def add_wsl_context_outputs(model, blobs_in, dim):
"""Add RoI classification and bounding box regression output ops."""
# Box classification layer
model.FC(
blobs_in[0],
'fc8c',
dim,
model.num_classes - 1,
weight_init=('XavierFill', {}),
bias_init=const_fill(0.0))
model.FC(
blobs_in[1],
'fc8d_frame',
dim,
model.num_classes - 1,
weight_init=('XavierFill', {}),
bias_init=const_fill(0.0))
model.net.FC([blobs_in[2], 'fc8d_frame_w', 'fc8d_frame_b'], 'fc8d_context')
model.net.Sub(['fc8d_frame', 'fc8d_context'], 'fc8d')
model.Softmax('fc8c', 'alpha_cls', axis=1)
model.Transpose('fc8d', 'fc8d_t', axes=(1, 0))
model.Softmax('fc8d_t', 'alpha_det_t', axis=1)
model.Transpose('alpha_det_t', 'alpha_det', axes=(1, 0))
model.net.Mul(['alpha_cls', 'alpha_det'], 'rois_pred')
if not model.train: # == if test
# model.net.Alias('rois_pred', 'cls_prob')
# Add BackGround predictions
model.net.Split(
'rois_pred', ['rois_bg_pred', 'notuse'],
split=[1, model.num_classes - 2],
axis=1)
model.net.Concat(['rois_bg_pred', 'rois_pred'],
['cls_prob', 'cls_prob_concat_dims'],
axis=1)
def add_wsl_outputs(model, blob_in, dim, prefix=''):
"""Add RoI classification and bounding box regression output ops."""
if cfg.WSL.CONTEXT:
return add_wsl_context_outputs(model, blob_in, dim, prefix=prefix)
# Box classification layer
model.FC(blob_in,
prefix + 'fc8c',
dim,
model.num_classes - 1,
weight_init=('XavierFill', {}),
bias_init=const_fill(0.0))
model.FC(blob_in,
prefix + 'fc8d',
dim,
model.num_classes - 1,
weight_init=('XavierFill', {}),
bias_init=const_fill(0.0))
model.Softmax(prefix + 'fc8c', prefix + 'alpha_cls', axis=1)
model.Transpose(prefix + 'fc8d', prefix + 'fc8d_t', axes=(1, 0))
model.Softmax(prefix + 'fc8d_t', prefix + 'alpha_det_t', axis=1)
model.Transpose(prefix + 'alpha_det_t', prefix + 'alpha_det', axes=(1, 0))
model.net.Mul([prefix + 'alpha_cls', prefix + 'alpha_det'],
prefix + 'rois_pred')
if not model.train: # == if test
# Add BackGround predictions
model.net.Split(prefix + 'rois_pred',
[prefix + 'rois_bg_pred', prefix + 'notuse'],
split=[1, model.num_classes - 2],
axis=1)
model.net.Concat(
[prefix + 'rois_bg_pred', prefix + 'rois_pred'],
[prefix + 'cls_prob', prefix + 'cls_prob_concat_dims'],
axis=1)
def _add_instance_level_classifier(model, blob_in, dim_in):
from detectron.utils.c2 import const_fill
from detectron.utils.c2 import gauss_fill
def negateGrad(inputs, outputs):
outputs[0].feed(inputs[0].data)
def grad_negateGrad(inputs, outputs):
scale = cfg.TRAIN.DA_INS_GRL_WEIGHT
grad_output = inputs[-1]
outputs[0].reshape(grad_output.shape)
outputs[0].data[...] = -1.0*scale*grad_output.data
model.GradientScalerLayer([blob_in], ['dc_grl'], -1.0*cfg.TRAIN.DA_INS_GRL_WEIGHT)
model.FC('dc_grl', 'dc_ip1', dim_in, 1024,
weight_init=gauss_fill(0.01), bias_init=const_fill(0.0))
model.Relu('dc_ip1', 'dc_relu_1')
model.Dropout('dc_relu_1', 'dc_drop_1', ratio=0.5, is_test=False)
model.FC('dc_drop_1', 'dc_ip2', 1024, 1024,
weight_init=gauss_fill(0.01), bias_init=const_fill(0.0))
model.Relu('dc_ip2', 'dc_relu_2')
model.Dropout('dc_relu_2', 'dc_drop_2', ratio=0.5, is_test=False)
dc_ip3 = model.FC('dc_drop_2', 'dc_ip3', 1024, 1,
weight_init=gauss_fill(0.05), bias_init=const_fill(0.0))
loss_gradient = None
if model.train:
dc_loss = model.net.SigmoidCrossEntropyLoss(
[dc_ip3, 'dc_label'],
'loss_dc',
scale=model.GetLossScale()
)
loss_gradient = blob_utils.get_loss_gradients(model, [dc_loss])
model.AddLosses('loss_dc')
return loss_gradient
def add_topdown_lateral_module(model, fpn_top, fpn_lateral, fpn_bottom,
dim_top, dim_lateral):
"""Add a top-down lateral module."""
# Lateral 1x1 conv
if cfg.FPN.USE_GN:
# use GroupNorm
lat = model.ConvGN(
fpn_lateral,
fpn_bottom + '_lateral',
dim_in=dim_lateral,
dim_out=dim_top,
group_gn=get_group_gn(dim_top),
kernel=1,
pad=0,
stride=1,
weight_init=(const_fill(0.0) if cfg.FPN.ZERO_INIT_LATERAL else
('XavierFill', {})),
bias_init=const_fill(0.0))
else:
lat = model.Conv(
fpn_lateral,
fpn_bottom + '_lateral',
dim_in=dim_lateral,
dim_out=dim_top,
kernel=1,
pad=0,
stride=1,
weight_init=(const_fill(0.0) if cfg.FPN.ZERO_INIT_LATERAL else
('XavierFill', {})),
bias_init=const_fill(0.0))
# Top-down 2x upsampling
td = model.net.UpsampleNearest(fpn_top, fpn_bottom + '_topdown', scale=2)
# Sum lateral and top-down
model.net.Sum([lat, td], fpn_bottom)
def add_fast_rcnn_outputs(model, blob_in, dim):
# Box classification layer
model.FC(
blob_in,
'cls_score',
dim,
model.num_classes,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0)
)
if not model.train:
# only add softmax when testing;during training the softmax is combined
model.Softmax('cls_score', 'cls_prob', engine='CUDNN')
# Box regression layer
num_box_reg_classes = (2 if cfg.MODEL.CLS_AGNOSTIC_BBOX_REG else model.num_classes)
model.FC(
blob_in,
'bbox_pred',
dim,
num_box_reg_classes * 4,
weight_init=gauss_fill(0.001),
bias_init=const_fill(0.0)
)
def add_bottomup_lateral_module(model, fpn_bottom, fpn_lateral, fpn_top,
fpn_dim):
"""Add a bottom-up lateral module."""
lat = model.Relu(fpn_lateral, fpn_lateral + '_relu')
# Bottom-up 2x downsampling
if cfg.FPN.USE_GN:
# use GroupNorm
bu = model.ConvGN(
fpn_bottom,
fpn_top + '_subsampled',
dim_in=fpn_dim,
dim_out=fpn_dim,
group_gn=get_group_gn(dim_top),
kernel=3,
pad=1,
stride=2,
weight_init=(const_fill(0.0) if cfg.FPN.ZERO_INIT_LATERAL else
('XavierFill', {})),
bias_init=const_fill(0.0))
else:
bu = model.Conv(
fpn_bottom,
fpn_top + '_subsampled',
dim_in=fpn_dim,
dim_out=fpn_dim,
kernel=3,
pad=1,
stride=2,
weight_init=(const_fill(0.0) if cfg.FPN.ZERO_INIT_LATERAL else
('XavierFill', {})),
bias_init=const_fill(0.0))
# Sum lateral and bottom-up
model.net.Sum([lat, bu], fpn_top)
def add_fast_rcnn_outputs(model, blob_in, dim):
"""Add RoI classification and bounding box regression output ops."""
model.FC(blob_in,
'cls_score',
dim,
model.num_classes,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0))
if not model.train: # == if test
# Only add softmax when testing; during training the softmax is combined
# with the label cross entropy loss for numerical stability
if not cfg.MODEL.WEIGHTED_LOSS:
model.Softmax('cls_score', 'cls_prob', engine='CUDNN')
else:
model.Softmax('cls_score', 'cls_prob1', engine='CUDNN')
model.net.Sigmoid('cls_score', 'cls_prob2', engine='CUDNN')
model.net.Mean(['cls_prob1', 'cls_prob2'], 'cls_prob')
model.FC(blob_in,
'bbox_pred',
dim,
model.num_classes * 4,
weight_init=gauss_fill(0.001),
bias_init=const_fill(0.0))
if cfg.MODEL.CASCADE_ON:
# add stage parameters to list
if '1' not in model.stage_params:
model.stage_params['1'] = []
for idx in range(-2, 0):
model.stage_params['1'].append(model.weights[idx])
model.stage_params['1'].append(model.biases[idx])
def add_fast_rcnn_outputs_test(model, blob_in, dim):
"""Add RoI classification and bounding box regression output ops."""
# Box classification layer
hidden_dim = cfg.FAST_RCNN.CONV_HEAD_DIM
roi_size = cfg.FAST_RCNN.ROI_XFORM_RESOLUTION
roi_resize = roi_size - cfg.FAST_RCNN.NUM_STACKED_CONVS * 2
model.FC(blob_in, 'fc6', hidden_dim * roi_size * roi_size, dim)
model.Relu('fc6', 'fc6')
model.FC('fc6', 'fc7', dim, dim)
model.Relu('fc7', 'fc7')
model.FC('fc7',
'cls_score',
dim,
model.num_classes,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0))
if not model.train: # == if test
# Only add softmax when testing; during training the softmax is combined
# with the label cross entropy loss for numerical stability
model.Softmax('cls_score', 'cls_prob', engine='CUDNN')
# Box regression layer
num_bbox_reg_classes = (2 if cfg.MODEL.CLS_AGNOSTIC_BBOX_REG else
model.num_classes)
#model.FC(current, 'fc6', dim_in * roi_size * roi_size, fc_dim)
#model.Relu('fc6', 'fc6')
model.FC(blob_in,
'bbox_pred',
hidden_dim * roi_size * roi_size,
num_bbox_reg_classes * 4,
weight_init=gauss_fill(0.001),
bias_init=const_fill(0.0))
def add_fast_rcnn_multilabel_outputs(model, blob_in, dim):
"""Add RoI classification and bounding box regression output ops."""
model.FC(
blob_in,
'cls_score',
dim,
2,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0)
)
model.FC(
blob_in,
'action_cls_logits',
dim,
15,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0)
)
if not model.train: # == if test
# Only add softmax when testing; during training the softmax is combined
# with the label cross entropy loss for numerical stability
model.Softmax('action_cls_logits', 'action_prob', engine='CUDNN')
model.Softmax('cls_score', 'cls_prob', engine='CUDNN')
model.FC(
blob_in,
'bbox_pred',
dim,
2 * 4,
weight_init=gauss_fill(0.001),
bias_init=const_fill(0.0)
)
def add_fast_rcnn_outputs(model, blob_in, dim):
"""Add RoI classification and bounding box regression output ops."""
# Box classification layer
model.FC(
blob_in,
'cls_score',
dim,
model.num_classes,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0)
)
if not model.train: # == if test
# Only add softmax when testing; during training the softmax is combined
# with the label cross entropy loss for numerical stability
model.Softmax('cls_score', 'cls_prob', engine='CUDNN')
# Box regression layer
num_bbox_reg_classes = (
2 if cfg.MODEL.CLS_AGNOSTIC_BBOX_REG else model.num_classes
)
model.FC(
blob_in,
'bbox_pred',
dim,
num_bbox_reg_classes * 4,
weight_init=gauss_fill(0.001),
bias_init=const_fill(0.0)
)
def add_cascade_rcnn_outputs(model, blob_in, dim, stage):
"""Add RoI classification and bounding box regression output ops."""
stage_name = "_{}".format(stage)
model.FC(
blob_in,
"cls_score" + stage_name,
dim,
model.num_classes,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0),
)
if not model.train: # == if test
# Only add softmax when testing; during training the softmax is combined
# with the label cross entropy loss for numerical stability
model.Softmax("cls_score" + stage_name,
"cls_prob" + stage_name,
engine="CUDNN")
num_bbox_reg_classes = 2 if cfg.MODEL.CLS_AGNOSTIC_BBOX_REG else model.num_classes
model.FC(
blob_in,
"bbox_pred" + stage_name,
dim,
num_bbox_reg_classes * 4,
weight_init=gauss_fill(0.001),
bias_init=const_fill(0.0),
)
# add stage parameters to list
if str(stage) not in model.stage_params:
model.stage_params[str(stage)] = []
for idx in range(-2, 0):
model.stage_params[str(stage)].append(model.weights[idx])
model.stage_params[str(stage)].append(model.biases[idx])
return "cls_prob" + stage_name, "bbox_pred" + stage_name
def add_fast_rcnn_outputs(model, blob_in, dim):
"""Add RoI classification and bounding box regression output ops."""
# Box classification layer
model.FC(blob_in,
'cls_score',
dim,
model.num_classes,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0))
if not model.train: # == if test
# Only add softmax when testing; during training the softmax is combined
# with the label cross entropy loss for numerical stability
model.Softmax('cls_score', 'cls_prob', engine='CUDNN')
# Box regression layer
num_bbox_reg_classes = (2 if cfg.MODEL.CLS_AGNOSTIC_BBOX_REG else
model.num_classes)
model.FC(blob_in,
'bbox_pred',
dim,
num_bbox_reg_classes * 4,
weight_init=gauss_fill(0.001),
bias_init=const_fill(0.0))
if cfg.MODEL.CASCADE_ON:
# add stage parameters to list
if '1' not in model.stage_params:
model.stage_params['1'] = []
for idx in range(-2, 0):
model.stage_params['1'].append(model.weights[idx])
model.stage_params['1'].append(model.biases[idx])
def add_fast_rcnn_outputs(model, blob_in, dim):
"""Add RoI classification and bounding box regression output ops."""
# Box classification layer
model.FC(blob_in,
'cls_score',
dim,
model.num_classes,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0))
if not model.train: # == if test
# Only add softmax when testing; during training the softmax is combined
# with the label cross entropy loss for numerical stability
model.Softmax('cls_score', 'cls_prob', engine='CUDNN')
# Box regression layer
num_bbox_reg_classes = (2 if cfg.MODEL.CLS_AGNOSTIC_BBOX_REG else
model.num_classes)
model.FC(blob_in,
'bbox_pred',
dim,
num_bbox_reg_classes * 4,
weight_init=gauss_fill(0.001),
bias_init=const_fill(0.0))
if cfg.PRED_STD:
model.FC(blob_in,
'bbox_pred_std',
dim,
num_bbox_reg_classes * 4,
weight_init=gauss_fill(0.0001),
bias_init=const_fill(1.0))
model.net.Abs('bbox_pred_std', 'bbox_pred_std_abs')
def add_fast_rcnn_outputs(model, blob_in, dim):
model.FC('fc7_newC',
'cls_score_toothbrush',
dim,
model.num_classes,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0))
num_bbox_reg_classes = (2 if cfg.MODEL.CLS_AGNOSTIC_BBOX_REG else
model.num_classes)
model.FC('fc7_newC',
'bbox_pred_toothbrush',
dim,
num_bbox_reg_classes * 4,
weight_init=gauss_fill(0.001),
bias_init=const_fill(0.0))
model.FC('fc7_oldC',
'cls_score',
dim,
model.num_classes,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0))
num_bbox_reg_classes = (2 if cfg.MODEL.CLS_AGNOSTIC_BBOX_REG else
model.num_classes)
model.FC('fc7_oldC',
'bbox_pred',
dim,
num_bbox_reg_classes * 4,
weight_init=gauss_fill(0.001),
bias_init=const_fill(0.0))
def add_single_scale_rpn_outputs(model, blob_in, dim_in, spatial_scale):
"""Add RPN outputs to a single scale model (i.e., no FPN)."""
anchors = generate_anchors(stride=1. / spatial_scale,
sizes=cfg.RPN.SIZES,
aspect_ratios=cfg.RPN.ASPECT_RATIOS)
num_anchors = anchors.shape[0]
dim_out = dim_in
# RPN hidden representation
model.Conv(blob_in,
'conv_rpn',
dim_in,
dim_out,
kernel=3,
pad=1,
stride=1,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0))
model.Relu('conv_rpn', 'conv_rpn')
# Proposal classification scores
model.Conv('conv_rpn',
'rpn_cls_logits',
dim_in,
num_anchors,
kernel=1,
pad=0,
stride=1,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0))
# Proposal bbox regression deltas
model.Conv('conv_rpn',
'rpn_bbox_pred',
dim_in,
4 * num_anchors,
kernel=1,
pad=0,
stride=1,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0))
if not model.train or cfg.MODEL.FASTER_RCNN:
# Proposals are needed during:
# 1) inference (== not model.train) for RPN only and Faster R-CNN
# OR
# 2) training for Faster R-CNN
# Otherwise (== training for RPN only), proposals are not needed
model.net.Sigmoid('rpn_cls_logits', 'rpn_cls_probs')
model.GenerateProposals(['rpn_cls_probs', 'rpn_bbox_pred', 'im_info'],
['rpn_rois', 'rpn_roi_probs'],
anchors=anchors,
spatial_scale=spatial_scale)
if cfg.MODEL.FASTER_RCNN:
if model.train:
# Add op that generates training labels for in-network RPN proposals
model.GenerateProposalLabels(['rpn_rois', 'roidb', 'im_info'])
else:
# Alias rois to rpn_rois for inference
model.net.Alias('rpn_rois', 'rois')
# Alias da_rois to rpn_rois for inference
model.net.Alias('rpn_rois', 'da_rois')
def add_keypoint_outputs(model, blob_in, dim):
"""Add Mask R-CNN keypoint specific outputs: keypoint heatmaps."""
# NxKxHxw
upsampling_heatmap = (cfg.KRCNN.UP_SCALE > 1)
if cfg.KRCNN.USE_DECONV:
#
blob_in = model.ConvTranspose(blob_in,
'kps_deconv',
dim,
cfg.KRCNN.DECONV_DIM,
kernel=cfg.KRCNN.DECONV_KERNEL,
pad=int(cfg.KRCNN.DECONV_KERNEL / 2 - 1),
stride=2,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0))
model.Relu('kps_deconv', 'kps_deconv')
dim = cfg.KRCNN.DECONV_DIM
if upsampling_heatmap:
blob_name = 'kps_score_lowres'
else:
blob_name = 'kps_score'
if cfg.KRCNN.USE_DECONV_OUTPUT:
# Use ConvTranspose to predict heatmaps; results in 2x upsampling
blob_out = model.ConvTranspose(blob_in,
blob_name,
dim,
cfg.KRCNN.NUM_KEYPOINTS,
kernel=cfg.KRCNN.DECONV_KERNEL,
pad=int(cfg.KRCNN.DECONV_KERNEL / 2 -
1),
stride=2,
weight_init=(cfg.KRCNN.CONV_INIT, {
'std': 0.001
}),
bias_init=const_fill(0.0))
else:
#
blob_out = model.Conv(blob_in,
blob_name,
dim,
cfg.KRCNN.NUM_KEYPOINTS,
kernel=1,
pad=0,
stride=1,
weight_init=(cfg.KRCNN.CONV_INIT, {
'std': 0.001
}),
bias_init=const_fill(0.0))
if upsampling_heatmap:
# Increase heatmap output size via bilinear upsampling
blob_out = model.BilinearInterpolation(blob_out, 'kps_score',
cfg.KRCNN.NUM_KEYPOINTS,
cfg.KRCNN.NUM_KEYPOINTS,
cfg.KRCNN.UP_SCALE)
return blob_out
def add_wsl_outputs(model, blob_in, dim, prefix=''):
"""Add RoI classification and bounding box regression output ops."""
if cfg.WSL.CONTEXT:
fc8c, fc8d = add_wsl_context_outputs(model,
blob_in,
dim,
prefix=prefix)
else:
# Box classification layer
fc8c = model.FC(
blob_in,
prefix + 'fc8c',
dim,
model.num_classes - 1,
weight_init=('XavierFill', {}),
# weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0),
)
fc8d = model.FC(
blob_in,
prefix + 'fc8d',
dim,
model.num_classes - 1,
weight_init=('XavierFill', {}),
# weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0),
)
if cfg.WSL.CMIL and model.train:
fc8c, fc8d = add_wsl_cmil(model, [fc8c, fc8d], dim, prefix=prefix)
model.Softmax(fc8c, prefix + 'alpha_cls', axis=1)
model.Transpose(fc8d, prefix + 'fc8d_t', axes=(1, 0))
model.Softmax(prefix + 'fc8d_t', prefix + 'alpha_det_t', axis=1)
model.Transpose(prefix + 'alpha_det_t', prefix + 'alpha_det', axes=(1, 0))
model.net.Mul([prefix + 'alpha_cls', prefix + 'alpha_det'],
prefix + 'rois_pred')
if not model.train: # == if test
# Add BackGround predictions
model.net.Split(prefix + 'rois_pred',
[prefix + 'rois_bg_pred', prefix + 'notuse'],
split=[1, model.num_classes - 2],
axis=1)
model.net.Concat(
[prefix + 'rois_bg_pred', prefix + 'rois_pred'],
[prefix + 'cls_prob', prefix + 'cls_prob_concat_dims'],
axis=1)
if cfg.WSL.CONTEXT:
blob_in = blob_in[0]
dim = dim
if cfg.WSL.CMIL:
add_wsl_cmil_outputs(model, blob_in, dim, prefix=prefix)
elif cfg.WSL.OICR:
add_wsl_oicr_outputs(model, blob_in, dim, prefix=prefix)
elif cfg.WSL.PCL:
add_wsl_pcl_outputs(model, blob_in, dim, prefix=prefix)
def add_rfcn_outputs(model, blob_in, dim_in, dim_reduce, spatial_scale):
if dim_reduce is not None:
# Optional dim reduction
blob_in = model.Conv(blob_in,
'conv_dim_reduce',
dim_in,
dim_reduce,
kernel=1,
pad=0,
stride=1,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0))
blob_in = model.Relu(blob_in, blob_in)
dim_in = dim_reduce
# Classification conv
model.Conv(blob_in,
'conv_cls',
dim_in,
model.num_classes * cfg.RFCN.PS_GRID_SIZE**2,
kernel=1,
pad=0,
stride=1,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0))
# Bounding-box regression conv
num_bbox_reg_classes = (2 if cfg.MODEL.CLS_AGNOSTIC_BBOX_REG else
model.num_classes)
model.Conv(blob_in,
'conv_bbox_pred',
dim_in,
4 * num_bbox_reg_classes * cfg.RFCN.PS_GRID_SIZE**2,
kernel=1,
pad=0,
stride=1,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0))
# Classification PS RoI pooling
model.net.PSRoIPool(['conv_cls', 'rois'],
['psroipooled_cls', '_mapping_channel_cls'],
group_size=cfg.RFCN.PS_GRID_SIZE,
output_dim=model.num_classes,
spatial_scale=spatial_scale)
model.AveragePool('psroipooled_cls',
'cls_score_4d',
kernel=cfg.RFCN.PS_GRID_SIZE)
model.net.Reshape('cls_score_4d', ['cls_score', '_cls_scores_shape'],
shape=(-1, cfg.MODEL.NUM_CLASSES))
if not model.train:
model.Softmax('cls_score', 'cls_prob', engine='CUDNN')
# Bbox regression PS RoI pooling
model.net.PSRoIPool(['conv_bbox_pred', 'rois'],
['psroipooled_bbox', '_mapping_channel_bbox'],
group_size=cfg.RFCN.PS_GRID_SIZE,
output_dim=4 * num_bbox_reg_classes,
spatial_scale=spatial_scale)
model.AveragePool('psroipooled_bbox',
'bbox_pred',
kernel=cfg.RFCN.PS_GRID_SIZE)
def add_topdown_lateral_module(model, fpn_top, fpn_lateral, fpn_bottom,
dim_top, dim_lateral, dim_bottom):
"""Add a top-down lateral module."""
# model.net.Copy(fpn_lateral, fpn_bottom)
# return
assert dim_top == dim_bottom or dim_lateral == dim_bottom
# Lateral 1x1 conv
if dim_lateral == dim_bottom:
lat = fpn_lateral
else:
lat = model.Conv(
fpn_lateral,
fpn_bottom + '_lateral',
dim_in=dim_lateral,
dim_out=dim_bottom,
kernel=1,
pad=0,
stride=1,
weight_init=(const_fill(0.0) if cfg.FPN.ZERO_INIT_LATERAL else
('XavierFill', {})),
bias_init=const_fill(0.0))
lat = model.SpatialBN(lat,
lat + '_bn',
dim_bottom,
is_test=not model.train)
lat = model.Relu(lat, lat)
# Top-down
if dim_top == dim_bottom:
td = fpn_top
else:
td = model.Conv(
fpn_top,
fpn_bottom + '_lateral',
dim_in=dim_top,
dim_out=dim_bottom,
kernel=1,
pad=0,
stride=1,
weight_init=(const_fill(0.0) if cfg.FPN.ZERO_INIT_LATERAL else
('XavierFill', {})),
bias_init=const_fill(0.0))
td = model.SpatialBN(td,
fpn_bottom + '_lateral_bn',
dim_bottom,
is_test=not model.train)
td = model.Relu(td, td)
# Top-down 2x upsampling
if 'res5_2_sum' not in str(fpn_top) or cfg.RESNETS.RES5_STRIDE == 2:
td = model.net.UpsampleNearest(td, fpn_bottom + '_topdown', scale=2)
# Sum lateral and top-down
model.net.Sum([lat, td], fpn_bottom)
def add_seg_outputs(model, blob_in, dim):
if 'deeplab' in cfg.MRCNN.ROI_MASK_HEAD:
return add_deeplab_outputs(model, blob_in, dim)
num_cls = cfg.MODEL.NUM_CLASSES if cfg.MRCNN.CLS_SPECIFIC_MASK else 1
if cfg.MRCNN.USE_FC_OUTPUT:
# Predict masks with a fully connected layer (ignore 'fcn' in the blob
# name)
blob_out = model.FC(blob_in,
'mask_fcn_logits',
dim,
num_cls * cfg.MRCNN.RESOLUTION**2,
weight_init=gauss_fill(0.001),
bias_init=const_fill(0.0))
else:
# Predict mask using Conv
# Use GaussianFill for class-agnostic mask prediction; fills based on
# fan-in can be too large in this case and cause divergence
fill = (cfg.MRCNN.CONV_INIT
if cfg.MRCNN.CLS_SPECIFIC_MASK else 'GaussianFill')
blob_out = model.Conv(blob_in,
'mask_fcn_logits',
dim,
num_cls - 1,
kernel=1,
pad=0,
stride=1,
weight_init=(fill, {
'std': 0.001
}),
bias_init=const_fill(0.0))
if cfg.MRCNN.UPSAMPLE_RATIO > 1:
blob_out = model.BilinearInterpolation('mask_fcn_logits',
'mask_fcn_logits_up',
num_cls, num_cls,
cfg.MRCNN.UPSAMPLE_RATIO)
if not model.train: # == if test
# blob_out = model.net.Sigmoid(blob_out, 'mask_fcn_probs')
# Add BackGround predictions
model.net.Split(blob_out, ['mask_fcn_logits_bg', 'mask_notuse'],
split=[1, model.num_classes - 2],
axis=1)
model.net.Concat(['mask_fcn_logits_bg', blob_out],
['mask_fcn_logits_', 'mask_fcn_logits_concat_dims'],
axis=1)
blob_out = model.net.Sigmoid('mask_fcn_logits_', 'mask_fcn_probs')
return blob_out
def _add_instance_level_classifier(model, blob_in, dim_in, spatial_scale):
from detectron.utils.c2 import const_fill
from detectron.utils.c2 import gauss_fill
# def negateGrad(inputs, outputs):
# outputs[0].feed(inputs[0].data)
# def grad_negateGrad(inputs, outputs):
# scale = cfg.TRAIN.DA_INS_GRL_WEIGHT
# grad_output = inputs[-1]
# outputs[0].reshape(grad_output.shape)
# outputs[0].data[...] = -1.0*scale*grad_output.data
model.RoIFeatureTransform(
blob_in,
'da_pool5',
blob_rois='da_rois',
method=cfg.FAST_RCNN.ROI_XFORM_METHOD,
resolution=7,
sampling_ratio=cfg.FAST_RCNN.ROI_XFORM_SAMPLING_RATIO,
spatial_scale=spatial_scale
)
model.FCShared('da_pool5', 'da_fc6', dim_in * 7 * 7, 4096,
weight='fc6_w', bias='fc6_b')
model.Relu('da_fc6', 'da_fc6')
model.FCShared('da_fc6', 'da_fc7', 4096, 4096,
weight='fc7_w', bias='fc7_b')
da_blobs = model.Relu('da_fc7', 'da_fc7')
model.GradientScalerLayer([da_blobs], ['dc_grl'], -1.0*cfg.TRAIN.DA_INS_GRL_WEIGHT)
model.FC('dc_grl', 'dc_ip1', 4096, 1024,
weight_init=gauss_fill(0.01), bias_init=const_fill(0.0))
model.Relu('dc_ip1', 'dc_relu_1')
model.Dropout('dc_relu_1', 'dc_drop_1', ratio=0.5, is_test=False)
model.FC('dc_drop_1', 'dc_ip2', 1024, 1024,
weight_init=gauss_fill(0.01), bias_init=const_fill(0.0))
model.Relu('dc_ip2', 'dc_relu_2')
model.Dropout('dc_relu_2', 'dc_drop_2', ratio=0.5, is_test=False)
dc_ip3 = model.FC('dc_drop_2', 'dc_ip3', 1024, 1,
weight_init=gauss_fill(0.05), bias_init=const_fill(0.0))
if cfg.TRAIN.PADA:
dc_ip3 = model.PADAbyGradientWeightingLayerD(dc_ip3,'pada_dc_ip3','pada_roi_weights')
loss_gradient = None
if model.train:
dc_loss = model.net.SigmoidCrossEntropyLoss(
[dc_ip3, 'dc_label'],
'loss_dc',
scale=model.GetLossScale()
)
loss_gradient = blob_utils.get_loss_gradients(model, [dc_loss])
model.AddLosses('loss_dc')
return loss_gradient, da_blobs, 4096
def add_mask_rcnn_outputs(model, blob_in, dim):
if cfg.MRCNN.DP_CASCADE_MASK_ON:
return add_dp_cascaded_mask_outputs(model, blob_in, dim)
if cfg.MRCNN.BBOX_CASCADE_MASK_ON:
if cfg.MRCNN.USE_CLS_EMBS:
return add_mask_emb_outputs(model, blob_in, dim)
return add_cascaded_mask_outputs(model, blob_in, dim)
"""Add Mask R-CNN specific outputs: either mask logits or probs."""
num_cls = cfg.MODEL.NUM_CLASSES if cfg.MRCNN.CLS_SPECIFIC_MASK else 1
if cfg.MRCNN.USE_CLS_EMBS:
return add_mask_emb_outputs(model, blob_in, dim)
if cfg.MRCNN.USE_FC_OUTPUT:
# Predict masks with a fully connected layer (ignore 'fcn' in the blob
# name)
blob_out = model.FC(blob_in,
'mask_fcn_logits',
dim,
num_cls * cfg.MRCNN.RESOLUTION**2,
weight_init=gauss_fill(0.001),
bias_init=const_fill(0.0))
else:
# Predict mask using Conv
# Use GaussianFill for class-agnostic mask prediction; fills based on
# fan-in can be too large in this case and cause divergence
fill = (cfg.MRCNN.CONV_INIT
if cfg.MRCNN.CLS_SPECIFIC_MASK else 'GaussianFill')
blob_out = model.Conv(blob_in,
'mask_fcn_logits',
dim,
num_cls,
kernel=1,
pad=0,
stride=1,
weight_init=(fill, {
'std': 0.001
}),
bias_init=const_fill(0.0))
if cfg.MRCNN.UPSAMPLE_RATIO > 1:
blob_out = model.BilinearInterpolation('mask_fcn_logits',
'mask_fcn_logits_up',
num_cls, num_cls,
cfg.MRCNN.UPSAMPLE_RATIO)
if not model.train: # == if test
blob_out = model.net.Sigmoid(blob_out, 'mask_fcn_probs')
return blob_out
def add_roi_Xconv2fc_head(model, blob_in, dim_in, spatial_scale):
"""Add a X conv + 2fc head"""
hidden_dim = cfg.FAST_RCNN.CONV_HEAD_DIM
roi_size = cfg.FAST_RCNN.ROI_XFORM_RESOLUTION
roi_feat = model.RoIFeatureTransform(
blob_in,
'roi_feat',
blob_rois='rois',
method=cfg.FAST_RCNN.ROI_XFORM_METHOD,
resolution=roi_size,
sampling_ratio=cfg.FAST_RCNN.ROI_XFORM_SAMPLING_RATIO,
spatial_scale=spatial_scale)
current = roi_feat
for i in range(cfg.FAST_RCNN.NUM_STACKED_CONVS):
current = model.Conv(current,
'head_conv' + str(i + 1),
dim_in,
hidden_dim,
3,
stride=1,
pad=1,
weight_init=('GaussianFill', {
'std': 0.01
}),
bias_init=('ConstantFill', {
'value': 0.
}),
no_bias=0)
current = model.Relu(current, current)
dim_in = hidden_dim
fc_dim = cfg.FAST_RCNN.MLP_HEAD_DIM
model.FC(current,
'fc6',
dim_in * roi_size * roi_size,
fc_dim,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0))
model.Relu('fc6', 'fc6')
model.FC('fc6',
'fc7',
fc_dim,
fc_dim,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0))
model.Relu('fc7', 'fc7')
return 'fc7', fc_dim
def add_apm_outputs2(model, blob_in, dim):
bpm_heads.add_bpm_outputs(model, blob_in, dim)
prefix_list = ['feature_' + str(i) for i in range(cfg.REID.BPM_STRIP_NUM)]
dim_inner = cfg.REID.BPM_DIM
im_per_batch = cfg.TRAIN.IMS_PER_BATCH if model.train else 1
fc_list = [prefix + '_fc' for prefix in prefix_list]
model.net.Concat(fc_list, ['fc8c', 'fc8c_split_info'], add_axis=1, axis=1)
feature_list = [prefix + '_conv' for prefix in prefix_list]
model.net.Concat(
feature_list, ['fc7', 'fc7_split_info'], add_axis=1, axis=1)
model.net.Reshape(
'fc7', ['fc7_', 'fc7_shape'],
shape=[im_per_batch * cfg.REID.BPM_STRIP_NUM, dim_inner])
model.FC(
'fc7_',
'fc8d_',
dim_inner,
model.num_classes - 1,
weight_init=('XavierFill', {}),
bias_init=const_fill(0.0))
model.net.Reshape(
'fc8d_', ['fc8d', 'fc8d__shape'],
shape=[im_per_batch, cfg.REID.BPM_STRIP_NUM, model.num_classes - 1])
model.Softmax('fc8c', 'alpha_cls', axis=2)
model.Transpose('fc8d', 'fc8d_t', axes=(0, 2, 1))
model.Softmax('fc8d_t', 'alpha_det_t', axis=2)
model.Transpose('alpha_det_t', 'alpha_det', axes=(0, 2, 1))
model.net.Mul(['alpha_cls', 'alpha_det'], 'rois_pred')
def mask_rcnn_fcn_head_v0upshare(model, blob_in, dim_in, spatial_scale):
"""Use a ResNet "conv5" / "stage5" head for mask prediction. Weights and
computation are shared with the conv5 box head. Computation can only be
shared during training, since inference is cascaded.
v0upshare design: conv5, convT 2x2.
"""
# Since box and mask head are shared, these must match
assert cfg.MRCNN.ROI_XFORM_RESOLUTION == cfg.FAST_RCNN.ROI_XFORM_RESOLUTION
if model.train: # share computation with bbox head at training time
dim_conv5 = 2048
blob_conv5 = model.net.SampleAs(['res5_2_sum', 'roi_has_mask_int32'],
['_[mask]_res5_2_sum_sliced'])
else: # re-compute at test time
blob_conv5, dim_conv5 = add_ResNet_roi_conv5_head_for_masks(
model, blob_in, dim_in, spatial_scale)
dim_reduced = cfg.MRCNN.DIM_REDUCED
blob_mask = model.ConvTranspose(
blob_conv5,
'conv5_mask',
dim_conv5,
dim_reduced,
kernel=2,
pad=0,
stride=2,
weight_init=(cfg.MRCNN.CONV_INIT, {
'std': 0.001
}), # std only for gauss
bias_init=const_fill(0.0))
model.Relu('conv5_mask', 'conv5_mask')
return blob_mask, dim_reduced
def mask_rcnn_fcn_head_v0up(model, blob_in, dim_in, spatial_scale):
"""v0up design: conv5, deconv 2x2 (no weight sharing with the box head)."""
blob_conv5, dim_conv5 = add_ResNet_roi_conv5_head_for_masks(
model,
blob_in,
dim_in,
spatial_scale
)
dim_reduced = cfg.MRCNN.DIM_REDUCED
model.ConvTranspose(
blob_conv5,
'conv5_mask',
dim_conv5,
dim_reduced,
kernel=2,
pad=0,
stride=2,
weight_init=('GaussianFill', {'std': 0.001}),
bias_init=const_fill(0.0)
)
blob_mask = model.Relu('conv5_mask', 'conv5_mask')
return blob_mask, dim_reduced
def add_body_uv_outputs(model, blob_in, dim):
"""Add DensePose body UV specific outputs: heatmaps of dense mask, patch index
and patch-specific UV coordinates. All dense masks are mapped to labels in
[0, ... S] for S semantically meaningful body parts.
"""
# Apply ConvTranspose to the feature representation; results in 2x upsampling
for name in ['AnnIndex', 'Index_UV', 'U', 'V']:
if name == 'AnnIndex':
dim_out = cfg.BODY_UV_RCNN.NUM_SEMANTIC_PARTS + 1
else:
dim_out = cfg.BODY_UV_RCNN.NUM_PATCHES + 1
model.ConvTranspose(blob_in,
name + '_lowres',
dim,
dim_out,
cfg.BODY_UV_RCNN.DECONV_KERNEL,
pad=int(cfg.BODY_UV_RCNN.DECONV_KERNEL / 2 - 1),
stride=2,
weight_init=(cfg.BODY_UV_RCNN.CONV_INIT, {
'std': 0.001
}),
bias_init=const_fill(0.0))
# Increase heatmap output size via bilinear upsampling
blob_outputs = []
for name in ['AnnIndex', 'Index_UV', 'U', 'V']:
blob_outputs.append(
model.BilinearInterpolation(
name + '_lowres',
name + '_estimated' if name in ['U', 'V'] else name,
cfg.BODY_UV_RCNN.NUM_PATCHES + 1,
cfg.BODY_UV_RCNN.NUM_PATCHES + 1, cfg.BODY_UV_RCNN.UP_SCALE))
return blob_outputs
def add_track_head(model, blob_in, dim_in, spatial_scale):
"""Add a Mask R-CNN track head."""
head_dim = cfg.TRCNN.MLP_HEAD_DIM
roi_size = cfg.TRCNN.ROI_XFORM_RESOLUTION
roi_feat = model.RoIFeatureTransform(
blob_in,
'track_roi_feat',
blob_rois='track_rois',
method=cfg.TRCNN.ROI_XFORM_METHOD,
resolution=roi_size,
sampling_ratio=cfg.TRCNN.ROI_XFORM_SAMPLING_RATIO,
spatial_scale=spatial_scale)
# Bottleneck operation
if cfg.TRCNN.MLP_HEAD_ON:
model.FC(
roi_feat,
"track_fc",
dim_in * roi_size * roi_size,
head_dim,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0),
)
track_fc = model.Relu("track_fc", "track_fc")
return track_fc, head_dim
# No bottleneck operation -> flattern feature vector
else:
model.Flatten(roi_feat, "track_fc")
track_fc = model.Relu("track_fc", "track_fc")
return roi_feat, dim_in * roi_size * roi_size
def mask_rcnn_fcn_head_v0up(model, blob_in, dim_in, spatial_scale):
"""v0up design: conv5, deconv 2x2 (no weight sharing with the box head)."""
blob_conv5, dim_conv5 = add_ResNet_roi_conv5_head_for_masks(
model,
blob_in,
dim_in,
spatial_scale
)
dim_reduced = cfg.MRCNN.DIM_REDUCED
model.ConvTranspose(
blob_conv5,
'conv5_mask',
dim_conv5,
dim_reduced,
kernel=2,
pad=0,
stride=2,
weight_init=('GaussianFill', {'std': 0.001}),
bias_init=const_fill(0.0)
)
blob_mask = model.Relu('conv5_mask', 'conv5_mask')
return blob_mask, dim_reduced
def add_roi_body_uv_head_v1convX(model, blob_in, dim_in, spatial_scale):
"""Add a DensePose body UV head. v1convX design: X * (conv)."""
hidden_dim = cfg.BODY_UV_RCNN.CONV_HEAD_DIM
kernel_size = cfg.BODY_UV_RCNN.CONV_HEAD_KERNEL
pad_size = kernel_size // 2
current = model.RoIFeatureTransform(
blob_in,
'_[body_uv]_roi_feat',
blob_rois='body_uv_rois',
method=cfg.BODY_UV_RCNN.ROI_XFORM_METHOD,
resolution=cfg.BODY_UV_RCNN.ROI_XFORM_RESOLUTION,
sampling_ratio=cfg.BODY_UV_RCNN.ROI_XFORM_SAMPLING_RATIO,
spatial_scale=spatial_scale)
for i in range(cfg.BODY_UV_RCNN.NUM_STACKED_CONVS):
current = model.Conv(current,
'body_conv_fcn' + str(i + 1),
dim_in,
hidden_dim,
kernel_size,
stride=1,
pad=pad_size,
weight_init=(cfg.BODY_UV_RCNN.CONV_INIT, {
'std': 0.01
}),
bias_init=const_fill(0.0))
current = model.Relu(current, current)
dim_in = hidden_dim
return current, hidden_dim
def add_mask_rcnn_outputs(model, blob_in, dim):
"""Add Mask R-CNN specific outputs: either mask logits or probs."""
num_cls = cfg.MODEL.NUM_CLASSES if cfg.MRCNN.CLS_SPECIFIC_MASK else 1
if cfg.MRCNN.USE_FC_OUTPUT:
# Predict masks with a fully connected layer (ignore 'fcn' in the blob
# name)
dim_fc = int(dim * (cfg.MRCNN.RESOLUTION / cfg.MRCNN.UPSAMPLE_RATIO)**2)
blob_out = model.FC(
blob_in,
'mask_fcn_logits',
dim_fc,
num_cls * cfg.MRCNN.RESOLUTION**2,
weight_init=gauss_fill(0.001),
bias_init=const_fill(0.0)
)
else:
# Predict mask using Conv
# Use GaussianFill for class-agnostic mask prediction; fills based on
# fan-in can be too large in this case and cause divergence
fill = (
cfg.MRCNN.CONV_INIT
if cfg.MRCNN.CLS_SPECIFIC_MASK else 'GaussianFill'
)
blob_out = model.Conv(
blob_in,
'mask_fcn_logits',
dim,
num_cls,
kernel=1,
pad=0,
stride=1,
weight_init=(fill, {'std': 0.001}),
bias_init=const_fill(0.0)
)
if cfg.MRCNN.UPSAMPLE_RATIO > 1:
blob_out = model.BilinearInterpolation(
'mask_fcn_logits', 'mask_fcn_logits_up', num_cls, num_cls,
cfg.MRCNN.UPSAMPLE_RATIO
)
if not model.train: # == if test
blob_out = model.net.Sigmoid(blob_out, 'mask_fcn_probs')
return blob_out
def add_attention_backbone_fpn_topdown_module(model, lateral_input_blobs,
output_blobs, fpn_dim_lateral,
fpn_dim, num_backbone_stages):
for index in range(num_backbone_stages - 1):
mul_blobs = []
fpn_top_up = model.net.UpsampleNearest(output_blobs[index],
output_blobs[index] +
'_topdown',
scale=2)
feature_lat = model.Conv(
lateral_input_blobs[index + 1],
lateral_input_blobs[index + 1] + '_lat',
dim_in=fpn_dim_lateral[index + 1],
dim_out=fpn_dim,
kernel=3,
pad=1,
stride=1,
weight_init=(const_fill(0.0) if cfg.FPN.ZERO_INIT_LATERAL else
('XavierFill', {})),
bias_init=const_fill(0.0))
mul_blobs.append(feature_lat)
if index == 0:
model.net.Sum([feature_lat, fpn_top_up], output_blobs[index + 1])
else:
for i in range(index):
re_scale = 2**(1 + index - i)
feature_c = model.Conv(
output_blobs[i],
output_blobs[i] + '_To_' + output_blobs[index + 1] + '_c',
dim_in=fpn_dim,
dim_out=fpn_dim,
kernel=3,
pad=1,
stride=1,
weight_init=(const_fill(0.0) if cfg.FPN.ZERO_INIT_LATERAL
else ('XavierFill', {})),
bias_init=const_fill(0.0))
feature_up = model.net.UpsampleNearest(feature_c,
feature_c + '_up',
scale=re_scale)
mul_blobs.append(feature_up)
feature_mul = model.net.MulAll(mul_blobs,
output_blobs[index + 1] + "_mulall")
model.net.Sum([feature_mul, fpn_top_up], output_blobs[index + 1])
def _add_image_level_classifier(model, blob_in, dim_in, spatial_scale_in):
from detectron.utils.c2 import const_fill
from detectron.utils.c2 import gauss_fill
def negateGrad(inputs, outputs):
outputs[0].feed(inputs[0].data)
def grad_negateGrad(inputs, outputs):
scale = cfg.TRAIN.DA_IMG_GRL_WEIGHT
grad_output = inputs[-1]
outputs[0].reshape(grad_output.shape)
outputs[0].data[...] = -1.0 * scale * grad_output.data
model.GradientScalerLayer([blob_in], ['da_grl'],
-1.0 * cfg.TRAIN.DA_IMG_GRL_WEIGHT)
model.Conv('da_grl',
'da_conv_1',
dim_in,
512,
kernel=1,
pad=0,
stride=1,
weight_init=gauss_fill(0.001),
bias_init=const_fill(0.0))
model.Relu('da_conv_1', 'da_conv_1')
model.Conv('da_conv_1',
'da_conv_2',
512,
1,
kernel=1,
pad=0,
stride=1,
weight_init=gauss_fill(0.001),
bias_init=const_fill(0.0))
if model.train:
model.net.SpatialNarrowAs(['da_label_wide', 'da_conv_2'], 'da_label')
loss_da = model.net.SigmoidCrossEntropyLoss(['da_conv_2', 'da_label'],
'loss_da',
scale=model.GetLossScale())
loss_gradient = blob_utils.get_loss_gradients(model, [loss_da])
model.AddLosses('loss_da')
return loss_gradient
else:
return None
def add_topdown_lateral_module(
model, fpn_top, fpn_lateral, fpn_bottom, dim_top, dim_lateral
):
"""Add a top-down lateral module."""
# Lateral 1x1 conv
if cfg.FPN.USE_GN:
# use GroupNorm
lat = model.ConvGN(
fpn_lateral,
fpn_bottom + '_lateral',
dim_in=dim_lateral,
dim_out=dim_top,
group_gn=get_group_gn(dim_top),
kernel=1,
pad=0,
stride=1,
weight_init=(
const_fill(0.0) if cfg.FPN.ZERO_INIT_LATERAL
else ('XavierFill', {})),
bias_init=const_fill(0.0)
)
else:
lat = model.Conv(
fpn_lateral,
fpn_bottom + '_lateral',
dim_in=dim_lateral,
dim_out=dim_top,
kernel=1,
pad=0,
stride=1,
weight_init=(
const_fill(0.0)
if cfg.FPN.ZERO_INIT_LATERAL else ('XavierFill', {})
),
bias_init=const_fill(0.0)
)
# Top-down 2x upsampling
td = model.net.UpsampleNearest(fpn_top, fpn_bottom + '_topdown', scale=2)
# Sum lateral and top-down
model.net.Sum([lat, td], fpn_bottom)
def mask_rcnn_fcn_head_v1upXconvs_gn(
model, blob_in, dim_in, spatial_scale, num_convs
):
"""v1upXconvs design: X * (conv 3x3), convT 2x2, with GroupNorm"""
current = model.RoIFeatureTransform(
blob_in,
blob_out='_mask_roi_feat',
blob_rois='mask_rois',
method=cfg.MRCNN.ROI_XFORM_METHOD,
resolution=cfg.MRCNN.ROI_XFORM_RESOLUTION,
sampling_ratio=cfg.MRCNN.ROI_XFORM_SAMPLING_RATIO,
spatial_scale=spatial_scale
)
dilation = cfg.MRCNN.DILATION
dim_inner = cfg.MRCNN.DIM_REDUCED
for i in range(num_convs):
current = model.ConvGN(
current,
'_mask_fcn' + str(i + 1),
dim_in,
dim_inner,
group_gn=get_group_gn(dim_inner),
kernel=3,
pad=1 * dilation,
stride=1,
weight_init=(cfg.MRCNN.CONV_INIT, {'std': 0.001}),
bias_init=('ConstantFill', {'value': 0.})
)
current = model.Relu(current, current)
dim_in = dim_inner
# upsample layer
model.ConvTranspose(
current,
'conv5_mask',
dim_inner,
dim_inner,
kernel=2,
pad=0,
stride=2,
weight_init=(cfg.MRCNN.CONV_INIT, {'std': 0.001}),
bias_init=const_fill(0.0)
)
blob_mask = model.Relu('conv5_mask', 'conv5_mask')
return blob_mask, dim_inner
def mask_rcnn_fcn_head_v0upshare(model, blob_in, dim_in, spatial_scale):
"""Use a ResNet "conv5" / "stage5" head for mask prediction. Weights and
computation are shared with the conv5 box head. Computation can only be
shared during training, since inference is cascaded.
v0upshare design: conv5, convT 2x2.
"""
# Since box and mask head are shared, these must match
assert cfg.MRCNN.ROI_XFORM_RESOLUTION == cfg.FAST_RCNN.ROI_XFORM_RESOLUTION
if model.train: # share computation with bbox head at training time
dim_conv5 = 2048
blob_conv5 = model.net.SampleAs(
['res5_2_sum', 'roi_has_mask_int32'],
['_[mask]_res5_2_sum_sliced']
)
else: # re-compute at test time
blob_conv5, dim_conv5 = add_ResNet_roi_conv5_head_for_masks(
model,
blob_in,
dim_in,
spatial_scale
)
dim_reduced = cfg.MRCNN.DIM_REDUCED
blob_mask = model.ConvTranspose(
blob_conv5,
'conv5_mask',
dim_conv5,
dim_reduced,
kernel=2,
pad=0,
stride=2,
weight_init=(cfg.MRCNN.CONV_INIT, {'std': 0.001}), # std only for gauss
bias_init=const_fill(0.0)
)
model.Relu('conv5_mask', 'conv5_mask')
return blob_mask, dim_reduced
def add_rfcn_outputs(model, blob_in, dim_in, dim_reduce, spatial_scale):
if dim_reduce is not None:
# Optional dim reduction
blob_in = model.Conv(
blob_in,
'conv_dim_reduce',
dim_in,
dim_reduce,
kernel=1,
pad=0,
stride=1,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0)
)
blob_in = model.Relu(blob_in, blob_in)
dim_in = dim_reduce
# Classification conv
model.Conv(
blob_in,
'conv_cls',
dim_in,
model.num_classes * cfg.RFCN.PS_GRID_SIZE**2,
kernel=1,
pad=0,
stride=1,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0)
)
# # Bounding-box regression conv
num_bbox_reg_classes = (
2 if cfg.MODEL.CLS_AGNOSTIC_BBOX_REG else model.num_classes
)
model.Conv(
blob_in,
'conv_bbox_pred',
dim_in,
4 * num_bbox_reg_classes * cfg.RFCN.PS_GRID_SIZE**2,
kernel=1,
pad=0,
stride=1,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0)
)
# Classification PS RoI pooling
model.net.PSRoIPool(
['conv_cls', 'rois'], ['psroipooled_cls', '_mapping_channel_cls'],
group_size=cfg.RFCN.PS_GRID_SIZE,
output_dim=model.num_classes,
spatial_scale=spatial_scale
)
model.AveragePool(
'psroipooled_cls', 'cls_score_4d', kernel=cfg.RFCN.PS_GRID_SIZE
)
model.net.Reshape(
'cls_score_4d', ['cls_score', '_cls_scores_shape'],
shape=(-1, cfg.MODEL.NUM_CLASSES)
)
if not model.train:
model.Softmax('cls_score', 'cls_prob', engine='CUDNN')
# Bbox regression PS RoI pooling
model.net.PSRoIPool(
['conv_bbox_pred', 'rois'],
['psroipooled_bbox', '_mapping_channel_bbox'],
group_size=cfg.RFCN.PS_GRID_SIZE,
output_dim=4 * num_bbox_reg_classes,
spatial_scale=spatial_scale
)
model.AveragePool(
'psroipooled_bbox', 'bbox_pred', kernel=cfg.RFCN.PS_GRID_SIZE
)
def add_fpn_rpn_outputs(model, blobs_in, dim_in, spatial_scales):
"""Add RPN on FPN specific outputs."""
num_anchors = len(cfg.FPN.RPN_ASPECT_RATIOS)
dim_out = dim_in
k_max = cfg.FPN.RPN_MAX_LEVEL # coarsest level of pyramid
k_min = cfg.FPN.RPN_MIN_LEVEL # finest level of pyramid
assert len(blobs_in) == k_max - k_min + 1
for lvl in range(k_min, k_max + 1):
bl_in = blobs_in[k_max - lvl] # blobs_in is in reversed order
sc = spatial_scales[k_max - lvl] # in reversed order
slvl = str(lvl)
if lvl == k_min:
# Create conv ops with randomly initialized weights and
# zeroed biases for the first FPN level; these will be shared by
# all other FPN levels
# RPN hidden representation
conv_rpn_fpn = model.Conv(
bl_in,
'conv_rpn_fpn' + slvl,
dim_in,
dim_out,
kernel=3,
pad=1,
stride=1,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0)
)
model.Relu(conv_rpn_fpn, conv_rpn_fpn)
# Proposal classification scores
rpn_cls_logits_fpn = model.Conv(
conv_rpn_fpn,
'rpn_cls_logits_fpn' + slvl,
dim_in,
num_anchors,
kernel=1,
pad=0,
stride=1,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0)
)
# Proposal bbox regression deltas
rpn_bbox_pred_fpn = model.Conv(
conv_rpn_fpn,
'rpn_bbox_pred_fpn' + slvl,
dim_in,
4 * num_anchors,
kernel=1,
pad=0,
stride=1,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0)
)
else:
# Share weights and biases
sk_min = str(k_min)
# RPN hidden representation
conv_rpn_fpn = model.ConvShared(
bl_in,
'conv_rpn_fpn' + slvl,
dim_in,
dim_out,
kernel=3,
pad=1,
stride=1,
weight='conv_rpn_fpn' + sk_min + '_w',
bias='conv_rpn_fpn' + sk_min + '_b'
)
model.Relu(conv_rpn_fpn, conv_rpn_fpn)
# Proposal classification scores
rpn_cls_logits_fpn = model.ConvShared(
conv_rpn_fpn,
'rpn_cls_logits_fpn' + slvl,
dim_in,
num_anchors,
kernel=1,
pad=0,
stride=1,
weight='rpn_cls_logits_fpn' + sk_min + '_w',
bias='rpn_cls_logits_fpn' + sk_min + '_b'
)
# Proposal bbox regression deltas
rpn_bbox_pred_fpn = model.ConvShared(
conv_rpn_fpn,
'rpn_bbox_pred_fpn' + slvl,
dim_in,
4 * num_anchors,
kernel=1,
pad=0,
stride=1,
weight='rpn_bbox_pred_fpn' + sk_min + '_w',
bias='rpn_bbox_pred_fpn' + sk_min + '_b'
)
if not model.train or cfg.MODEL.FASTER_RCNN:
# Proposals are needed during:
# 1) inference (== not model.train) for RPN only and Faster R-CNN
# OR
# 2) training for Faster R-CNN
# Otherwise (== training for RPN only), proposals are not needed
lvl_anchors = generate_anchors(
stride=2.**lvl,
sizes=(cfg.FPN.RPN_ANCHOR_START_SIZE * 2.**(lvl - k_min), ),
aspect_ratios=cfg.FPN.RPN_ASPECT_RATIOS
)
rpn_cls_probs_fpn = model.net.Sigmoid(
rpn_cls_logits_fpn, 'rpn_cls_probs_fpn' + slvl
)
model.GenerateProposals(
[rpn_cls_probs_fpn, rpn_bbox_pred_fpn, 'im_info'],
['rpn_rois_fpn' + slvl, 'rpn_roi_probs_fpn' + slvl],
anchors=lvl_anchors,
spatial_scale=sc
)
def add_keypoint_outputs(model, blob_in, dim):
"""Add Mask R-CNN keypoint specific outputs: keypoint heatmaps."""
# NxKxHxW
upsample_heatmap = (cfg.KRCNN.UP_SCALE > 1)
if cfg.KRCNN.USE_DECONV:
# Apply ConvTranspose to the feature representation; results in 2x
# upsampling
blob_in = model.ConvTranspose(
blob_in,
'kps_deconv',
dim,
cfg.KRCNN.DECONV_DIM,
kernel=cfg.KRCNN.DECONV_KERNEL,
pad=int(cfg.KRCNN.DECONV_KERNEL / 2 - 1),
stride=2,
weight_init=gauss_fill(0.01),
bias_init=const_fill(0.0)
)
model.Relu('kps_deconv', 'kps_deconv')
dim = cfg.KRCNN.DECONV_DIM
if upsample_heatmap:
blob_name = 'kps_score_lowres'
else:
blob_name = 'kps_score'
if cfg.KRCNN.USE_DECONV_OUTPUT:
# Use ConvTranspose to predict heatmaps; results in 2x upsampling
blob_out = model.ConvTranspose(
blob_in,
blob_name,
dim,
cfg.KRCNN.NUM_KEYPOINTS,
kernel=cfg.KRCNN.DECONV_KERNEL,
pad=int(cfg.KRCNN.DECONV_KERNEL / 2 - 1),
stride=2,
weight_init=(cfg.KRCNN.CONV_INIT, {'std': 0.001}),
bias_init=const_fill(0.0)
)
else:
# Use Conv to predict heatmaps; does no upsampling
blob_out = model.Conv(
blob_in,
blob_name,
dim,
cfg.KRCNN.NUM_KEYPOINTS,
kernel=1,
pad=0,
stride=1,
weight_init=(cfg.KRCNN.CONV_INIT, {'std': 0.001}),
bias_init=const_fill(0.0)
)
if upsample_heatmap:
# Increase heatmap output size via bilinear upsampling
blob_out = model.BilinearInterpolation(
blob_out, 'kps_score', cfg.KRCNN.NUM_KEYPOINTS,
cfg.KRCNN.NUM_KEYPOINTS, cfg.KRCNN.UP_SCALE
)
return blob_out
def add_fpn(model, fpn_level_info):
"""Add FPN connections based on the model described in the FPN paper."""
# FPN levels are built starting from the highest/coarest level of the
# backbone (usually "conv5"). First we build down, recursively constructing
# lower/finer resolution FPN levels. Then we build up, constructing levels
# that are even higher/coarser than the starting level.
fpn_dim = cfg.FPN.DIM
min_level, max_level = get_min_max_levels()
# Count the number of backbone stages that we will generate FPN levels for
# starting from the coarest backbone stage (usually the "conv5"-like level)
# E.g., if the backbone level info defines stages 4 stages: "conv5",
# "conv4", ... "conv2" and min_level=2, then we end up with 4 - (2 - 2) = 4
# backbone stages to add FPN to.
num_backbone_stages = (
len(fpn_level_info.blobs) - (min_level - LOWEST_BACKBONE_LVL)
)
lateral_input_blobs = fpn_level_info.blobs[:num_backbone_stages]
output_blobs = [
'fpn_inner_{}'.format(s)
for s in fpn_level_info.blobs[:num_backbone_stages]
]
fpn_dim_lateral = fpn_level_info.dims
xavier_fill = ('XavierFill', {})
# For the coarsest backbone level: 1x1 conv only seeds recursion
if cfg.FPN.USE_GN:
# use GroupNorm
c = model.ConvGN(
lateral_input_blobs[0],
output_blobs[0], # note: this is a prefix
dim_in=fpn_dim_lateral[0],
dim_out=fpn_dim,
group_gn=get_group_gn(fpn_dim),
kernel=1,
pad=0,
stride=1,
weight_init=xavier_fill,
bias_init=const_fill(0.0)
)
output_blobs[0] = c # rename it
else:
model.Conv(
lateral_input_blobs[0],
output_blobs[0],
dim_in=fpn_dim_lateral[0],
dim_out=fpn_dim,
kernel=1,
pad=0,
stride=1,
weight_init=xavier_fill,
bias_init=const_fill(0.0)
)
#
# Step 1: recursively build down starting from the coarsest backbone level
#
# For other levels add top-down and lateral connections
for i in range(num_backbone_stages - 1):
add_topdown_lateral_module(
model,
output_blobs[i], # top-down blob
lateral_input_blobs[i + 1], # lateral blob
output_blobs[i + 1], # next output blob
fpn_dim, # output dimension
fpn_dim_lateral[i + 1] # lateral input dimension
)
# Post-hoc scale-specific 3x3 convs
blobs_fpn = []
spatial_scales = []
for i in range(num_backbone_stages):
if cfg.FPN.USE_GN:
# use GroupNorm
fpn_blob = model.ConvGN(
output_blobs[i],
'fpn_{}'.format(fpn_level_info.blobs[i]),
dim_in=fpn_dim,
dim_out=fpn_dim,
group_gn=get_group_gn(fpn_dim),
kernel=3,
pad=1,
stride=1,
weight_init=xavier_fill,
bias_init=const_fill(0.0)
)
else:
fpn_blob = model.Conv(
output_blobs[i],
'fpn_{}'.format(fpn_level_info.blobs[i]),
dim_in=fpn_dim,
dim_out=fpn_dim,
kernel=3,
pad=1,
stride=1,
weight_init=xavier_fill,
bias_init=const_fill(0.0)
)
blobs_fpn += [fpn_blob]
spatial_scales += [fpn_level_info.spatial_scales[i]]
#
# Step 2: build up starting from the coarsest backbone level
#
# Check if we need the P6 feature map
if not cfg.FPN.EXTRA_CONV_LEVELS and max_level == HIGHEST_BACKBONE_LVL + 1:
# Original FPN P6 level implementation from our CVPR'17 FPN paper
P6_blob_in = blobs_fpn[0]
P6_name = P6_blob_in + '_subsampled_2x'
# Use max pooling to simulate stride 2 subsampling
P6_blob = model.MaxPool(P6_blob_in, P6_name, kernel=1, pad=0, stride=2)
blobs_fpn.insert(0, P6_blob)
spatial_scales.insert(0, spatial_scales[0] * 0.5)
# Coarser FPN levels introduced for RetinaNet
if cfg.FPN.EXTRA_CONV_LEVELS and max_level > HIGHEST_BACKBONE_LVL:
fpn_blob = fpn_level_info.blobs[0]
dim_in = fpn_level_info.dims[0]
for i in range(HIGHEST_BACKBONE_LVL + 1, max_level + 1):
fpn_blob_in = fpn_blob
if i > HIGHEST_BACKBONE_LVL + 1:
fpn_blob_in = model.Relu(fpn_blob, fpn_blob + '_relu')
fpn_blob = model.Conv(
fpn_blob_in,
'fpn_' + str(i),
dim_in=dim_in,
dim_out=fpn_dim,
kernel=3,
pad=1,
stride=2,
weight_init=xavier_fill,
bias_init=const_fill(0.0)
)
dim_in = fpn_dim
blobs_fpn.insert(0, fpn_blob)
spatial_scales.insert(0, spatial_scales[0] * 0.5)
return blobs_fpn, fpn_dim, spatial_scales
