Platform For AI:Optimize a TensorFlow model

Procedure

This topic demonstrates how to optimize a TensorFlow model using a public ResNet50 model. Adapt these steps for your own TensorFlow models.

1. Import PAI-Blade and dependencies.

   import os
   import numpy as np
   import tensorflow.compat.v1 as tf
   import blade

2. Write a function to download the model and test data.

   PAI-Blade supports optimization without test data (zero-input optimization), but results are more accurate with real input data. Provide test data when possible. The following sample code downloads the model and test data.

   def _wget_demo_tgz():
       # Download a public ResNet50 model.
       url = 'http://pai-blade.oss-cn-zhangjiakou.aliyuncs.com/demo/mask_rcnn_resnet50_atrous_coco_2018_01_28.tar.gz'
       local_tgz = os.path.basename(url)
       local_dir = local_tgz.split('.')[0]
       if not os.path.exists(local_dir):
           blade.util.wget_url(url, local_tgz)
           blade.util.unpack(local_tgz)
       model_path = os.path.abspath(os.path.join(local_dir, "frozen_inference_graph.pb"))
       graph_def = tf.GraphDef()
       with open(model_path, 'rb') as f:
           graph_def.ParseFromString(f.read())
       # Use random numbers as test data.
       test_data = np.random.rand(1, 800,1000, 3)
       return graph_def, {'image_tensor:0': test_data}
   
   graph_def, test_data = _wget_demo_tgz()

3. Call blade.optimize to optimize the model. For parameter details, see Python API reference.

   input_nodes=['image_tensor']
   output_nodes = ['detection_boxes', 'detection_scores', 'detection_classes', 'num_detections', 'detection_masks']
   
   optimized_model, opt_spec, report = blade.optimize(
       graph_def,                 # Model to optimize. In this example, a tf.GraphDef object. Can also be a SavedModel path.
       'o1',                      # Optimization level. Valid values: o1 and o2.
       device_type='gpu',         # Target device. Valid values: gpu, cpu, and edge.
       inputs=input_nodes,        # Input nodes. Optional. PAI-Blade automatically infers if not specified.
       outputs=output_nodes,      # Output nodes.
       test_data=[test_data]      # Test data.
   )

   blade.optimize returns three objects:

   • optimized_model: The optimized model. A tf.GraphDef object in this example.

   • opt_spec: Configuration information, environment variables, and resource files required to reproduce optimization results. Apply using a with statement.

   • report: Optimization report. Print directly to view. For parameter details, see Optimization report.

   Progress information during optimization:

   [Progress] 5%, phase: user_test_data_validation.
   [Progress] 10%, phase: test_data_deduction.
   [Progress] 15%, phase: CombinedSwitch_1.
   [Progress] 24%, phase: TfStripUnusedNodes_22.
   [Progress] 33%, phase: TfStripDebugOps_23.
   [Progress] 42%, phase: TfFoldConstants_24.
   [Progress] 51%, phase: CombinedSequence_7.
   [Progress] 59%, phase: TfCudnnrnnBilstm_25.
   [Progress] 68%, phase: TfFoldBatchNorms_26.
   [Progress] 77%, phase: TfNonMaxSuppressionOpt_27.
   [Progress] 86%, phase: CombinedSwitch_20.
   [Progress] 95%, phase: model_collecting.
   [Progress] 100%, Finished!

4. Print the optimization report.

   print("Report: {}".format(report))

   The report shows which optimization items contribute most to performance improvement.

   Report: {
     // ......
     "optimizations": [
       // ......
       {
         "name": "TfNonMaxSuppressionOpt",
         "status": "effective",
         "speedup": "1.58",        // Acceleration ratio.
         "pre_run": "522.74 ms",   // Latency before optimization.
         "post_run": "331.45 ms"   // Latency after optimization.
       },
       {
         "name": "TfAutoMixedPrecisionGpu",
         "status": "effective",
         "speedup": "2.43",
         "pre_run": "333.30 ms",
         "post_run": "136.97 ms"
       }
       // ......
     ],
     // End-to-end optimization results.
     "overall": {
       "baseline": "505.91 ms",    // Latency of the original model.
       "optimized": "136.83 ms",   // Latency of the optimized model.
       "speedup": "3.70"           // Acceleration ratio.
     },
     // ......
   }

5. Compare the performance before and after optimization.

   import time
   
   def benchmark(model):
       tf.reset_default_graph()
       with tf.Session() as sess:
           sess.graph.as_default()
           tf.import_graph_def(model, name="")
           # Warmup!
           for i in range(0, 1000):
               sess.run(['image_tensor:0'], test_data)
           # Benchmark!
           num_runs = 1000
           start = time.time()
           for i in range(0, num_runs):
               sess.run(['image_tensor:0'], test_data)
           elapsed = time.time() - start
           rt_ms = elapsed / num_runs * 1000.0
           # Show the result!
           print("Latency of model: {:.2f} ms.".format(rt_ms))
   
   # Test the speed of the original model.
   benchmark(graph_def)
   
   # Test the speed of the optimized model.
   with opt_spec:
       benchmark(optimized_model)

   Performance test results. The results match the values in the optimization report.

   Latency of model: 530.26 ms.
   Latency of model: 148.40 ms.

Extensions

Troubleshooting

Next steps

After optimizing a model, execute it directly in Python or deploy it as an EAS service. PAI-Blade also provides a C++ SDK to integrate optimized models into applications. For more information, see Use an SDK to deploy a TensorFlow model for inference.