Runner and model

As mentioned in basic dataflow, the dataflow between DataLoader, model and evaluator follows some rules. Don’t remember clearly? Let’s review it:

# Training process
for data_batch in train_dataloader:
    data_batch = model.data_preprocessor(data_batch, training=True)
    if isinstance(data_batch, dict):
        losses = model(**data_batch, mode='loss')
    elif isinstance(data_batch, (list, tuple)):
        losses = model(*data_batch, mode='loss')
        raise TypeError()
# Validation process
for data_batch in val_dataloader:
    data_batch = model.data_preprocessor(data_batch, training=False)
    if isinstance(data_batch, dict):
        outputs = model(**data_batch, mode='predict')
    elif isinstance(data_batch, (list, tuple)):
        outputs = model(**data_batch, mode='predict')
        raise TypeError()
    evaluator.process(data_samples=outputs, data_batch=data_batch)
metrics = evaluator.evaluate(len(val_dataloader.dataset))

In runner tutorial, we simply mentioned the relationship between DataLoader, model and evaluator, and introduced the concept of data_preprocessor. You may have a certain understanding of the model. However, during the running of Runner, the situation is far more complex than the above pseudo-code.

In order to focus your attention on the algorithm itself, and ignore the complex relationship between the model, DataLoader and evaluator, we designed BaseModel. In most cases, the only thing you need to do is to make your model inherit from BaseModel, and implement the forward as required to perform the training, testing, and validation process.

Before continuing reading the model tutorial, let’s throw out two questions that we hope you will find the answers after reading the model tutorial:

  1. When do we update the parameters of model? and how to update the parameters by a custom optimization process?

  2. Why is the concept of data_preprocessor necessary? What functions can it perform?

Interface introduction

Usually, we should define a model to implement the body of the algorithm. In MMEngine, model will be managed by Runner, and need to implement some interfaces, such as train_step, val_step, and test_step. For high-level tasks like detection, classification, and segmentation, the interfaces mentioned above commonly implement a standard workflow. For example, train_step will calculate the loss and update the parameters of the model, and val_step/test_step will calculate the metrics and return the predictions. Therefore, MMEnine abstracts the BaseModel to implement the common workflow.

Benefits from the BaseModel, we only need to make the model inherit from BaseModel, and implement the forward function to perform the training, testing, and validation process.


BaseModel inherits from BaseModule,which can be used to initialize the model parameters dynamically.

forward: The arguments of forward need to match with the data given by DataLoader. If the DataLoader samples a tuple data, forward needs to accept the value of unpacked *data. If DataLoader returns a dict data, forward needs to accept the key-value of unpacked **data. forward also accepts mode parameter, which is used to control the running branch:

  • mode='loss': loss mode is enabled in training process, and forward returns a differentiable loss dict. Each key-value pair in loss dict will be used to log the training status and optimize the parameters of model. This branch will be called by train_step

  • mode='predict': predict mode is enabled in validation/testing process, and forward will return predictions, which matches with arguments of process. Repositories of OpenMMLab have a more strict rules. The predictions must be a list and each element of it must be a BaseDataElement. This branch will be called by val_step

  • mode='tensor': In tensor and predict modes, forward will return the predictions. The difference is that forward will return a tensor or a container or tensor which has not been processed by a series of post-process methods, such as non-maximum suppression (NMS). You can customize your post-process method after getting the result of tensor mode.

train_step: Get the loss dict by calling forward with loss mode. BaseModel implements a standard optimization process as follows:

def train_step(self, data, optim_wrapper):
    # See details in the next section
    data = self.data_preprocessor(data, training=True)
    # `loss` mode, return a loss dict. Actually train_step accepts
    #  both tuple  dict input, and unpack it with ** or *
    loss = self(**data, mode='loss')
    # Parse the loss dict and return the parsed losses for optimization
    # and log_vars for logging
    parsed_losses, log_vars = self.parse_losses()
    optim_wrapper.update_params(parsed_losses)  # 更新参数
    return log_vars

val_step: Get the predictions by calling forward with predict mode.

def val_step(self, data, optim_wrapper):
    data = self.data_preprocessor(data, training=False)
    outputs = self(**data, mode='predict')
    return outputs

test_step: There is no difference between val_step and test_step in BaseModel. But we can customize it in the subclasses, for example, you can get validation loss in val_step.

Understand the interfaces of BaseModel, now we are able to come up with a more complete pseudo-code:

# training
for data_batch in train_dataloader:
    loss_dict = model.train_step(data_batch)
# validation
for data_batch in val_dataloader:
    preds = model.test_step(data_batch)
    evaluator.process(data_samples=outputs, data_batch=data_batch)
metrics = evaluator.evaluate(len(val_dataloader.dataset))

Great!, ignoring Hook, the pseudo-code above almost implements the main logic in loop! Let’s go back to 15 minutes to get started with MMEngine, we may truly understand what MMResNet has done:

import torch.nn.functional as F
import torchvision
from mmengine.model import BaseModel

class MMResNet50(BaseModel):
    def __init__(self):
        self.resnet = torchvision.models.resnet50()

    def forward(self, imgs, labels, mode):
        x = self.resnet(imgs)
        if mode == 'loss':
            return {'loss': F.cross_entropy(x, labels)}
        elif mode == 'predict':
            return x, labels

    # train_step, val_step and test_step have been implemented in BaseModel.
    # We list the equivalent code here for better understanding
    def train_step(self, data, optim_wrapper):
        data = self.data_preprocessor(data)
        loss = self(*data, mode='loss')
        parsed_losses, log_vars = self.parse_losses()
        return log_vars

    def val_step(self, data, optim_wrapper):
        data = self.data_preprocessor(data)
        outputs = self(*data, mode='predict')
        return outputs

    def test_step(self, data, optim_wrapper):
        data = self.data_preprocessor(data)
        outputs = self(*data, mode='predict')
        return outputs

Now, you may have a deeper understanding of dataflow, and can answer the first question in Runner and model.

BaseModel.train_step implements the standard optimization, and if we want to customize a new optimization process, we can override it in the subclass. However, it is important to note that we need to make sure that train_step returns a loss dict.


If your computer is equipped with a GPU (or other hardware that can accelerate training, such as MPS, IPU, etc.), when you run the 15 minutes tutorial, you will see that the program is running on the GPU, but, when does MMEngine move the data and model from the CPU to the GPU?

In fact, the Runner will move the model to the specified device during the construction, while the data will be moved to the specified device at the self.data_preprocessor(data) mentioned in the code snippet of the previous section. The moved data will be further passed to the model.

Makes sense but it’s weird, isn’t it? At this point you may be wondering:

  1. MMResNet50 does not define data_preprocessor, but why it can still access data_preprocessor and move data to GPU?

  2. Why BaseModel does not move data by data =, but needs the DataPreprocessor to move data?

The answer to the first question is that: MMResNet50 inherit from BaseModel, and super().__init__ will build a default data_preprocessor for it. The equivalent implementation of the default one is like this:

class BaseDataPreprocessor(nn.Module):
    def forward(self, data, training=True):  # ignore the training parameter here
        # suppose data given by CIFAR10 is a tuple. Actually
        # BaseDataPreprocessor could move various type of data
        # to target device.
        return tuple(_data.cuda() for _data in data)

BaseDataPreprocessor will move the data to the specified device.

Before answering the second question, let’s think about a few more questions

  1. Where should we perform normalization? transform or Model?

    It sounds reasonable to put it in transform to take advantage of Dataloader’s multi-process acceleration, and in the model to move it to GPU to use GPU resources to accelerate normalization. However, while we are debating whether CPU normalization is faster than GPU normalization, the time of data moving from CPU to GPU is much longer than the former.

    In fact, for less computationally intensive operations like normalization, it takes much less time than data transferring, which has a higher priority for being optimized. If I could move the data to the specified device while it is still in uint8 and before it is normalized (the size of normalized float data is 4 times larger than that of unit8), it would reduce the bandwidth and greatly improve the efficiency of data transferring. This “lagged” normalization behavior is one of the main reasons why we designed the DataPreprocessor. The data preprocessor moves the data first and then normalizes it.

  2. How we implement the data augmentation like MixUp and Mosaic?

    Although it seems that MixUp and Mosaic are just special data transformations that should be implemented in transform. However, considering that these two transformations involve fusing multiple images into one, it would be very difficult to implement them in transform since the current paradigm of transform is to do various enhancements on one image. It would be hard to read additional images from dataset because the dataset is not accessible in the transform. However, if we implement Mosaic or Mixup based on the batch_data sampled from Dataloader, everything becomes easy. We can access multiple images at the same time, and we can easily perform the image fusion operation.

    class MixUpDataPreprocessor(nn.Module):
        def __init__(self, num_class, alpha):
            self.alpha = alpha
        def forward(self, data, training=True):
            data = tuple(_data.cuda() for _data in data)
            # Only perform MixUp in training mode
            if not training:
                return data
            label = F.one_hot(label)  # label to OneHot
            batch_size = len(label)
            index = torch.randperm(batch_size)  # Get the index of fused image
            img, label = data
            lam = np.random.beta(self.alpha, self.alpha)  # Fusion factor
            # MixUp
            img = lam * img + (1 - lam) * img[index, :]
            label = lam * batch_scores + (1 - lam) * batch_scores[index, :]
            # Since the returned label is onehot encoded, the `forward` of the
            # model should also be adjusted.
            return tuple(img, label)

    Therefore, besides data transferring and normalization, another major function of data_preprocessor is BatchAugmentation. The modularity of the data preprocessor also helps us to achieve a free combination between algorithms and data augmentation.

  3. What should we do if the data sampled from the DataLoader does not match the model input, should I modify the DataLoader or the model interface?

    The answer is: neither is appropriate. The ideal solution is to do the adaptation without breaking the existing interface between the model and the DataLoader. DataPreprocessor could also handle this, you can customize your DataPreprocessor to convert the incoming to the target type.

By now, You must understand the rationale of the data preprocessor and can confidently answer the two questions posed at the beginning of the tutorial! But you may still wonder what is the optim_wrapper passed to train_step, and how do the predictions returned by test_step and val_step relate to the evaluator. You will find more introduction in the evaluation tutorial and the optimizer wrapper tutorial.

Read the Docs v: latest
On Read the Docs
Project Home

Free document hosting provided by Read the Docs.