fix issue 207

Author: nicholas-leonard

SeqLSTM.lua

@@ -359,8 +359,8 @@ function SeqLSTM:accGradParameters(input, gradOutput, scale)
 end
 
 function SeqLSTM:forget()
-   self.c0:zero()
-   self.h0:zero()
+   self.c0:resize(0)
+   self.h0:resize(0)
 end
 
 -- Toggle to feed long sequences using multiple forwards.

test/test.lua

@@ -4898,6 +4898,17 @@ function rnntest.FastLSTM_issue203()
    mytester:assert(err < 0.000001, "error "..err)
 end
 
+function rnntest.SeqLSTM_issue207()
+   local lstm = nn.SeqLSTM(10, 10)
+   lstm.batchfirst = true
+   lstm:remember('both')
+   lstm:training()
+   lstm:forward(torch.Tensor(32, 20, 10))
+   lstm:evaluate()
+   lstm:forget()
+   lstm:forward(torch.Tensor(1, 20, 10))
+end
+
 function rnn.test(tests, benchmark_)
    mytester = torch.Tester()
    benchmark = benchmark_