is widely used for object-detection, segmentation and image captioning.

!curl -LO http://images.cocodataset.org/zips/val2017.zip
!curl -LO http://images.cocodataset.org/annotations/annotations_trainval2017.zip
!unzip -q val2017.zip
!unzip annotations_trainval2017.zip

!ls

Check required package versions

Here are the minimum required versions of AWS Neuron packages. We run a check.

import pkg_resources
from distutils.version import LooseVersion

assert LooseVersion(pkg_resources.get_distribution('neuron-cc').version) > LooseVersion('1.0.20000')
assert LooseVersion(pkg_resources.get_distribution('tensorflow-neuron').version) > LooseVersion('1.15.3.1.0.2000')
print('passed package version checks')

Generate YOLO v4 tensorflow SavedModel (pretrained on COCO 2017 dataset)

Script yolo_v4_coco_saved_model.py will generate a tensorflow SavedModel using pretrained weights from https://github.com/Tianxiaomo/pytorch-YOLOv4.

!python3 yolo_v4_coco_saved_model.py

This tensorflow SavedModel can be loaded as a tensorflow predictor. When a JPEG format image is provided as input, the output result of the tensorflow predictor contains information for drawing bounding boxes and classification results.

import json
import tensorflow as tf
from PIL import Image
import matplotlib.pyplot as plt
import matplotlib.patches as patches

# launch predictor and run inference on an arbitrary image in the validation dataset
yolo_pred_cpu = tf.contrib.predictor.from_saved_model('./yolo_v4_coco_saved_model')
image_path = './val2017/000000581781.jpg'
with open(image_path, 'rb') as f:
    feeds = {'image': [f.read()]}
results = yolo_pred_cpu(feeds)

# load annotations to decode classification result
with open('./annotations/instances_val2017.json') as f:
    annotate_json = json.load(f)
label_info = {idx+1: cat['name'] for idx, cat in enumerate(annotate_json['categories'])}

# draw picture and bounding boxes
fig, ax = plt.subplots(figsize=(10, 10))
ax.imshow(Image.open(image_path).convert('RGB'))
wanted = results['scores'][0] > 0.1
for xyxy, label_no_bg in zip(results['boxes'][0][wanted], results['classes'][0][wanted]):
    xywh = xyxy[0], xyxy[1], xyxy[2] - xyxy[0], xyxy[3] - xyxy[1]
    rect = patches.Rectangle((xywh[0], xywh[1]), xywh[2], xywh[3], linewidth=1, edgecolor='g', facecolor='none')
    ax.add_patch(rect)
    rx, ry = rect.get_xy()
    rx = rx + rect.get_width() / 2.0
    ax.annotate(label_info[label_no_bg + 1], (rx, ry), color='w', backgroundcolor='g', fontsize=10,
                ha='center', va='center', bbox=dict(boxstyle='square,pad=0.01', fc='g', ec='none', alpha=0.5))
plt.show()

Part 2: Compile the Pretrained SavedModel for Inferentia

We make use of the Python compilation API tfn.saved_model.compile that is avaiable in tensorflow-neuron<2. For the purpose of reducing Neuron runtime overhead, it is necessary to make use of arguments no_fuse_ops and minimum_segment_size.

import shutil
import tensorflow as tf
import tensorflow.neuron as tfn


def no_fuse_condition(op):
    return any(op.name.startswith(pat) for pat in ['reshape', 'lambda_1/Cast', 'lambda_2/Cast', 'lambda_3/Cast'])

with tf.Session(graph=tf.Graph()) as sess:
    tf.saved_model.loader.load(sess, ['serve'], './yolo_v4_coco_saved_model')
    no_fuse_ops = [op.name for op in sess.graph.get_operations() if no_fuse_condition(op)]
shutil.rmtree('./yolo_v4_coco_saved_model_neuron', ignore_errors=True)
result = tfn.saved_model.compile(
    './yolo_v4_coco_saved_model', './yolo_v4_coco_saved_model_neuron',
    # we partition the graph before casting from float16 to float32, to help reduce the output tensor size by 1/2
    no_fuse_ops=no_fuse_ops,
    # to enforce trivial compilable subgraphs to run on CPU
    minimum_segment_size=100,
    batch_size=1,
    dynamic_batch_size=True,
)
print(result)

Part 3: Evaluate Model Quality after Compilation

Define evaluation functions

We first define some handy helper functions for running evaluation on the COCO 2017 dataset.

import os
import json
import time
import numpy as np
import tensorflow as tf
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval


def cocoapi_eval(jsonfile,
                 style,
                 coco_gt=None,
                 anno_file=None,
                 max_dets=(100, 300, 1000)):
    """
    Args:
        jsonfile: Evaluation json file, eg: bbox.json, mask.json.
        style: COCOeval style, can be `bbox` , `segm` and `proposal`.
        coco_gt: Whether to load COCOAPI through anno_file,
                 eg: coco_gt = COCO(anno_file)
        anno_file: COCO annotations file.
        max_dets: COCO evaluation maxDets.
    """
    assert coco_gt is not None or anno_file is not None

    if coco_gt is None:
        coco_gt = COCO(anno_file)
    print("Start evaluate...")
    coco_dt = coco_gt.loadRes(jsonfile)
    if style == 'proposal':
        coco_eval = COCOeval(coco_gt, coco_dt, 'bbox')
        coco_eval.params.useCats = 0
        coco_eval.params.maxDets = list(max_dets)
    else:
        coco_eval = COCOeval(coco_gt, coco_dt, style)
    coco_eval.evaluate()
    coco_eval.accumulate()
    coco_eval.summarize()
    return coco_eval.stats


def bbox_eval(anno_file, bbox_list):
    coco_gt = COCO(anno_file)

    outfile = 'bbox_detections.json'
    print('Generating json file...')
    with open(outfile, 'w') as f:
        json.dump(bbox_list, f)

    map_stats = cocoapi_eval(outfile, 'bbox', coco_gt=coco_gt)
    return map_stats


def get_image_as_bytes(images, eval_pre_path):
    batch_im_id_list = []
    batch_im_name_list = []
    batch_img_bytes_list = []
    n = len(images)
    batch_im_id = []
    batch_im_name = []
    batch_img_bytes = []
    for i, im in enumerate(images):
        im_id = im['id']
        file_name = im['file_name']
        if i % eval_batch_size == 0 and i != 0:
            batch_im_id_list.append(batch_im_id)
            batch_im_name_list.append(batch_im_name)
            batch_img_bytes_list.append(batch_img_bytes)
            batch_im_id = []
            batch_im_name = []
            batch_img_bytes = []
        batch_im_id.append(im_id)
        batch_im_name.append(file_name)

        with open(os.path.join(eval_pre_path, file_name), 'rb') as f:
            batch_img_bytes.append(f.read())
    return batch_im_id_list, batch_im_name_list, batch_img_bytes_list


def analyze_bbox(results, batch_im_id, _clsid2catid):
    bbox_list = []
    k = 0
    for boxes, scores, classes in zip(results['boxes'], results['scores'], results['classes']):
        if boxes is not None:
            im_id = batch_im_id[k]
            n = len(boxes)
            for p in range(n):
                clsid = classes[p]
                score = scores[p]
                xmin, ymin, xmax, ymax = boxes[p]
                catid = (_clsid2catid[int(clsid)])
                w = xmax - xmin + 1
                h = ymax - ymin + 1

                bbox = [xmin, ymin, w, h]
                # Round to the nearest 10th to avoid huge file sizes, as COCO suggests
                bbox = [round(float(x) * 10) / 10 for x in bbox]
                bbox_res = {
                    'image_id': im_id,
                    'category_id': catid,
                    'bbox': bbox,
                    'score': float(score),
                }
                bbox_list.append(bbox_res)
        k += 1
    return bbox_list

Here is the actual evaluation loop. To fully utilize all four cores on one Inferentia, the optimal setup is to run multi-threaded inference using a ThreadPoolExecutor. The following cell is a multi-threaded adaptation of the evaluation routine at https://github.com/miemie2013/Keras-YOLOv4/blob/910c4c6f7265f5828fceed0f784496a0b46516bf/tools/cocotools.py#L97.

from concurrent import futures

NUM_THREADS = 4

def evaluate(yolo_predictor, images, eval_pre_path, anno_file, eval_batch_size, _clsid2catid):
    batch_im_id_list, batch_im_name_list, batch_img_bytes_list = get_image_as_bytes(images, eval_pre_path)

    # warm up
    yolo_predictor({'image': np.array(batch_img_bytes_list[0], dtype=object)})

    def yolo_predictor_timer(yolo_pred, image):
        begin = time.time()
        result = yolo_pred(image)
        delta = time.time() - begin
        return result, delta

    latency = []
    with futures.ThreadPoolExecutor(NUM_THREADS) as exe:
        fut_im_list = []
        fut_list = []

        start_time = time.time()
        for batch_im_id, batch_im_name, batch_img_bytes in zip(batch_im_id_list, batch_im_name_list, batch_img_bytes_list):
            if len(batch_img_bytes) != eval_batch_size:
                continue
            fut = exe.submit(yolo_predictor_timer, yolo_predictor, {'image': np.array(batch_img_bytes, dtype=object)})
            fut_im_list.append((batch_im_id, batch_im_name))
            fut_list.append(fut)
        bbox_list = []
        sum_time = 0.0
        count = 0
        for (batch_im_id, batch_im_name), fut in zip(fut_im_list, fut_list):
            results, times = fut.result()
            # Adjust latency since we are in batch
            latency.append(times / eval_batch_size)
            sum_time += times
            bbox_list.extend(analyze_bbox(results, batch_im_id, _clsid2catid))
            for _ in batch_im_id:
                count += 1
                if count % 1000 == 0:
                    print('Test iter {}'.format(count))

        throughput = len(images) / (sum_time / NUM_THREADS)


    print('Average Images Per Second:', throughput)
    print("Latency P50: {:.1f} ms".format(np.percentile(latency, 50)*1000.0))
    print("Latency P90: {:.1f} ms".format(np.percentile(latency, 90)*1000.0))
    print("Latency P95: {:.1f} ms".format(np.percentile(latency, 95)*1000.0))
    print("Latency P99: {:.1f} ms".format(np.percentile(latency, 99)*1000.0))

    # start evaluation
    box_ap_stats = bbox_eval(anno_file, bbox_list)
    return box_ap_stats

Evaluate mean average precision (mAP) score

Here is the code to calculate mAP scores of the YOLO v4 model. The expected mAP score is around 0.487 if we use the pretrained weights.

yolo_pred = tf.contrib.predictor.from_saved_model('./yolo_v4_coco_saved_model_neuron')

val_coco_root = './val2017'
val_annotate = './annotations/instances_val2017.json'
clsid2catid = {0: 1, 1: 2, 2: 3, 3: 4, 4: 5, 5: 6, 6: 7, 7: 8, 8: 9, 9: 10, 10: 11, 11: 13, 12: 14, 13: 15, 14: 16,
               15: 17, 16: 18, 17: 19, 18: 20, 19: 21, 20: 22, 21: 23, 22: 24, 23: 25, 24: 27, 25: 28, 26: 31,
               27: 32, 28: 33, 29: 34, 30: 35, 31: 36, 32: 37, 33: 38, 34: 39, 35: 40, 36: 41, 37: 42, 38: 43,
               39: 44, 40: 46, 41: 47, 42: 48, 43: 49, 44: 50, 45: 51, 46: 52, 47: 53, 48: 54, 49: 55, 50: 56,
               51: 57, 52: 58, 53: 59, 54: 60, 55: 61, 56: 62, 57: 63, 58: 64, 59: 65, 60: 67, 61: 70, 62: 72,
               63: 73, 64: 74, 65: 75, 66: 76, 67: 77, 68: 78, 69: 79, 70: 80, 71: 81, 72: 82, 73: 84, 74: 85,
               75: 86, 76: 87, 77: 88, 78: 89, 79: 90}
eval_batch_size = 8
with open(val_annotate, 'r', encoding='utf-8') as f2:
    for line in f2:
        line = line.strip()
        dataset = json.loads(line)
        images = dataset['images']
box_ap = evaluate(yolo_pred, images, val_coco_root, val_annotate, eval_batch_size, clsid2catid)