For more details, please refer to arXiv paper and Keras implementation.
$ pip install git+https://github.com/monochromegane/starchart.git
$ git clone https://github.com/monochromegane/ssd_mlengine.git
$ starchart train -m ssd_mlengine \
-M trainer.task \
-s BASIC_GPU \
-- \
--annotation_path=gs://PATH_TO_ANNOTATION_FILE \
--prior_path=gs://PATH_TO_PRIOR_FILE \
--weight_path=gs://PATH_TO_WEIGHT_FILE \
--images_path=gs://PATH_TO_IMAGES_FILE \
--model_dir=TRAIN_PATH/model
When the training is over,
$ starchart expose -m ssd_mlengine
When the prediction API is published,
$ python predict.py -k 1 -c 0.4 -i image.jpg
# {'predictions': [{'key': '1',
# 'objects': [[8.0, # class
# 0.45196059346199036, # confidence
# 104.90727233886719, # xmin
# 97.99836730957031, # ymin
# 212.12222290039062, # xmax
# 315.3045349121094]]}]} # ymax
The code of predict.py
is like the following:
import argparse
from oauth2client.client import GoogleCredentials
from googleapiclient import discovery
from googleapiclient import errors
from PIL import Image
parser = argparse.ArgumentParser()
parser.add_argument('-k', '--keep_top_k', type=int, default=10)
parser.add_argument('-c', '--confidence_threshold', type=float, default=0.8)
parser.add_argument('-i', '--image', required=True)
args = parser.parse_args()
project = YOUR_PROJECT_ID
model = 'ssd_mlengine'
version = YOUR_MODEL_VERSION
credentials = GoogleCredentials.get_application_default()
ml = discovery.build('ml', 'v1', credentials=credentials)
size = (300, 300)
img = Image.open(args.image)
original_size = img.size
resized_img = img.resize(size)
data = []
for i in range(size[1]):
data.append([])
for j in range(size[0]):
data[i].append(resized_img.getpixel((j, i)))
body = {'instances': [{'key': '1',
'data': data,
'keep_top_k': args.keep_top_k,
'original_size': original_size,
'confidence_threshold': args.confidence_threshold
}]}
request = ml.projects().predict(name='projects/{}/models/{}/versions/{}'.format(project, model, version), body=body)
try:
response = request.execute()
output = response['predictions']
print(output)
except errors.HttpError as err:
print(err._get_reason())
- A account for Google Cloud Machine Learning Engine
- Install monochromegane/starchart
- Put these files to Google Cloud Storage.
- Ground truth file. You can generate by the code
- Images file. The source images of ground truth (tar.gz).
weights_SSD300.hdf5
. Pre-trained weight file. Download from hereprior_boxes_ssd300.pkl
. Pre-calcuration default boxes. Download from here. And re-packaging for this repository.import pickle pickle.load(open('prior_boxes_ssd300.pkl', 'rb'))[4332:].dump('prior_boxes_ssd300_min.pkl')
In this implementation, the 38 x 38 split default box via the pool 4 layer has been removed. The Google Cloud ML Engine limits the size of the saved model to 256 MB, as using the full default box exceeds the limit. The default box of 38x38 splitting at the size of 300x300 is small and it seems that there is no influence on object detection.