我正在使用 cnn + lstm + ctc 网络(基于 https://arxiv.org/pdf/1507.05717.pdf )进行中文场景文本识别。对于大量的类(3500+)，网络很难训练。我听说使用 Group LSTM(https://arxiv.org/abs/1703.10722，O. Kuchaiev 和 B. Ginsburg“LSTM 网络的分解技巧”，ICLR 2017 研讨会。)可以减少参数数量并加速训练，所以我尝试使用它在我的代码中。
我使用双层双向 lstm。这是使用 tf.contrib.rnn.LSTMCell 的原始代码
rnn_outputs, _, _ =
tf.contrib.rnn.stack_bidirectional_dynamic_rnn(
[tf.contrib.rnn.LSTMCell(num_units=self.num_hidden, state_is_tuple=True) for _ in range(self.num_layers)],
[tf.contrib.rnn.LSTMCell(num_units=self.num_hidden, state_is_tuple=True) for _ in range(self.num_layers)],
self.rnn_inputs, dtype=tf.float32, sequence_length=self.rnn_seq_len, scope='BDDLSTM')
训练很慢。 100 小时后，在测试集上的预测准确率仍然是 39%。
现在我想使用 tf.contrib.rnn.GLSTMCell。当我用这个 GLSTMCell 替换 LSTMCell 时
rnn_outputs, _, _ = tf.contrib.rnn.stack_bidirectional_dynamic_rnn(
[tf.contrib.rnn.GLSTMCell(num_units=self.num_hidden, num_proj=self.num_proj, number_of_groups=4) for _ in range(self.num_layers)],
[tf.contrib.rnn.GLSTMCell(num_units=self.num_hidden, num_proj=self.num_proj, number_of_groups=4) for _ in range(self.num_layers)],
self.rnn_inputs, dtype=tf.float32, sequence_length=self.rnn_seq_len, scope='BDDLSTM')
出现以下错误
/home/frisasz/miniconda2/envs/dl/bin/python "/media/frisasz/DATA/FSZ_Work/deep learning/IDOCR_/work/train.py"
Traceback (most recent call last):
File "/media/frisasz/DATA/FSZ_Work/deep learning/IDOCR_/work/train.py", line 171, in <module>
train(train_dir='/media/frisasz/Windows/40T/', val_dir='../../0000/40V/')
File "/media/frisasz/DATA/FSZ_Work/deep learning/IDOCR_/work/train.py", line 41, in train
FLAGS.momentum)
File "/media/frisasz/DATA/FSZ_Work/deep learning/IDOCR_/work/model.py", line 61, in __init__
self.logits = self.rnn_net()
File "/media/frisasz/DATA/FSZ_Work/deep learning/IDOCR_/work/model.py", line 278, in rnn_net
self.rnn_inputs, dtype=tf.float32, sequence_length=self.rnn_seq_len, scope='BDDLSTM')
File "/home/frisasz/miniconda2/envs/dl/lib/python2.7/site-packages/tensorflow/contrib/rnn/python/ops/rnn.py", line 220, in stack_bidirectional_dynamic_rnn
dtype=dtype)
File "/home/frisasz/miniconda2/envs/dl/lib/python2.7/site-packages/tensorflow/python/ops/rnn.py", line 375, in bidirectional_dynamic_rnn
time_major=time_major, scope=fw_scope)
File "/home/frisasz/miniconda2/envs/dl/lib/python2.7/site-packages/tensorflow/python/ops/rnn.py", line 574, in dynamic_rnn
dtype=dtype)
File "/home/frisasz/miniconda2/envs/dl/lib/python2.7/site-packages/tensorflow/python/ops/rnn.py", line 737, in _dynamic_rnn_loop
swap_memory=swap_memory)
File "/home/frisasz/miniconda2/envs/dl/lib/python2.7/site-packages/tensorflow/python/ops/control_flow_ops.py", line 2770, in while_loop
result = context.BuildLoop(cond, body, loop_vars, shape_invariants)
File "/home/frisasz/miniconda2/envs/dl/lib/python2.7/site-packages/tensorflow/python/ops/control_flow_ops.py", line 2599, in BuildLoop
pred, body, original_loop_vars, loop_vars, shape_invariants)
File "/home/frisasz/miniconda2/envs/dl/lib/python2.7/site-packages/tensorflow/python/ops/control_flow_ops.py", line 2549, in _BuildLoop
body_result = body(*packed_vars_for_body)
File "/home/frisasz/miniconda2/envs/dl/lib/python2.7/site-packages/tensorflow/python/ops/rnn.py", line 720, in _time_step
skip_conditionals=True)
File "/home/frisasz/miniconda2/envs/dl/lib/python2.7/site-packages/tensorflow/python/ops/rnn.py", line 206, in _rnn_step
new_output, new_state = call_cell()
File "/home/frisasz/miniconda2/envs/dl/lib/python2.7/site-packages/tensorflow/python/ops/rnn.py", line 708, in <lambda>
call_cell = lambda: cell(input_t, state)
File "/home/frisasz/miniconda2/envs/dl/lib/python2.7/site-packages/tensorflow/python/ops/rnn_cell_impl.py", line 180, in __call__
return super(RNNCell, self).__call__(inputs, state)
File "/home/frisasz/miniconda2/envs/dl/lib/python2.7/site-packages/tensorflow/python/layers/base.py", line 441, in __call__
outputs = self.call(inputs, *args, **kwargs)
File "/home/frisasz/miniconda2/envs/dl/lib/python2.7/site-packages/tensorflow/contrib/rnn/python/ops/rnn_cell.py", line 2054, in call
R_k = _linear(x_g_id, 4 * self._group_shape[1], bias=False)
File "/home/frisasz/miniconda2/envs/dl/lib/python2.7/site-packages/tensorflow/python/ops/rnn_cell_impl.py", line 1005, in _linear
"but saw %s" % (shape, shape[1]))
ValueError: linear expects shape[1] to be provided for shape (?, ?), but saw ?
Process finished with exit code 1
我不确定 GLSTMCell 是否可以简单地替换 tf.contrib.rnn.stack_bidirectional_dynamic_rnn()(或其他有助于构建 rnn 的函数)中的 LSTMCell。我没有找到任何使用 GLSTMCell 的示例。有人知道用 GLSTMCell 构建双向 rnn 的正确方法吗？
请您参考如下方法:
我在尝试使用 bidirectional_dynamic_rnn 构建双向 GLSTM 时遇到了完全相同的错误。
在我的例子中，问题来自于 GLSTM 只能在以静态方式定义时使用:计算图形时不能有未定义的形状参数(例如 batch_size)。
因此，尝试在图形中定义所有将在 GLSTM 单元中的某个点结束的形状，它应该可以正常工作。
