Workflow

This section describes the layers supported by 2:4 structured sparsity, and API call sequence and example.

Due to hardware restrictions, the Atlas inference product and Atlas training product, and Atlas 350 Accelerator Card do not support the 2:4 structured sparsity feature.

AMCT supports retraining-based 2:4 structured sparsity. The supported layers and their restrictions are as follows. For details about the sparsity sample, see Sample List.

**Table 1** Supported layers and restrictions
Technique	Supported Layer Type	Restriction
2:4 structured sparsity	torch.nn.Linear	Layers sharing the weight do not support sparsity.
	torch.nn.Conv2d	Layers sharing the weight do not support sparsity. The shape of the input data must be (N, Cin, Hin, Win).
	torch.nn.ConvTranspose2d	Layers sharing the weight do not support sparsity. The shape of the input data must be (N, Cin, Hin, Win).

API Call Sequence

The following figure shows the API call sequence of 2:4 structured sparsity. The training runs on the CPU or NPU environment of the PyTorch framework. Based on the inference script of the open-source framework, the AMCT API is called to compress the model. The compressed model needs to be converted into an offline model that adapts to the AI processor using the ATC before it can be used for inference on the AI processor.

Figure 1 API call sequence for 2:4 structured sparsity

The user implements the operations in blue, while those in gray are implemented by using AMCT APIs.

Build an original PyTorch model, call the create_prune_retrain_model API to modify the model, and replace the to-be-sparsified operator with the 2:4 structured sparsity operator.
Train the modified model until the accuracy meets your requirement. If the training is interrupted, call the restore_prune_retrain_model API to prune the original model again based on the sparsity record file and perform QAT until the accuracy meets your requirement.
Restore the replaced operator and prune the weight data by calling the save_prune_retrain_model API based on the retrained 2:4 structured sparsity model, thus to generate the final ONNX fake-quantized model and deployable model.

Example

Take the following steps to get started. Update the sample code based on your situation.
Tweak the arguments passed to AMCT API calls as required. Sparsity relies on the user training result. Ensure that a PyTorch training script that yields satisfactory training accuracy is available.

(Optional) Run inference on the original model in the PyTorch environment based on the test dataset to validate the environment setup and inference script. (Update the sample code based on your situation.)
This step is recommended as it guarantees a properly functioning original model for inference with acceptable accuracy. You can use a subset from the test dataset to improve the efficiency.
1 2 3
ori_model.load() # Test the model. user_test_model(ori_model, test_data, test_iterations)

Run AMCT to execute the 2:4 structured sparsity process.

Modify the original model and replace the to-be-sparsified operator with the 2:4 structured sparsity operator.

Before performing this step, restore the already trained parameters, for example, ori_model.load() in 1.

simple_cfg = './retrain.cfg'
record_file = './tmp/record.txt'
prune_retrain_model = amct.create_prune_retrain_model(model=ori_model,
                                input_data=ori_model_input_data,
                                config_defination=simple_cfg,
                                record_file=record_file)

Implement gradient descent optimization on the modified graph and train the graph on the training dataset. (Update the sample code based on your situation.)
1. Implement gradient descent optimization on the modified graph.
  Perform this step after the model is trimmed.
  1
  optimizer = user_create_optimizer(prune_retrain_model)
2. Restore the model from existing checkpoints and train the model.
  Note: Restore model parameters from checkpoints before training.
  1 2
  quant_pth = './ckpt/user_model' user_train_model(optimizer, prune_retrain_model, train_data)
Call the save_prune_retrain_model API to save the model, restore the replaced operator, and perform structured sparsification on the weight data, thus to generate the final ONNX fake-quantized model and deployable model.
1 2 3 4
prune_retrain_model = amct.save_prune_retrain_model( model=pruned_retrain_model, save_path=save_path, input_data=input_data)

(Optional) Run inference on the sparsified model (prune_retrain_model) in the ONNX Runtime environment based on the test dataset (test_data) to test the accuracy. (Update the sample code based on your situation.)
Check the accuracy drop (from 2:4 structured sparsity) of the fake-quantized model after sparsity by comparing with that of the original model in 1.
1 2
prune_retrain_model = './results/user_model_fake_prune_model.onnx' user_do_inference_onnx(prune_retrain_model, test_data, test_iterations)

If the training is interrupted, restore data from the checkpoints to resume the training.

Prepare an original model.
1
ori_model= user_create_model()

Run AMCT to resume the QAT process.

Modify the model, replace the to-be-sparsified operator with the 2:4 structured sparsity operator, and save the model as a new prune_model.

model = ori_model
input_data = ori_model_input_data
record_file = './tmp/record.txt'
config_defination = './prune_cfg.cfg'
save_pth_path = /your/path/to/save/tmp.pth
model.load_state_dict(torch.load(state_dict_path))
prune_retrain_model = amct.restore_prune_retrain_model(model=ori_model,
                                                       input_data=ori_model_input_data,
                                                       record_file=record_file,
                                                       config_defination='./prune_cfg.cfg',
                                                       save_pth_path=/your/path/to/save/tmp.pth,
                                                       'state_dict')

Implement gradient descent optimization on the modified model, restore parameters from the checkpoints, and train the model on the training dataset. (Update the sample code based on your situation.)
1. Restore the model parameters from the checkpoints after sparsity.
  1 2
  quant_pth = './ckpt/user_prune_model' user_train_model(optimizer, prune_retrain_model, train_data)
2. Implement gradient descent optimization on the modified graph.
  Perform this step after model parameters are restored.
  1
  optimizer = user_create_optimizer(prune_retrain_model)
3. Restore the model from existing checkpoints and train the model.
  Note: Restore model parameters from checkpoints before training.
  1
  user_train_model(optimizer, prune_retrain_model, train_data)
Call the save_prune_retrain_model API to save the model, restore the replaced operator, and perform structured sparsification on the weight data, thus to generate the final ONNX fake-quantized model and deployable model.
1 2 3 4
prune_retrain_model = amct.save_prune_retrain_model( model=pruned_retrain_model, save_path=save_path, input_data=input_data)

(Optional) Run inference on the sparsified model (prune_retrain_model) in the ONNX Runtime environment based on the test dataset (test_data) to test the accuracy. (Update the sample code based on your situation.)
Check the accuracy drop (from 2:4 structured sparsity) of the fake-quantized model after sparsity by comparing with that of the original model in 1.
1 2
prune_retrain_model = './results/user_model_fake_prune_model.onnx' user_do_inference_onnx(prune_retrain_model, test_data, test_iterations)

Parent topic: 2:4 Structured Sparsity