为什么 tensorflow 模型文件的大小取决于数据集的大小？

Question

我保存的模型在 10K 句子的数据集上训练后的 .index、.meta 和 .data 文件的大小分别为 3KB、58MB 和 375MB

保持网络架构不变，并在 100K 句子的数据集上进行训练，文件大小分别为 3KB、139MB 和 860MB

我认为这表明大小取决于数据集的大小。 According to this answer，文件的大小应该与数据集的大小无关，因为神经网络的架构是相同的。

为什么大小差别这么大？

我还想知道除了链接答案中提到的文件之外，这些文件还包含什么。

这些文件是否包含与训练历史相关的信息，例如每一步的损失值等？

Answer 1

import tensorflow as tf
from tensorflow.python.training import checkpoint_utils as cp
cp.list_variables('./model.ckpt-12520')

运行上面的 sn -p 会得到以下输出

[('Variable', []), ('decoder/attention_wrapper/attention_layer/kernel', [600, 300]), ('decoder/attention_wrapper/attention_layer/kernel/Adam', [600, 300]), ('decoder/attention_wrapper/attention_layer/kernel/Adam_1', [600, 300]), ('decoder/attention_wrapper/bahdanau_attention/attention_b', [300]), ('decoder/attention_wrapper/bahdanau_attention/attention_b/Adam', [300]), ('decoder/attention_wrapper/bahdanau_attention/attention_b/Adam_1', [300]), ('decoder/attention_wrapper/bahdanau_attention/attention_g', []), ('decoder/attention_wrapper/bahdanau_attention/attention_g/Adam', []), ('decoder/attention_wrapper/bahdanau_attention/attention_g/Adam_1', []), ('decoder/attention_wrapper/bahdanau_attention/attention_v', [300]), ('decoder/attention_wrapper/bahdanau_attention/attention_v/Adam', [300]), ('decoder/attention_wrapper/bahdanau_attention/attention_v/Adam_1', [300]), ('decoder/attention_wrapper/bahdanau_attention/query_layer/kernel', [300, 300]), ('decoder/attention_wrapper/bahdanau_attention/query_layer/kernel/Adam', [300, 300]), ('decoder/attention_wrapper/bahdanau_attention/query_layer/kernel/Adam_1', [300, 300]), ('decoder/attention_wrapper/basic_lstm_cell/bias', [1200]), ('decoder/attention_wrapper/basic_lstm_cell/bias/Adam', [1200]), ('decoder/attention_wrapper/basic_lstm_cell/bias/Adam_1', [1200]), ('decoder/attention_wrapper/basic_lstm_cell/kernel', [900, 1200]), ('decoder/attention_wrapper/basic_lstm_cell/kernel/Adam', [900, 1200]), ('decoder/attention_wrapper/basic_lstm_cell/kernel/Adam_1', [900, 1200]), ('decoder/dense/kernel', [300, 49018]), ('decoder/dense/kernel/Adam', [300, 49018]), ('decoder/dense/kernel/Adam_1', [300, 49018]), ('decoder/memory_layer/kernel', [300, 300]), ('decoder/memory_layer/kernel/Adam', [300, 300]), ('decoder/memory_layer/kernel/Adam_1', [300, 300]), ('embeddings', [49018, 300]), ('embeddings/Adam', [49018, 300]), ('embeddings/Adam_1', [49018, 300]), ('loss/beta1_power', []), ('loss/beta2_power', []), ('stack_bidirectional_rnn/cell_0/bidirectional_rnn/bw/basic_lstm_cell/bias', [600]), ('stack_bidirectional_rnn/cell_0/bidirectional_rnn/bw/basic_lstm_cell/bias/Adam', [600]), ('stack_bidirectional_rnn/cell_0/bidirectional_rnn/bw/basic_lstm_cell/bias/Adam_1', [600]), ('stack_bidirectional_rnn/cell_0/bidirectional_rnn/bw/basic_lstm_cell/kernel', [450, 600]), ('stack_bidirectional_rnn/cell_0/bidirectional_rnn/bw/basic_lstm_cell/kernel/Adam', [450, 600]), ('stack_bidirectional_rnn/cell_0/bidirectional_rnn/bw/basic_lstm_cell/kernel/Adam_1', [450, 600]), ('stack_bidirectional_rnn/cell_0/bidirectional_rnn/fw/basic_lstm_cell/bias', [600]), ('stack_bidirectional_rnn/cell_0/bidirectional_rnn/fw/basic_lstm_cell/bias/Adam', [600]), ('stack_bidirectional_rnn/cell_0/bidirectional_rnn/fw/basic_lstm_cell/bias/Adam_1', [600]), ('stack_bidirectional_rnn/cell_0/bidirectional_rnn/fw/basic_lstm_cell/kernel', [450, 600]), ('stack_bidirectional_rnn/cell_0/bidirectional_rnn/fw/basic_lstm_cell/kernel/Adam', [450, 600]), ('stack_bidirectional_rnn/cell_0/bidirectional_rnn/fw/basic_lstm_cell/kernel/Adam_1', [450, 600]), ('stack_bidirectional_rnn/cell_1/bidirectional_rnn/bw/basic_lstm_cell/bias', [600]), ('stack_bidirectional_rnn/cell_1/bidirectional_rnn/bw/basic_lstm_cell/bias/Adam', [600]), ('stack_bidirectional_rnn/cell_1/bidirectional_rnn/bw/basic_lstm_cell/bias/Adam_1', [600]), ('stack_bidirectional_rnn/cell_1/bidirectional_rnn/bw/basic_lstm_cell/kernel', [450, 600]), ('stack_bidirectional_rnn/cell_1/bidirectional_rnn/bw/basic_lstm_cell/kernel/Adam', [450, 600]), ('stack_bidirectional_rnn/cell_1/bidirectional_rnn/bw/basic_lstm_cell/kernel/Adam_1', [450, 600]), ('stack_bidirectional_rnn/cell_1/bidirectional_rnn/fw/basic_lstm_cell/bias', [600]), ('stack_bidirectional_rnn/cell_1/bidirectional_rnn/fw/basic_lstm_cell/bias/Adam', [600]), ('stack_bidirectional_rnn/cell_1/bidirectional_rnn/fw/basic_lstm_cell/bias/Adam_1', [600]), ('stack_bidirectional_rnn/cell_1/bidirectional_rnn/fw/basic_lstm_cell/kernel', [450, 600]), ('stack_bidirectional_rnn/cell_1/bidirectional_rnn/fw/basic_lstm_cell/kernel/Adam', [450, 600]), ('stack_bidirectional_rnn/cell_1/bidirectional_rnn/fw/basic_lstm_cell/kernel/Adam_1', [450, 600])]

我意识到 embeddings 变量正在存储单词嵌入，这会导致这些文件的大小增加

cp.load_variable('./model.ckpt-12520', 'embeddings')

Answer 2

培训摘要包含在您的活动文件中。