mxnet gluon import export predict

export_gluon.py

from mxnet.gluon import nn
from gluoncv.model_zoo import get_model
from mxnet import image, cpu, init
from gluoncv.data.transforms.presets.imagenet import transform_eval

context = [cpu()]

net = get_model("network_prefix", ctx=context, pretrained=True)
with net.name_scope():
    net.output = nn.Dense(4)
net.output.initialize(init.Xavier(), ctx=context)
net.hybridize(static_alloc=True, static_shape=True)
net.load_parameters("model-best.params", ctx=context)

img = image.imread("Hamster.jpeg")
img = transform_eval(img)
_ = net(img)

net.export("model")

import_model.py

#!/usr/bin/env python3
from glob import glob
from mxnet import nd, image, gluon
from gluoncv.data.transforms.presets.imagenet import transform_eval

net = gluon.SymbolBlock.imports('model-symbol.json', ['data'], 'model-0000.params')
classes = ['LA', 'B', 'E', 'LS']


def p(i):
    img = image.imread(i)
    img = transform_eval(img)
    pred = net(img)

    topK = 4
    ind = nd.topk(pred, k=topK)[0].astype('int')
    out = ''
    for j in range(topK):
        out += '\t%s %.3f' % (classes[ind[j].asscalar()], nd.softmax(pred)[0][ind[j]].asscalar())
    print('%s predicted as %s \t| %s' % (i, classes[ind[0].asscalar()], out))


for i in glob("*.jpeg"):
    p(i)