python - 对 im2txt 微调 inception v3 时损失并没有减少

Question

我在微调 im2txt 的预训练 Inception v3 模型时遇到问题。由于某种原因，初始训练并没有减少太多损失，并且微调 Inception v3 也没有减少我的训练数据的任何损失。我正在尝试找出原因，任何见解都会有所帮助。

im2txt 是一个模型，它接受图像输入并打印出标题列表作为输出。最初，im2txt 打印出一个标题作为连贯的句子，描述图像。为了适合我的项目，我更改了训练数据中的代码和标签，以便它打印出与图像相关的单词列表。

例如，我的图像看起来像这样。请注意，图像中的对象比平均 Imagenet 图像要多:

我的标签标题如下所示:

 female lady woman clothes shop customer

我总共有 400,000 张图像和相应的标签说明。我初始训练了13万步，微调了17万步。词汇共有750个单词。初始训练+微调的损失曲线(从步骤 130,000 开始)如下:

准确率和召回率约为0.35~40。

训练的配置文件如下:

# Copyright 2016 The TensorFlow 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.
# ==============================================================================

"""Image-to-text model and training configurations."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function


class ModelConfig(object):
  """Wrapper class for model hyperparameters."""

  def __init__(self):
    """Sets the default model hyperparameters."""
    # File pattern of sharded TFRecord file containing SequenceExample protos.
    # Must be provided in training and evaluation modes.
    self.input_file_pattern = None

    # Image format ("jpeg" or "png").
    self.image_format = "jpeg"

    # Approximate number of values per input shard. Used to ensure sufficient
    # mixing between shards in training.
    self.values_per_input_shard = 2300
    # Minimum number of shards to keep in the input queue.
    self.input_queue_capacity_factor = 2
    # Number of threads for prefetching SequenceExample protos.
    self.num_input_reader_threads = 1

    # Name of the SequenceExample context feature containing image data.
    self.image_feature_name = "image/data"
    # Name of the SequenceExample feature list containing integer captions.
    self.caption_feature_name = "image/caption_ids"

    # Number of unique words in the vocab (plus 1, for <UNK>).
    # The default value is larger than the expected actual vocab size to allow
    # for differences between tokenizer versions used in preprocessing. There is
    # no harm in using a value greater than the actual vocab size, but using a
    # value less than the actual vocab size will result in an error.
    self.vocab_size = 750

    # Number of threads for image preprocessing. Should be a multiple of 2.
    self.num_preprocess_threads = 4

    # Batch size.
    self.batch_size = 32

    # File containing an Inception v3 checkpoint to initialize the variables
    # of the Inception model. Must be provided when starting training for the
    # first time.
    self.inception_checkpoint_file = None

    # Dimensions of Inception v3 input images.
    self.image_height = 299
    self.image_width = 299

    # Scale used to initialize model variables.
    self.initializer_scale = 0.08

    # LSTM input and output dimensionality, respectively.
    self.embedding_size = 512
    self.num_lstm_units = 512

    # If < 1.0, the dropout keep probability applied to LSTM variables.
    self.lstm_dropout_keep_prob = 0.7


class TrainingConfig(object):
  """Wrapper class for training hyperparameters."""

  def __init__(self):
    """Sets the default training hyperparameters."""
    # Number of examples per epoch of training data.
    self.num_examples_per_epoch = 100000

    # Optimizer for training the model.
    self.optimizer = "SGD"

    # Learning rate for the initial phase of training.
    self.initial_learning_rate = 2.0
    self.learning_rate_decay_factor = 0.5
    self.num_epochs_per_decay = 1.0

    # Learning rate when fine tuning the Inception v3 parameters.
    self.train_inception_learning_rate = 0.005

    # If not None, clip gradients to this value.
    self.clip_gradients = 5.0

    # How many model checkpoints to keep.
    self.max_checkpoints_to_keep = 5

任何建议、见解或观察都会很棒。

Answer 1

请注意，im2txt 的功能非常强大，因为它可以生成可读的句子。在句子中，一个单词与相邻的单词相关，这就是它起作用的原因。在您的情况下，您正在更改模型以生成一组顺序不相关的标签。实际上，在im2txt模型中，单词“female”、“woman”和“lady”基本上是相同的概念，im2txt可以创建滑动这些单词的句子。例如，在im2txt中:“这位女士穿着一条漂亮的粉红色裙子”与“这位女士的裙子是粉红色的”相同，或者应该非常相似。在你的情况下，如果你不提供一些单词顺序规则，它会让你的模型变得很困惑，并且可能无法学习。

如果您想从图像中获取标签列表，您应该仅使用具有多标签分类的初始模型(通过 sigmoid 层更改 softmax 层并使用 sigmoid 交叉熵作为损失函数)。