Source code for flash.text.seq2seq.core.model
# Copyright The PyTorch Lightning team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import warnings
from typing import Any, Dict, Iterable, List, Optional, Type, Union
import torch
from pytorch_lightning import Callback
from pytorch_lightning.utilities import rank_zero_info
from torch import Tensor
from torch.nn import Module
from flash.core.data.io.input import DataKeys, ServeInput
from flash.core.data.io.input_transform import InputTransform
from flash.core.data.io.output import Output
from flash.core.data.io.output_transform import OutputTransform
from flash.core.model import Task
from flash.core.registry import ExternalRegistry, FlashRegistry
from flash.core.serve import Composition
from flash.core.utilities.imports import _TOPIC_TEXT_AVAILABLE, requires
from flash.core.utilities.providers import _HUGGINGFACE
from flash.core.utilities.types import (
INPUT_TRANSFORM_TYPE,
LOSS_FN_TYPE,
LR_SCHEDULER_TYPE,
METRICS_TYPE,
OPTIMIZER_TYPE,
)
from flash.text.input import TextDeserializer
from flash.text.ort_callback import ORTCallback
from flash.text.seq2seq.core.collate import TextSeq2SeqCollate
if _TOPIC_TEXT_AVAILABLE:
from transformers import AutoModelForSeq2SeqLM
HUGGINGFACE_BACKBONES = ExternalRegistry(
AutoModelForSeq2SeqLM.from_pretrained,
"backbones",
_HUGGINGFACE,
)
else:
AutoModelForSeq2SeqLM, PreTrainedTokenizerBase = None, None
HUGGINGFACE_BACKBONES = FlashRegistry("backbones")
def _pad_tensors_to_max_len(model_cfg, tensor, max_length):
pad_token_id = model_cfg.pad_token_id if model_cfg.pad_token_id else model_cfg.eos_token_id
if pad_token_id is None:
raise ValueError(
f"Make sure that either `config.pad_token_id` or `config.eos_token_id` "
f"is defined if tensor has to be padded to `max_length`={max_length}"
)
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
[docs]class Seq2SeqTask(Task):
"""General Task for Sequence2Sequence.
Args:
max_source_length: The maximum length to pad / truncate input sequences to.
max_target_length: The maximum length to pad / truncate target sequences to.
padding: The type of padding to apply. One of: "longest" or ``True``, "max_length", "do_not_pad" or
``False``.
loss_fn: Loss function for training
optimizer: Optimizer to use for training.
lr_scheduler: The LR scheduler to use during training.
metrics: Metrics to compute for training and evaluation. Changing this argument currently has no effect
learning_rate: Learning rate to use for training, defaults to `3e-4`
num_beams: Number of beams to use in validation when generating predictions. Defaults to `4`
enable_ort: Enable Torch ONNX Runtime Optimization: https://onnxruntime.ai/docs/#onnx-runtime-for-training
"""
required_extras: str = "text"
backbones: FlashRegistry = FlashRegistry("backbones") + HUGGINGFACE_BACKBONES
def __init__(
self,
backbone: str = "t5-small",
tokenizer_kwargs: Optional[Dict[str, Any]] = None,
max_source_length: int = 128,
max_target_length: int = 128,
padding: Union[str, bool] = "max_length",
loss_fn: LOSS_FN_TYPE = None,
optimizer: OPTIMIZER_TYPE = "Adam",
lr_scheduler: LR_SCHEDULER_TYPE = None,
metrics: METRICS_TYPE = None,
learning_rate: Optional[float] = None,
num_beams: Optional[int] = None,
enable_ort: bool = False,
output_transform: Optional[OutputTransform] = None,
):
os.environ["TOKENIZERS_PARALLELISM"] = "TRUE"
# disable HF thousand warnings
warnings.simplefilter("ignore")
# set os environ variable for multiprocesses
os.environ["PYTHONWARNINGS"] = "ignore"
super().__init__(
loss_fn=loss_fn,
optimizer=optimizer,
lr_scheduler=lr_scheduler,
metrics=metrics,
learning_rate=learning_rate,
output_transform=output_transform,
)
self.collate_fn = TextSeq2SeqCollate(
backbone=backbone,
tokenizer_kwargs=tokenizer_kwargs,
max_source_length=max_source_length,
max_target_length=max_target_length,
padding=padding,
)
self.model = self.backbones.get(backbone)()
self.enable_ort = enable_ort
self.max_source_length = max_source_length
self.max_target_length = max_target_length
self.num_beams = num_beams
self._initialize_model_specific_parameters()
def forward(self, x: Any) -> Any:
max_length = self.max_target_length
num_beams = self.num_beams if self.num_beams else self.model.config.num_beams
generated_tokens = self.model.generate(
input_ids=x["input_ids"],
attention_mask=x.get("attention_mask", None),
max_length=max_length,
num_beams=num_beams,
)
# in case the batch is shorter than max length, the output should be padded
if generated_tokens.shape[-1] < max_length:
generated_tokens = _pad_tensors_to_max_len(
model_cfg=self.model.config, tensor=generated_tokens, max_length=max_length
)
return generated_tokens
def training_step(self, batch: Any, batch_idx: int) -> Tensor:
batch["labels"] = batch.pop(DataKeys.TARGET)
outputs = self.model(**batch)
loss = outputs[0]
self.log("train_loss", loss)
return loss
def common_step(self, prefix: str, batch: Any) -> Tensor:
batch["labels"] = batch.pop(DataKeys.TARGET)
generated_tokens = self(batch)
self.compute_metrics(generated_tokens, batch, prefix)
def validation_step(self, batch: Any, batch_idx: int, dataloader_idx: int = 0):
self.common_step("val", batch)
def test_step(self, batch: Any, batch_idx: int, dataloader_idx: int = 0):
self.common_step("test", batch)
def predict_step(self, batch: Any, batch_idx: int, dataloader_idx: int = 0) -> Any:
output = super().predict_step(batch, batch_idx, dataloader_idx)
return self.decode(output)
def compute_metrics(self, generated_tokens, batch, prefix):
pass
@property
def task(self) -> Optional[str]:
"""Override to define AutoConfig task specific parameters stored within the model."""
return
def _initialize_model_specific_parameters(self):
task_specific_params = self.model.config.task_specific_params
if task_specific_params:
pars = task_specific_params.get(self.task, {})
rank_zero_info(f"Overriding model paramameters for {self.task} as defined within the model:\n {pars}")
self.model.config.update(pars)
def decode(self, tokens: Tensor) -> List[str]:
decoded_str = self.collate_fn.tokenizer.batch_decode(tokens, skip_special_tokens=True)
return [str.strip(s) for s in decoded_str]
[docs] def modules_to_freeze(self) -> Union[Module, Iterable[Union[Module, Iterable]]]:
"""Return the module attributes of the model to be frozen."""
model_type = self.model.config.model_type
_modules = []
is_t5 = model_type in ["t5", "mt5"]
model = self.model if is_t5 else self.model.model
_modules.append(model.shared)
for layer in (model.encoder, model.decoder):
_modules.append(layer.embed_tokens)
if not is_t5:
_modules.append(layer.embed_positions)
return _modules
def configure_callbacks(self) -> List[Callback]:
callbacks = super().configure_callbacks() or []
if self.enable_ort:
callbacks.append(ORTCallback())
return callbacks
@requires("serve")
def serve(
self,
host: str = "127.0.0.1",
port: int = 8000,
sanity_check: bool = True,
input_cls: Optional[Type[ServeInput]] = TextDeserializer,
transform: INPUT_TRANSFORM_TYPE = InputTransform,
transform_kwargs: Optional[Dict] = None,
output: Optional[Union[str, Output]] = None,
) -> Composition:
return super().serve(host, port, sanity_check, input_cls, transform, transform_kwargs, output)
