def _do_broadcast(all_blobs):
assert len(all_blobs) % cfg.NUM_GPUS == 0, \
('Unexpected value for NUM_GPUS. Make sure you are not '
'running single-GPU inference with NUM_GPUS > 1.')
blobs_per_gpu = int(len(all_blobs) / cfg.NUM_GPUS)
for i in range(blobs_per_gpu):
blobs = [p for p in all_blobs[i::blobs_per_gpu]]
data = workspace.FetchBlob(blobs[0])
logger.debug('Broadcasting {} to'.format(str(blobs[0])))
for i, p in enumerate(blobs[1:]):
logger.debug(' |-> {}'.format(str(p)))
with c2_utils.CudaScope(i + 1):
workspace.FeedBlob(p, data)
def InitializeLossWeight(self):
weight_cls1 = np.array([0.5]).astype(np.float32)
weight_cls2 = np.array([0.5]).astype(np.float32)
weight_bbox1 = np.array([0.5]).astype(np.float32)
weight_bbox2 = np.array([0.5]).astype(np.float32)
for i in range(cfg.NUM_GPUS):
with c2_utils.CudaScope(i):
workspace.FeedBlob('gpu_{}/weight_cls1'.format(i), weight_cls1)
workspace.FeedBlob('gpu_{}/weight_cls2'.format(i), weight_cls2)
workspace.FeedBlob('gpu_{}/weight_bbox1'.format(i),
weight_bbox1)
workspace.FeedBlob('gpu_{}/weight_bbox2'.format(i),
weight_bbox2)
def _add_allreduce_graph(model):
"""Construct the graph that performs Allreduce on the gradients."""
# Need to all-reduce the per-GPU gradients if training with more than 1 GPU
all_params = model.TrainableParams()
assert len(all_params) % cfg.NUM_GPUS == 0
# The model parameters are replicated on each GPU, get the number
# distinct parameter blobs (i.e., the number of parameter blobs on
# each GPU)
params_per_gpu = int(len(all_params) / cfg.NUM_GPUS)
with c2_utils.CudaScope(0):
# Iterate over distinct parameter blobs
for i in range(params_per_gpu):
# Gradients from all GPUs for this parameter blob
gradients = [
model.param_to_grad[p] for p in all_params[i::params_per_gpu]
]
if len(gradients) > 0:
if cfg.USE_NCCL:
model.net.NCCLAllreduce(gradients, gradients)
else:
muji.Allreduce(model.net, gradients, reduced_affix='')
def _CorrectMomentum(self, correction):
"""The MomentumSGDUpdate op implements the update V as
V := mu * V + lr * grad,
where mu is the momentum factor, lr is the learning rate, and grad is
the stochastic gradient. Since V is not defined independently of the
learning rate (as it should ideally be), when the learning rate is
changed we should scale the update history V in order to make it
compatible in scale with lr * grad.
"""
logger.info(
'Scaling update history by {:.6f} (new lr / old lr)'.format(
correction))
for i in range(cfg.NUM_GPUS):
with c2_utils.CudaScope(i):
for param in self.TrainableParams(gpu_id=i):
op = core.CreateOperator('Scale', [param + '_momentum'],
[param + '_momentum'],
scale=correction)
workspace.RunOperatorOnce(op)
def UpdateLossWeight(self):
scale = 10
# set bias for constraint for weight >0
bias = 0.5
lr = workspace.FetchBlob('gpu_0/lr').astype(np.float32)
weight_cls1 = workspace.FetchBlob('gpu_0/weight_cls1').astype(
np.float32)
weight_cls2 = workspace.FetchBlob('gpu_0/weight_cls2').astype(
np.float32)
#loss_cls1 = workspace.FetchBlob('gpu_0/loss_cls1')
#loss_cls2 = workspace.FetchBlob('gpu_0/loss_cls2')
weight_cls1 -= lr * scale * workspace.FetchBlob(
'gpu_0/weight_cls1_grad') + bias
weight_cls2 -= lr * scale * workspace.FetchBlob(
'gpu_0/weight_cls2_grad') + bias
weight_cls1 = weight_cls1 / (weight_cls1 + weight_cls2)
weight_cls2 = weight_cls2 / (weight_cls1 + weight_cls2)
weight_bbox1 = workspace.FetchBlob('gpu_0/weight_bbox1').astype(
np.float32)
weight_bbox2 = workspace.FetchBlob('gpu_0/weight_bbox2').astype(
np.float32)
#loss_bbox1 = workspace.FetchBlob('gpu_0/loss_bbox1')
#loss_bbox2 = workspace.FetchBlob('gpu_0/loss_bbox2')
weight_bbox1 -= lr * scale * workspace.FetchBlob(
'gpu_0/weight_bbox1_grad') + bias
weight_bbox2 -= lr * scale * workspace.FetchBlob(
'gpu_0/weight_bbox2_grad') + bias
weight_bbox1 = weight_bbox1 / (weight_bbox1 + weight_bbox2)
weight_bbox2 = weight_bbox2 / (weight_bbox1 + weight_bbox2)
for i in range(cfg.NUM_GPUS):
with c2_utils.CudaScope(i):
workspace.FeedBlob('gpu_{}/weight_cls1'.format(i), weight_cls1)
workspace.FeedBlob('gpu_{}/weight_cls2'.format(i), weight_cls2)
workspace.FeedBlob('gpu_{}/weight_bbox1'.format(i),
weight_bbox1)
workspace.FeedBlob('gpu_{}/weight_bbox2'.format(i),
weight_bbox2)
def _SetNewLr(self, cur_lr, new_lr):
"""Do the actual work of updating the model and workspace blobs.
"""
for i in range(cfg.NUM_GPUS):
with c2_utils.CudaScope(i):
workspace.FeedBlob('gpu_{}/lr'.format(i),
np.array([new_lr], dtype=np.float32))
lr_scale_new_param = cfg.SOLVER.LR_SCALE_NEW_PARAM
workspace.FeedBlob(
'gpu_{}/lr_new_param'.format(i),
np.array([new_lr * lr_scale_new_param], dtype=np.float32))
lr_scale_new_fc = cfg.SOLVER.LR_SCALE_NEW_FC
workspace.FeedBlob(
'gpu_{}/lr_new_fc'.format(i),
np.array([new_lr * lr_scale_new_fc], dtype=np.float32))
ratio = _get_lr_change_ratio(cur_lr, new_lr)
if cfg.SOLVER.SCALE_MOMENTUM and cur_lr > 1e-7 and \
ratio > cfg.SOLVER.SCALE_MOMENTUM_THRESHOLD:
self._CorrectMomentum(new_lr / cur_lr)
def GenerateProposals(self, blobs_in, blobs_out, anchors, spatial_scale):
"""Op for generating RPN porposals.
blobs_in:
- 'rpn_cls_probs': 4D tensor of shape (N, A, H, W), where N is the
number of minibatch images, A is the number of anchors per
locations, and (H, W) is the spatial size of the prediction grid.
Each value represents a "probability of object" rating in [0, 1].
- 'rpn_bbox_pred': 4D tensor of shape (N, 4 * A, H, W) of predicted
deltas for transformation anchor boxes into RPN proposals.
- 'im_info': 2D tensor of shape (N, 3) where the three columns encode
the input image's [height, width, scale]. Height and width are
for the input to the network, not the original image; scale is the
scale factor used to scale the original image to the network input
size.
blobs_out:
- 'rpn_rois': 2D tensor of shape (R, 5), for R RPN proposals where the
five columns encode [batch ind, x1, y1, x2, y2]. The boxes are
w.r.t. the network input, which is a *scaled* version of the
original image; these proposals must be scaled by 1 / scale (where
scale comes from im_info; see above) to transform it back to the
original input image coordinate system.
- 'rpn_roi_probs': 1D tensor of objectness probability scores
(extracted from rpn_cls_probs; see above).
"""
cfg_key = 'TRAIN' if self.train else 'TEST'
if cfg[cfg_key].GENERATE_PROPOSALS_ON_GPU:
rpn_pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
rpn_post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
rpn_nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
rpn_min_size = float(cfg[cfg_key].RPN_MIN_SIZE)
input_name = str(blobs_in[0])
lvl = int(input_name[-1]) if input_name[-1].isdigit() else None
anchors_name = 'anchors{}'.format(lvl) if lvl else 'anchors'
for i in range(cfg.NUM_GPUS):
with c2_utils.CudaScope(i):
workspace.FeedBlob(
'gpu_{}/{}'.format(i, anchors_name),
anchors.astype(np.float32))
self.net.GenerateProposals(
blobs_in + [anchors_name],
blobs_out,
spatial_scale=spatial_scale,
pre_nms_topN=rpn_pre_nms_topN,
post_nms_topN=rpn_post_nms_topN,
nms_thresh=rpn_nms_thresh,
min_size=rpn_min_size,
)
else:
name = 'GenerateProposalsOp:' + ','.join([str(b) for b in blobs_in])
# spatial_scale passed to the Python op is only used in
# convert_pkl_to_pb
self.net.Python(
GenerateProposalsOp(anchors, spatial_scale, self.train).forward
)(blobs_in, blobs_out, name=name, spatial_scale=spatial_scale)
return blobs_out
def softmax_surgery(model):
print('softmax surgery')
gpu_prefixs = ['gpu_' + str(i) for i in range(cfg.NUM_GPUS)]
old_ops = model.net._net.op[:]
num_op = len(model.net._net.op)
is_end = False
del model.net._net.op[:]
gpu_point = {gpu_prefix: -1 for gpu_prefix in gpu_prefixs}
while (True):
for gpu_prefix in gpu_prefixs:
for i, op in enumerate(old_ops):
if i <= gpu_point[gpu_prefix]:
continue
gpu = op.input[0].split('/')[0]
if gpu == gpu_prefix:
pass
else:
continue
if op.type == 'Softmax' and 'fc8d_t' in op.input[0]:
gpu_point[gpu_prefix] = i
# print(op)
print('find softmax: ', op.input[0], '\t-->\t',
op.output[0])
break
model.net._net.op.extend([op])
if i == num_op - 1:
is_end = True
if is_end:
break
if gpu_point[gpu_prefixs[0]] == -1 or gpu_point[
gpu_prefixs[1]] == -1 or gpu_point[
gpu_prefixs[2]] == -1 or gpu_point[gpu_prefixs[3]] == -1:
break
assert old_ops[gpu_point[gpu_prefixs[0]]].input[0].split('/')[
1] == old_ops[gpu_point[gpu_prefixs[1]]].input[0].split('/')[1]
assert old_ops[gpu_point[gpu_prefixs[0]]].input[0].split('/')[
1] == old_ops[gpu_point[gpu_prefixs[2]]].input[0].split('/')[1]
assert old_ops[gpu_point[gpu_prefixs[0]]].input[0].split('/')[
1] == old_ops[gpu_point[gpu_prefixs[3]]].input[0].split('/')[1]
in_blobs = []
out_blobs = []
for gpu_prefix in gpu_prefixs:
in_blob = old_ops[gpu_point[gpu_prefix]].input[0]
in_blobs.append(in_blob)
out_blob = old_ops[gpu_point[gpu_prefix]].output[0]
out_blobs.append(out_blob)
in_blob_name = in_blobs[0].split('/')[1]
out_blob_name = out_blobs[0].split('/')[1]
for gpu_prefix in gpu_prefixs:
gpu_id = int(gpu_prefix.split('_')[1])
with c2_utils.CudaScope(gpu_id):
for i in range(cfg.NUM_GPUS):
if gpu_id == i:
continue
model.net.Copy(
in_blobs[i],
gpu_prefix + '/' + in_blob_name + '_gpu_' + str(i))
model.net.StopGradient(
gpu_prefix + '/' + in_blob_name + '_gpu_' + str(i),
gpu_prefix + '/' + in_blob_name + '_gpu_' + str(i))
concat_in_blobs = [
gpu_prefix + '/' + in_blob_name + '_gpu_' + str(i)
for i in range(cfg.NUM_GPUS)
]
concat_in_blobs[gpu_id] = in_blobs[gpu_id]
model.net.Concat(concat_in_blobs, [
gpu_prefix + '/' + in_blob_name + '_cross',
gpu_prefix + '/' + in_blob_name + '_cross_split_info'
],
axis=1)
op = old_ops[gpu_point[gpu_prefix]]
op.input[0] = gpu_prefix + '/' + in_blob_name + '_cross'
op.output[0] = gpu_prefix + '/' + out_blob_name + '_cross'
model.net._net.op.extend([op])
split_out_blobs = [
gpu_prefix + '/' + str(i) + '_useless'
for i in range(len(out_blobs))
]
split_out_blobs[gpu_id] = out_blobs[gpu_id]
model.net.Split([
gpu_prefix + '/' + out_blob_name + '_cross',
gpu_prefix + '/' + in_blob_name + '_cross_split_info'
],
split_out_blobs,
axis=1)
return
num_op = len(model.net._net.op)
for i, op in enumerate(model.net._net.op):
print(op)
exit(0)
