def add_ResNet_roi_conv5_head(model, blob_in, dim_in, spatial_scale):
"""Adds an RoI feature transformation (e.g., RoI pooling) followed by a
res5/conv5 head applied to each RoI."""
# TODO(rbg): This contains Fast R-CNN specific config options making it non-
# reusable; make this more generic with model-specific wrappers
l = model.RoIFeatureTransform(
blob_in,
'pool5',
blob_rois='rois',
method=cfg.FAST_RCNN.ROI_XFORM_METHOD,
resolution=cfg.FAST_RCNN.ROI_XFORM_RESOLUTION,
sampling_ratio=cfg.FAST_RCNN.ROI_XFORM_SAMPLING_RATIO,
spatial_scale=spatial_scale)
l = model.net.RoIFeatureBoost([l, 'obn_scores'], l)
# save memory
if cfg.TRAIN.FREEZE_CONV_BODY:
model.StopGradient(l, l)
dim_bottleneck = cfg.RESNETS.NUM_GROUPS * cfg.RESNETS.WIDTH_PER_GROUP
stride_init = int(cfg.FAST_RCNN.ROI_XFORM_RESOLUTION / 7)
s, dim_in = add_stage(model, 'res5', 'pool5', 3, dim_in, 2048,
dim_bottleneck * 8, 1, stride_init)
s = model.AveragePool(s, 'res5_pool', kernel=7)
return s, 2048
def add_bpm_pse_outputs_c4(model, blob_in, dim):
current = add_attr_outputs(model, blob_in, dim)
# ReID stream
num_attr = cfg.REID.PSE_VIEW
attr_pred_list = ['attr_pred_' + str(i) for i in range(num_attr)]
model.net.Split(
current, attr_pred_list, split=[1 for i in range(num_attr)], axis=1)
attr_feature_list = []
for attr_id in range(num_attr):
prefix = '_[v{}]'.format(attr_id)
shape_0 = cfg.TRAIN.IMS_PER_BATCH if model.train else 1
model.net.Reshape(
attr_pred_list[attr_id],
[attr_pred_list[attr_id], 'attr_old_shape_' + str(attr_id)],
shape=[shape_0, 1, 1, 1])
dim_bottleneck = cfg.RESNETS.NUM_GROUPS * cfg.RESNETS.WIDTH_PER_GROUP
current, dim_out = add_stage(
model,
prefix + '_res5',
blob_in,
3,
dim,
2048,
dim_bottleneck * 8,
cfg.RESNETS.RES5_DILATION,
stride_init=cfg.RESNETS.RES5_STRIDE)
current = model.net.Mul(
[current, attr_pred_list[attr_id]],
prefix + '_scale',
broadcast=True,
)
attr_feature_list.append(current)
current = model.net.Sum(
[attr_feature for attr_feature in attr_feature_list], 'v_scale')
add_bpm_outputs(model, current, dim_out)