Basic training functionality¶

basic_train wraps together the data (in a DataBunch object) with a PyTorch model to define a Learner object. Here the basic training loop is defined for the fit method. The Learner object is the entry point of most of the Callback objects that will customize this training loop in different ways. Some of the most commonly used customizations are available through the train module, notably:

• Learner.lr_find will launch an LR range test that will help you select a good learning rate.
• Learner.fit_one_cycle will launch a training using the 1cycle policy to help you train your model faster.
• Learner.to_fp16 will convert your model to half precision and help you launch a training in mixed precision.

The main purpose of Learner is to train model using Learner.fit. After every epoch, all metrics will be printed and also made available to callbacks.

The default weight decay will be wd, which will be handled using the method from Fixing Weight Decay Regularization in Adam if true_wd is set (otherwise it's L2 regularization). If true_wd is set it will affect all optimizers, not only Adam. If bn_wd is False, then weight decay will be removed from batchnorm layers, as recommended in Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour. If train_bn, batchnorm layer learnable params are trained even for frozen layer groups.

To use discriminative layer training, pass a list of nn.Module as layer_groups; each nn.Module will be used to customize the optimization of the corresponding layer group.

If path is provided, all the model files created will be saved in path/model_dir; if not, then they will be saved in data.path/model_dir.

You can pass a list of callbacks that you have already created, or (more commonly) simply pass a list of callback functions to callback_fns and each function will be called (passing self) on object initialization, with the results stored as callback objects. For a walk-through, see the training overview page. You may also want to use an application specific model. For example, if you are dealing with a vision dataset, here the MNIST, you might want to use the cnn_learner method:

Model fitting methods¶

Runs the learning rate finder defined in LRFinder, as discussed in Cyclical Learning Rates for Training Neural Networks.

Uses discriminative layer training if multiple learning rates or weight decay values are passed. To control training behaviour, use the callback system or one or more of the pre-defined callbacks.

Use cycle length cyc_len, a per cycle maximal learning rate max_lr, momentum moms, division factor div_factor, weight decay wd, and optional callbacks callbacks. Uses the OneCycleScheduler callback. Please refer to What is 1-cycle for a conceptual background of 1-cycle training policy and more technical details on what do the method's arguments do.

See results¶

predict can be used to get a single prediction from the trained learner on one specific piece of data you are interested in.

Each element of the dataset is a tuple, where the first element is the data itself, while the second element is the target label. So to get the data, we need to index one more time.

The first two elements of the tuple are, respectively, the predicted class and label. Label here is essentially an internal representation of each class, since class name is a string and cannot be used in computation. To check what each label corresponds to, run:

So category 0 is 3 while category 1 is 7.

The last element in the tuple is the predicted probabilities. For a categorization dataset, the number of probabilities returned is the same as the number of classes; probs[i] is the probability that the item belongs to learn.data.classes[i].

You could always check yourself if the probabilities given make sense.

It will run inference using the learner on all the data in the ds_type dataset and return the predictions; if n_batch is not specified, it will run the predictions on the default batch size. If with_loss, it will also return the loss on each prediction.

Here is how you check the default batch size.

The first element of the tuple is a tensor that contains all the predictions.

While the second element of the tuple is a tensor that contains all the target labels.

For more details about what each number mean, refer to the documentation of predict.

Since get_preds gets predictions on all the data in the ds_type dataset, here the number of predictions will be equal to the number of data in the validation dataset.

To get predictions on the entire training dataset, simply set the ds_type argument accordingly.

To also get prediction loss along with the predictions and the targets, set with_loss=True in the arguments.

Note that the third tensor in the output tuple contains the losses.

Return the calculated loss and the metrics of the current model on the given data loader dl. The default data loader dl is the validation dataloader.

You can check the default metrics of the learner using:

Note that the text number on the top is the ground truth, or the target label, the one in the middle is the prediction, while the image number on the bottom is the image data itself.

Note that the number of predictions given equals to the batch size.

Since the total number of predictions is too large, we will only look at a part of them.

For more details, refer to ClassificationInterpretation

Model summary¶

Test time augmentation¶

Applies Test Time Augmentation to learn on the dataset ds_type. We take the average of our regular predictions (with a weight beta) with the average of predictions obtained through augmented versions of the training set (with a weight 1-beta). The transforms decided for the training set are applied with a few changes scale controls the scale for zoom (which isn't random), the cropping isn't random but we make sure to get the four corners of the image. Flipping isn't random but applied once on each of those corner images (so that makes 8 augmented versions total).

Gradient clipping¶

Mixed precision training¶

Uses the MixedPrecision callback to train in mixed precision (i.e. forward and backward passes using fp16, with weight updates using fp32), using all NVIDIA recommendations for ensuring speed and accuracy.

Distributed training¶

If you want to use ditributed training or torch.nn.DataParallel these will directly wrap the model for you.

Discriminative layer training¶

When fitting a model you can pass a list of learning rates (and/or weight decay amounts), which will apply a different rate to each layer group (i.e. the parameters of each module in self.layer_groups). See the Universal Language Model Fine-tuning for Text Classification paper for details and experimental results in NLP (we also frequently use them successfully in computer vision, but have not published a paper on this topic yet). When working with a Learner on which you've called split, you can set hyperparameters in four ways:

1. param = [val1, val2 ..., valn] (n = number of layer groups)
2. param = val
3. param = slice(start,end)
4. param = slice(end)

If we chose to set it in way 1, we must specify a number of values exactly equal to the number of layer groups. If we chose to set it in way 2, the chosen value will be repeated for all layer groups. See Learner.lr_range for an explanation of the slice syntax).

Here's an example of how to use discriminative learning rates (note that you don't actually need to manually call Learner.split in this case, since fastai uses this exact function as the default split for resnet18; this is just to show how to customize it):

Rather than manually setting an LR for every group, it's often easier to use Learner.lr_range. This is a convenience method that returns one learning rate for each layer group. If you pass slice(start,end) then the first group's learning rate is start, the last is end, and the remaining are evenly geometrically spaced.

If you pass just slice(end) then the last group's learning rate is end, and all the other groups are end/10. For instance (for our learner that has 3 layer groups):

Sets every layer group to trainable (i.e. requires_grad=True).

Sets every layer group except the last to untrainable (i.e. requires_grad=False).

What does 'the last layer group' mean?

In the case of transfer learning, such as learn = cnn_learner(data, models.resnet18, metrics=error_rate), learn.modelwill print out two large groups of layers: (0) Sequential and (1) Sequental in the following structure. We can consider the last conv layer as the break line between the two groups.

Sequential(
  (0): Sequential(
    (0): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
    (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU(inplace)
    ...

            (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
             (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
  )
  (1): Sequential(
    (0): AdaptiveConcatPool2d(
      (ap): AdaptiveAvgPool2d(output_size=1)
      (mp): AdaptiveMaxPool2d(output_size=1)
    )
    (1): Flatten()
    (2): BatchNorm1d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (3): Dropout(p=0.25)
    (4): Linear(in_features=1024, out_features=512, bias=True)
    (5): ReLU(inplace)
    (6): BatchNorm1d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (7): Dropout(p=0.5)
    (8): Linear(in_features=512, out_features=12, bias=True)
  )
)

learn.freeze freezes the first group and keeps the second or last group free to train, including multiple layers inside (this is why calling it 'group'), as you can see in learn.summary() output. How to read the table below, please see model summary docs.

======================================================================
Layer (type)         Output Shape         Param #    Trainable 
======================================================================
...
...
...
______________________________________________________________________
Conv2d               [1, 512, 4, 4]       2,359,296  False     
______________________________________________________________________
BatchNorm2d          [1, 512, 4, 4]       1,024      True      
______________________________________________________________________
AdaptiveAvgPool2d    [1, 512, 1, 1]       0          False     
______________________________________________________________________
AdaptiveMaxPool2d    [1, 512, 1, 1]       0          False     
______________________________________________________________________
Flatten              [1, 1024]            0          False     
______________________________________________________________________
BatchNorm1d          [1, 1024]            2,048      True      
______________________________________________________________________
Dropout              [1, 1024]            0          False     
______________________________________________________________________
Linear               [1, 512]             524,800    True      
______________________________________________________________________
ReLU                 [1, 512]             0          False     
______________________________________________________________________
BatchNorm1d          [1, 512]             1,024      True      
______________________________________________________________________
Dropout              [1, 512]             0          False     
______________________________________________________________________
Linear               [1, 12]              6,156      True      
______________________________________________________________________

Total params: 11,710,540
Total trainable params: 543,628
Total non-trainable params: 11,166,912

From above we know what is layer group, but what exactly does freeze_to do behind the scenes?

The freeze_to source code can be understood as the following pseudo-code:

def freeze_to(self, n:int)->None:
    for g in self.layer_groups[:n]: freeze 
    for g in self.layer_groups[n:]: unfreeze

In other words, for example, freeze_to(1) is to freeze layer group 0 and unfreeze the rest layer groups, and freeze_to(3) is to freeze layer groups 0, 1, and 2 but unfreeze the rest layer groups (if there are more layer groups left).

Both freeze and unfreeze sources are defined using freeze_to:

• When we say freeze, we mean that in the specified layer groups the requires_grad of all layers with weights (except BatchNorm layers) are set False, so the layer weights won't be updated during training.
• when we say unfreeze, we mean that in the specified layer groups the requires_grad of all layers with weights (except BatchNorm layers) are set True, so the layer weights will be updated during training.

A convenience method that sets layer_groups based on the result of split_model. If split_on is a function, it calls that function and passes the result to split_model (see above for example).

Saving and loading models¶

Simply call Learner.save and Learner.load to save and load models. Only the parameters are saved, not the actual architecture (so you'll need to create your model in the same way before loading weights back in). Models are saved to the path/model_dir directory.

If argument name is a pathlib object that's an absolute path, it'll override the default base directory (learn.path), otherwise the model will be saved in a file relative to learn.path.

This method only works after save (don't confuse with export/load_learner pair).

If the purge argument is True (default) load internally calls purge with clear_opt=False to presever learn.opt.

Deploying your model¶

When you are ready to put your model in production, export the minimal state of your Learner with:

If argument fname is a pathlib object that's an absolute path, it'll override the default base directory (learn.path), otherwise the model will be saved in a file relative to learn.path.

Passing destroy=True will destroy the Learner, freeing most of its memory consumption. For specifics see Learner.destroy.

This method only works with the Learner whose data was created through the data block API.

Otherwise, you will have to create a Learner yourself at inference and load the model with Learner.load.

This function only works after export (don't confuse with save/load pair).

The db_kwargs will be passed to the call to databunch so you can specify a bs for the test set, or num_workers.

WARNING: If you used any customized classes when creating your learner, you must first define these classes first before executing load_learner.

You can find more information and multiple examples in this tutorial.

Freeing memory¶

If you want to be able to do more without needing to restart your notebook, the following methods are designed to free memory when it's no longer needed.

Refer to this tutorial to learn how and when to use these methods.

If learn.path is read-only, you can set model_dir attribute in Learner to a full libpath path that is writable (by setting learn.model_dir or passing model_dir argument in the Learner constructor).

If you need to free the memory consumed by the Learner object, call this method.

It can also be automatically invoked through Learner.export via its destroy=True argument.

Other methods¶

Initializes all weights (except batchnorm) using function init, which will often be from PyTorch's nn.init module.

Uses MixUpCallback.

You generally won't need to call this yourself - it's used to create the optim optimizer before fitting the model.

A Learner creates a Recorder object automatically - you do not need to explicitly pass it to callback_fns - because other callbacks rely on it being available. It stores the smoothed loss, hyperparameter values, and metrics for each batch, and provides plotting methods for each. Note that Learner automatically sets an attribute with the snake-cased name of each callback, so you can access this through Learner.recorder, as shown below.

Plotting methods¶

This is mainly used with the learning rate finder, since it shows a scatterplot of loss vs learning rate.

Note that validation losses are only calculated once per epoch, whereas training losses are calculated after every batch.

Note that metrics are only collected at the end of each epoch, so you'll need to train at least two epochs to have anything to show here.

Callback methods¶

You don't call these yourself - they're called by fastai's Callback system automatically to enable the class's functionality. Refer to Callback for more details.

Inner functions¶

The following functions are used along the way by the Recorder or can be called by other callbacks.

Module functions¶

Generally you'll want to use a Learner to train your model, since they provide a lot of functionality and make things easier. However, for ultimate flexibility, you can call the same underlying functions that Learner calls behind the scenes:

Note that you have to create the Optimizer yourself if you call this function, whereas Learn.fit creates it for you automatically.

You won't generally need to call this yourself - it's what fit calls for each epoch.

This is what fit calls after each epoch. You can call it if you want to run inference on a DataLoader manually.

You won't generally need to call this yourself - it's what fit and validate call for each batch. It only does a backward pass if you set opt.

Other classes¶

Open This Notebook¶

Open in GCP Notebooks

Undocumented Methods - Methods moved below this line will intentionally be hidden¶

New Methods - Please document or move to the undocumented section¶