Help Center> ModelArts> User Guide (Senior AI Engineers)> Training Management> Auto Search Jobs> Example: Using Backbone Replacement to Optimize Built-in Algorithms on Ascend 910 Without Modifying Code
View PDF

Example: Using Backbone Replacement to Optimize Built-in Algorithms on Ascend 910 Without Modifying Code

On Ascend 910, ModelArts finds some better structures than the original ResNet-50. This example describes how to use the found structures to optimize the ResNet-50 of the built-in algorithms without modifying code.

Sample Code

The following sample code uses a built-in algorithm to train the flower dataset on the ResNet-50 network.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
 
import argparse
import os
import json
import sys
import moxing as mox
from moxing.tensorflow.builtin_algorithms.algorithm_common import AlgorithmLego
app_url = os.path.dirname(os.path.dirname(__file__))
sys.path.insert(0, os.path.dirname(__file__))
from algo_util import data_util

parser = argparse.ArgumentParser(description='')
parser.add_argument('--task_type', type=str, default=None, help='')
parser.add_argument('--train_url', type=str, default=None, help='')

def main():
  args, _ = parser.parse_known_args()
  print ("args", args)
  flag_dict = data_util.do_preprocess()

  alego = AlgorithmLego()
  alego.add(args.task_type, flag_dict=flag_dict)
  alego.start()

  mox.file.shift("os", "mox")
  with open("{}/metric.json".format(args.train_url)) as f:
    metric = json.load(f)
    accuracy = metric["total_metric"]["total_metric_values"]["accuracy"]
    print("accuracy:%d." % accuracy)

if __name__ == "__main__":
  main()

Compiling the Configuration File

Compile the YAML configuration file. For details, see Modifying Reporter Without Modifying Code in Code Compilation Specifications and Example: Replacing the Original ResNet-50 with a Better Network Architecture.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
 
general:
  cpu_per_instance: 1
  gpu_per_instance: 0
  npu_per_instance: 1

search_space:
  builtin: ResNet50

search_algorithm:
  type: grid_search
  reward_attr: accuracy
  report_keys:
    - name: accuracy
      regex: (?<=accuracy:).+(?=.)
scheduler:
  type: FIFOScheduler

Starting a Search Job

In this example, you need to prepare a dataset based on the input requirements of the built-in algorithm, upload and configure the dataset following the instructions in Example: Replacing the Original ResNet-50 with a Better Network Architecture, upload the Python script and YAML configuration file, and select the Ascend chip-powered search engine based on the section "Creating an Auto Search Job". Then start the search job on the Ascend chip.

Parent topic: Auto Search Jobs