forward_all with error objective as output

[kevin-w-li]

This might be a misuse, but at least there should be some warning.

The PyCaffe function forward_all assumes that, in each batch, the blobs have the same outermost dimension as the input. But if the network has the error objective as the output, which will be singleton returned by forward not matter how many input there is, then the line at line 204

all_outs[out].extend(out_blob.copy())

will complain out_blob is not an array. Also, when checking for padding before this function returns, six.next does not look for the correct output (error output or blob value).

I guess this pad value should be calculated using the batch size of the network, and deal with the error value differently...

The same problem is also in other 'all' functions