mytrainer.py

'''
ZFJ 2022
Originated From: huggingface/transformers/trainer.py
modified it to use top_p sampling
'''
import torch
from torch.utils.data import Dataset
from torch.utils.data import DataLoader
from torch import nn
import numpy as np

from typing import Optional, List, Dict, Union, Any

from transformers import Seq2SeqTrainer
from transformers.deepspeed import is_deepspeed_zero3_enabled, deepspeed_init
from transformers.trainer_utils import EvalLoopOutput, has_length, denumpify_detensorize, EvalPrediction
from transformers.file_utils import is_torch_tpu_available
from transformers.trainer_pt_utils import find_batch_size, nested_concat, nested_numpify, nested_truncate, IterableDatasetShard
from transformers.utils import logging

logger = logging.get_logger(__name__)

if is_torch_tpu_available():
    import torch_xla.core.xla_model as xm
    import torch_xla.distributed.parallel_loader as pl

class MySeq2SeqTrainer(Seq2SeqTrainer):
    def predict(
        self,
        test_dataset: Dataset,
        ignore_keys: Optional[List[str]] = None,
        metric_key_prefix: str = "test",
        max_length: Optional[int] = None,
        num_beams: Optional[int] = None,
        do_sample: bool = False,
        top_p: Optional[float] = None,
        num_return_sequences: int = 1,
    ):
        self._max_length = max_length if max_length is not None else None
        self._num_beams = num_beams
        self._do_sample = do_sample
        self._top_p = top_p
        self._num_return_sequences = num_return_sequences
        return super().predict(test_dataset, ignore_keys=ignore_keys, metric_key_prefix=metric_key_prefix)

    def evaluate(
        self,
        eval_dataset: Optional[Dataset] = None,
        ignore_keys: Optional[List[str]] = None,
        metric_key_prefix: str = "eval",
        max_length: Optional[int] = None,
        num_beams: Optional[int] = None,
    ):
        self._do_sample = False
        self._num_return_sequences = 1
        self._top_p = None
        return super().evaluate(eval_dataset, ignore_keys=ignore_keys, metric_key_prefix=metric_key_prefix, max_length=max_length, num_beams=num_beams)

    def evaluation_loop(
        self,
        dataloader: DataLoader,
        description: str,
        prediction_loss_only: Optional[bool] = None,
        ignore_keys: Optional[List[str]] = None,
        metric_key_prefix: str = "eval",
    ) -> EvalLoopOutput:
        """
        Prediction/evaluation loop, shared by `Trainer.evaluate()` and `Trainer.predict()`.

        Works both with or without labels.
        """
        args = self.args

        prediction_loss_only = prediction_loss_only if prediction_loss_only is not None else args.prediction_loss_only

        # if eval is called w/o train init deepspeed here
        if args.deepspeed and not self.deepspeed:

            # XXX: eval doesn't have `resume_from_checkpoint` arg but we should be able to do eval
            # from the checkpoint eventually
            deepspeed_engine, _, _ = deepspeed_init(
                self, num_training_steps=0, resume_from_checkpoint=None, inference=True
            )
            self.model = deepspeed_engine.module
            self.model_wrapped = deepspeed_engine
            self.deepspeed = deepspeed_engine

        model = self._wrap_model(self.model, training=False)

        # if full fp16 or bf16 eval is wanted and this ``evaluation`` or ``predict`` isn't called
        # while ``train`` is running, cast it to the right dtype first and then put on device
        if not self.is_in_train:
            if args.fp16_full_eval:
                model = model.to(dtype=torch.float16, device=args.device)
            elif args.bf16_full_eval:
                model = model.to(dtype=torch.bfloat16, device=args.device)

        batch_size = dataloader.batch_size

        logger.info(f"***** Running {description} *****")
        if has_length(dataloader.dataset):
            logger.info(f"  Num examples = {self.num_examples(dataloader)}")
        else:
            logger.info("  Num examples: Unknown")
        logger.info(f"  Batch size = {batch_size}")

        model.eval()

        self.callback_handler.eval_dataloader = dataloader
        # Do this before wrapping.
        eval_dataset = dataloader.dataset

        if is_torch_tpu_available():
            dataloader = pl.ParallelLoader(dataloader, [args.device]).per_device_loader(args.device)

        if args.past_index >= 0:
            self._past = None

        # Initialize containers
        # losses/preds/labels on GPU/TPU (accumulated for eval_accumulation_steps)
        losses_host = None
        preds_host = None
        labels_host = None
        # losses/preds/labels on CPU (final containers)
        all_losses = None
        all_preds = None
        all_labels = None
        # Will be useful when we have an iterable dataset so don't know its length.

        observed_num_examples = 0
        # Main evaluation loop
        for step, inputs in enumerate(dataloader):
            # Update the observed num examples
            observed_batch_size = find_batch_size(inputs)
            if observed_batch_size is not None:
                observed_num_examples += observed_batch_size
                # For batch samplers, batch_size is not known by the dataloader in advance.
                if batch_size is None:
                    batch_size = observed_batch_size

            # Prediction step
            loss, logits, labels = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)

            if is_torch_tpu_available():
                xm.mark_step()

            # Update containers on host
            if loss is not None:
                losses = self._nested_gather(loss.repeat(batch_size))
                losses_host = losses if losses_host is None else torch.cat((losses_host, losses), dim=0)
            if labels is not None:
                labels = self._pad_across_processes(labels)
                labels = self._nested_gather(labels)
                labels_host = labels if labels_host is None else nested_concat(labels_host, labels, padding_index=-100)
            if logits is not None:
                logits = self._pad_across_processes(logits)
                logits = self._nested_gather(logits)
                if self.preprocess_logits_for_metrics is not None:
                    logits = self.preprocess_logits_for_metrics(logits, labels)
                preds_host = logits if preds_host is None else nested_concat(preds_host, logits, padding_index=-100)
            self.control = self.callback_handler.on_prediction_step(args, self.state, self.control)

            # Gather all tensors and put them back on the CPU if we have done enough accumulation steps.
            if args.eval_accumulation_steps is not None and (step + 1) % args.eval_accumulation_steps == 0:
                if losses_host is not None:
                    losses = nested_numpify(losses_host)
                    all_losses = losses if all_losses is None else np.concatenate((all_losses, losses), axis=0)
                if preds_host is not None:
                    logits = nested_numpify(preds_host)
                    all_preds = logits if all_preds is None else nested_concat(all_preds, logits, padding_index=-100)
                if labels_host is not None:
                    labels = nested_numpify(labels_host)
                    all_labels = (
                        labels if all_labels is None else nested_concat(all_labels, labels, padding_index=-100)
                    )

                # Set back to None to begin a new accumulation
                losses_host, preds_host, labels_host = None, None, None

        if args.past_index and hasattr(self, "_past"):
            # Clean the state at the end of the evaluation loop
            delattr(self, "_past")

        # Gather all remaining tensors and put them back on the CPU
        if losses_host is not None:
            losses = nested_numpify(losses_host)
            all_losses = losses if all_losses is None else np.concatenate((all_losses, losses), axis=0)
        if preds_host is not None:
            logits = nested_numpify(preds_host)
            all_preds = logits if all_preds is None else nested_concat(all_preds, logits, padding_index=-100)
        if labels_host is not None:
            labels = nested_numpify(labels_host)
            all_labels = labels if all_labels is None else nested_concat(all_labels, labels, padding_index=-100)

        # Number of samples
        if has_length(eval_dataset):
            num_samples = len(eval_dataset)
        # The instance check is weird and does not actually check for the type, but whether the dataset has the right
        # methods. Therefore we need to make sure it also has the attribute.
        elif isinstance(eval_dataset, IterableDatasetShard) and hasattr(eval_dataset, "num_examples"):
            num_samples = eval_dataset.num_examples
        else:
            num_samples = observed_num_examples

        # Number of losses has been rounded to a multiple of batch_size and in a distributed training, the number of
        # samplers has been rounded to a multiple of batch_size, so we truncate.
        if all_losses is not None:
            all_losses = all_losses[:num_samples]
        if all_preds is not None:
            all_preds = nested_truncate(all_preds, num_samples) if not self._do_sample else nested_truncate(all_preds, num_samples*self._num_return_sequences)
        if all_labels is not None:
            all_labels = nested_truncate(all_labels, num_samples)

        # Metrics!
        if self.compute_metrics is not None and all_preds is not None and all_labels is not None:
            metrics = self.compute_metrics(EvalPrediction(predictions=all_preds, label_ids=all_labels))
        else:
            metrics = {}

        # To be JSON-serializable, we need to remove numpy types or zero-d tensors
        metrics = denumpify_detensorize(metrics)

        if all_losses is not None:
            metrics[f"{metric_key_prefix}_loss"] = all_losses.mean().item()

        # Prefix all keys with metric_key_prefix + '_'
        for key in list(metrics.keys()):
            if not key.startswith(f"{metric_key_prefix}_"):
                metrics[f"{metric_key_prefix}_{key}"] = metrics.pop(key)

        return EvalLoopOutput(predictions=all_preds, label_ids=all_labels, metrics=metrics, num_samples=num_samples)

    def prediction_step(
        self,
        model: nn.Module,
        inputs: Dict[str, Union[torch.Tensor, Any]],
        prediction_loss_only: bool,
        ignore_keys: Optional[List[str]] = None,
    ):
        if not self.args.predict_with_generate or prediction_loss_only:
            return super().prediction_step(
                model, inputs, prediction_loss_only=prediction_loss_only, ignore_keys=ignore_keys
            )

        has_labels = "labels" in inputs
        inputs = self._prepare_inputs(inputs)

        # XXX: adapt synced_gpus for fairscale as well
        gen_kwargs = {
            "max_length": self._max_length if self._max_length is not None else self.model.config.max_length,
            "num_beams": self._num_beams,
            "do_sample": self._do_sample,
            "top_p": self._top_p if self._do_sample else None,
            "num_return_sequences": self._num_return_sequences,
            "synced_gpus": True if is_deepspeed_zero3_enabled() else False,
        }

        if "attention_mask" in inputs:
            gen_kwargs["attention_mask"] = inputs.get("attention_mask", None)

        # prepare generation inputs
        # some encoder-decoder models can have varying encder's and thus
        # varying model input names
        if hasattr(self.model, "encoder") and self.model.encoder.main_input_name != self.model.main_input_name:
            generation_inputs = inputs[self.model.encoder.main_input_name]
        else:
            generation_inputs = inputs[self.model.main_input_name]

        generated_tokens = self.model.generate(
            generation_inputs,
            **gen_kwargs,
        )

        if self._do_sample:
            return (None, generated_tokens, None)

        # in case the batch is shorter than max length, the output should be padded
        if generated_tokens.shape[-1] < gen_kwargs["max_length"]:
            generated_tokens = self._pad_tensors_to_max_len(generated_tokens, gen_kwargs["max_length"])

        with torch.no_grad():
            with self.autocast_smart_context_manager():
                outputs = model(**inputs)
            if has_labels:
                if self.label_smoother is not None:
                    loss = self.label_smoother(outputs, inputs["labels"]).mean().detach()
                else:
                    loss = (outputs["loss"] if isinstance(outputs, dict) else outputs[0]).mean().detach()
            else:
                loss = None

        if self.args.prediction_loss_only:
            return (loss, None, None)

        if has_labels:
            labels = inputs["labels"]
            if labels.shape[-1] < gen_kwargs["max_length"]:
                labels = self._pad_tensors_to_max_len(labels, gen_kwargs["max_length"])
        else:
            labels = None

        return (loss, generated_tokens, labels)

    def _pad_tensors_to_max_len(self, tensor, max_length):
        if self.tokenizer is not None and hasattr(self.tokenizer, "pad_token_id"):
            # If PAD token is not defined at least EOS token has to be defined
            pad_token_id = (
                self.tokenizer.pad_token_id if self.tokenizer.pad_token_id is not None else self.tokenizer.eos_token_id
            )
        else:
            if self.model.config.pad_token_id is not None:
                pad_token_id = self.model.config.pad_token_id
            else:
                raise ValueError("Pad_token_id must be set in the configuration of the model, in order to pad tensors")

        padded_tensor = pad_token_id * torch.ones(
            (tensor.shape[0], max_length), dtype=tensor.dtype, device=tensor.device
        )
        padded_tensor[:, : tensor.shape[-1]] = tensor
        return padded_tensor