TopSpin 1.1.5

A PyTorch based high-level Deep Learning training framework. Seamlessly switch between single-GPU and multi-server; Supports gradient accumulation, learning rate warmup and decay, early stop, mixed precision training, and AutoML.

TopSpin - BERT-style Deep Learning Training Framework

TopSpin, is a PyTorch based high-level Deep Learning training framework, designed for BERT-style deep learning models, as opposed to GPT3 based deep learning. User defined model can be launched in either single-GPU, multi-GPU, or multi-server, without changing one line of codes.

TopSpin supports many useful techiniques in Deep Learning Training, such as gradient accumulation, learning rate warmup and decay, early stop, mixed precision training, parameter regularizers like VAT automatic validation data evaluation in training, TensorBoard, True gradient for variable batch size in some NLP applications, and provide rich statistics in training log, like network time statistics, traininig time prediction.

TopSpin, besides, provides inherent AutoML support that searches possible model parameter combinations in one or mutliple servers in parallel.

1. Install

python3 -m pip install TopSpin

2. Example

• NLP
  • Intent detection Task
• ASR
  • EDSR
• CV
  • Speaker Change detection Task

3. Core Components

4. Training Run Modes

mode 1: Debug Run

mode 2: Auto ML

mode 3: Distributed

5. Stop and Restore training

6. Miscellaneous Functions

1. Learning Rate Warmup and Decay
2. Early Stop
3. Mixed Precision Training
4. Gradient Accumulation
5. Parameter Regularization
6. Automatic Validation Data Evaluation in Training
7. TensorBoard
8. True gradient in Training
9. Rich statistics in Training Log

7. Config Your Servers

8. Common Questions and Issues

Version 5 What's New?

1. More safe

Before excuting GPU running, TopSpin would examine the availability of desired GPUs and accessibility of data folder.

2. Supportive of more rich information returned in train_one_batch.

Now the user defined function can return a dict,

 return {
   "loss": loss,
   "batch_num": b_x.size(0), # depends on user's application.
   "figure": figure          # optional
 }

If your batch_num is just the number of sentences, then you could ignore this value; otherwise, in some cases this number actually equating to word number, you have to set it exactly.

Espeically, the key figure records all information to draw by TopSpin automatically.

3. Automatic loss drawing in training

figure = {
 "F1-score": 0.34,
 "accuracy": 0.46
}

as well as any information desired to show in a figure. TopSpin would append your training loss, as well as validation and testing data errors by default.

4. Rigid definition on model initialization

class ModelWrapperBase:
 def __init__(self, param: ParamBase, user_model_cls):
  ...
  model = user_model_cls(param)
  ...

In comparison, lower versions permit passing a user model instance into ModelWrapperBase.

===============================================================================

Version 4

Section 1 - About TopSpin

Since starting to develop 2 years ago, it has been evolved for 5 versions , including one for tensorflow 1.X, one for tensorflow 2.X, three versions for PyTorch. This README mainly introduce the most recent version, namely estimator4 , as it contains the richest functions among all versions.

1. For managers, TLs, and team members

We list only interfaces of main modules in estimator4, with their private functions omitted.

a) palframe.pytorch.estimator4.param.ParamBase

class ParamBase(abc.ABC):

  def __init__(self, run_tag: str, restore_training_from_path_work=None): 
    # Many default parameters are defined
    ...
    self.path_initial_model = None
    self.path_inference_model = None

    self.optimizer_name = "Adam"
    self.lr = 0.001
    self.weight_decay = 0
    self.param_norm = 1
    self.seed = 0     # 0 means random.

    self.servers_file = None  # for distributed training
    self.gpu_num = 1          # gpus required by your task
    self.use_gpu = True
    self.gpus = [0]           # do NOT set it, unless you use debug mode. 
    self.use_amp = True       # mixed precision based training

    self.batch_dim = 0

    self.iter_num_update_optimizer = 1  # gradient accumulation

    self.train_files = None     # file num >= 1
    self.vali_file = None       # file_num <= 1
    self.test_files = None      # file_num >= 0

    self.train_sample_num = None
    self.epoch_num = None       

    self.eval_gap_sample_num = None   # frequentcy to evaluate dev set.

    self.warmup_ratio = 0.1           # learning rate warm up.
    self.ending_lr_ratio = 0.001      # 1 means no lr decay

    self.model_saved_num = 3

    self.num_workers_loading_data = 4

    # None value denotes no limits on the maximum model size, or you should
    # set a value.
    self.automl_max_model_size = None

    self.true_gradient = False  # if data in each GPU is unbalenced.

    self.debug_level = 1        # debug=0, info=1, warning=2, error=3

    self.detect_anomaly = False # only for debugging.

b) palframe.pytorch.estimator4.model_wrapper.ModelWrapper

class ModelWrapperBase:
  def __init__(self, param: ParamBase, model: torch.nn.Module):
    ...

  def evaluate_file(self, data_file) -> float:
    # return an evaluation result. Smaller, better.
    ...

  def predict(self, *batch):
    # Just invoke it
    return self._model(*batch)
    ...

c) palframe.pytorch.estimator4.train.TrainerBase

class TrainerBase:
  def __init__(self,
               model_wrapper: ModelWrapperBase,
               train_data_iter,  
               optimizer: typing.Union[Optimizer, None, str]=None):
    ...

  def train_one_batch(self, *batch)-> tensor:
    # Given a batch=[batch_x, batch_y], return a loss tensor
    ...

  def early_stop(self, epoch_id, loss_history: list, vali_error_history: list):
    # override it when necessary
    return False  # default

  def train(self):
    # Just invoke it
    ...

d) palframe.pytorch.estimator4.predict.PredictorBase

class PredictorBase:
  def __init__(self,
               model_wrapper: ModelWrapperBase):
    # You do NOT need to change it.

  def evaluate_file(self, data_file) -> float:
    # You do NOT need to change it.
    with torch.no_grad():
      return self._model_wrapper.evaluate_file(data_file)

  def predict(self, batch_data):
    # You do NOT need to change it.
    with torch.no_grad():
      return self._model_wrapper.predict(batch_data)

2. AutoML

Use ParameterRange to mark all parameters and set their candidate values, then TopSpin would search all combinations for you in the backhand.

    self.lr = ParameterRange([5e-4, 7e-4])
    self.warmup_ratio = ParameterRange([0.1, 0.2, 0.3])
    self.hidden_dim = ParameterRange([100, 200, 300])

3. Automatic GPUs and servers allocation

We assume a single task requires 2 GPUs, then we run it on my current server and use GPU 2 and 3.

param = MyParam.get_instance(),
param.gpu_num = 2    

starter.start_train(
  param,
  "nlp_tasks/task00_intent/ver_2_2_7_9/train.py",
  [starter.Server(None, [2, 3])]
)

Let continue with MyParam definition that uses AutoML and it finally generates 2x3x3=18 parameter variants. Then we have only one server with GPU 2, 3 available. Let's assume one task takes 1 hour, then we have to wait for 18 hours. The waiting looks boring ...

Luckily, Shuang brother say, my task is done, and you can use GPU 0, 1 of that server. We update our starting script like this.

param = MyParam.get_instance(),
param.gpu_num = 2    

starter.start_train(
  param,
  "nlp_tasks/task00_intent/ver_2_2_7_9/train.py",
  [starter.Server(None, [0, 1, 2, 3])]
)

In this case, two tasks can be run in parallel. Hence, our waiting time is reduced to 18/(4/2)=9 hours. Life looks a bit better.

More luckily, xiaozhao sister said, my task is also done, you can use RTX8000 that has 8 GPUs. We further update our script like this,

param = MyParam.get_instance(),
param.gpu_num = 2    

starter.start_train(
  param,
  "nlp_tasks/task00_intent/ver_2_2_7_9/train.py",
  [starter.Server(None, [0, 1, 2, 3]),
   starter.Server("192.168.1.228", [0, 1, 2, 3, 4, 5, 6, 7])]
)

Then our waiting time becomes only 18/((4+8)/2)=3 hours. Life should be much better.

4. Built-in support for distributed training

User: How to update my codes to a distributed version?

TopSpin: Sir, you are all set. Try a different starting command.

param = MyParam.get_instance(),
param.servers_file = "my_cluster_servers.txt"   # You tell TopSpin this.
param.gpu_num = 8

starter.start_distributed_train(
  param,
  "patbert/ver_2_2_7_9/train.py",
)

Section 2 - An step-by-step example

Step 1. Configure your environment

Configure your PyCharam

Add PAL_Frame folder to your Content Root.

Set PYTHONPATH in Mac

If you run in your Mac terminal, then you need to add to ~/.profile.

export PYTHONPATH={your-PAL-frame-folder}/PAL_frame:$PYTHONPATH
export PYTHONPATH=./:$PYTHONPATH

Test the new settings

>> source ~/.profile
>> echo $PYTHONPATH

You should see the correct values.

Set PYTHONPATH in Linux

Add the above codes to ~/.bashrc, and repeat the testing.

>> source ~/.bashrc
>> echo $PYTHONPATH

Step 2. Define param.py

class Param(ParamBase):
  def __init__(self):
    super(Param, self).__init__("nlp_example.intent_dection")

    # standard parameters derived from ParamBase
    self.gpu_num = 2                    # required GPU number.
    self.use_gpu = False
    self.use_amp = False                # mixed precision based training
    self.optimizer_name = "Adam"
    self.lr = 5e-4
    self.batch_size = 16                # for one worker
    self.iter_num_update_optimizer = 1  # gradient accumulation
    self.warmup_ratio = 0.1             # lr warm percentage of all steps.
    self.model_kept_num = 5             # save you models
    self.path_feat = "feat/nlp/intent_dection"
    self.train_files = [f"{self.path_feat}/train.pkl"]
    self.vali_file = f"{self.path_feat}/vali.pkl"
    self.test_files = []
    self.train_sample_num = 100
    self.eval_gap_sample_num = 100      # sample number between two evaluations.
    self.epoch_num = 5

    # user model specific parameters 
    self.max_seq_len = 128
    self.class_number = 13
    self.embedding_size = 128
    self.kernel_sizes = [3, 4, 5]
    self.kernel_number = 128
    self.dropout_ratio = 0.3
    self.vocab_size = 60000

Step 3. Define model.py

class Model(nn.Module):
  def __init__(self, param: Param):
    super(Model, self).__init__()

    self._embedding = nn.Embedding(param.vocab_size, param.embedding_size)

    self._textcnn = nlp_torch.TextCNN(
      kernels=param.kernel_sizes,
      in_channel=1,
      out_channel=param.kernel_number,
      max_seq_len=param.max_seq_len,
      dim=param.embedding_size,
      dropout=param.dropout_ratio
    )
    self._textcnn_output_size = len(param.kernel_sizes) * param.kernel_number

    self._dense = nlp_torch.Dense(
      nn.Linear(self._textcnn_output_size, param.class_number)
    )

  def forward(self, word_ids):
    word_ids = word_ids.unsqueeze(1)
    embedding_out = self._embedding(word_ids)
    textcnn_out = self._textcnn(embedding_out)
    out = self._dense(textcnn_out)
    pred_labels = torch.argmax(out, 1)

    return out, pred_labels

Step 4. Define model_wrapper.py

class ModelWrapper(ModelWrapperBase):
  def __init__(self, param: Param):
    super(ModelWrapper, self).__init__(
      param, Model(param)
    )

  def evaluate_file(self, data_file: str):
    '''
    :param data_file:
    :return: smaller better, such as WER, -F1-value, -Accuracy.
    '''
    start_time = time.time()
    all_true_labels = []
    all_pred_labels = []
    for _, batch in get_batch_data(self._param, [data_file], 1, False):
      batch = [e.to(self._device) for e in batch]

      b_word_ids, b_labels = batch
      logits, pred_labels = self.predict(b_word_ids)

      all_true_labels.extend(b_labels.tolist())
      all_pred_labels.extend(pred_labels.tolist())

    result = Measure.calc_precision_recall_fvalue(
      all_true_labels, all_pred_labels
    )
    total_time = time.time() - start_time
    avg_time = total_time / (len(all_true_labels) + 1e-6)
    weighted_f = result["weighted_f"]
    Logger.info(
      f"eval: "
      f"file={data_file} weighted_f={weighted_f} result={result} "
      f"total_time={total_time:.4f} secs avg_time={avg_time:.4f} sec/sample "
    )
    Logger.info(f"WEIGHTED_F : {weighted_f}")

    return -weighted_f

  def predict(self, b_word_ids):
    logits, pred_labels = self._model(b_word_ids)
    return logits, pred_labels

Step 5. Define train.py

class Trainer(TrainerBase):
  def __init__(self, param):

    super(Trainer, self).__init__(
      ModelWrapper(param),
      get_batch_data(param, param.train_files, param.epoch_num, True),
      None
    )

  def train_one_batch(self, b_word_ids, b_label):
    logits, pred_labels = self._model_wrapper.predict(b_word_ids)
    return nn.functional.cross_entropy(logits, b_label, reduction="mean")

Step 6. Run mode1: Debug Run

>> python3 example/nlp/intent_dection/train.py 

By design, quck run mode is for debugging, though you could run for simple tasks which require only 1 GPU. If you use GPU in a server, you should make sure param.gpus = [0] is available.

This mode is not supportive of GPU allocation, server allocation, task parallelization. Strongly encourage you guys to use mode2, auto run.

Step 7. AutoML

    # batch size does influence the final performance.
    self.iter_num_update_optimizer = ParameterRange([1, 2, 3]) 


    self.embedding_size = ParameterRange([128, 256, 512]) 
    self.kernel_sizes = ParameterRange([[3, 4, 5], [1, 2, 3, 4, 5], [5, 6, 7]])
    self.kernel_number = ParameterRange([128, 256]) 
    self.dropout_ratio = ParameterRange([0.1, 0.3])

    # even you could AutoML training data.
    self.train_files = ParameterRange(["train.1.pkl", "train.2.pkl"])

Step 8. Run mode2: Auto Run - for AutoML or batch tasks

By convention, we write a train_starter.py

  starter.start_train(
    Param.get_instance(), 
    "example/nlp/intent_dection/train.py",
    [starter.Server(None, [1, 3, 4, 6, 7])]
  )

Step 9. Run mode3: Dist Run - for large training

Set your server_file in your param.py, and use a different starting function

  starter.start_distributed_train(
    Param.get_instance(),
    "example/nlp/intent_dection/train.py",
  )

Step 10. Read your log

Run.

>> nohup python3 example/nlp/intent_dection/train_starter.py > log.any_name &

In current folder, log.any_name only tell you if your running is normal. You should check work/{run_dir}/log/log.rank_0 to read outputs of your model. {run_dir} can be found in log.any_name.

When your running fails, you can check work/{run_dir}/log/log.node_0, and the problematic batch data and model are automatically saved into work/{run_dir }/bug/ for your bug replay.

Step 11. Stop a training

There is a unique method to stop your training, regardless of any starting mode.

python3 palframe/pytorch/estimator4/stopper.py {ParamBase.path_work}

Step 12. More advanced usage

You could have different models run together in a server pool to use your resources to an extreme. Use RunManager class.

  run_manager = RunManager(tasks, servers)
  run_manager.run()

See definition of Task and Server in palframe/pytorch/estimator4/starter.py.

class Server:
  def __init__(self, ip, gpus):
    ...

class Task:
  def __init__(self, param: ParamBase, source_script_and_params):
    ...

In the PAT-BERT project, one pretrained model would be tested in 10 downstream NLP tasks, and each task has multiple set of parameter configurations. We typically use RunManager to finish 216 tasks in 180 GPUs in just 16 hours.

Step 13. Full operations for the above NLP example

>> python3 example/nlp/intent_dection/make_features.py 
>> nohup python3 example/nlp/intent_dection/train_starter.py > log.any_name &

Section 3. Miscellaneous functions

1. Single-node multi-GPU support

2. Distributed training

3. Automaic GPUs and servers allocation for task parallelization

4. AutoML

5. Three running modes

No any change for your codes.

1. Debug mode, for debugging only, which uses single GPU (or CPU).
2. Auto mode, your common choice, which supports automatic GPUs and servers allocation.
3. Distributed mode, for quite complex tasks that need more than one server.

6. Learning rate warmup and decay

7. Training restoration after server is down

So far we only support Debug mode, Distributed mode, and a single task in Auto mode. Todo: to support Auto mode fully.

In the construction function of the user's Parameter class, set the restore_training_from_path_work as the running path of your last run. Its default value is None.

8. Early stop

9. Mixed precision training

Capable of speeding up training by 80% without a performance degradation.

10. Four predefined datasets for small and large data loading

Effectively support extremely large dataset.

11. Gradient accumulation

In the case when Distributed mode is unavailable.

12 Dev set evaluation

The best evaluation score and its corresponding model are showed and stored.

13. TensorBoard

>> tensorboard --logdir work/{run-log}/tensorboard --bind_all

14. True gradient

If your GPUs have unbalanced running tasks, such as different batch sizes, you could active this function to improve your final performance. Besides setting param.true_gradient = True, in train_one_batch, you should return (loss, batch num), instead of just loss.

15. Rich information in log file

1. You model trainable parameter information

   Besides total number of parameters are displayed, parameters are sorted by their sizees, percentage on overall parameters.

2. Training time prediction

3. Network efficiency estimate

16. Deployment

A default PredictorBase is provided and we have preset some enviroment for you, such as set inference and no gradient.

Each deployment supports only GPU and is defined as param.gpus[0]. Thus, you should set proper param.gpus values.

Section 4. Config your linux server

Remove password in login

1. >> ssh-keygen -t rsa
    生成的过程中提示输入，直接回车，接受默认值就行了。
    其中公共密钥保存在 ~/.ssh/id_rsa.pub， 私有密钥保存在 ~/.ssh/id_rsa
2. 然后改一下 .ssh 目录的权限，使用命令 "chmod 755 ~/.ssh"
3. 之后把这个密钥对中的公共密钥复制到你要访问的机器上去，并保存为~/.ssh/authorized_keys.
4. 设置权限: chmod 644 ~/.ssh/authorized_keys

Permit net ports

Share your data disk to other servers

>> sshfs {user}@{ip}:{data-folder} {local-folder} -o idmap=user -o allow_other

Install required python packages

>> python3 -m pip install -r requirements.txt  --user `whoami` 

Section 5. Common errors

E1. Not running in your working directory.

The working directory is an essential concept in our framework, so is the same in Java. When your project has many self-defined modules, do NOT run them in the deepest folder, but in this working directory.

E2. GPU is not available. TopSpin detects your designated GPUs are unvailable, and refuse to run.

E3. Data path is not available. TopSpin detects your designated data path are unvailable from servers to run, and refuse to run.

E4. Can not automatically set ParamBase.net_name.

In the most cases, TopSpin would set this for you. Yet in some very rare case, TopSpin can not set a correct value. One workable solution is ls /sys/class/net) to try those value one by one. A typical net_name is en0, eth0.

Section 6. Useful functions in TopSpin

1. from palframe.measure import Measure

Section 7. Popular Questions

Q1. As I do not need distributed training in foreseeable future, do I need TopSpin?

There are looots of functionalities besides the distributed training.

Q2. Is debugging in TopSpin different from others?

Nooo any difference. In the debugging mode, TopSpin would set your program to using only ONE worker as well as one thread in the background.

Q3. Who should I contact when encountering an error?

Q4. If, in training, your evaluation on dev and testing datasets takes more than 30 minutes, the training would halt.

This is because the training is distributed while the evaluation is running on GPU (rank_0), and other GPUs are still waiting for the evaluation to be done. If the evaluation takes more than the default maximum 30 minutes, the pytorch distribution system thinks there may have some exception occured and halts.

You may ask why you donot set the evaluation in a distributed style too. I tried and did not find an elegant design pattern, though I can techinically solve this issue that users would have more burden to use TopSpin.

One short-cut of this issue is, sample a moderate-size data from your true dev data and use as dev in training; after training is finished, test the whole true dev data in one time.

Q5. In order to speedup the data loading, could I copy the training data to a local path in each server, and how to set TopSpin?

Sure. You can do it, and it indeed speedups the data loading as well as the training. There are two ways to do it.

First, copy ALL data to each server. Then, you do not need to change any code. It runs perfectly.

Second, copy 1/n data to n servers respectively to save disk space. Then in the user's train.py, you need to update the parameters of get_batch_data(...) in Trainer.__init__(...). Change dist.get_rank() and dist.get_world_size() to dist.get_local_rank() and self._param.gpu_num.