# HG changeset patch
# User James Bergstra <bergstrj@iro.umontreal.ca>
# Date 1282683669 14400
# Node ID c6d08a76096058b71f32ace3a9846144abeb51ea
# Parent  8ba8b08e0442daedac3dcacf5d47fa7cfdd6e211
added sgd_updates to gd/sgd.py. Modif mcRBM to use it.

diff -r 8ba8b08e0442 -r c6d08a760960 pylearn/algorithms/mcRBM.py
--- a/pylearn/algorithms/mcRBM.py	Tue Aug 24 16:51:53 2010 -0400
+++ b/pylearn/algorithms/mcRBM.py	Tue Aug 24 17:01:09 2010 -0400
@@ -203,6 +203,7 @@
 import pylearn
 from pylearn.sampling.hmc import HMC_sampler
 from pylearn.io import image_tiling
+from pylearn.gd.sgd import sgd_updates
 
 #TODO: This should be in the datasets folder
 import pylearn.datasets.config
@@ -215,15 +216,6 @@
 
 
 #TODO: This should be in the nnet part of the library
-def sgd_updates(params, grads, lr):
-    try:
-        float(lr)
-        lr = [lr for p in params]
-    except TypeError:
-        pass
-    updates = [(p, p - plr * gp) for (plr, p, gp) in zip(lr, params, grads)]
-    return updates
-
 def hidden_cov_units_preactivation_given_v(rbm, v, small=0.5):
     """Return argument to the sigmoid that would give mean of covariance hid units
 
@@ -453,7 +445,7 @@
     sgd_ups = sgd_updates(
                 rbm.params,
                 grads,
-                lr=[2*s_lr, .2*s_lr, .02*s_lr, .1*s_lr, .02*s_lr ])
+                stepsizes=[2*s_lr, .2*s_lr, .02*s_lr, .1*s_lr, .02*s_lr ])
     learn_fn = function([batch_idx, s_lr, s_l1_penalty], 
             outputs=[ 
                 grads[0].norm(2),
diff -r 8ba8b08e0442 -r c6d08a760960 pylearn/gd/sgd.py
--- a/pylearn/gd/sgd.py	Tue Aug 24 16:51:53 2010 -0400
+++ b/pylearn/gd/sgd.py	Tue Aug 24 17:01:09 2010 -0400
@@ -1,8 +1,27 @@
-"""A stochastic gradient descent minimizer. (Possibly the simplest minimizer.)
+"""A stochastic gradient descent minimizer.
 """
 
 import theano
 
+def sgd_updates(params, grads, stepsizes):
+    """Return a list of (pairs) that can be used as updates in theano.function to implement
+    stochastic gradient descent.
+
+    :param params: variables to adjust in order to minimize some cost
+    :type params: a list of variables (theano.function will require shared variables)
+    :param grads: the gradient on each param (with respect to some cost)
+    :type grads: list of theano expressions
+    :param stepsizes: step by this amount times the negative gradient on each iteration
+    :type stepsizes: [symbolic] scalar or list of one [symbolic] scalar per param
+    """
+    try:
+        iter(stepsizes)
+    except:
+        stepsizes = [stepsizes for p in params]
+    updates = [(p, p - step * gp) for (step, p, gp) in zip(stepsizes, params, grads)]
+    return updates
+
+
 class StochasticGradientDescent(theano.Module):
     """Fixed stepsize gradient descent