提交 2d41daaf authored 作者: Seton Steven Bocco's avatar Seton Steven Bocco 提交者: notoraptor

We get error just with batch size > 2**16.

No need of big input.
上级 0fa0aa36
......@@ -1070,49 +1070,44 @@ def get_conv3d_test_cases():
def run_conv_small_batched_vs_multicall(inputs_shape, filters_shape, batch_sub, subsample):
# Run function for issue $5985 (see tests below): https://github.com/Theano/Theano/issues/5985
# We use GPU RNG to help create big arrays on GPU directly and then avoid transfer.
from ..rng_mrg import GPUA_mrg_uniform
# Inspired from gpuarray/tests/tesr_rng_mrg.py
seed = 12345
curr_rstate = np.array([[seed] * 6], dtype='int32')
rstate = gpuarray_shared_constructor(curr_rstate)
algo = 'small'
batch_size = inputs_shape[0]
rstate_inputs, inputs = GPUA_mrg_uniform.new(rstate, dtype='float32',
size=inputs_shape, ndim=len(inputs_shape))
rstate_filters, filters = GPUA_mrg_uniform.new(rstate_inputs, dtype='float32',
size=filters_shape, ndim=len(filters_shape))
inputs_size = 4.0 # sizeof(float32)
for i in inputs_shape:
inputs_size *= i
print('(input size', (inputs_size / 1024 / 1024 / 1024), 'Gb)', end=' ')
utt.seed_rng()
inputs_val = np.random.random(inputs_shape).astype('float32')
filters_val = np.random.random(filters_shape).astype('float32')
inputs_val /= 10
filters_val /= 10
inputs = theano.shared(inputs_val)
filters = theano.shared(filters_val)
conv = dnn.dnn_conv(img=inputs, kerns=filters, algo=algo, subsample=subsample)
# Just compute last inputs to reduce execution time.
size = inputs_shape[0]
batched_outputs = [dnn.dnn_conv(img=inputs[i:(i + 1)], kerns=filters, algo=algo, subsample=subsample)
for i in range(size - batch_sub, size)]
f = theano.function([], [conv] + batched_outputs, mode=mode_with_gpu)
outputs = f()
res_all = outputs[0]
res_batch = outputs[1:]
for i in range(batch_sub):
utt.assert_allclose(res_batch[i], res_all[size - batch_sub + i])
# Just compute firt and last outputs to reduce execution time.
sub_conv_top = dnn.dnn_conv(img=inputs[:batch_sub],
kerns=filters, algo=algo, subsample=subsample)
sub_conv_bottom = dnn.dnn_conv(img=inputs[(batch_size - batch_sub):],
kerns=filters, algo=algo, subsample=subsample)
f = theano.function([], [conv, sub_conv_top, sub_conv_bottom], mode=mode_with_gpu)
res_all, res_batch_top, res_batch_bottom = f()
for i in range(0, batch_sub):
utt.assert_allclose(res_all[i], res_batch_top[i])
p = batch_size - batch_sub + i
# It seems there is a liimit batch size of 65536 for a good computation
# with algorithm `small`.
checked_limit = 2**16
if p >= checked_limit:
# It seems results are repeated in the entire conv.
# It should not happen.
if np.allclose(res_all[p % checked_limit], res_all[p]):
print('\nconv[%d] == conv[%d] == %s' % (p % checked_limit, p, res_all[p]))
utt.assert_allclose(res_all[p], res_batch_bottom[i])
def test_batched_conv_small():
# Tested on TITAN X:
# pass up to 65536 inputs (inputs size exactly 1Gb), fail with 65536 + 1 inputs and upper.
# Is there any limitation around number of elements, or input size ?
# But all dimensions and strides for following tensors are under int32 limits.
# Maybe the problem is with the internal pointer used by cuDNN to iterate over input
# (could this pointer not be able to manage more than 1 Gb?).
# NB: Subsample is set to (3, 3) to reduce output size.
yield (run_conv_small_batched_vs_multicall, (65535, 4, 32, 32), (1, 4, 16, 16), 25, (3, 3)) # OK
yield (run_conv_small_batched_vs_multicall, (65536, 4, 32, 32), (1, 4, 16, 16), 25, (3, 3)) # OK
yield (run_conv_small_batched_vs_multicall, (65537, 4, 32, 32), (1, 4, 16, 16), 25, (3, 3)) # ERROR
yield (run_conv_small_batched_vs_multicall, (65534, 2, 2, 2), (1, 2, 2, 2), 5, (1, 1)) # OK
yield (run_conv_small_batched_vs_multicall, (65535, 2, 2, 2), (1, 2, 2, 2), 5, (1, 1)) # OK
yield (run_conv_small_batched_vs_multicall, (65536, 2, 2, 2), (1, 2, 2, 2), 5, (1, 1)) # OK
yield (run_conv_small_batched_vs_multicall, (65537, 2, 2, 2), (1, 2, 2, 2), 5, (1, 1)) # ERROR
def test_conv3d_fwd():
......
Markdown 格式
0%
您添加了 0 到此讨论。请谨慎行事。
请先完成此评论的编辑!
注册 或者 后发表评论