use float32 reference and raise rtol for test_conv3d in float16

theano/gpuarray/tests/test_dnn.py

@@ -33,6 +33,13 @@ def set_precision(floatX):
     return precision
 
 
+# If using float16, cast reference input to float32
+def ref_cast(x):
+    if theano.config.floatX == 'float16':
+        x = T.cast(x, 'float32')
+    return x
+
+
 def test_dnn_conv_desc_merge():
     if not dnn.dnn_available(test_ctx_name):
         raise SkipTest(dnn.dnn_available.msg)
@@ -1044,13 +1051,18 @@ def test_conv3d_fwd():
         # Compile a theano function for the reference implementation
         conv_ref = theano.tensor.nnet.corr3d.Corr3dMM(border_mode=border_mode,
                                                       subsample=subsample
-                                                      )(inputs, flipped_filters)
+                                                      )(ref_cast(inputs), flipped_filters)
         f_ref = theano.function([], conv_ref, mode="FAST_RUN")
 
         # Compare the results of the two implementations
         res_ref = f_ref()
         res = f()
-        utt.assert_allclose(res_ref, res)
+        # raise rtol to make the test pass with more seed.
+        rtol = None
+        # Raise tolerance for float16
+        if theano.config.floatX == 'float16':
+            rtol = 6e-2
+        utt.assert_allclose(res_ref, res, rtol=rtol)
 
     test_cases = get_conv3d_test_cases()
     for (i_shape, f_shape, subsample), border_mode, conv_mode in test_cases:
@@ -1091,7 +1103,7 @@ def test_conv3d_bwd():
         # Compile a theano function for the reference implementation
         conv_ref = theano.tensor.nnet.corr3d.Corr3dMM(border_mode=border_mode,
                                                       subsample=subsample
-                                                      )(inputs, flipped_filters)
+                                                      )(ref_cast(inputs), flipped_filters)
         (grad_i_ref,
          grad_w_ref) = theano.tensor.grad(conv_ref.sum(),
                                           [inputs, filters])
@@ -1102,8 +1114,12 @@ def test_conv3d_bwd():
         res = f()
         # Needed for big size for some seed
         # raise rtol to make the test pass with more seed.
-        utt.assert_allclose(res_ref[0], res[0], rtol=2e-5)
-        utt.assert_allclose(res_ref[1], res[1], rtol=2e-5)
+        rtol = None
+        # Raise tolerance for float16
+        if theano.config.floatX == 'float16':
+            rtol = 5e-2
+        utt.assert_allclose(res_ref[0], res[0], rtol=rtol)
+        utt.assert_allclose(res_ref[1], res[1], rtol=rtol)
 
     test_cases = get_conv3d_test_cases()
     for (i_shape, f_shape, subsample), border_mode, conv_mode in test_cases: