diff --git a/sgd.lua b/sgd.lua
index e21c696..a93c5cf 100644
--- a/sgd.lua
+++ b/sgd.lua
@@ -36,6 +36,7 @@ function optim.sgd(opfunc, x, config, state)
    local nesterov = config.nesterov or false
    local lrs = config.learningRates
    local wds = config.weightDecays
+   local clip = config.clipval
    state.evalCounter = state.evalCounter or 0
    local nevals = state.evalCounter
    assert(not nesterov or (mom > 0 and damp == 0), "Nesterov momentum requires a momentum and zero dampening")
@@ -71,7 +72,13 @@ function optim.sgd(opfunc, x, config, state)
    -- (4) learning rate decay (annealing)
    local clr = lr / (1 + nevals*lrd)
 
-   -- (5) parameter update with single or individual learning rates
+   -- (5) gradient clipping
+   ndfdx = dfdx:norm()
+   if clip and ndfdx >= clip then
+      dfdx:div(ndfdx):mul(clip)
+   end 
+
+   -- (6) parameter update with single or individual learning rates
    if lrs then
       if not state.deltaParameters then
          state.deltaParameters = torch.Tensor():typeAs(x):resizeAs(dfdx)
@@ -82,7 +89,7 @@ function optim.sgd(opfunc, x, config, state)
       x:add(-clr, dfdx)
    end
 
-   -- (6) update evaluation counter
+   -- (7) update evaluation counter
    state.evalCounter = state.evalCounter + 1
 
    -- return x*, f(x) before optimization