def get_config(is_train):
class General:
# number of iteration for print the metrics to stdout
log_frequency = 10
# the directory name for the experiment, the default is the name of config
name = __name__.rsplit("/")[-1].rsplit(".")[-1]
# batch size per GPU
batch_image = 2 if is_train else 1
# use FP16 for weight and activation
# recommend to toggle when you are training on Volta or later GPUs
fp16 = False
# number of threads used for the data loader
# this term affects both the CPU utilization and the MEM usage
# lower this if you are training on Desktop
loader_worker = 8
# switch the built in profile to find the bottleneck of network
profile = False
class KvstoreParam:
# the type of communicator used to sync model parameters
kvstore = "nccl" # "local", "aggregated"
batch_image = General.batch_image
# GPUs to use
gpus = [0, 1, 2, 3, 4, 5, 6, 7]
fp16 = General.fp16
class NormalizeParam:
# the type of normalizer used for network
# see also ModelParam.pretrain.fixed_param for the freeze of gamma/beta
normalizer = normalizer_factory(type="fixbn") # freeze bn stats
normalizer = normalizer_factory(
type="localbn") # use bn stats in one GPU
normalizer = normalizer_factory(
type="syncbn",
ndev=len(KvstoreParam.gpus)) # use bn stats across GPUs
normalizer = normalizer_factory(type="gn") # use GroupNorm
class BackboneParam:
# you can control the FP16 option and normalizer for each individual component
fp16 = General.fp16
normalizer = NormalizeParam.normalizer
# some backbone component accept additional configs, like the depth for ResNet
depth = 50
class NeckParam:
# you can control the FP16 option and normalizer for each individual component
fp16 = General.fp16
normalizer = NormalizeParam.normalizer
class RpnParam:
# you can control the FP16 option and normalizer for each individual component
fp16 = General.fp16
normalizer = NormalizeParam.normalizer
batch_image = General.batch_image
# use ONNX-compatible proposal operator instead of the one written in C++/CUDA
nnvm_proposal = True
# use in-network rpn target operator instead of the label generated by data loader
# if your network is quite fast, the CPU might not feed the labels fast enough
# else you can offload the rpn target generation to CPU to save GPU resources
nnvm_rpn_target = False
# anchor grid generated are used in the rpn target assign and proposal decoding
class anchor_generate:
scale = (8, )
ratio = (0.5, 1.0, 2.0)
stride = (4, 8, 16, 32, 64)
# number of anchors per image
image_anchor = 256
# to avoid generate the same anchor grid more than once
# we cache an anchor grid in the arg_params
# max_side specify the max side of resized input image
# 3000 is a safe bet, increase it if necessary
max_side = 1400
# valid when use nnvm_rpn_target, controls the rpn target assign
class anchor_assign:
# number of pixels the anchor box could extend out of the image border
allowed_border = 0
# iou lower bound with groundtruth box for foreground anchor
pos_thr = 0.7
# iou upper bound with groundtruth box for background anchor
neg_thr = 0.3
# every groundtruth box will match the anchors overlaps most with it by default
# increase the threshold to avoid matching low quality anchors
min_pos_thr = 0.0
# number of anchors per image
image_anchor = 256
# fraction of foreground anchors per image
pos_fraction = 0.5
# rpn head structure
class head:
# number of channels for the 3x3 conv in rpn head
conv_channel = 256
# mean and std for rpn regression target
mean = (0, 0, 0, 0)
std = (1, 1, 1, 1)
# the proposal generation for RCNN
class proposal:
# number of top-scored proposals to take before NMS
pre_nms_top_n = 2000 if is_train else 1000
# number of top-scored proposals to take after NMS
post_nms_top_n = 2000 if is_train else 1000
# proposal NMS threshold
nms_thr = 0.7
# min proposal box to keep, 0 means keep all
min_bbox_side = 0
# the proposal sampling for RCNN during training
class subsample_proposal:
# add gt to proposals
proposal_wo_gt = False
# number of proposals sampled per image during training
image_roi = 512
# the maxinum fraction of foreground proposals
fg_fraction = 0.25
# iou lower bound with gt bbox for foreground proposals
fg_thr = 0.5
# iou upper bound with gt bbox for background proposals
bg_thr_hi = 0.5
# iou lower bound with gt bbox for background proposals
# set to non-zero value could remove some trivial background proposals
bg_thr_lo = 0.0
# the target encoding for RCNN bbox head
class bbox_target:
# 1(background) + num_class
# could be num_class if using sigmoid activition instead of softmax one
num_reg_class = 1 + 80
# share the regressor for all classes
class_agnostic = False
# the mean, std, and weight for bbox head regression target
weight = (1.0, 1.0, 1.0, 1.0)
mean = (0.0, 0.0, 0.0, 0.0)
std = (0.1, 0.1, 0.2, 0.2)
class BboxParam:
# you can control the FP16 option and normalizer for each individual component
fp16 = General.fp16
normalizer = NormalizeParam.normalizer
# num_class may be different from RpnParam.bbox_target.num_reg_class
# if the class_agnostic regressor is adopted
num_class = 1 + 80
image_roi = RpnParam.subsample_proposal.image_roi
batch_image = General.batch_image
class regress_target:
class_agnostic = RpnParam.bbox_target.class_agnostic
mean = RpnParam.bbox_target.mean
std = RpnParam.bbox_target.std
class MaskParam:
# you can control the FP16 option and normalizer for each individual component
fp16 = General.fp16
normalizer = NormalizeParam.normalizer
# output resolution of mask head
resolution = 28
# number of channels for 3x3 convs in mask head
dim_reduced = 256
# mask head only trains on foreground proposals
# so we discard all the background proposals to save computation
num_fg_roi = int(RpnParam.subsample_proposal.image_roi *
RpnParam.subsample_proposal.fg_fraction)
class RoiParam:
# you can control the FP16 option and normalizer for each individual component
fp16 = General.fp16
normalizer = NormalizeParam.normalizer
# Each RoI is pooled into an out_size x out_size fixed-length representation
out_size = 7
# the total stride of the feature map to pool from
stride = (4, 8, 16, 32)
# FPN specific configs
# objects of size in [224^2, 448^2) will be assgin to P4
roi_canonical_scale = 224
roi_canonical_level = 4
class MaskRoiParam:
# you can control the FP16 option and normalizer for each individual component
fp16 = General.fp16
normalizer = NormalizeParam.normalizer
# Each RoI is pooled into an out_size x out_size fixed-length representation
out_size = 14
# the total stride of the feature map to pool from
stride = (4, 8, 16, 32)
# FPN specific configs
# objects of size in [224^2, 448^2) will be assgin to P4
roi_canonical_scale = 224
roi_canonical_level = 4
class DatasetParam:
# specify the roidbs to read for training/validation
if is_train:
# == coco_train2017
image_set = ("coco_train2014", "coco_valminusminival2014")
else:
# == coco_val2017
image_set = ("coco_minival2014", )
class OptimizeParam:
class optimizer:
type = "sgd"
# learning rate will automaticly adapt to different batch size
# the base learning rate is 0.02 for 16 images
lr = 0.01 / 8 * len(KvstoreParam.gpus) * KvstoreParam.batch_image
momentum = 0.9
wd = 0.0001
clip_gradient = None
class schedule:
# correspond to the 1x, 2x, ... training schedule
mult = 2
begin_epoch = 0
end_epoch = 6 * mult
lr_mode = "step" # or "cosine"
# lr step factor
lr_factor = 0.1
# lr step iterations
if mult <= 1:
lr_iter = [
60000 * mult * 16 //
(len(KvstoreParam.gpus) * KvstoreParam.batch_image),
80000 * mult * 16 //
(len(KvstoreParam.gpus) * KvstoreParam.batch_image)
]
else:
# follow the practice in arXiv:1811.08883
# reduce the lr in the last 60k and 20k iterations
lr_iter = [
-60000 * 16 //
(len(KvstoreParam.gpus) * KvstoreParam.batch_image),
-20000 * 16 //
(len(KvstoreParam.gpus) * KvstoreParam.batch_image)
]
# follow the practice in arXiv:1706.02677
class warmup:
type = "gradual"
lr = 0.01 / 8 * len(
KvstoreParam.gpus) * KvstoreParam.batch_image / 3
iter = 500
class TestParam:
# detection below min_det_score will be removed in the evaluation
min_det_score = 0.05
# only the top max_det_per_image detecitons will be evaluated
max_det_per_image = 100
# callback, useful in multi-scale testing
process_roidb = lambda x: x
# callback, useful in scale-aware post-processing
process_output = lambda x, y: process_output(x, y)
# the model name and epoch used during test
# by default the last checkpoint is employed
# user can override this with --epoch N when invoking script
class model:
prefix = "experiments/{}/checkpoint".format(General.name)
epoch = OptimizeParam.schedule.end_epoch
class nms:
type = "nms" # or "softnms"
thr = 0.5
# we make use of the coco test toolchain
# if no coco format annotation file is specified
# test script will generate one on the fly from roidb
class coco:
annotation = "data/coco/annotations/instances_minival2014.json"
# compose the components to for a detector
backbone = Backbone(BackboneParam)
neck = Neck(NeckParam)
rpn_head = RpnHead(RpnParam, MaskParam)
roi_extractor = RoiExtractor(RoiParam)
mask_roi_extractor = RoiExtractor(MaskRoiParam)
bbox_head = BboxHead(BboxParam)
mask_head = MaskHead(BboxParam, MaskParam, MaskRoiParam)
bbox_post_processer = BboxPostProcessor(TestParam)
detector = Detector()
if is_train:
train_sym = detector.get_train_symbol(backbone, neck, rpn_head,
roi_extractor,
mask_roi_extractor, bbox_head,
mask_head)
test_sym = None
else:
train_sym = None
test_sym = detector.get_test_symbol(backbone, neck, rpn_head,
roi_extractor, mask_roi_extractor,
bbox_head, mask_head,
bbox_post_processer)
class ModelParam:
train_symbol = train_sym
test_symbol = test_sym
# training model from scratch
from_scratch = False
# use random seed when initializating
random = True
# sublinear memory checkpointing
memonger = False
# checkpointing up to a layer
# recompute early stage of a network is cheaper
memonger_until = "stage3_unit21_plus"
class pretrain:
# the model name and epoch used for initialization
prefix = "pretrain_model/resnet%s_v1b" % BackboneParam.depth
epoch = 0
# any params partially match the fixed_param will be fixed
# fixed params will not be updated
fixed_param = ["conv0", "stage1", "gamma", "beta"]
# any params partially match the excluded_param will not be fixed
excluded_param = ["mask_fcn"]
# callback, useful for adding cached anchor or complex initialization
def process_weight(sym, arg, aux):
for stride in RpnParam.anchor_generate.stride:
add_anchor_to_arg(sym, arg, aux,
RpnParam.anchor_generate.max_side, stride,
RpnParam.anchor_generate.scale,
RpnParam.anchor_generate.ratio)
# data processing
class NormParam:
# mean/std for input image
mean = tuple(i * 255 for i in (0.485, 0.456, 0.406)) # RGB order
std = tuple(i * 255 for i in (0.229, 0.224, 0.225))
# data processing
class ResizeParam:
# the input is resized to a short side not exceeding short
# and a long side not exceeding long
short = 800
long = 1333
# SimpleDet is written in MXNet symbolic API which features the fastest
# execution while requires static input shape
# All the inputs are padded to the maximum shape item on the dataset
class PadParam:
# the resized input is padded to short x long with 0 in bottom-right corner
short = 800
long = 1333
max_num_gt = 100
max_len_gt_poly = 2500
# this control the rpn target generation offloaded to CPU data loader
# refer to RpnParam.anchor_generate for more infos
class AnchorTarget2DParam:
def __init__(self):
self.generate = self._generate()
class _generate:
def __init__(self):
self.stride = (4, 8, 16, 32, 64)
# the shorts and longs have to be pre-computed since the
# loader knows nothing of the network
# the downsampled side can be calculated as ceil(side / 2)
self.short = (200, 100, 50, 25, 13)
self.long = (334, 167, 84, 42, 21)
scales = (8, )
aspects = (0.5, 1.0, 2.0)
class assign:
allowed_border = 0
pos_thr = 0.7
neg_thr = 0.3
min_pos_thr = 0.0
class sample:
image_anchor = 256
pos_fraction = 0.5
# align blobs name between loader and network
class RenameParam:
mapping = dict(image="data")
from core.detection_input import ReadRoiRecord, Resize2DImageBbox, \
ConvertImageFromHwcToChw, Flip2DImageBbox, Pad2DImageBbox, \
RenameRecord, Norm2DImage
from models.maskrcnn.input import PreprocessGtPoly, EncodeGtPoly, \
Resize2DImageBboxMask, Flip2DImageBboxMask, Pad2DImageBboxMask
from models.FPN.input import PyramidAnchorTarget2D
# modular data augmentation design
if is_train:
transform = [
ReadRoiRecord(None),
Norm2DImage(NormParam),
PreprocessGtPoly(),
Resize2DImageBboxMask(ResizeParam),
Flip2DImageBboxMask(),
EncodeGtPoly(PadParam),
Pad2DImageBboxMask(PadParam),
ConvertImageFromHwcToChw(),
RenameRecord(RenameParam.mapping)
]
data_name = ["data"]
label_name = ["im_info", "gt_bbox", "gt_poly"]
if not RpnParam.nnvm_rpn_target:
transform.append(PyramidAnchorTarget2D(AnchorTarget2DParam()))
label_name += ["rpn_cls_label", "rpn_reg_target", "rpn_reg_weight"]
else:
transform = [
ReadRoiRecord(None),
Norm2DImage(NormParam),
Resize2DImageBbox(ResizeParam),
ConvertImageFromHwcToChw(),
RenameRecord(RenameParam.mapping)
]
data_name = ["data", "im_info", "im_id", "rec_id"]
label_name = []
import core.detection_metric as metric
from models.maskrcnn.metric import SigmoidCELossMetric
from mxboard import SummaryWriter
# summary writer logs metric to tensorboard for a better track of training
sw = SummaryWriter(logdir="./tflogs", flush_secs=5)
rpn_acc_metric = metric.AccWithIgnore(
name="RpnAcc",
output_names=["rpn_cls_loss_output", "rpn_cls_label_blockgrad_output"],
label_names=[],
summary=sw)
rpn_l1_metric = metric.L1(
name="RpnL1",
output_names=["rpn_reg_loss_output", "rpn_cls_label_blockgrad_output"],
label_names=[],
summary=sw)
box_acc_metric = metric.AccWithIgnore(
name="RcnnAcc",
output_names=["bbox_cls_loss_output", "bbox_label_blockgrad_output"],
label_names=[],
summary=sw)
box_l1_metric = metric.L1(
name="RcnnL1",
output_names=["bbox_reg_loss_output", "bbox_label_blockgrad_output"],
label_names=[],
summary=sw)
mask_cls_metric = SigmoidCELossMetric(name="MaskCE",
output_names=["mask_loss_output"],
label_names=[],
summary=sw)
metric_list = [
rpn_acc_metric, rpn_l1_metric, box_acc_metric, box_l1_metric,
mask_cls_metric
]
return General, KvstoreParam, RpnParam, RoiParam, BboxParam, DatasetParam, \
ModelParam, OptimizeParam, TestParam, \
transform, data_name, label_name, metric_list
def get_config(is_train):
class General:
log_frequency = 10
name = __name__.rsplit("/")[-1].rsplit(".")[-1]
batch_image = 2 if is_train else 1
fp16 = False
loader_worker = 8
class KvstoreParam:
kvstore = "nccl"
batch_image = General.batch_image
gpus = [0, 1, 2, 3, 4, 5, 6, 7]
fp16 = General.fp16
class NormalizeParam:
normalizer = normalizer_factory(type="gn")
class BackboneParam:
fp16 = General.fp16
normalizer = NormalizeParam.normalizer
depth = 50
class NeckParam:
fp16 = General.fp16
normalizer = NormalizeParam.normalizer
class RpnParam:
fp16 = General.fp16
normalizer = NormalizeParam.normalizer
batch_image = General.batch_image
nnvm_proposal = True
nnvm_rpn_target = False
class anchor_generate:
scale = (8, )
ratio = (0.5, 1.0, 2.0)
stride = (4, 8, 16, 32, 64)
image_anchor = 256
max_side = 1400
class anchor_assign:
allowed_border = 0
pos_thr = 0.7
neg_thr = 0.3
min_pos_thr = 0.0
image_anchor = 256
pos_fraction = 0.5
class head:
conv_channel = 256
mean = (0, 0, 0, 0)
std = (1, 1, 1, 1)
class proposal:
pre_nms_top_n = 2000 if is_train else 1000
post_nms_top_n = 2000 if is_train else 1000
nms_thr = 0.7
min_bbox_side = 0
class subsample_proposal:
proposal_wo_gt = False
image_roi = 512
fg_fraction = 0.25
fg_thr = 0.5
bg_thr_hi = 0.5
bg_thr_lo = 0.0
class bbox_target:
num_reg_class = 81
class_agnostic = False
weight = (1.0, 1.0, 1.0, 1.0)
mean = (0.0, 0.0, 0.0, 0.0)
std = (0.1, 0.1, 0.2, 0.2)
class BboxParam:
fp16 = General.fp16
normalizer = NormalizeParam.normalizer
num_class = 1 + 80
image_roi = 512
batch_image = General.batch_image
class regress_target:
class_agnostic = False
mean = (0.0, 0.0, 0.0, 0.0)
std = (0.1, 0.1, 0.2, 0.2)
class MaskParam:
fp16 = General.fp16
normalizer = NormalizeParam.normalizer
resolution = 28
dim_reduced = 256
num_fg_roi = int(RpnParam.subsample_proposal.image_roi *
RpnParam.subsample_proposal.fg_fraction)
class RoiParam:
fp16 = General.fp16
normalizer = NormalizeParam.normalizer
out_size = 7
stride = (4, 8, 16, 32)
roi_canonical_scale = 224
roi_canonical_level = 4
class MaskRoiParam:
fp16 = General.fp16
normalizer = NormalizeParam.normalizer
out_size = 14
stride = (4, 8, 16, 32)
roi_canonical_scale = 224
roi_canonical_level = 4
class DatasetParam:
if is_train:
image_set = ("coco_train2017", )
else:
image_set = ("coco_val2017", )
class OptimizeParam:
class optimizer:
type = "sgd"
lr = 0.02 / 8 * len(KvstoreParam.gpus) * KvstoreParam.batch_image
momentum = 0.9
wd = 0.0001
clip_gradient = None
class schedule:
mult = 2
begin_epoch = 0
end_epoch = 6 * mult
lr_iter = [
60000 * mult * 16 //
(len(KvstoreParam.gpus) * KvstoreParam.batch_image),
80000 * mult * 16 //
(len(KvstoreParam.gpus) * KvstoreParam.batch_image)
]
class warmup:
type = "gradual"
lr = 0.01 / 8 * len(
KvstoreParam.gpus) * KvstoreParam.batch_image / 3.0
iter = 500
class TestParam:
min_det_score = 0.05
max_det_per_image = 100
process_roidb = lambda x: x
process_output = lambda x, y: process_output(x, y)
class model:
prefix = "experiments/{}/checkpoint".format(General.name)
epoch = OptimizeParam.schedule.end_epoch
class nms:
type = "nms"
thr = 0.5
class coco:
annotation = "data/coco/annotations/instances_minival2014.json"
backbone = Backbone(BackboneParam)
neck = Neck(NeckParam)
rpn_head = RpnHead(RpnParam, MaskParam)
roi_extractor = RoiExtractor(RoiParam)
mask_roi_extractor = RoiExtractor(MaskRoiParam)
bbox_head = BboxHead(BboxParam)
mask_head = MaskHead(BboxParam, MaskParam, MaskRoiParam)
bbox_post_processer = BboxPostProcessor(TestParam)
detector = Detector()
if is_train:
train_sym = detector.get_train_symbol(backbone, neck, rpn_head,
roi_extractor,
mask_roi_extractor, bbox_head,
mask_head)
test_sym = None
else:
train_sym = None
test_sym = detector.get_test_symbol(backbone, neck, rpn_head,
roi_extractor, mask_roi_extractor,
bbox_head, mask_head,
bbox_post_processer)
class ModelParam:
train_symbol = train_sym
test_symbol = test_sym
from_scratch = True
random = True
memonger = False
memonger_until = "stage3_unit21_plus"
class pretrain:
prefix = "pretrain_model/resnet%s_v1b" % BackboneParam.depth
epoch = 0
fixed_param = []
def process_weight(sym, arg, aux):
for stride in RpnParam.anchor_generate.stride:
add_anchor_to_arg(sym, arg, aux,
RpnParam.anchor_generate.max_side, stride,
RpnParam.anchor_generate.scale,
RpnParam.anchor_generate.ratio)
# data processing
class NormParam:
mean = tuple(i * 255 for i in (0.485, 0.456, 0.406)) # RGB order
std = tuple(i * 255 for i in (0.229, 0.224, 0.225))
# data processing
class ResizeParam:
short = 800
long = 1333
class PadParam:
short = 800
long = 1333
max_num_gt = 100
max_len_gt_poly = 2500
class AnchorTarget2DParam:
def __init__(self):
self.generate = self._generate()
class _generate:
def __init__(self):
self.stride = (4, 8, 16, 32, 64)
self.short = (200, 100, 50, 25, 13)
self.long = (334, 167, 84, 42, 21)
scales = (8)
aspects = (0.5, 1.0, 2.0)
class assign:
allowed_border = 0
pos_thr = 0.7
neg_thr = 0.3
min_pos_thr = 0.0
class sample:
image_anchor = 256
pos_fraction = 0.5
class RenameParam:
mapping = dict(image="data")
from core.detection_input import ReadRoiRecord, Resize2DImageBbox, \
ConvertImageFromHwcToChw, Flip2DImageBbox, Pad2DImageBbox, \
RenameRecord, Norm2DImage
from models.maskrcnn.input import PreprocessGtPoly, EncodeGtPoly, \
Resize2DImageBboxMask, Flip2DImageBboxMask, Pad2DImageBboxMask
from models.FPN.input import PyramidAnchorTarget2D
if is_train:
transform = [
ReadRoiRecord(None),
Norm2DImage(NormParam),
PreprocessGtPoly(),
Resize2DImageBboxMask(ResizeParam),
Flip2DImageBboxMask(),
EncodeGtPoly(PadParam),
Pad2DImageBboxMask(PadParam),
ConvertImageFromHwcToChw(),
RenameRecord(RenameParam.mapping)
]
data_name = ["data"]
label_name = ["im_info", "gt_bbox", "gt_poly"]
if not RpnParam.nnvm_rpn_target:
transform.append(PyramidAnchorTarget2D(AnchorTarget2DParam()))
label_name += ["rpn_cls_label", "rpn_reg_target", "rpn_reg_weight"]
else:
transform = [
ReadRoiRecord(None),
Norm2DImage(NormParam),
Resize2DImageBbox(ResizeParam),
ConvertImageFromHwcToChw(),
RenameRecord(RenameParam.mapping)
]
data_name = ["data", "im_info", "im_id", "rec_id"]
label_name = []
import core.detection_metric as metric
from models.maskrcnn.metric import SigmoidCELossMetric
rpn_acc_metric = metric.AccWithIgnore(
"RpnAcc", ["rpn_cls_loss_output", "rpn_cls_label_blockgrad_output"],
[])
rpn_l1_metric = metric.L1(
"RpnL1", ["rpn_reg_loss_output", "rpn_cls_label_blockgrad_output"], [])
# for bbox, the label is generated in network so it is an output
box_acc_metric = metric.AccWithIgnore(
"RcnnAcc", ["bbox_cls_loss_output", "bbox_label_blockgrad_output"], [])
box_l1_metric = metric.L1(
"RcnnL1", ["bbox_reg_loss_output", "bbox_label_blockgrad_output"], [])
mask_cls_metric = SigmoidCELossMetric("MaskCE", ["mask_loss_output"], [])
metric_list = [
rpn_acc_metric,
rpn_l1_metric,
box_acc_metric,
box_l1_metric,
]
return General, KvstoreParam, RpnParam, RoiParam, BboxParam, DatasetParam, \
ModelParam, OptimizeParam, TestParam, \
transform, data_name, label_name, metric_list
pos_thr = 0.7
neg_thr = 0.3
min_pos_thr = 0.0
class sample:
image_anchor = 256
pos_fraction = 0.5
class RenameParam:
mapping = dict(image="data")
transform = [
ReadRoiRecord(None),
PreprocessGtPoly(),
Resize2DImageBboxMask(ResizeParam),
Flip2DImageBboxMask(),
EncodeGtPoly(PadParam),
Pad2DImageBboxMask(PadParam),
ConvertImageFromHwcToChw(),
AnchorTarget2D(AnchorTarget2DParam),
RenameRecord(RenameParam.mapping)
]
DEBUG = True
with open("data/cache/coco_valminusminival2014.roidb", "rb") as fin:
roidb = pkl.load(fin)
roidb = [rec for rec in roidb if rec["gt_bbox"].shape[0] > 0]
roidb = [
roidb[i] for i in np.random.choice(len(roidb), 20, replace=False)
]