littletomatodonkey
commented
4 years ago 此外，shape强制指定之后，运行时会报错(1-2个minibatch后)，麻烦也帮忙看下~是用的cuda9，paddle1.6.1，官网提供的docker环境，将该op替换成roialign是可以正常运行的 代码如下：
def _conv_offset(self,
input,
filter_size=3,
stride=1,
padding=1,
act=None,
name=None):
out_channel = input.shape[1] * 2
out = fluid.layers.conv2d(
input,
num_filters=out_channel,
filter_size=filter_size,
stride=stride,
padding=padding,
param_attr=ParamAttr(
initializer=Constant(0.0), name=name + ".w_0"),
bias_attr=ParamAttr(
initializer=Constant(0.0), name=name + ".b_0"),
act=act,
name=name)
return out
def __call__(self, head_inputs, rois, spatial_scale, is_mask=False):
"""
Adopt RoI align onto several level of feature maps to get RoI features.
Distribute RoIs to different levels by area and get a list of RoI
features by distributed RoIs and their corresponding feature maps.
Returns:
roi_feat(Variable): RoI features with shape of [M, C, R, R],
where M is the number of RoIs and R is RoI resolution
"""
k_min = self.min_level
k_max = self.max_level
num_roi_lvls = k_max - k_min + 1
name_list = list(head_inputs.keys())
input_name_list = name_list[-num_roi_lvls:]
spatial_scale = spatial_scale[-num_roi_lvls:]
rois_dist, restore_index = fluid.layers.distribute_fpn_proposals(
rois, k_min, k_max, self.canconical_level, self.canonical_size)
# rois_dist is in ascend order
roi_out_list = []
resolution = is_mask and self.mask_resolution or self.box_resolution
for lvl in range(num_roi_lvls):
name_index = num_roi_lvls - lvl - 1
rois_input = rois_dist[lvl]
head_input = head_inputs[input_name_list[name_index]]
sc = spatial_scale[name_index]
offset_name = input_name_list[name_index] + "_dpooling_offset"
trans = self._conv_offset( head_input, name=offset_name )
print( "[gry debug]in dpooling, offset name: {}, trans shape: {}".format(offset_name, trans.shape) )
roi_out = fluid.layers.deformable_roi_pooling(input=head_input,
rois=rois_input,
trans=trans,
no_trans=False,
spatial_scale=sc,
group_size=(1, 1),
pooled_height=resolution,
pooled_width=resolution,
sample_per_part=4,
trans_std=0.1,
position_sensitive=False)
roi_out_list.append(roi_out)
roi_feat_shuffle = fluid.layers.concat(roi_out_list)
roi_feat_ = fluid.layers.gather(roi_feat_shuffle, restore_index)
roi_feat = fluid.layers.lod_reset(roi_feat_, rois)
return roi_feat报错信息如下：
2019-12-25 10:37:47,071-INFO: iter: 0, lr: 0.006667, 'loss_bbox': '0.007775', 'loss_rpn_cls': '0.695010', 'loss_rpn_bbox': '0.015740', 'loss_cls': '4.402408', 'loss': '5.120934', time: 0.004, eta: 0:12:54
F1225 10:37:47.355821 28440 device_context.cc:334] cudaGetLastError invalid configuration argument errno: 9
*** Check failure stack trace: ***
@ 0x7f78b1f5438d google::LogMessage::Fail()
@ 0x7f78b1f57e3c google::LogMessage::SendToLog()
@ 0x7f78b1f53eb3 google::LogMessage::Flush()
@ 0x7f78b1f5934e google::LogMessageFatal::~LogMessageFatal()
@ 0x7f78b470c557 paddle::platform::CUDADeviceContext::Wait()
@ 0x7f78b46a62d1 paddle::framework::TransDataDevice()
@ 0x7f78b46a4c9e paddle::framework::TransformData()
@ 0x7f78b46833eb paddle::framework::OperatorWithKernel::PrepareData()
@ 0x7f78b4684838 paddle::framework::OperatorWithKernel::RunImpl()
@ 0x7f78b4684f11 paddle::framework::OperatorWithKernel::RunImpl()
@ 0x7f78b467eb9c paddle::framework::OperatorBase::Run()
@ 0x7f78b4463e26 paddle::framework::details::ComputationOpHandle::RunImpl()
@ 0x7f78b441b0c6 paddle::framework::details::FastThreadedSSAGraphExecutor::RunOpSync()
@ 0x7f78b4419e0f paddle::framework::details::FastThreadedSSAGraphExecutor::RunOp()
@ 0x7f78b441a0d4 _ZNSt17_Function_handlerIFvvESt17reference_wrapperISt12_Bind_simpleIFS1_ISt5_BindIFZN6paddle9framework7details28FastThreadedSSAGraphExecutor10RunOpAsyncEPSt13unordered_mapIPNS6_12OpHandleBaseESt6atomicIiESt4hashISA_ESt8equal_toISA_ESaISt4pairIKSA_SC_EEESA_RKSt10shared_ptrINS5_13BlockingQueueImEEEEUlvE_vEEEvEEEE9_M_invokeERKSt9_Any_data
@ 0x7f78b20733d3 std::_Function_handler<>::_M_invoke()
@ 0x7f78b1ebf5b7 std::__future_base::_State_base::_M_do_set()
@ 0x7f795ed24a99 __pthread_once_slow
@ 0x7f78b4415c22 _ZNSt13__future_base11_Task_stateISt5_BindIFZN6paddle9framework7details28FastThreadedSSAGraphExecutor10RunOpAsyncEPSt13unordered_mapIPNS4_12OpHandleBaseESt6atomicIiESt4hashIS8_ESt8equal_toIS8_ESaISt4pairIKS8_SA_EEES8_RKSt10shared_ptrINS3_13BlockingQueueImEEEEUlvE_vEESaIiEFvvEE6_M_runEv
@ 0x7f78b1ec0d74 _ZZN10ThreadPoolC1EmENKUlvE_clEv
@ 0x7f78f712ac80 (unknown)
@ 0x7f795ed1d6ba start_thread
@ 0x7f795ea5341d clone
@ (nil) (unknown)
Aborted
cjt222
这个assert在编译时可以删去，但是设定part_height的值时一定要小于等于trans_dims[2]，不然会发生越界
在deformable_psroi_pooling_op里面，添加了编译shape的assertion，但是我得到的tensor的shape在编译时可能是-1，想问下这个shape检查的assertion是否可以删去呢？（大部分tensor在reshape之后，都会是-1）
https://github.com/PaddlePaddle/Paddle/blob/70c073a0019f3405f9074579a8c48a43281dfdea/paddle/fluid/operators/deformable_psroi_pooling_op.cc#L184