def initialize(self, context):
self._context = context
self._batch_size = context.system_properties["batch_size"]
self.initialized = True
properties = context.system_properties
model_dir = properties.get("model_dir")
n_data = 11
inputs_list = []
for i in range(n_data):
inputs_list.append("data{}".format(i))
self.batch_size = 50
data_shape = [('data0', (self.batch_size, 23)),
('data1', (self.batch_size, 1)),
('data2', (self.batch_size, 1)),
('data3', (self.batch_size, 1)),
('data4', (self.batch_size, 1)),
('data5', (self.batch_size, 15)),
('data6', (self.batch_size, 15)),
('data7', (self.batch_size, 15)),
('data8', (self.batch_size, 15)),
('data9', (self.batch_size, 15)),
('data10', (self.batch_size, 15))]
self.mxnet_ctx = mx.neuron()
load_symbol, args, auxs = mx.model.load_checkpoint(
os.path.join(model_dir, "compiled"), 0)
self.mod = mx.mod.Module(symbol=load_symbol,
context=mx.neuron(),
data_names=inputs_list,
label_names=None)
self.mod.bind(for_training=False,
data_shapes=data_shape,
label_shapes=self.mod._label_shapes)
self.mod.set_params(args, auxs)
def transform_fn(mod, payload, input_content_type, output_content_type):
logging.info("Invoking user-defined transform_fn")
Batch = namedtuple("Batch", ["data"])
ctx = mx.neuron()
if input_content_type != "application/x-npy":
raise RuntimeError("Input content type must be application/x-npy")
# pre-processing
io_bytes_obj = io.BytesIO(payload)
npy_payload = np.load(io_bytes_obj)
mx_ndarray = mx.nd.array(npy_payload.reshape((1, 28, 28)))
print(mx_ndarray)
mx_ndarray = mx_ndarray.expand_dims(axis=0)
print(mx_ndarray)
inference_payload = mx_ndarray.astype(dtype="float32")
# prediction/inference
mod.forward(Batch([inference_payload]))
# post-processing
result = mod.get_outputs()[0].asnumpy()
result = np.squeeze(result)
result_exp = np.exp(result - np.max(result))
result = result_exp / np.sum(result_exp)
output_json = json.dumps(result.tolist())
output_content_type = "application/json"
return output_json, output_content_type
def get_context():
mxnet_version = version.parse(mx.__version__)
if mxnet_version >= version.parse("1.8"):
import mx_neuron as neuron
return mx.cpu()
else:
return mx.neuron()
def _load_model(self):
"""
Load MXNet Model
"""
#------- Compile and set context based on device -----------
if self.dev_type == 'inf':
ctx = mx.neuron() #inferentia context
elif self.dev_type == 'gpu':
ctx = mx.gpu(self.dev_id)
else:
ctx = mx.cpu()
#-------Standard MXNet checkpoint load -----------
sym, self.args, self.aux = mx.model.load_checkpoint(os.path.join(self.model_dir_path,
self.model_prefix), int(self.epoch))
#-------Standard MXNet Module instantiation code-----------
self.mod = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
print(self.mod.data_names)
def model_fn(model_dir):
logging.info('Invoking user-defined model_fn')
ctx = mx.neuron()
Batch = namedtuple('Batch', ['data'])
dtype='float32'
# print("param {}".format(os.environ.get('MODEL_NAME_CUSTOM')))
print("ctx {}".format(ctx))
sym, arg_params, aux_params = mx.model.load_checkpoint(os.path.join(model_dir, 'compiled'), 0)
mod = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
for arg in arg_params:
arg_params[arg] = arg_params[arg].astype(dtype)
for arg in aux_params:
aux_params[arg] = aux_params[arg].astype(dtype)
exe = mod.bind(for_training=False,
data_shapes=[('data', (1, 1, 28, 28))],
label_shapes=mod._label_shapes)
mod.set_params(arg_params, aux_params, allow_missing=True)
return mod
import mxnet as mx
#import neomxnet
import os
import json
import numpy as np
from collections import namedtuple
import os
dtype = 'float32'
Batch = namedtuple('Batch', ['data'])
ctx = mx.neuron()
is_gpu = False
def model_fn(model_dir):
print("param {}".format(os.environ.get('MODEL_NAME_CUSTOM')))
print("ctx {}".format(ctx))
sym, arg_params, aux_params = mx.model.load_checkpoint(
os.path.join(model_dir, os.environ.get('MODEL_NAME_CUSTOM')), 0)
mod = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
for arg in arg_params:
arg_params[arg] = arg_params[arg].astype(dtype)
for arg in aux_params:
aux_params[arg] = aux_params[arg].astype(dtype)
exe = mod.bind(for_training=False,
data_shapes=[('data', (1, 3, 224, 224))],
label_shapes=mod._label_shapes)
mod.set_params(arg_params, aux_params, allow_missing=True)
return mod