Utilities for Trainer

This page lists all the utility functions used by Trainer.

Most of those are only useful if you are studying the code of the Trainer in the library.


class transformers.EvalPrediction(predictions: Union[numpy.ndarray, Tuple[numpy.ndarray]], label_ids: numpy.ndarray)[source]

Evaluation output (always contains labels), to be used to compute metrics.

  • predictions (np.ndarray) – Predictions of the model.

  • label_ids (np.ndarray) – Targets to be matched.

class transformers.IntervalStrategy(value)[source]

An enumeration.

transformers.set_seed(seed: int)[source]

Helper function for reproducible behavior to set the seed in random, numpy, torch and/or tf (if installed).


seed (int) – The seed to set.

transformers.torch_distributed_zero_first(local_rank: int)[source]

Decorator to make all processes in distributed training wait for each local_master to do something.


local_rank (int) – The rank of the local process.

Callbacks internals

class transformers.trainer_callback.CallbackHandler(callbacks, model, tokenizer, optimizer, lr_scheduler)[source]

Internal class that just calls the list of callbacks in order.

Distributed Evaluation

class transformers.trainer_pt_utils.DistributedTensorGatherer(world_size, num_samples, make_multiple_of=None, padding_index=- 100)[source]

A class responsible for properly gathering tensors (or nested list/tuple of tensors) on the CPU by chunks.

If our dataset has 16 samples with a batch size of 2 on 3 processes and we gather then transfer on CPU at every step, our sampler will generate the following indices:

[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0, 1]

to get something of size a multiple of 3 (so that each process gets the same dataset length). Then process 0, 1 and 2 will be responsible of making predictions for the following samples:

  • P0: [0, 1, 2, 3, 4, 5]

  • P1: [6, 7, 8, 9, 10, 11]

  • P2: [12, 13, 14, 15, 0, 1]

The first batch treated on each process will be

  • P0: [0, 1]

  • P1: [6, 7]

  • P2: [12, 13]

So if we gather at the end of the first batch, we will get a tensor (nested list/tuple of tensor) corresponding to the following indices:

[0, 1, 6, 7, 12, 13]

If we directly concatenate our results without taking any precautions, the user will then get the predictions for the indices in this order at the end of the prediction loop:

[0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15, 4, 5, 10, 11, 0, 1]

For some reason, that’s not going to roll their boat. This class is there to solve that problem.

  • world_size (int) – The number of processes used in the distributed training.

  • num_samples (int) – The number of samples in our dataset.

  • make_multiple_of (int, optional) – If passed, the class assumes the datasets passed to each process are made to be a multiple of this argument (by adding samples).

  • padding_index (int, optional, defaults to -100) – The padding index to use if the arrays don’t all have the same sequence length.


Add arrays to the internal storage, Will initialize the storage to the full size at the first arrays passed so that if we’re bound to get an OOM, it happens at the beginning.


Return the properly gathered arrays and truncate to the number of samples (since the sampler added some extras to get each process a dataset of the same length).

Distributed Evaluation

class transformers.HfArgumentParser(dataclass_types: Union[NewType.<locals>.new_type, Iterable[NewType.<locals>.new_type]], **kwargs)[source]

This subclass of argparse.ArgumentParser uses type hints on dataclasses to generate arguments.

The class is designed to play well with the native argparse. In particular, you can add more (non-dataclass backed) arguments to the parser after initialization and you’ll get the output back after parsing as an additional namespace.

Debug Utilities

class transformers.debug_utils.DebugUnderflowOverflow(model, max_frames_to_save=21, trace_batch_nums=[], abort_after_batch_num=None)[source]

This debug class helps detect and understand where the model starts getting very large or very small, and more importantly nan or inf weight and activation elements.

There are 2 working modes:

  1. Underflow/overflow detection (default)

  2. Specific batch absolute min/max tracing without detection

Mode 1: Underflow/overflow detection

To activate the underflow/overflow detection, initialize the object with the model

debug_overflow = DebugUnderflowOverflow(model)

then run the training as normal and if nan or inf gets detected in at least one of the weight, input or output elements this module will throw an exception and will print max_frames_to_save frames that lead to this event, each frame reporting

  1. the fully qualified module name plus the class name whose forward was run

  2. the absolute min and max value of all elements for each module weights, and the inputs and output

For example, here is the header and the last few frames in detection report for google/mt5-small run in fp16 mixed precision

Detected inf/nan during batch_number=0
Last 21 forward frames:
abs min  abs max  metadata
                  encoder.block.2.layer.1.DenseReluDense.wi_0 Linear
2.17e-07 4.50e+00 weight
1.79e-06 4.65e+00 input[0]
2.68e-06 3.70e+01 output
                  encoder.block.2.layer.1.DenseReluDense.wi_1 Linear
8.08e-07 2.66e+01 weight
1.79e-06 4.65e+00 input[0]
1.27e-04 2.37e+02 output
                  encoder.block.2.layer.1.DenseReluDense.wo Linear
1.01e-06 6.44e+00 weight
0.00e+00 9.74e+03 input[0]
3.18e-04 6.27e+04 output
                  encoder.block.2.layer.1.DenseReluDense T5DenseGatedGeluDense
1.79e-06 4.65e+00 input[0]
3.18e-04 6.27e+04 output
                  encoder.block.2.layer.1.dropout Dropout
3.18e-04 6.27e+04 input[0]
0.00e+00      inf output

You can see here, that T5DenseGatedGeluDense.forward resulted in output activations, whose absolute max value was around 62.7K, which is very close to fp16’s top limit of 64K. In the next frame we have Dropout which renormalizes the weights, after it zeroed some of the elements, which pushes the absolute max value to more than 64K, and we get an overlow.

As you can see it’s the previous frames that we need to look into when the numbers start going into very large for fp16 numbers.

The tracking is done in a forward hook, which gets invoked immediately after forward has completed.

By default the last 21 frames are printed. You can change the default to adjust for your needs. For example

debug_overflow = DebugUnderflowOverflow(model, max_frames_to_save=100)

Mode 2. Specific batch absolute min/max tracing without detection

The second work mode is per-batch tracing with the underflow/overflow detection feature turned off.

Let’s say you want to watch the absolute min and max values for all the ingredients of each forward call of a given batch, and only do that for batches 1 and 3. Then you instantiate this class as

debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[1,3])

And now full batches 1 and 3 will be traced using the same format as explained above. Batches are 0-indexed.

This is helpful if you know that the program starts misbehaving after a certain batch number, so you can fast-forward right to that area.

You can also specify the batch number after which to stop the training, with

debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[1,3], abort_after_batch_num=3)

This feature is mainly useful in the tracing mode, but you can use it for any more.

  • model (nn.Module) – The model to debug.

  • max_frames_to_save (int, optional, defaults to 21) – How many frames back to record

  • trace_batch_nums (List[int], optional, defaults to []) – Which batch numbers to trace (turns detection off)

  • abort_after_batch_num (int, optional) – Whether to abort after a certain batch number has finished