Support variable-dimension input feature for 2D convolution operation. #2215
Conversation
qingqing01
requested review from
reyoung,
lcy-seso,
hedaoyuan,
Xreki and
xinghai-sun
| self.__height__ = 0 | ||
|
|
||
| def pre_scan(self, dat): | ||
| self.__height__ += 1 | ||
|
|
||
| def finish_pre_scan(self, argument): | ||
| def finish_pre_scan(self, argument, dat=None): |
There was a problem hiding this comment.
It is not necessary to change finish_pre_scan interface. The self.shape could be set in pre_scan routine. and should we check shape are same for each data?
There was a problem hiding this comment.
Done. I move the shape settings to pre_scan function and also check the shape for each data.
| raise ValueError("The dimension of input is greater than 3.") | ||
| dim = reduce(lambda x, y: x * y, self.__shape__) | ||
| if len(self.__shape__) == 1: | ||
| assert dim == self.input_type.dim |
There was a problem hiding this comment.
if len(self.__shape__) == 1 and dim != self.input_type.dim:
raise ValueError(...)Please use exception instead of assert.
And it seems that this check should be in pre_scan method.
| self.__height__ = 0 | ||
|
|
||
| def scan(self, dat): | ||
| if isinstance(dat, numpy.ndarray): | ||
| assert self.__shape__ == dat.shape |
There was a problem hiding this comment.
It is no need to validate data if you validate data shape in pre_scan method.
| self.__height__ = 0 | ||
|
|
||
| def scan(self, dat): | ||
| if isinstance(dat, numpy.ndarray): | ||
| assert self.__shape__ == dat.shape | ||
| dat = dat.flatten() |
There was a problem hiding this comment.
what about not a numpy array?
There was a problem hiding this comment.
Handled this case. I test the numpy.array() time and it's better to use NumPy array.
qingqing01
changed the title
| self.__height__ = 0 | ||
|
|
||
| def pre_scan(self, dat): | ||
| self.__height__ += 1 | ||
|
|
||
| def finish_pre_scan(self, argument): | ||
| def finish_pre_scan(self, argument, dat=None): |
There was a problem hiding this comment.
Done. I move the shape settings to pre_scan function and also check the shape for each data.
| raise ValueError("The dimension of input is greater than 3.") | ||
| dim = reduce(lambda x, y: x * y, self.__shape__) | ||
| if len(self.__shape__) == 1: | ||
| assert dim == self.input_type.dim |
| self.__height__ = 0 | ||
|
|
||
| def scan(self, dat): | ||
| if isinstance(dat, numpy.ndarray): | ||
| assert self.__shape__ == dat.shape |
| self.__height__ = 0 | ||
|
|
||
| def scan(self, dat): | ||
| if isinstance(dat, numpy.ndarray): | ||
| assert self.__shape__ == dat.shape | ||
| dat = dat.flatten() |
There was a problem hiding this comment.
Handled this case. I test the numpy.array() time and it's better to use NumPy array.
qingqing01
force-pushed
the
variable_input
branch
from
7069654
to
8e06f73
Compare
dense_arraytype instead of usingdense_vector.The data feeder will set height and width according to the shape of input data each mini-batch.
Fix #2198