I did indeed have a topology issue (I'd set up two layers both named
input, and didn't get a conx-related error, which seems a bit
unfortunate :-). Thanks for the pointer.
I was aware of the syntax you discuss below. The issue that made it
difficult for me to just use that stuff was that bias is not treated
as a connection. Personally, it would have been easier for me to get
my head around some of this stuff had it not been treated differently
in the API, as many text just refer to the bias units as the 0th
input to a unit. :-)
--b
On Apr 24, 2007, at 11:12 AM, Douglas S. Blank wrote:
Belinda,
I don't think that a layer should be Undefined. Sounds like that
layer hasn't been .connect() ed to another layer yet (when the
layers get connected, there is logic there to determine if each of
the layers are Input, Hidden, or Output types). Could there be a
mistake in the topology?
FYI, there are bias units (aka, .weight[]) on Input and Context
layers, even though they are not used.
Also, did you know that you can refer to weights with this syntax:
A single weight:
net["fromLayerName", "toLayerName"][fromPos][toPos]
net["input", "hidden"][0][1]
and the entire matrix:
net["fromLayerName", "toLayerName"].weight
net["input", "hidden"].weight
If none of these bits help, feel free to send me some code and I'll
take a look at it.
-Doug
belinda thom wrote:
Hi,
I'd borrowed and then modified some code to store weights in a
dictionary:
for layer in layers:
if layer.type != 'Input':
for i in range(layer.size):
debug_here()
d['%s[%d][0]'%(layer.name[0:3],i+1)]=layer.weight[i]
for connection in connections:
for i in range(connection.fromLayer.size):
for j in range(connection.toLayer.size):
d['%s[%d]%s[%d]'%(connection.fromLayer.name[0:3],i
+1, \
connection.toLayer.name[0:3],j
+1)] = connection.weight[i][j]
w/the hope that putting these things in a dictionary would make
the actual weight values easier to understand/explain in class.
My plan was to have connection weights available via, e.g. d['inp
[1]hid[1]'] where this would access the 1st input to 1st hidden
unit weight. For biases, I would do, e.g. d['hid[1][0]'] which
would access 1st hidden unit's bias.
My problem is with the "if layer.type != 'Input'" check (part of
the code I inherited), as I'm getting more bias weights than make
sense. So I started poking around in the debugger and noticed an
'Undefined' layer type. What is this? I think its the thing that's
producing, e.g. d['inp[1][0]'], where I would expect the input
layers to have no useful bias.
Thanks,
--b
_______________________________________________
Pyro-users mailing list
[email protected]
http://emergent.brynmawr.edu/mailman/listinfo/pyro-users
_______________________________________________
Pyro-users mailing list
[email protected]
http://emergent.brynmawr.edu/mailman/listinfo/pyro-users
