def __init__(self, input_tensors, output_tensors,
return_input=False, sess=None):
"""
Args:
input_tensors (list): list of names.
output_tensors (list): list of names.
return_input (bool): same as :attr:`PredictorBase.return_input`.
sess (tf.Session): the session this predictor runs in. If None,
will use the default session at the first call.
"""
self.return_input = return_input
self.input_tensors = input_tensors
self.output_tensors = output_tensors
self.sess = sess
self._use_callable = get_tf_version_number() >= 1.2
if self._use_callable:
if sess is not None:
self._callable = sess.make_callable(
fetches=output_tensors,
feed_list=input_tensors,
accept_options=self.ACCEPT_OPTIONS)
else:
self._callable = None
else:
log_once(
"TF>=1.2 is recommended for better performance of predictor!", 'warn')
def get_model():
if config.MODE_FPN:
if get_tf_version_number() < 1.6:
logger.warn("FPN has chances to crash in TF<1.6, due to a TF issue.")
return ResNetFPNModel()
else:
return ResNetC4Model()
def get_model():
if config.MODE_FPN:
if get_tf_version_number() < 1.6:
logger.warn("FPN has chances to crash in TF<1.6, due to a TF issue.")
return ResNetFPNModel()
else:
return ResNetC4Model()
def monkeypatch_tf_layers():
if get_tf_version_number() < 1.4:
if not hasattr(tf.layers, 'Dense'):
from tensorflow.python.layers.core import Dense
tf.layers.Dense = Dense
from tensorflow.python.layers.normalization import BatchNormalization
tf.layers.BatchNormalization = BatchNormalization
from tensorflow.python.layers.convolutional import Conv2DTranspose, Conv2D
tf.layers.Conv2DTranspose = Conv2DTranspose
tf.layers.Conv2D = Conv2D
from tensorflow.python.layers.pooling import MaxPooling2D, AveragePooling2D
tf.layers.MaxPooling2D = MaxPooling2D
tf.layers.AveragePooling2D = AveragePooling2D
def Conv3D(
inputs,
filters,
kernel_size,
strides=(1, 1, 1),
padding='same',
data_format='channels_last',
dilation_rate=(1, 1, 1),
activation=None,
use_bias=True,
kernel_initializer=tf.contrib.layers.variance_scaling_initializer(2.0),
bias_initializer=tf.zeros_initializer(),
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
split=1):
"""
A wrapper around `tf.layers.Conv2D`.
Some differences to maintain backward-compatibility:
1. Default kernel initializer is variance_scaling_initializer(2.0).
2. Default padding is 'same'.
3. Support 'split' argument to do group conv.
Variable Names:
* ``W``: weights
* ``b``: bias
"""
if split == 1:
with rename_get_variable({'kernel': 'W', 'bias': 'b'}):
layer = tf.layers.Conv3D(filters,
kernel_size,
strides=strides,
padding=padding,
data_format=data_format,
dilation_rate=dilation_rate,
activation=activation,
use_bias=use_bias,
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer,
kernel_regularizer=kernel_regularizer,
bias_regularizer=bias_regularizer,
activity_regularizer=activity_regularizer)
ret = layer.apply(inputs, scope=tf.get_variable_scope())
ret = tf.identity(ret, name='output')
ret.variables = VariableHolder(W=layer.kernel)
if use_bias:
ret.variables.b = layer.bias
else:
# group conv implementation
data_format = get_data_format(data_format, tfmode=False)
in_shape = inputs.get_shape().as_list()
channel_axis = 3 if data_format == 'NHWC' else 1
in_channel = in_shape[channel_axis]
assert in_channel is not None, "[Conv3D] Input cannot have unknown channel!"
assert in_channel % split == 0
assert kernel_regularizer is None and bias_regularizer is None and activity_regularizer is None, \
"Not supported by group conv now!"
out_channel = filters
assert out_channel % split == 0
assert dilation_rate == (1, 1) or get_tf_version_number(
) >= 1.5, 'TF>=1.5 required for group dilated conv'
kernel_shape = shape2d(kernel_size)
filter_shape = kernel_shape + [in_channel / split, out_channel]
stride = shape4d(strides, data_format=data_format)
kwargs = dict(data_format=data_format)
if get_tf_version_number() >= 1.5:
kwargs['dilations'] = shape4d(dilation_rate,
data_format=data_format)
W = tf.get_variable('W', filter_shape, initializer=kernel_initializer)
if use_bias:
b = tf.get_variable('b', [out_channel],
initializer=bias_initializer)
inputs = tf.split(inputs, split, channel_axis)
kernels = tf.split(W, split, 3)
outputs = [
tf.nn.conv2d(i, k, stride, padding.upper(), **kwargs)
for i, k in zip(inputs, kernels)
]
conv = tf.concat(outputs, channel_axis)
if activation is None:
activation = tf.identity
ret = activation(tf.nn.bias_add(conv, b, data_format=data_format)
if use_bias else conv,
name='output')
ret.variables = VariableHolder(W=W)
if use_bias:
ret.variables.b = b
return ret
self._eval()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--load', help='load a model for evaluation or training. Can overwrite BACKBONE.WEIGHTS')
parser.add_argument('--logdir', help='log directory', default='train_log/maskrcnn')
parser.add_argument('--visualize', action='store_true', help='visualize intermediate results')
parser.add_argument('--evaluate', help="Run evaluation on COCO. "
"This argument is the path to the output json evaluation file")
parser.add_argument('--predict', help="Run prediction on a given image. "
"This argument is the path to the input image file")
parser.add_argument('--config', help="A list of KEY=VALUE to overwrite those defined in config.py",
nargs='+')
if get_tf_version_number() < 1.6:
# https://github.com/tensorflow/tensorflow/issues/14657
logger.warn("TF<1.6 has a bug which may lead to crash in FasterRCNN training if you're unlucky.")
args = parser.parse_args()
cfg.update_args(args.config)
MODEL = ResNetFPNModel() if cfg.MODE_FPN else ResNetC4Model()
if args.visualize or args.evaluate or args.predict:
assert args.load
finalize_configs(is_training=False)
if args.predict or args.visualize:
cfg.TEST.RESULT_SCORE_THRESH = cfg.TEST.RESULT_SCORE_THRESH_VIS
def BatchNorm3d(inputs,
axis=None,
training=None,
momentum=0.9,
epsilon=1e-5,
center=True,
scale=True,
beta_initializer=tf.zeros_initializer(),
gamma_initializer=tf.ones_initializer(),
virtual_batch_size=None,
data_format='channels_last',
internal_update=False,
sync_statistics=None):
"""
Almost equivalent to `tf.layers.batch_normalization`, but different (and more powerful)
in the following:
1. Accepts an alternative `data_format` option when `axis` is None. For 2D input, this argument will be ignored.
2. Default value for `momentum` and `epsilon` is different.
3. Default value for `training` is automatically obtained from tensorpack's `TowerContext`, but can be overwritten.
4. Support the `internal_update` option, which enables the use of BatchNorm layer inside conditionals.
5. Support the `sync_statistics` option, which is very useful in small-batch models.
Args:
internal_update (bool): if False, add EMA update ops to
`tf.GraphKeys.UPDATE_OPS`. If True, update EMA inside the layer by control dependencies.
They are very similar in speed, but `internal_update=True` can be used
when you have conditionals in your model, or when you have multiple networks to train.
Corresponding TF issue: https://github.com/tensorflow/tensorflow/issues/14699
sync_statistics: either None or "nccl". By default (None), it uses statistics of the input tensor to normalize.
When set to "nccl", this layer must be used under tensorpack multi-gpu trainers,
and it then uses per-machine (multiple GPU) statistics to normalize.
Note that this implementation averages the per-tower E[x] and E[x^2] among towers to compute
global mean&variance. The result is the global mean&variance only if each tower has the same batch size.
This option has no effect when not training.
This option is also known as "Cross-GPU BatchNorm" as mentioned in https://arxiv.org/abs/1711.07240.
Corresponding TF issue: https://github.com/tensorflow/tensorflow/issues/18222
Variable Names:
* ``beta``: the bias term. Will be zero-inited by default.
* ``gamma``: the scale term. Will be one-inited by default.
* ``mean/EMA``: the moving average of mean.
* ``variance/EMA``: the moving average of variance.
Note:
Combinations of ``training`` and ``ctx.is_training``:
* ``training == ctx.is_training``: standard BN, EMA are maintained during training
and used during inference. This is the default.
* ``training and not ctx.is_training``: still use batch statistics in inference.
* ``not training and ctx.is_training``: use EMA to normalize in
training. This is useful when you load a pre-trained BN and
don't want to fine tune the EMA. EMA will not be updated in
this case.
"""
# parse shapes
data_format = get_data_format(data_format, tfmode=False)
shape = inputs.get_shape().as_list()
ndims = len(shape)
# in 3d conv, we have 5d dim [batch, c, d, h, w]
# assert ndims in [2, 4], ndims
if sync_statistics is not None:
sync_statistics = sync_statistics.lower()
assert sync_statistics in [None, 'nccl', 'horovod'], sync_statistics
if axis is None:
if ndims == 2:
data_format = 'NHWC'
axis = 1
elif ndims == 5:
axis = 1 if data_format == 'NCHW' else 4
else:
axis = 1 if data_format == 'NCHW' else 3
else:
data_format = 'NCHW' if axis == 1 else 'NHWC'
num_chan = shape[axis]
# parse training/ctx
ctx = get_current_tower_context()
if training is None:
training = ctx.is_training
training = bool(training)
TF_version = get_tf_version_number()
if not training and ctx.is_training:
assert TF_version >= 1.4, \
"Fine tuning a BatchNorm model with fixed statistics is only " \
"supported after https://github.com/tensorflow/tensorflow/pull/12580 "
if ctx.is_main_training_tower: # only warn in first tower
logger.warn(
"[BatchNorm] Using moving_mean/moving_variance in training.")
# Using moving_mean/moving_variance in training, which means we
# loaded a pre-trained BN and only fine-tuning the affine part.
if sync_statistics is None or not (training and ctx.is_training):
coll_bk = backup_collection([tf.GraphKeys.UPDATE_OPS])
with rename_get_variable({
'moving_mean': 'mean/EMA',
'moving_variance': 'variance/EMA'
}):
tf_args = dict(axis=axis,
momentum=momentum,
epsilon=epsilon,
center=center,
scale=scale,
beta_initializer=beta_initializer,
gamma_initializer=gamma_initializer,
fused=True,
_reuse=tf.get_variable_scope().reuse)
if TF_version >= 1.5:
tf_args['virtual_batch_size'] = virtual_batch_size
else:
assert virtual_batch_size is None, "Feature not supported in this version of TF!"
layer = tf.layers.BatchNormalization(**tf_args)
xn = layer.apply(inputs,
training=training,
scope=tf.get_variable_scope())
# maintain EMA only on one GPU is OK, even in replicated mode.
# because during training, EMA isn't used
if ctx.is_main_training_tower:
for v in layer.non_trainable_variables:
add_model_variable(v)
if not ctx.is_main_training_tower or internal_update:
restore_collection(coll_bk)
if training and internal_update:
assert layer.updates
with tf.control_dependencies(layer.updates):
ret = tf.identity(xn, name='output')
else:
ret = tf.identity(xn, name='output')
vh = ret.variables = VariableHolder(
moving_mean=layer.moving_mean,
mean=layer.moving_mean, # for backward-compatibility
moving_variance=layer.moving_variance,
variance=layer.moving_variance) # for backward-compatibility
if scale:
vh.gamma = layer.gamma
if center:
vh.beta = layer.beta
else:
red_axis = [0] if ndims == 2 else (
[0, 2, 3] if axis == 1 else [0, 1, 2])
if ndims == 5:
red_axis = [0, 2, 3, 4] if axis == 1 else [0, 1, 2, 3]
new_shape = None # don't need to reshape unless ...
if ndims == 4 and axis == 1:
new_shape = [1, num_chan, 1, 1]
if ndims == 5 and axis == 1:
new_shape = [1, num_chan, 1, 1, 1]
batch_mean = tf.reduce_mean(inputs, axis=red_axis)
batch_mean_square = tf.reduce_mean(tf.square(inputs), axis=red_axis)
if sync_statistics == 'nccl':
if six.PY3 and TF_version <= 1.8 and ctx.is_main_training_tower:
logger.warn(
"A TensorFlow bug will cause cross-GPU BatchNorm to fail. "
"Apply this patch: https://github.com/tensorflow/tensorflow/pull/20360"
)
from tensorflow.contrib.nccl.ops import gen_nccl_ops
shared_name = re.sub('tower[0-9]+/', '',
tf.get_variable_scope().name)
num_dev = ctx.total
batch_mean = gen_nccl_ops.nccl_all_reduce(
input=batch_mean,
reduction='sum',
num_devices=num_dev,
shared_name=shared_name + '_NCCL_mean') * (1.0 / num_dev)
batch_mean_square = gen_nccl_ops.nccl_all_reduce(
input=batch_mean_square,
reduction='sum',
num_devices=num_dev,
shared_name=shared_name + '_NCCL_mean_square') * (1.0 /
num_dev)
elif sync_statistics == 'horovod':
# Require https://github.com/uber/horovod/pull/331
# Proof-of-concept, not ready yet.
import horovod.tensorflow as hvd
batch_mean = hvd.allreduce(batch_mean, average=True)
batch_mean_square = hvd.allreduce(batch_mean_square, average=True)
batch_var = batch_mean_square - tf.square(batch_mean)
batch_mean_vec = batch_mean
batch_var_vec = batch_var
beta, gamma, moving_mean, moving_var = get_bn_variables(
num_chan, scale, center, beta_initializer, gamma_initializer)
if new_shape is not None:
batch_mean = tf.reshape(batch_mean, new_shape)
batch_var = tf.reshape(batch_var, new_shape)
# Using fused_batch_norm(is_training=False) is actually slightly faster,
# but hopefully this call will be JITed in the future.
xn = tf.nn.batch_normalization(inputs, batch_mean, batch_var,
tf.reshape(beta, new_shape),
tf.reshape(gamma, new_shape),
epsilon)
else:
xn = tf.nn.batch_normalization(inputs, batch_mean, batch_var, beta,
gamma, epsilon)
if ctx.is_main_training_tower:
ret = update_bn_ema(xn, batch_mean_vec, batch_var_vec, moving_mean,
moving_var, momentum, internal_update)
else:
ret = tf.identity(xn, name='output')
vh = ret.variables = VariableHolder(
moving_mean=moving_mean,
mean=moving_mean, # for backward-compatibility
moving_variance=moving_var,
variance=moving_var) # for backward-compatibility
if scale:
vh.gamma = gamma
if center:
vh.beta = beta
return ret
def build_graph(self, *inputs):
is_training = get_current_tower_context().is_training
if config.MODE_MASK:
image, anchor_labels, anchor_boxes, gt_boxes, gt_labels, gt_masks = inputs
else:
image, anchor_labels, anchor_boxes, gt_boxes, gt_labels = inputs
image = self.preprocess(image) # 1CHW
featuremap = resnet_c4_backbone(image, config.RESNET_NUM_BLOCK[:3])
rpn_label_logits, rpn_box_logits = rpn_head('rpn', featuremap, 1024,
config.NUM_ANCHOR)
fm_anchors, anchor_labels, anchor_boxes = self.narrow_to_featuremap(
featuremap, get_all_anchors(), anchor_labels, anchor_boxes)
anchor_boxes_encoded = encode_bbox_target(anchor_boxes, fm_anchors)
image_shape2d = tf.shape(image)[2:] # h,w
pred_boxes_decoded = decode_bbox_target(
rpn_box_logits, fm_anchors) # fHxfWxNAx4, floatbox
proposal_boxes, proposal_scores = generate_rpn_proposals(
tf.reshape(pred_boxes_decoded, [-1, 4]),
tf.reshape(rpn_label_logits,
[-1]), image_shape2d, config.TRAIN_PRE_NMS_TOPK
if is_training else config.TEST_PRE_NMS_TOPK,
config.TRAIN_POST_NMS_TOPK
if is_training else config.TEST_POST_NMS_TOPK)
if is_training:
# sample proposal boxes in training
rcnn_boxes, rcnn_labels, fg_inds_wrt_gt = sample_fast_rcnn_targets(
proposal_boxes, gt_boxes, gt_labels)
else:
# The boxes to be used to crop RoIs.
# Use all proposal boxes in inference
rcnn_boxes = proposal_boxes
boxes_on_featuremap = rcnn_boxes * (1.0 / config.ANCHOR_STRIDE)
roi_resized = roi_align(featuremap, boxes_on_featuremap, 14)
# HACK to work around https://github.com/tensorflow/tensorflow/issues/14657
# which was fixed in TF 1.6
def ff_true():
feature_fastrcnn = resnet_conv5(
roi_resized, config.RESNET_NUM_BLOCK[-1]) # nxcx7x7
feature_gap = GlobalAvgPooling('gap',
feature_fastrcnn,
data_format='channels_first')
fastrcnn_label_logits, fastrcnn_box_logits = fastrcnn_outputs(
'fastrcnn', feature_gap, config.NUM_CLASS)
# Return C5 feature to be shared with mask branch
return feature_fastrcnn, fastrcnn_label_logits, fastrcnn_box_logits
def ff_false():
ncls = config.NUM_CLASS
return tf.zeros([0, 2048, 7,
7]), tf.zeros([0,
ncls]), tf.zeros([0, ncls - 1, 4])
if get_tf_version_number() >= 1.6:
feature_fastrcnn, fastrcnn_label_logits, fastrcnn_box_logits = ff_true(
)
else:
logger.warn("This example may drop support for TF < 1.6 soon.")
feature_fastrcnn, fastrcnn_label_logits, fastrcnn_box_logits = tf.cond(
tf.size(boxes_on_featuremap) > 0, ff_true, ff_false)
if is_training:
# rpn loss
rpn_label_loss, rpn_box_loss = rpn_losses(anchor_labels,
anchor_boxes_encoded,
rpn_label_logits,
rpn_box_logits)
# fastrcnn loss
matched_gt_boxes = tf.gather(gt_boxes, fg_inds_wrt_gt)
fg_inds_wrt_sample = tf.reshape(tf.where(rcnn_labels > 0),
[-1]) # fg inds w.r.t all samples
fg_sampled_boxes = tf.gather(rcnn_boxes, fg_inds_wrt_sample)
fg_fastrcnn_box_logits = tf.gather(fastrcnn_box_logits,
fg_inds_wrt_sample)
fastrcnn_label_loss, fastrcnn_box_loss = self.fastrcnn_training(
image, rcnn_labels, fg_sampled_boxes, matched_gt_boxes,
fastrcnn_label_logits, fg_fastrcnn_box_logits)
if config.MODE_MASK:
# maskrcnn loss
fg_labels = tf.gather(rcnn_labels, fg_inds_wrt_sample)
# In training, mask branch shares the same C5 feature.
fg_feature = tf.gather(feature_fastrcnn, fg_inds_wrt_sample)
mask_logits = maskrcnn_upXconv_head(
'maskrcnn', fg_feature, config.NUM_CLASS,
num_convs=0) # #fg x #cat x 14x14
matched_gt_masks = tf.gather(gt_masks,
fg_inds_wrt_gt) # nfg x H x W
target_masks_for_fg = crop_and_resize(
tf.expand_dims(matched_gt_masks, 1),
fg_sampled_boxes,
tf.range(tf.size(fg_inds_wrt_gt)),
14,
pad_border=False) # nfg x 1x14x14
target_masks_for_fg = tf.squeeze(target_masks_for_fg, 1,
'sampled_fg_mask_targets')
mrcnn_loss = maskrcnn_loss(mask_logits, fg_labels,
target_masks_for_fg)
else:
mrcnn_loss = 0.0
wd_cost = regularize_cost(
'(?:group1|group2|group3|rpn|fastrcnn|maskrcnn)/.*W',
l2_regularizer(1e-4),
name='wd_cost')
total_cost = tf.add_n([
rpn_label_loss, rpn_box_loss, fastrcnn_label_loss,
fastrcnn_box_loss, mrcnn_loss, wd_cost
], 'total_cost')
add_moving_summary(total_cost, wd_cost)
return total_cost
else:
final_boxes, final_labels = self.fastrcnn_inference(
image_shape2d, rcnn_boxes, fastrcnn_label_logits,
fastrcnn_box_logits)
if config.MODE_MASK:
# HACK to work around https://github.com/tensorflow/tensorflow/issues/14657
def f1():
roi_resized = roi_align(
featuremap, final_boxes * (1.0 / config.ANCHOR_STRIDE),
14)
feature_maskrcnn = resnet_conv5(
roi_resized, config.RESNET_NUM_BLOCK[-1])
mask_logits = maskrcnn_upXconv_head(
'maskrcnn', feature_maskrcnn, config.NUM_CLASS,
0) # #result x #cat x 14x14
indices = tf.stack([
tf.range(tf.size(final_labels)),
tf.to_int32(final_labels) - 1
],
axis=1)
final_mask_logits = tf.gather_nd(mask_logits,
indices) # #resultx14x14
return tf.sigmoid(final_mask_logits)
final_masks = tf.cond(
tf.size(final_labels) > 0, f1,
lambda: tf.zeros([0, 14, 14]))
tf.identity(final_masks, name='final_masks')
def BatchNorm_SplitGPU(x, use_local_stat=None, decay=0.9, epsilon=1e-5,
use_scale=True, use_bias=True,
gamma_init=tf.constant_initializer(1.0), data_format='NHWC',
internal_update=False, split_num = 1):
"""
"""
print split_num
if data_format == 'channels_last':
data_format = 'NHWC'
assert data_format == 'NHWC'
shape = x.get_shape().as_list()
ndims = len(shape)
assert ndims in [2, 4]
if ndims == 2:
data_format = 'NHWC'
if data_format == 'NCHW':
n_out = shape[1]
else:
n_out = shape[-1] # channel
assert n_out is not None, "Input to BatchNorm cannot have unknown channels!"
beta, gamma, moving_mean, moving_var = get_bn_variables(n_out, use_scale, use_bias, gamma_init)
ctx = get_current_tower_context()
if use_local_stat is None:
use_local_stat = ctx.is_training
use_local_stat = bool(use_local_stat)
if use_local_stat:
if ndims == 2:
x = tf.reshape(x, [-1, 1, 1, n_out]) # fused_bn only takes 4D input
# fused_bn has error using NCHW? (see #190)
inputs = tf.concat(tf.split(x, split_num, 0), -1) # N/S_n x H x W x C*S_n
beta_, gamma_ = None, None
beta_ = tf.reshape([beta]*split_num, [-1])
gamma_ = tf.reshape([gamma]*split_num, [-1])
xn, batch_mean, batch_var = tf.nn.fused_batch_norm(inputs, gamma_, beta_,epsilon=epsilon,is_training=True, data_format=data_format)
xn = tf.concat(tf.split(xn, split_num, 3), 0)
"""
"""
# inputs = tf.concat(tf.split(x, split_num, 0), -1) # N/split_num x H x W x C*split_num
# axis = [0, 1, 2]
# batch_mean, batch_var = tf.nn.moments(inputs, axis) # C*split_num
# beta_, gamma_ = None, None
# beta_ = tf.reshape([beta]*split_num, [-1])
# gamma_ = tf.reshape([gamma]*split_num, [-1])
# xn = tf.nn.batch_normalization(inputs, batch_mean, batch_var, beta_, gamma_, epsilon)
# xn = tf.concat(tf.split(xn, split_num, 3), 0)
if ndims == 2:
xn = tf.squeeze(xn, [1, 2])
else:
if ctx.is_training:
assert get_tf_version_number() < 1.4, \
"Fine tuning a BatchNorm model with fixed statistics is only " \
"supported after https://github.com/tensorflow/tensorflow/pull/12580 "
if ctx.is_main_training_tower: # only warn in first tower
logger.warn("[BatchNorm] Using moving_mean/moving_variance in training.")
# Using moving_mean/moving_variance in training, which means we
# loaded a pre-trained BN and only fine-tuning the affine part.
xn, batch_mean, batch_var = tf.nn.fused_batch_norm(
x, gamma, beta,
mean=moving_mean, variance=moving_var, epsilon=epsilon,
data_format=data_format, is_training=False)
else:
if ndims == 4 and data_format == 'NCHW':
[g, b, mm, mv] = [reshape_for_bn(_, ndims, n_out, data_format)
for _ in [gamma, beta, moving_mean, moving_var]]
xn = tf.nn.batch_normalization(x, mm, mv, b, g, epsilon)
batch_mean = tf.concat([moving_mean] * split_num, 0)
batch_var = tf.concat([moving_var] * split_num, 0)
else:
# avoid the reshape if possible (when channel is the last dimension)
xn = tf.nn.batch_normalization(
x, moving_mean, moving_var, beta, gamma, epsilon)
batch_mean = tf.concat([moving_mean] * split_num, 0)
batch_var = tf.concat([moving_var] * split_num, 0)
# maintain EMA only on one GPU is OK, even in replicated mode.
# because training time doesn't use EMA
if ctx.is_main_training_tower:
add_model_variable(moving_mean)
add_model_variable(moving_var)
if ctx.is_main_training_tower and use_local_stat:
# print (xn)
ret = update_bn_ema(xn, batch_mean[:n_out], batch_var[:n_out], moving_mean, moving_var, decay, internal_update)
else:
ret = tf.identity(xn, name='output')
ret = tf.identity(xn, name='output')
vh = ret.variables = VariableHolder(mean=moving_mean, variance=moving_var)
if use_scale:
vh.gamma = gamma
if use_bias:
vh.beta = beta
assert batch_mean is not None, 'batch_mean outputs is None'
return ret
def build_graph(self, *inputs):
is_training = get_current_tower_context().is_training
if config.MODE_MASK:
image, anchor_labels, anchor_boxes, gt_boxes, gt_labels, gt_masks = inputs
else:
image, anchor_labels, anchor_boxes, gt_boxes, gt_labels = inputs
image = self.preprocess(image) # 1CHW
featuremap = resnet_c4_backbone(image, config.RESNET_NUM_BLOCK[:3])
rpn_label_logits, rpn_box_logits = rpn_head('rpn', featuremap, 1024, config.NUM_ANCHOR)
fm_anchors, anchor_labels, anchor_boxes = self.narrow_to_featuremap(
featuremap, get_all_anchors(), anchor_labels, anchor_boxes)
anchor_boxes_encoded = encode_bbox_target(anchor_boxes, fm_anchors)
image_shape2d = tf.shape(image)[2:] # h,w
pred_boxes_decoded = decode_bbox_target(rpn_box_logits, fm_anchors) # fHxfWxNAx4, floatbox
proposal_boxes, proposal_scores = generate_rpn_proposals(
tf.reshape(pred_boxes_decoded, [-1, 4]),
tf.reshape(rpn_label_logits, [-1]),
image_shape2d,
config.TRAIN_PRE_NMS_TOPK if is_training else config.TEST_PRE_NMS_TOPK,
config.TRAIN_POST_NMS_TOPK if is_training else config.TEST_POST_NMS_TOPK)
if is_training:
# sample proposal boxes in training
rcnn_boxes, rcnn_labels, fg_inds_wrt_gt = sample_fast_rcnn_targets(
proposal_boxes, gt_boxes, gt_labels)
else:
# The boxes to be used to crop RoIs.
# Use all proposal boxes in inference
rcnn_boxes = proposal_boxes
boxes_on_featuremap = rcnn_boxes * (1.0 / config.ANCHOR_STRIDE)
roi_resized = roi_align(featuremap, boxes_on_featuremap, 14)
# HACK to work around https://github.com/tensorflow/tensorflow/issues/14657
# which was fixed in TF 1.6
def ff_true():
feature_fastrcnn = resnet_conv5(roi_resized, config.RESNET_NUM_BLOCK[-1]) # nxcx7x7
feature_gap = GlobalAvgPooling('gap', feature_fastrcnn, data_format='channels_first')
fastrcnn_label_logits, fastrcnn_box_logits = fastrcnn_outputs('fastrcnn', feature_gap, config.NUM_CLASS)
# Return C5 feature to be shared with mask branch
return feature_fastrcnn, fastrcnn_label_logits, fastrcnn_box_logits
def ff_false():
ncls = config.NUM_CLASS
return tf.zeros([0, 2048, 7, 7]), tf.zeros([0, ncls]), tf.zeros([0, ncls - 1, 4])
if get_tf_version_number() >= 1.6:
feature_fastrcnn, fastrcnn_label_logits, fastrcnn_box_logits = ff_true()
else:
logger.warn("This example may drop support for TF < 1.6 soon.")
feature_fastrcnn, fastrcnn_label_logits, fastrcnn_box_logits = tf.cond(
tf.size(boxes_on_featuremap) > 0, ff_true, ff_false)
if is_training:
# rpn loss
rpn_label_loss, rpn_box_loss = rpn_losses(
anchor_labels, anchor_boxes_encoded, rpn_label_logits, rpn_box_logits)
# fastrcnn loss
matched_gt_boxes = tf.gather(gt_boxes, fg_inds_wrt_gt)
fg_inds_wrt_sample = tf.reshape(tf.where(rcnn_labels > 0), [-1]) # fg inds w.r.t all samples
fg_sampled_boxes = tf.gather(rcnn_boxes, fg_inds_wrt_sample)
fg_fastrcnn_box_logits = tf.gather(fastrcnn_box_logits, fg_inds_wrt_sample)
fastrcnn_label_loss, fastrcnn_box_loss = self.fastrcnn_training(
image, rcnn_labels, fg_sampled_boxes,
matched_gt_boxes, fastrcnn_label_logits, fg_fastrcnn_box_logits)
if config.MODE_MASK:
# maskrcnn loss
fg_labels = tf.gather(rcnn_labels, fg_inds_wrt_sample)
# In training, mask branch shares the same C5 feature.
fg_feature = tf.gather(feature_fastrcnn, fg_inds_wrt_sample)
mask_logits = maskrcnn_upXconv_head(
'maskrcnn', fg_feature, config.NUM_CLASS, num_convs=0) # #fg x #cat x 14x14
target_masks_for_fg = crop_and_resize(
tf.expand_dims(gt_masks, 1),
fg_sampled_boxes,
fg_inds_wrt_gt, 14,
pad_border=False) # nfg x 1x14x14
target_masks_for_fg = tf.squeeze(target_masks_for_fg, 1, 'sampled_fg_mask_targets')
mrcnn_loss = maskrcnn_loss(mask_logits, fg_labels, target_masks_for_fg)
else:
mrcnn_loss = 0.0
wd_cost = regularize_cost(
'(?:group1|group2|group3|rpn|fastrcnn|maskrcnn)/.*W',
l2_regularizer(1e-4), name='wd_cost')
total_cost = tf.add_n([
rpn_label_loss, rpn_box_loss,
fastrcnn_label_loss, fastrcnn_box_loss,
mrcnn_loss,
wd_cost], 'total_cost')
add_moving_summary(total_cost, wd_cost)
return total_cost
else:
final_boxes, final_labels = self.fastrcnn_inference(
image_shape2d, rcnn_boxes, fastrcnn_label_logits, fastrcnn_box_logits)
if config.MODE_MASK:
# HACK to work around https://github.com/tensorflow/tensorflow/issues/14657
def f1():
roi_resized = roi_align(featuremap, final_boxes * (1.0 / config.ANCHOR_STRIDE), 14)
feature_maskrcnn = resnet_conv5(roi_resized, config.RESNET_NUM_BLOCK[-1])
mask_logits = maskrcnn_upXconv_head(
'maskrcnn', feature_maskrcnn, config.NUM_CLASS, 0) # #result x #cat x 14x14
indices = tf.stack([tf.range(tf.size(final_labels)), tf.to_int32(final_labels) - 1], axis=1)
final_mask_logits = tf.gather_nd(mask_logits, indices) # #resultx14x14
return tf.sigmoid(final_mask_logits)
final_masks = tf.cond(tf.size(final_labels) > 0, f1, lambda: tf.zeros([0, 14, 14]))
tf.identity(final_masks, name='final_masks')
self._eval()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--load', help='load a model for evaluation or training')
parser.add_argument('--logdir', help='log directory', default='train_log/maskrcnn')
parser.add_argument('--config', help="A list of KEY=VALUE to overwrite those defined in config.py",
nargs='+')
parser.add_argument('--visualize', action='store_true', help='visualize intermediate results')
parser.add_argument('--evaluate', help="Run evaluation on COCO. "
"This argument is the path to the output json evaluation file")
parser.add_argument('--predict', help="Run prediction on a given image. "
"This argument is the path to the input image file")
if get_tf_version_number() < 1.6:
# https://github.com/tensorflow/tensorflow/issues/14657
logger.warn("TF<1.6 has a bug which may lead to crash in FasterRCNN training if you're unlucky.")
args = parser.parse_args()
cfg.update_args(args.config)
MODEL = ResNetFPNModel() if cfg.MODE_FPN else ResNetC4Model()
if args.visualize or args.evaluate or args.predict:
assert args.load
finalize_configs(is_training=False)
if args.predict or args.visualize:
cfg.TEST.RESULT_SCORE_THRESH = cfg.TEST.RESULT_SCORE_THRESH_VIS
def BatchNorm(inputs,
axis=None,
training=None,
momentum=0.9,
epsilon=1e-5,
center=True,
scale=True,
beta_initializer=tf.zeros_initializer(),
gamma_initializer=tf.ones_initializer(),
virtual_batch_size=None,
internal_update=False):
"""
Mostly equivalent to `tf.layers.batch_normalization`, but different in
the following:
1. Accepts `data_format` when `axis` is None. For 2D input, this argument will be ignored.
2. Default value for `momentum` and `epsilon` is different.
3. Default value for `training` is automatically obtained from `TowerContext`.
4. Support the `internal_update` option.
Args:
internal_update (bool): if False, add EMA update ops to
`tf.GraphKeys.UPDATE_OPS`. If True, update EMA inside the layer
by control dependencies.
Variable Names:
* ``beta``: the bias term. Will be zero-inited by default.
* ``gamma``: the scale term. Will be one-inited by default. Input will be transformed by ``x * gamma + beta``.
* ``mean/EMA``: the moving average of mean.
* ``variance/EMA``: the moving average of variance.
Note:
1. About multi-GPU training: moving averages across GPUs are not aggregated.
Batch statistics are computed independently. This is consistent with most frameworks.
2. Combinations of ``training`` and ``ctx.is_training``:
* ``training == ctx.is_training``: standard BN, EMA are
maintained during training and used during inference. This is
the default.
* ``training and not ctx.is_training``: still use batch statistics in inference.
* ``not training and ctx.is_training``: use EMA to normalize in
training. This is useful when you load a pre-trained BN and
don't want to fine tune the EMA. EMA will not be updated in
this case.
"""
# parse shapes
shape = inputs.get_shape().as_list()
ndims = len(shape)
assert axis is not None
# parse training/ctx
ctx = get_current_tower_context()
if training is None:
training = ctx.is_training
training = bool(training)
TF_version = get_tf_version_number()
if not training and ctx.is_training:
assert TF_version >= 1.4, \
"Fine tuning a BatchNorm model with fixed statistics is only " \
"supported after https://github.com/tensorflow/tensorflow/pull/12580 "
if ctx.is_main_training_tower: # only warn in first tower
logger.warn(
"[BatchNorm] Using moving_mean/moving_variance in training.")
# Using moving_mean/moving_variance in training, which means we
# loaded a pre-trained BN and only fine-tuning the affine part.
coll_bk = backup_collection([tf.GraphKeys.UPDATE_OPS])
with rename_get_variable({
'moving_mean': 'mean/EMA',
'moving_variance': 'variance/EMA'
}):
if TF_version >= 1.5:
layer = tf.layers.BatchNormalization(
axis=axis,
momentum=momentum,
epsilon=epsilon,
center=center,
scale=scale,
beta_initializer=beta_initializer,
gamma_initializer=gamma_initializer,
virtual_batch_size=virtual_batch_size,
fused=True,
_reuse=tf.get_variable_scope().reuse)
else:
assert virtual_batch_size is None, "Feature not supported in this version of TF!"
layer = tf.layers.BatchNormalization(
axis=axis,
momentum=momentum,
epsilon=epsilon,
center=center,
scale=scale,
beta_initializer=beta_initializer,
gamma_initializer=gamma_initializer,
fused=True,
_reuse=tf.get_variable_scope().reuse)
xn = layer.apply(inputs,
training=training,
scope=tf.get_variable_scope())
# maintain EMA only on one GPU is OK, even in replicated mode.
# because training time doesn't use EMA
if ctx.is_main_training_tower:
for v in layer.non_trainable_variables:
add_model_variable(v)
if not ctx.is_main_training_tower or internal_update:
restore_collection(coll_bk)
if training and internal_update:
assert layer.updates
with tf.control_dependencies(layer.updates):
ret = tf.identity(xn, name='output')
else:
ret = tf.identity(xn, name='output')
vh = ret.variables = VariableHolder(
moving_mean=layer.moving_mean,
mean=layer.moving_mean, # for backward-compatibility
moving_variance=layer.moving_variance,
variance=layer.moving_variance) # for backward-compatibility
if scale:
vh.gamma = layer.gamma
if center:
vh.beta = layer.beta
return ret
def test(self):
if get_tf_version_number() < 1.4:
return True # requires leaky_relu
self.assertSurvive(self.script, args=None)
