def compare_grid_sample():
# do gradcheck
N = random.randint(1, 8)
C = 2 # random.randint(1, 8)
H = 5 # random.randint(1, 8)
W = 4 # random.randint(1, 8)
input = Variable(torch.randn(N, C, H, W).cuda(), requires_grad=True)
input_p = input.clone().data.contiguous()
grid = Variable(torch.randn(N, H, W, 2).cuda(), requires_grad=True)
grid_clone = grid.clone().contiguous()
out_offcial = F.grid_sample(input, grid)
grad_outputs = Variable(torch.rand(out_offcial.size()).cuda())
grad_outputs_clone = grad_outputs.clone().contiguous()
grad_inputs = torch.autograd.grad(out_offcial, (input, grid), grad_outputs.contiguous())
grad_input_off = grad_inputs[0]
crf = RoICropFunction()
grid_yx = torch.stack([grid_clone.data[:,:,:,1], grid_clone.data[:,:,:,0]], 3).contiguous().cuda()
out_stn = crf.forward(input_p, grid_yx)
grad_inputs = crf.backward(grad_outputs_clone.data)
grad_input_stn = grad_inputs[0]
pdb.set_trace()
delta = (grad_input_off.data - grad_input_stn).sum()
def roi_feature_transform(self,
blobs_in,
rois,
method='RoIPoolF',
resolution=7,
spatial_scale=1. / 16.,
sampling_ratio=0):
"""Add the specified RoI pooling method. The sampling_ratio argument
is supported for some, but not all, RoI transform methods.
RoIFeatureTransform abstracts away:
- Use of FPN or not
- Specifics of the transform method
"""
assert method in {'RoIPoolF', 'RoICrop', 'RoIAlign'}, \
'Unknown pooling method: {}'.format(method)
# Single feature level
# rois: holds R regions of interest, each is a 5-tuple
# (batch_idx, x1, y1, x2, y2) specifying an image batch index and a
# rectangle (x1, y1, x2, y2)
if method == 'RoIPoolF':
xform_out = RoIPoolFunction(resolution, resolution,
spatial_scale)(blobs_in, rois)
elif method == 'RoICrop':
grid_xy = net_utils.affine_grid_gen(rois,
blobs_in.size()[2:],
self.grid_size)
grid_yx = torch.stack(
[grid_xy.data[:, :, :, 1], grid_xy.data[:, :, :, 0]],
3).contiguous()
xform_out = RoICropFunction()(blobs_in, Variable(grid_yx).detach())
if cfg.CROP_RESIZE_WITH_MAX_POOL:
xform_out = F.max_pool2d(xform_out, 2, 2)
elif method == 'RoIAlign':
xform_out = RoIAlignFunction(resolution, resolution, spatial_scale,
sampling_ratio)(blobs_in, rois)
return xform_out
def roi_feature_transform(self,
blobs_in,
rpn_ret,
blob_rois='rois',
method='RoIPoolF',
resolution=7,
spatial_scale=1. / 16.,
sampling_ratio=0):
"""Add the specified RoI pooling method. The sampling_ratio argument
is supported for some, but not all, RoI transform methods.
RoIFeatureTransform abstracts away:
- Use of FPN or not
- Specifics of the transform method
"""
assert method in {'RoIPoolF', 'RoICrop', 'RoIAlign'}, \
'Unknown pooling method: {}'.format(method)
if isinstance(blobs_in, list):
# FPN case: add RoIFeatureTransform to each FPN level
device_id = blobs_in[0].get_device()
k_max = cfg.FPN.ROI_MAX_LEVEL # coarsest level of pyramid
k_min = cfg.FPN.ROI_MIN_LEVEL # finest level of pyramid
assert len(blobs_in) == k_max - k_min + 1
bl_out_list = []
for lvl in range(k_min, k_max + 1):
bl_in = blobs_in[k_max - lvl] # blobs_in is in reversed order
sc = spatial_scale[k_max - lvl] # in reversed order
bl_rois = blob_rois + '_fpn' + str(lvl)
if len(rpn_ret[bl_rois]):
rois = Variable(torch.from_numpy(
rpn_ret[bl_rois])).cuda(device_id)
if method == 'RoIPoolF':
# Warning!: Not check if implementation matches Detectron
xform_out = RoIPoolFunction(resolution, resolution,
sc)(bl_in, rois)
elif method == 'RoICrop':
# Warning!: Not check if implementation matches Detectron
grid_xy = net_utils.affine_grid_gen(
rois,
bl_in.size()[2:], self.grid_size)
grid_yx = torch.stack([
grid_xy.data[:, :, :, 1], grid_xy.data[:, :, :, 0]
], 3).contiguous()
xform_out = RoICropFunction()(
bl_in, Variable(grid_yx).detach())
if cfg.CROP_RESIZE_WITH_MAX_POOL:
xform_out = F.max_pool2d(xform_out, 2, 2)
elif method == 'RoIAlign':
xform_out = RoIAlignFunction(resolution, resolution,
sc, sampling_ratio)(bl_in,
rois)
bl_out_list.append(xform_out)
# The pooled features from all levels are concatenated along the
# batch dimension into a single 4D tensor.
xform_shuffled = torch.cat(bl_out_list, dim=0)
# Unshuffle to match rois from dataloader
device_id = xform_shuffled.get_device()
restore_bl = rpn_ret[blob_rois + '_idx_restore_int32']
restore_bl = Variable(
torch.from_numpy(restore_bl.astype(
'int64', copy=False))).cuda(device_id)
xform_out = xform_shuffled[restore_bl]
else:
# Single feature level
# rois: holds R regions of interest, each is a 5-tuple
# (batch_idx, x1, y1, x2, y2) specifying an image batch index and a
# rectangle (x1, y1, x2, y2)
device_id = blobs_in.get_device()
rois = Variable(torch.from_numpy(
rpn_ret[blob_rois])).cuda(device_id)
if method == 'RoIPoolF':
xform_out = RoIPoolFunction(resolution, resolution,
spatial_scale)(blobs_in, rois)
elif method == 'RoICrop':
grid_xy = net_utils.affine_grid_gen(rois,
blobs_in.size()[2:],
self.grid_size)
grid_yx = torch.stack(
[grid_xy.data[:, :, :, 1], grid_xy.data[:, :, :, 0]],
3).contiguous()
xform_out = RoICropFunction()(blobs_in,
Variable(grid_yx).detach())
if cfg.CROP_RESIZE_WITH_MAX_POOL:
xform_out = F.max_pool2d(xform_out, 2, 2)
elif method == 'RoIAlign':
xform_out = RoIAlignFunction(resolution, resolution,
spatial_scale,
sampling_ratio)(blobs_in, rois)
return xform_out
