- Download the zhwiki dataset for pre-training. Extract and clean text in the dataset with [WikiExtractor](https://github.com/attardi/wikiextractor). Convert the dataset to TFRecord format and move the files to a specified path.
- Download dataset for fine-tuning and evaluation such as CLUENER, TNEWS, SQuAD v1.1, etc.
-[Quick Start](#Quick-Start)
> Notes:
-[Script Description](#Script-Description)
If you are running a fine-tuning or evaluation task, prepare a checkpoint from pre-train.
-[Script and Sample Code](#Script-Code-Structure)
-[Script Parameters](#Script-Parameters)
## Running the Example
-[Training Process](#Training-Process)
### Pre-Training
-[Evaluation Process](#Evaluation-Process)
- Set options in `config.py`, including optimizer and network. Click [here](https://www.mindspore.cn/tutorial/zh-CN/master/use/data_preparation/loading_the_datasets.html#tfrecord) for more information about dataset and the json schema file.
- Run `run_standalone_pretrain.sh` for non-distributed pre-training of BERT-base, BERT-NEZHA and BERT-large model.
-[Description of Random Situation](#Description-of-Random-Situation)
-[ModelZoo Homepage](#ModelZoo-Homepage)
``` bash
sh scripts/run_standalone_pretrain.sh DEVICE_ID EPOCH_SIZE DATA_DIR SCHEMA_DIR
## Description
```
- Run `run_distribute_pretrain.sh` for distributed pre-training of BERT-base, BERT-NEZHA and BERT-large model.
This is an example of training Bert with MLPerf v0.7 dataset by second-order optimizer THOR. THOR is a novel approximate seond-order optimization method in MindSpore. With fewer iterations, THOR can finish Bert-Large training in 14 minutes to a masked lm accuracy of 71.3% using 8 Ascend 910, which is much faster than SGD with Momentum.
``` bash
## Model Architecture
sh scripts/run_distribute_pretrain.sh DEVICE_NUM EPOCH_SIZE DATA_DIR SCHEMA_DIR RANK_TABLE_FILE
The architecture of Bert contains 3 embedding layers which are used to look up token embeddings, position embeddings and segmentation embeddings; Then BERT basically consists of a stack of Transformer encoder blocks; finally bert are trained for two tasks: Masked Language Model and Next Sentence Prediction.
--enable_save_ckpt enable save checkpoint: "true" | "false", default is "true"
> ```
--enable_lossscale enable lossscale: "true" | "false", default is "true"
--do_shuffle enable shuffle: "true" | "false", default is "true"
--enable_data_sink enable data sink: "true" | "false", default is "true"
## Features
--data_sink_steps set data sink steps: N, default is 1
The classical first-order optimization algorithm, such as SGD, has a small amount of computation, but the convergence speed is slow and requires lots of iterations. The second-order optimization algorithm uses the second-order derivative of the target function to accelerate convergence, can converge faster to the optimal value of the model and requires less iterations. But the application of the second-order optimization algorithm in deep neural network training is not common because of the high computation cost. The main computational cost of the second-order optimization algorithm lies in the inverse operation of the second-order information matrix (Hessian matrix, FIM, etc.), and the time complexity is about O (n^3). On the basis of the existing natural gradient algorithm, we developed the available second-order optimizer THOR in MindSpore by adopting approximation and shearing of FIM information matrix to reduce the computational complexity of the inverse matrix. With eight Ascend 910 chips, THOR can complete Bert-Large training in 14min.
--save_checkpoint_path path to save checkpoint files: PATH, default is ""
--save_checkpoint_steps steps for saving checkpoint files: N, default is 1000
## Environment Requirements
--save_checkpoint_num number for saving checkpoint files: N, default is 1
- Hardware（Ascend/GPU）
--data_dir path to dataset directory: PATH, default is ""
- Prepare hardware environment with Ascend or GPU processor. If you want to try Ascend , please send the [application form](https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/file/other/Ascend%20Model%20Zoo%E4%BD%93%E9%AA%8C%E8%B5%84%E6%BA%90%E7%94%B3%E8%AF%B7%E8%A1%A8.docx) to ascend@huawei.com. Once approved, you can get the resources.
--schema_dir path to schema.json file, PATH, default is ""
- For more information, please check the resources below：
It contains of parameters of BERT model and options for training, which is set in file `config.py`, `bert_net_config.py` and `evaluation_config.py` respectively.
Parameters for dataset and network (Pre-Training/Evaluation):
# run evaluation example
batch_size batch size of input dataset: N, default is 12
pythonpretrain_eval.py
seq_length length of input sequence: N, default is 128
```
vocab_size size of each embedding vector: N, must be consistant with the dataset you use. Default is 21136
> For distributed training, a hccl configuration file with JSON format needs to be created in advance. About the configuration file, you can refer to the [HCCL_TOOL](https://gitee.com/mindspore/mindspore/tree/master/model_zoo/utils/hccl_tools).
hidden_size size of bert encoder layers: N, default is 768
num_hidden_layers number of hidden layers: N, default is 12
## Script Description
num_attention_heads number of attention heads: N, default is 12
intermediate_size size of intermediate layer: N, default is 3072
### Script Code Structure
hidden_act activation function used: ACTIVATION, default is "gelu"
hidden_dropout_prob dropout probability for BertOutput: Q, default is 0.1
```shell
attention_probs_dropout_prob dropout probability for BertAttention: Q, default is 0.1
├── model_zoo
max_position_embeddings maximum length of sequences: N, default is 512
├──official
type_vocab_size size of token type vocab: N, default is 16
├──nlp
initializer_range initialization value of TruncatedNormal: Q, default is 0.02
├── bert_thor
use_relative_positions use relative positions or not: True | False, default is False
├── README.md # descriptions bert_thor
input_mask_from_dataset use the input mask loaded form dataset or not: True | False, default is True
├── scripts
token_type_ids_from_dataset use the token type ids loaded from dataset or not: True | False, default is True
├── run_distribute_pretrain.sh # launch distributed training for Ascend
dtype data type of input: mstype.float16 | mstype.float32, default is mstype.float32
└── run_standalone_pretrain.sh # launch single training for Ascend
compute_type compute type in BertTransformer: mstype.float16 | mstype.float32, default is mstype.float16
├──src
├── bert_for_pre_training.py # Bert for pretraining
Parameters for optimizer:
├── bert_model.py # Bert model
Thor:
├── bert_net_config.py # network config setting
momentum momentum for the moving average: Q
├── config.py # config setting used in dataset.py
weight_decay weight decay: Q
├── dataset.py # Data operations used in run_pretrain.py
loss_scale loss scale: N
├── dataset_helper.py # Dataset help for minddata dataset
frequency the step interval to update second-order information matrix: N, default is 100
├── evaluation_config.py # config settings, will be used in finetune.py
batch_size batch size of input dataset: N, default is 12
├── fused_layer_norm.py # fused layernorm
```
├── grad_reducer_thor.py # grad_reducer_thor
├── lr_generator.py # learning rate generator
├── model_thor.py # Model
├── thor_for_bert.py # thor_for_bert
├── thor_for_bert_arg.py # thor_for_bert_arg
├── thor_layer.py # thor_layer
└── utils.py # utils
├── pretrain_eval.py # infer script
└── run_pretrain.py # train script
```
### Script Parameters
Parameters for both training and inference can be set in config.py.
```
"device_target": 'Ascend', # device where the code will be implemented
"distribute": "false", # Run distribute
"epoch_size": "1", # Epoch size
"enable_save_ckpt": "true", # Enable save checkpoint
"enable_lossscale": "false", # Use lossscale or not
"do_shuffle": "true", # Enable shuffle for dataset
"save_checkpoint_path": "", # Save checkpoint path
"train_steps": -1, # meaning run all steps according to epoch number
"device_id": 4, # Device id, default is 4
"enable_data_sink": "true", # Enable data sink, default is true
"data_sink_steps": "100", # Sink steps for each epoch, default is 100
"save_checkpoint_steps",: 1000, # Save checkpoint steps
"save_checkpoint_num": 1, # Save checkpoint numbers, default is 1
```
### Training Process
#### Ascend 910
```
sh run_distribute_pretrain.sh [DEVICE_NUM] [EPOCH_SIZE] [DATA_DIR] [SCHEMA_DIR] [RANK_TABLE_FILE]
```
We need three parameters for this scripts.
-`DEVICE_NUM`: the device number for distributed train.
-`EPOCH_SIZE`: Epoch size used in the model
-`DATA_DIR`：Data path, it is better to use absolute path.
-`SCHEMA_DIR `：Schema path, it is better to use absolute path
-`RANK_TABLE_FILE`: the path of rank_table.json
Training result will be stored in the current path, whose folder name begins with the file name that the user defines. Under this, you can find checkpoint file together with result like the followings in log.
```
...
epoch: 1, step: 1, outputs are [5.0842705], total_time_span is 795.4807660579681, step_time_span is 795.4807660579681
epoch: 1, step: 100, outputs are [4.4550357], total_time_span is 579.6836116313934, step_time_span is 5.855390016478721
epoch: 1, step: 101, outputs are [4.804837], total_time_span is 0.6697461605072021, step_time_span is 0.6697461605072021
epoch: 1, step: 200, outputs are [4.453913], total_time_span is 26.3735454082489, step_time_span is 0.2663994485681707
epoch: 1, step: 201, outputs are [4.6619444], total_time_span is 0.6340286731719971, step_time_span is 0.6340286731719971
epoch: 1, step: 300, outputs are [4.251204], total_time_span is 26.366267919540405, step_time_span is 0.2663259385812162
epoch: 1, step: 301, outputs are [4.1396527], total_time_span is 0.6269843578338623, step_time_span is 0.6269843578338623
epoch: 1, step: 400, outputs are [4.3717675], total_time_span is 26.37460947036743, step_time_span is 0.2664101966703781
epoch: 1, step: 401, outputs are [4.9887424], total_time_span is 0.6313872337341309, step_time_span is 0.6313872337341309
epoch: 1, step: 500, outputs are [4.7275505], total_time_span is 26.377585411071777, step_time_span is 0.2664402566774927
......
epoch: 3, step: 2001, outputs are [1.5040319], total_time_span is 0.6242287158966064, step_time_span is 0.6242287158966064
epoch: 3, step: 2100, outputs are [1.232682], total_time_span is 26.37802791595459, step_time_span is 0.26644472642378375
epoch: 3, step: 2101, outputs are [1.1442064], total_time_span is 0.6277685165405273, step_time_span is 0.6277685165405273
epoch: 3, step: 2200, outputs are [1.8860981], total_time_span is 26.378745555877686, step_time_span is 0.2664519753118958
epoch: 3, step: 2201, outputs are [1.4248213], total_time_span is 0.6273438930511475, step_time_span is 0.6273438930511475
epoch: 3, step: 2300, outputs are [1.2741681], total_time_span is 26.374130964279175, step_time_span is 0.2664053632755472
epoch: 3, step: 2301, outputs are [1.2470423], total_time_span is 0.6276984214782715, step_time_span is 0.6276984214782715
epoch: 3, step: 2400, outputs are [1.2646998], total_time_span is 26.37843370437622, step_time_span is 0.2664488252967295
epoch: 3, step: 2401, outputs are [1.2794371], total_time_span is 0.6266779899597168, step_time_span is 0.6266779899597168
epoch: 3, step: 2500, outputs are [1.265375], total_time_span is 26.374578714370728, step_time_span is 0.2664098860037447
...
```
### Evaluation Process
Before running the command below, please check the checkpoint path used for evaluation. Please set the checkpoint path to be the absolute full path, e.g., "username/bert_thor/LOG0/checkpoint_bert-3_1000.ckpt".
#### Ascend 910
```
python pretrain_eval.py
```
We need two parameters in evaluation_config.py for this scripts.
-`DATA_FILE`：the file of evaluation dataset.
-`FINETUNE_CKPT`: the absolute path for checkpoint file.
> checkpoint can be produced in training process.
Inference result will be stored in the example path, you can find result like the followings in log.
