如何解决无效参数:没有足够的时间用于目标转换序列
我试图运行这个 HTR 模型 https://github.com/arthurflor23/handwritten-text-recognition
,但它给了我这个错误 Invalid argument: Not enough time for target transition sequence
。问题,我认为在ctc_batch_cost
。我的图片尺寸是 (137,518),文本的 max_len 是 137。我知道如何解决这个问题吗?
解决方法
我解决了这个问题,这是由于输入的大小。
Layer (type) Output Shape Param #
=================================================================
input (InputLayer) [(None,1024,128,1)] 0
_________________________________________________________________
conv2d (Conv2D) (None,64,16) 160
_________________________________________________________________
p_re_lu (PReLU) (None,16) 16
_________________________________________________________________
batch_normalization (BatchNo (None,16) 112
_________________________________________________________________
full_gated_conv2d (FullGated (None,16) 4640
_________________________________________________________________
conv2d_1 (Conv2D) (None,32) 4640
_________________________________________________________________
p_re_lu_1 (PReLU) (None,32) 32
_________________________________________________________________
batch_normalization_1 (Batch (None,32) 224
_________________________________________________________________
full_gated_conv2d_1 (FullGat (None,32) 18496
_________________________________________________________________
conv2d_2 (Conv2D) (None,512,16,40) 10280
_________________________________________________________________
p_re_lu_2 (PReLU) (None,40) 40
_________________________________________________________________
batch_normalization_2 (Batch (None,40) 280
_________________________________________________________________
full_gated_conv2d_2 (FullGat (None,40) 28880
_________________________________________________________________
dropout (Dropout) (None,40) 0
_________________________________________________________________
conv2d_3 (Conv2D) (None,48) 17328
_________________________________________________________________
p_re_lu_3 (PReLU) (None,48) 48
_________________________________________________________________
batch_normalization_3 (Batch (None,48) 336
_________________________________________________________________
full_gated_conv2d_3 (FullGat (None,48) 41568
_________________________________________________________________
dropout_1 (Dropout) (None,48) 0
_________________________________________________________________
conv2d_4 (Conv2D) (None,256,4,56) 21560
_________________________________________________________________
p_re_lu_4 (PReLU) (None,56) 56
_________________________________________________________________
batch_normalization_4 (Batch (None,56) 392
_________________________________________________________________
full_gated_conv2d_4 (FullGat (None,56) 56560
_________________________________________________________________
dropout_2 (Dropout) (None,56) 0
_________________________________________________________________
conv2d_5 (Conv2D) (None,64) 32320
_________________________________________________________________
p_re_lu_5 (PReLU) (None,64) 64
_________________________________________________________________
batch_normalization_5 (Batch (None,64) 448
_________________________________________________________________
reshape (Reshape) (None,256) 0
_________________________________________________________________
bidirectional (Bidirectional (None,256) 296448
_________________________________________________________________
dense (Dense) (None,256) 65792
_________________________________________________________________
bidirectional_1 (Bidirection (None,256) 296448
_________________________________________________________________
dense_1 (Dense) (None,332) 85324
=================================================================
look at the final layer ( dense_1 ) the second dimension is 256,so your text label should be <=256,not more. The problem comes from here.
版权声明:本文内容由互联网用户自发贡献,该文观点与技术仅代表作者本人。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容, 请发送邮件至 dio@foxmail.com 举报,一经查实,本站将立刻删除。