You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@mxnet.apache.org by GitBox <gi...@apache.org> on 2018/06/25 23:35:25 UTC
[GitHub] szha closed pull request #11259: [MXNET-184] Fix flaky test
test_operator.test_binary_op due to numerical errors
szha closed pull request #11259: [MXNET-184] Fix flaky test test_operator.test_binary_op due to numerical errors
URL: https://github.com/apache/incubator-mxnet/pull/11259
This is a PR merged from a forked repository.
As GitHub hides the original diff on merge, it is displayed below for
the sake of provenance:
As this is a foreign pull request (from a fork), the diff is supplied
below (as it won't show otherwise due to GitHub magic):
diff --git a/tests/python/unittest/test_operator.py b/tests/python/unittest/test_operator.py
index 67426693436..c1f6ba0dcf9 100644
--- a/tests/python/unittest/test_operator.py
+++ b/tests/python/unittest/test_operator.py
@@ -60,14 +60,14 @@ def check_rnn_consistency(cell1, cell2, T, N, I, H, grad_req):
dy = mx.random.uniform(shape=mod1.get_outputs()[0].shape)
mod1.backward(out_grads=[dy])
- mod2.backward(out_grads=[dy])
+ mod2.backward(out_grads=[dy])
if grad_req != 'null':
assert_allclose(mod1.get_input_grads()[0].asnumpy(), mod2.get_input_grads()[0].asnumpy(), rtol=1e-2, atol=1e-4)
else:
assert(mod1.get_input_grads()[0] == None)
assert(mod2.get_input_grads()[0] == None)
-
-
+
+
@with_seed()
def test_lstm_sym():
@@ -77,7 +77,7 @@ def test_lstm_sym():
stack.add(mx.rnn.LSTMCell(H, prefix='l0_'))
stack.add(mx.rnn.LSTMCell(H, prefix='l1_'))
stack.add(mx.rnn.LSTMCell(H, prefix='l2_'))
-
+
check_rnn_consistency(fused, stack, T, N, I, H, 'write')
check_rnn_consistency(fused, stack, T, N, I, H, 'add')
check_rnn_consistency(fused, stack, T, N, I, H, 'null')
@@ -120,21 +120,21 @@ def test_gru_sym():
@with_seed()
def test_gru_bidirectional():
T, N, I, H = 5, 20, 800, 800
-
+
fused = mx.rnn.FusedRNNCell(H, num_layers=2, mode='gru',
bidirectional=True, get_next_state=True, prefix='')
-
+
stack = mx.rnn.SequentialRNNCell()
stack.add(mx.rnn.BidirectionalCell(
mx.rnn.GRUCell(H, prefix='l0_'),
mx.rnn.GRUCell(H, prefix='r0_'),
- output_prefix='bi_gru_0_'))
-
+ output_prefix='bi_gru_0_'))
+
stack.add(mx.rnn.BidirectionalCell(
mx.rnn.GRUCell(H, prefix='l1_'),
mx.rnn.GRUCell(H, prefix='r1_'),
output_prefix='bi_gru_1_'))
-
+
check_rnn_consistency(fused, stack, T, N, I, H, 'write')
check_rnn_consistency(fused, stack, T, N, I, H, 'add')
check_rnn_consistency(fused, stack, T, N, I, H, 'null')
@@ -1553,6 +1553,7 @@ def gen_broadcast_data_int(idx):
def gen_binary_data(dummy):
ndim = np.random.randint(1, 6)
shape = np.random.randint(1, 6, size=(ndim,))
+ #print("gen shape {}".format(shape))
return [np.random.random(shape), np.random.random(shape)]
@@ -1565,27 +1566,46 @@ def check_binary_op_forward(symbol, baseline, gen_data, rtol=1e-3, atol=1e-5, mx
sample_num = 200
for i in range(sample_num):
d = gen_data(i)
- x = baseline(d[0], d[1])
y = symbol.bind(default_context(), args={'a': mx.nd.array(d[0]), 'b': mx.nd.array(d[1])})
y.forward(is_train=True)
y = y.outputs[0].asnumpy()
+ x = baseline(d[0], d[1]).astype(y.dtype)
+
+ #np.set_printoptions(precision=20)
+
+ a = d[0]
+ b = d[1]
+ #print("a: {} {}".format(a.dtype, a))
+ #print("a: {} {}".format(b.dtype, b))
+
+ #print("x: {} {}".format(x.dtype, x))
+ #print("y: {} {}".format(y.dtype, y))
if mx_nd_func is not None:
d0 = mx.nd.array(d[0], dtype=d[0].dtype)
d1 = mx.nd.array(d[1], dtype=d[1].dtype)
assert_almost_equal(y, mx_nd_func(d0, d1).asnumpy(), rtol=rtol, atol=atol)
idx = np.abs(x-y) > atol+rtol*np.abs(x)
if idx.any():
- print('found precision problem')
+ import binascii
+ np.set_printoptions(precision=20)
+ logging.error('found precision problem:')
d[0] = np.broadcast_to(d[0], x.shape)
d[1] = np.broadcast_to(d[1], x.shape)
- print('a: {}'.format(d[0][idx]))
- print('b: {}'.format(d[1][idx]))
- import struct
- print('a hex: {}'.format(struct.pack('d', d[0][idx]).encode('hex')))
- print('b hex: {}'.format(struct.pack('d', np.broadcast_to(d[1], x.shape)[idx]).encode('hex')))
- print('in baseline(a, b): {}'.format(x[idx]))
- print('in symbol(a, b): {}'.format(y[idx]))
- print('diff: {}'.format(np.abs(x-y)[idx] - atol-rtol*np.abs(x)[idx]))
+ logging.error('input a: {}'.format(d[0][idx]))
+ logging.error('input b: {}'.format(d[1][idx]))
+ logging.error("output x: {} {}".format(x.dtype, x))
+ logging.error("output y: {} {}".format(y.dtype, y))
+ def ftohex(xs):
+ import struct
+ return list(map(lambda x: binascii.hexlify(struct.pack('d', x)), xs.flatten()))
+ logging.error('output x in baseline(a, b): {}'.format(x[idx]))
+ logging.error('output y in symbol(a, b): {}'.format(y[idx]))
+ logging.error('output x in baseline(a,b) hex: {}'.format(ftohex(x[idx])))
+ logging.error('output y in symbol(a,b) hex: {}'.format(ftohex(y[idx])))
+ logging.error('input a hex: {}'.format(ftohex(d[0][idx])))
+ logging.error('input a hex: {}'.format(ftohex(d[1][idx])))
+
+ logging.error('diff: {}'.format(np.abs(x-y)[idx] - atol-rtol*np.abs(x)[idx]))
assert_allclose(y, x, rtol=rtol, atol=atol)
@@ -1644,10 +1664,13 @@ def test_bdiv(a, b):
check_binary_op_backward(c, lambda g_out, a, b: (g_out / b, - g_out * a / (b * b)), gen_binary_data)
def test_bmod(a, b):
- c = a % b
+ # Python and numpy operate only in double so to avoid numerical errors we have to use
+ # doubles as well. This was a flaky test before when using float32. seed 1688524483, 1768433044
+ #c = a % b
+ c = mx.sym.cast(a, dtype='float64') % mx.sym.cast(b, dtype='float64')
# '%' is sensitive to the precision of the calculation. Force numpy to match mxnet's float32.
- # Issue exposed with seed 1768433044
- check_binary_op_forward(c, lambda a, b: np.float32(a) % np.float32(b), gen_binary_data)
+ #check_binary_op_forward(c, lambda a, b: np.float32(a) % np.float32(b), gen_binary_data)
+ check_binary_op_forward(c, lambda a, b: np.float32(a) % np.float32(b), gen_binary_data, rtol=0, atol=0)
check_binary_op_backward(c,
lambda g_out, a, b: (g_out, - g_out * (np.float32(a) // np.float32(b))), gen_binary_data)
@@ -1678,7 +1701,6 @@ def test_bneq(a, b):
test_bpow(a, b)
test_bneq(a, b)
-
@with_seed()
def test_broadcast_binary_op():
def check_bmaxmin_gradient(test_sym, x, y, delta, rtol, atol):
@@ -4761,7 +4783,7 @@ def test_laop():
check_fw(test_gemm, [a2, b2, c2], [r2])
if grad_check == 1:
check_grad(test_gemm, [a2, b2, c2])
-
+
# Check gemm2 operator same way as gemm.
res_gemm = 4. * np.dot(data_in1, data_in2)
test_gemm = mx.sym.linalg.gemm2(data1, data2, alpha=4.)
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services