paddlenlp.trainer.utils.helper 源代码

# Copyright 2020-present the HuggingFace Inc. team.
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# 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.

# This file is modified from
#  https://github.com/huggingface/transformers/blob/main/src/transformers

from typing import Any, Optional

import numpy as np
import paddle
import paddle.distributed as dist

__all__ = [
    "distributed_concat",
    "paddle_pad_and_concatenate",
    "nested_concat",
    "nested_detach",
    "nested_numpify",
    "nested_truncate",
]


def distributed_concat(tensor: Any,
                       num_total_examples: Optional[int] = None) -> Any:
    try:
        if isinstance(tensor, (tuple, list)):
            return type(tensor)(distributed_concat(t, num_total_examples)
                                for t in tensor)
        output_tensors = []
        # output_tensors = [tensor.clone() for _ in range(dist.get_world_size())]
        # output_tensors = [
        #     t if len(t.shape) > 0 else t[None] for t in output_tensors
        # ]
        dist.all_gather(output_tensors, tensor)
        concat = paddle.concat(output_tensors, axis=0)

        # truncate the dummy elements added by SequentialDistributedSampler
        if num_total_examples is not None:
            concat = concat[:num_total_examples]
        return concat
    except AssertionError:
        raise AssertionError("Not currently using distributed training")


[文档]def paddle_pad_and_concatenate(tensor1, tensor2, padding_index=-100): """Concatenates `tensor1` and `tensor2` on first axis, applying padding on the second if necessary.""" if len(tensor1.shape) == 1 or tensor1.shape[1] == tensor2.shape[1]: return paddle.concat((tensor1, tensor2), axis=0) # raise ValueError("Error") # Let's figure out the new shape new_shape = (tensor1.shape[0] + tensor2.shape[0], max(tensor1.shape[1], tensor2.shape[1])) + tuple( tensor1.shape[2:]) # Now let's fill the result tensor # result = tensor1.new_full(new_shape, padding_index) result = paddle.full(new_shape, padding_index, dtype=tensor1.dtype) result[:tensor1.shape[0], :tensor1.shape[1]] = tensor1 result[tensor1.shape[0]:, :tensor2.shape[1]] = tensor2 return result
def numpy_pad_and_concatenate(array1, array2, padding_index=-100): """Concatenates `array1` and `array2` on first axis, applying padding on the second if necessary.""" if len(array1.shape) == 1 or array1.shape[1] == array2.shape[1]: return np.concatenate((array1, array2), axis=0) # Let's figure out the new shape new_shape = (array1.shape[0] + array2.shape[0], max(array1.shape[1], array2.shape[1])) + array1.shape[2:] # Now let's fill the result tensor result = np.full_like(array1, padding_index, shape=new_shape) result[:array1.shape[0], :array1.shape[1]] = array1 result[array1.shape[0]:, :array2.shape[1]] = array2 return result
[文档]def nested_concat(tensors, new_tensors, padding_index=-100): """ Concat the `new_tensors` to `tensors` on the first dim and pad them on the second if needed. Works for tensors or nested list/tuples of tensors. """ assert type(tensors) == type( new_tensors ), f"Expected `tensors` and `new_tensors` to have the same type but found {type(tensors)} and {type(new_tensors)}." if isinstance(tensors, (list, tuple)): return type(tensors)(nested_concat(t, n, padding_index=padding_index) for t, n in zip(tensors, new_tensors)) elif isinstance(tensors, paddle.Tensor): return paddle_pad_and_concatenate(tensors, new_tensors, padding_index=padding_index) elif isinstance(tensors, np.ndarray): return numpy_pad_and_concatenate(tensors, new_tensors, padding_index=padding_index) else: raise TypeError( f"Unsupported type for concatenation: got {type(tensors)}")
[文档]def nested_detach(tensors): "Detach `tensors` (even if it's a nested list/tuple of tensors)." if isinstance(tensors, (list, tuple)): return type(tensors)(nested_detach(t) for t in tensors) return tensors.detach()
[文档]def nested_numpify(tensors): "Numpify `tensors` (even if it's a nested list/tuple of tensors)." if isinstance(tensors, (list, tuple)): return type(tensors)(nested_numpify(t) for t in tensors) t = tensors.cpu() if t.dtype == paddle.float16: t = t.cast(paddle.float32) return t.numpy()
[文档]def nested_truncate(tensors, limit): "Truncate `tensors` at `limit` (even if it's a nested list/tuple of tensors)." if isinstance(tensors, (list, tuple)): return type(tensors)(nested_truncate(t, limit) for t in tensors) return tensors[:limit]