Quantization Example

Prerequisites

• Model

  Upload a TensorFlow model and its weight file to any directory on the Linux server as the AMCT installation user.

  The following uses the sample/resnet_v1_50 model as an example. The pre_model/ResNet50_train.meta file is used for retrain. The pre_model/ResNet50_eval.meta file is used for validation.

• Image dataset

  Because retrain needs to use a large amount of data to further optimize the quantization parameters, the retrain data needs to be consistent with that used to train the model. The ResNet-50 dataset is trained on the subset ILSVRC-2012-CLS of ImageNet. Therefore, you need to prepare a dataset in TFRecord format. To use a different dataset, you need to preprocess the data by yourself.

Procedure

1. Retrain the resnet_v1_50 model.
   Go to the sample/resnet_v1_50 directory and run the following command to retrain the resnet_v1_50 model:

   python3.7.5 resnet_v1_50_retrain_sample.py --train_set TRAIN_SET [--train_keyword TRAIN_KEYWORD] --eval_set EVAL_SET [--eval_keyword EVAL_KEYWORD] --train_model TRAIN_MODEL --eval_model EVAL_MODEL

   The following table describes the available command-line options.

   Table 4-18 Command-line options for retrain

   Option	Description
   -h	Optional. Displays help information.
   --train_set TRAIN_SET	Required. A string. Default: None Sets the directory of the image dataset.
   --train_keyword TRAIN_KEYWORD	Optional. A string. Default: None Selects files containing the keyword in the training dataset directory. If not specified, all files in the training dataset directory are used by default.
   --eval_set EVAL_SET	Required. A string. Default: None Sets the validation dataset directory.
   --eval_keyword EVAL_KEYWORD	Optional. A string. Default: None Selects files containing the keyword in the validation dataset directory. If not specified, all files in the validation dataset directory are used by default.
   --train_model TRAIN_MODEL	Required. A string. Default: ./pre_model/resnet_v1_50_train.meta Sets the directory of the model to retrain.
   --eval_model EVAL_MODEL	Required. A string. Default: ./pre_model/resnet_v1_50_eval.meta Sets the directory of the model for validation.
   --num_parallel_reads NUM_PARALLEL_READS	Optional. An int. Default: 4 Sets the number of threads for reading datasets. Set the argument according to the hardware compute capability.
   --buffer_size BUFFER_SIZE	Optional. An int. Default: 1000 Sets the size for dataset out-of-order buffer. Set this option according to the memory capacity.
   --repeat_count REPEAT_COUNT	Optional. An int. Default: 0 Sets the number of repetitions of the dataset. The argument 0 indicates infinite repetition.
   --batch_size BATCH_SIZE	Optional. An int. Default: 32 Sets the batch size of TensorFlow execution. Set the argument according to the memory or video RAM capacity.
   --ckpt CKPT_PATH	Optional. A string. Default: None. Sets the directory of the benchmark weight checkpoint file of the ResNet-50 v1 model. If not specified, the file is automatically downloaded.
   --learning_rate LEARNING_RATE	Optional. A float. Default: 1e-6 Sets the learning rate.
   --save_interval SAVE_INTERVAL	Optional. An int. Default: 1000 Sets the retrain save interval.
   --momentum MOMENTUM	Optional. A float. Default: 0.9 Sets the momentum of the RMSPropOptimizer.
   --train_iter TRAIN_ITER	Optional. An int. Default: 1000 Sets the number of training iterations.

   An example is as follows:

   python3.7.5 resnet_v1_50_retrain_sample.py  --train_set /data/Datasets/imagenet_TF_trainset/  --eval_set /data/Datasets/imagenet_TF_evalset/ 

   If the upload directory of the test dataset is the same as that of the validation dataset, add the --train_keyword TRAIN_KEYWORD and --eval_keyword EVAL_KEYWORD options to the quantization command to ensure that the --train_set option points to the test dataset and the --eval_set option points to the validation dataset.

   If messages similar to the following are displayed, the retrain is successful:

   The model after retraining top 1 accuracy = 68.14%.
   The model after retraining top 5 accuracy = 88.03%.

2. View the retrain result.
   After the retrain is successful, find the following folders and files generated to the sample/resnet_v1_50 directory:

   • amct_log: AMCT log folder, including the retrain log amct_tensorflow.log.
   • results: retrain result folder, containing the quantized model file resnet_v1_50_quantized.pb.

     This model can be used for accuracy simulation in the TensorFlow environment and can also be deployed on the Ascend AI Processor.

   • tmp: retrain temporary folder, containing:
     • checkpoint: checkpoints saved during TensorFlow training.
     • config.json: retrain-based quantization configuration file that describes how to retrain each layer in the model. If a retrain configuration file already exists in the directory where the retrain script is located, when create_quant_retrain_config is called again, the existing retrain configuration file is overwritten if the name of the new retrain configuration file is the same as that of the existing file. Otherwise, a new retrain configuration file is generated.
     • events.out.tfevents.xxxxxxxxxx.xxx: file that describes the retrain loss information, which can be viewed on the TensorBoard.
     • record.txt: file that records quantization factors. For details about the prototype definition of the file, see Quantization Factor Record File.
     • resnet_v1_50.pb: retrained model file.
     • resnet_v1_50_retrain-0.data-00000-of-00001: weight file generated during retrain.
     • resnet_v1_50_retrain-0.index: index of the weight file generated during retrain.
     • resnet_v1_50_retrain-0.meta: model file to be retrained.
     • resnet_v1_50_retrain-1.data-00000-of-00001: weight file generated during retrain.
     • resnet_v1_50_retrain-1.index: index of the weight file generated during retrain.
     • resnet_v1_50_retrain-1.meta: model file to be retrained.

   When the model is retrained, the existing result files will be overwritten.