BasicConvLSTM batch size can't be different between training and inference about rnn HOT 4 CLOSED

Actually, any model in Tensorflow, as far as I know, different batch size can not be setted whatever in training or in inference process. Hence, I suggest that you can set the batch size in tf.placeholder with None, which will avoid this problem you suffered.

from rnn.

Thanks for your quick reply. More specifically, starting from Line 90 in ConvLSTM.py (peephole part), 'w_ci' is defined by vs.get_variable("w_ci", cell.shape, inputs.dtype), which results in shape [512, 7, 7, 128] during training, and 512 is the batch size I set in this case. However if I set batch size to 1000 during inference, the shape of 'w_ci' will change to [1000, 7, 7, 128], which, will fail because this is not consistent with the training shape.

Apparently the first dimension of 'w_ci' cannot be None or -1 in this case otherwise the variable cannot be created. However once the variable is created, the shape is also fixed, which leads to unchangeable batch size.

If I set off the peephole mode, this problem will not appear which means I can set different batch size. So this problem is a little bit weird to me. I will try fixing it.

from rnn.

Thanks for your reminder. After check carefully, it should be a vital bug. And it has been debuged as follow:

            kernel_shape = cell.shape.as_list()[-3:]
            w_ci = vs.get_variable(
                "w_ci", kernel_shape, inputs.dtype)
            w_cf = vs.get_variable(
                "w_cf", kernel_shape, inputs.dtype)
            w_co = vs.get_variable(
                "w_co", kernel_shape, inputs.dtype)

You can change your code in your project, and if there is any problem, please feel free to contact me.

from rnn.

Everything looks well! Thanks for your help!

from rnn.