leezu commented on issue #15856:
URL:
https://github.com/apache/incubator-mxnet/issues/15856#issuecomment-622617937
```
[2020-05-01T23:59:35.381Z] ________________
test_parallel_random_seed_setting_for_context _________________
[2020-05-01T23:59:35.381Z]
[2020-05-01T23:59:35.381Z] @with_seed()
[2020-05-01T23:59:35.381Z] def
test_parallel_random_seed_setting_for_context():
[2020-05-01T23:59:35.381Z] seed_to_test = 1234
[2020-05-01T23:59:35.381Z] dev_type =
mx.context.current_context().device_type
[2020-05-01T23:59:35.381Z] for dtype in ['float16', 'float32',
'float64']:
[2020-05-01T23:59:35.381Z] samples_imp = []
[2020-05-01T23:59:35.381Z] samples_sym = []
[2020-05-01T23:59:35.381Z] # Collect random number samples from
the generators of all devices, each seeded with the same number.
[2020-05-01T23:59:35.381Z] for dev_id in range(0,
mx.context.num_gpus() if dev_type == 'gpu' else 1):
[2020-05-01T23:59:35.381Z] with mx.Context(dev_type, dev_id):
[2020-05-01T23:59:35.381Z] ctx =
mx.context.current_context()
[2020-05-01T23:59:35.381Z] # Avoid excessive test cpu
runtimes.
[2020-05-01T23:59:35.381Z] num_temp_seeds = 25 if
dev_type == 'gpu' else 1
[2020-05-01T23:59:35.381Z] # To flush out a possible
race condition, run multiple times.
[2020-05-01T23:59:35.381Z] for _ in range(20):
[2020-05-01T23:59:35.381Z] # Create enough samples
such that we get a meaningful distribution.
[2020-05-01T23:59:35.381Z] shape = (200, 200)
[2020-05-01T23:59:35.381Z] params = { 'low': -1.5,
'high': 3.0 }
[2020-05-01T23:59:35.381Z]
params.update(shape=shape, dtype=dtype)
[2020-05-01T23:59:35.381Z]
[2020-05-01T23:59:35.381Z] # Check imperative.
`uniform` uses parallel rng.
[2020-05-01T23:59:35.381Z] seed =
set_seed_variously_for_context(ctx, 1, num_temp_seeds, seed_to_test)
[2020-05-01T23:59:35.381Z] rnds =
mx.nd.random.uniform(**params)
[2020-05-01T23:59:35.381Z]
samples_imp.append(rnds.asnumpy())
[2020-05-01T23:59:35.381Z]
[2020-05-01T23:59:35.381Z] # Check symbolic.
`uniform` uses parallel rng.
[2020-05-01T23:59:35.381Z] X = mx.sym.Variable("X")
[2020-05-01T23:59:35.381Z] Y =
mx.sym.random.uniform(**params) + X
[2020-05-01T23:59:35.381Z] x = mx.nd.zeros(shape,
dtype=dtype)
[2020-05-01T23:59:35.381Z] xgrad =
mx.nd.zeros(shape, dtype=dtype)
[2020-05-01T23:59:35.381Z] yexec = Y.bind(ctx, {'X'
: x}, {'X': xgrad})
[2020-05-01T23:59:35.381Z]
set_seed_variously_for_context(ctx, seed, num_temp_seeds, seed_to_test)
[2020-05-01T23:59:35.381Z]
yexec.forward(is_train=True)
[2020-05-01T23:59:35.381Z]
yexec.backward(yexec.outputs[0])
[2020-05-01T23:59:35.381Z]
samples_sym.append(yexec.outputs[0].asnumpy())
[2020-05-01T23:59:35.381Z] # The samples should be identical
across different gpu devices.
[2020-05-01T23:59:35.381Z] for i in range(1, len(samples_imp)):
[2020-05-01T23:59:35.381Z] assert same(samples_imp[i - 1],
samples_imp[i])
[2020-05-01T23:59:35.381Z] for i in range(1, len(samples_sym)):
[2020-05-01T23:59:35.381Z] > assert same(samples_sym[i - 1],
samples_sym[i])
[2020-05-01T23:59:35.381Z] E assert False
[2020-05-01T23:59:35.381Z] E + where False = same(array([[
5.83024027e-01, 1.67935735e+00, 8.77294878e-01,\n 2.51641507e+00,
2.95223944e+00, 2.75676270e+00,\n...-1.33459247e-01,\n -7.95632855e-01,
1.66822078e+00, -1.47758483e-01,\n 1.39249882e+00, -2.99302394e-01]]),
array([[ 5.83024027e-001, 1.67935735e+000, 8.77294878e-001,\n
2.51641507e+000, 2.95223944e+000, 2.75676270e...59247e-001,\n
-7.95632855e-001, 1.66822078e+000, -1.47758483e-001,\n
1.39249882e+000, -2.99302394e-001]]))
[2020-05-01T23:59:35.381Z]
[2020-05-01T23:59:35.381Z] tests/python/unittest/test_random.py:560:
AssertionError
```
http://jenkins.mxnet-ci.amazon-ml.com/blue/rest/organizations/jenkins/pipelines/mxnet-validation/pipelines/unix-gpu/branches/master/runs/1928/nodes/349/steps/510/log/?start=0
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
