retinaNet pmml predict

[zj463261929]

1. h5-> pmml:
   model = load_model('resnet50_coco_best_v2.1.0.h5', backbone_name='resnet101')

backbone = 'resnet'
RetinanetToPmml(
model,
input_shape=(224,224,3),
backbone_name=backbone,
pmml_file_name="retinanet_with_coco_1.pmml",
model_name="RetinaNet model",
description="A Retinanet model with resnet as backbone"
)
2. pmml predict:
from pypmml import Model
import cv2
import pandas as pd

model = Model.load('retinanet_with_coco_1.pmml')
img = cv2.imread("20190506053526137.jpg", 0)
data = pd.DataFrame(img)
result = model.predict(data)
print (result.shape)

why result shape = (0, 0)??? or The prediction method is wrong???

[Nirmal-Neel]

Hi @zj463261929, I am not sure how pypmml works. They will be able to explain it better. Again, do they support the DeepNetwork element as this is not part of the official schema yet?

[zj463261929]

thank you, Can you tell me how I should predict my model, e.g "retinanet_with_coco_1.pmml"
input: image data、class name
output: bbox、score、class information

[Nirmal-Neel]

Talking about the PMML, it takes an image as input. The resource "KerasRetinaNet.jar" contains the utility functions for pre-processing and converting the image into required format (base64 string in this case). I hope it answers your question.