def _decode_cfg_value(v):
"""Decodes a raw config value (e.g., from a yaml config files or command
line argument) into a Python object.
"""
# Configs parsed from raw yaml will contain dictionary keys that need to be
# converted to AttrDict objects
if isinstance(v, dict):
return AttrDict(v)
# All remaining processing is only applied to strings
if not isinstance(v, basestring):
return v
# Try to interpret `v` as a:
# string, number, tuple, list, dict, boolean, or None
try:
v = literal_eval(v)
# The following two excepts allow v to pass through when it represents a
# string.
#
# Longer explanation:
# The type of v is always a string (before calling literal_eval), but
# sometimes it *represents* a string and other times a data structure, like
# a list. In the case that v represents a string, what we got back from the
# yaml parser is 'foo' *without quotes* (so, not '"foo"'). literal_eval is
# ok with '"foo"', but will raise a ValueError if given 'foo'. In other
# cases, like paths (v = 'foo/bar' and not v = '"foo/bar"'), literal_eval
# will raise a SyntaxError.
except ValueError:
pass
except SyntaxError:
pass
return v
def merge_cfg_from_file(cfg_filename, config):
"""Load a yaml config file and merge it into the global config."""
with open(cfg_filename, 'r') as f:
yaml_cfg = AttrDict(yaml.load(f))
_config = _cls2dict(config)
_merge_a_into_b(yaml_cfg, _config)
_dict2cls(_config, config)
def _cls2dict(config):
output = AttrDict()
for a in dir(config):
value = getattr(config, a)
if not a.startswith("__") and not callable(value):
assert isinstance(value, AttrDict)
output[a] = value
return output
from tools.collections import AttrDict __C = AttrDict() cfg = __C __C.net_type = 'mbv2' # mbv2 / res __C.net_config = """[[16, 16], 'mbconv_k3_t1', [], 0, 1]| [[16, 24], 'mbconv_k3_t6', [], 0, 2]| [[24, 48], 'mbconv_k7_t6', ['mbconv_k3_t3'], 1, 2]| [[48, 72], 'mbconv_k5_t6', ['mbconv_k3_t6', 'mbconv_k3_t3'], 2, 2]| [[72, 128], 'mbconv_k3_t6', ['mbconv_k3_t3', 'mbconv_k3_t3'], 2, 1]| [[128, 160], 'mbconv_k3_t6', ['mbconv_k7_t3', 'mbconv_k5_t6', 'mbconv_k7_t3'], 3, 2]| [[160, 176], 'mbconv_k3_t3', ['mbconv_k3_t6', 'mbconv_k3_t6', 'mbconv_k3_t6'], 3, 1]| [[176, 384], 'mbconv_k7_t6', [], 0, 1]| [[384, 1984], 'conv1_1']""" __C.train_params = AttrDict() __C.train_params.epochs = 240 __C.train_params.use_seed = True __C.train_params.seed = 0 __C.optim = AttrDict() __C.optim.init_lr = 0.5 __C.optim.min_lr = 1e-5 __C.optim.lr_schedule = 'cosine' # cosine poly __C.optim.momentum = 0.9 __C.optim.weight_decay = 4e-5 __C.optim.use_grad_clip = False __C.optim.grad_clip = 10 __C.optim.label_smooth = True
from tools.collections import AttrDict __C = AttrDict() cfg = __C __C.net_type = 'mbv2' # mbv2 / res __C.net_config = "" __C.train_params = AttrDict() __C.train_params.epochs = 240 __C.train_params.use_seed = False __C.train_params.seed = 0 __C.optim = AttrDict() __C.optim.init_lr = 0.5 __C.optim.min_lr = 1e-5 __C.optim.lr_schedule = 'cosine' # cosine poly __C.optim.momentum = 0.9 __C.optim.weight_decay = 4e-5 __C.optim.use_grad_clip = False __C.optim.grad_clip = 10 __C.optim.label_smooth = True __C.optim.smooth_alpha = 0.1 __C.optim.if_resume = False __C.optim.resume = AttrDict() __C.optim.resume.load_path = '' __C.optim.resume.load_epoch = 0 __C.data = AttrDict()
class Config(object):
# ==================================
MODEL = AttrDict()
# Path to pretrained imagenet model
MODEL.PRETRAIN_IMAGENET_MODEL = os.path.join('datasets/pretrain_model', "resnet50_imagenet.pth")
# Path to pretrained weights file
MODEL.PRETRAIN_COCO_MODEL = os.path.join('datasets/pretrain_model', 'mask_rcnn_coco.pth')
MODEL.INIT_FILE_CHOICE = 'last' # or file (xxx.pth)
MODEL.INIT_MODEL = None # set in 'utils.py'
MODEL.BACKBONE = 'resnet101'
MODEL.BACKBONE_STRIDES = []
MODEL.BACKBONE_SHAPES = []
# ==================================
DATASET = AttrDict()
# Number of classification classes (including background)
DATASET.NUM_CLASSES = 81
DATASET.YEAR = '2014'
DATASET.PATH = 'datasets/coco'
# ==================================
RPN = AttrDict()
# Length of square anchor side in pixels
RPN.ANCHOR_SCALES = (32, 64, 128, 256, 512)
# Ratios of anchors at each cell (width/height)
# A value of 1 represents a square anchor, and 0.5 is a wide anchor
RPN.ANCHOR_RATIOS = [0.5, 1, 2]
# Anchor stride
# If 1 then anchors are created for each cell in the backbone feature map.
# If 2, then anchors are created for every other cell, and so on (stride=2,3,4...).
RPN.ANCHOR_STRIDE = 1
# Non-max suppression threshold to filter RPN proposals.
# You can reduce this during training to generate more proposals.
RPN.NMS_THRESHOLD = 0.7
# How many anchors per image to use for RPN training
RPN.TRAIN_ANCHORS_PER_IMAGE = 256
# ROIs kept after non-maximum suppression for RPN part
RPN.PRE_NMS_LIMIT = 6000
RPN.POST_NMS_ROIS_TRAINING = 2000
RPN.POST_NMS_ROIS_INFERENCE = 1000
RPN.TARGET_POS_THRES = .7
RPN.TARGET_NEG_THRES = .3
# ==================================
MRCNN = AttrDict()
# If enabled, resize instance masks to a smaller size to reduce
# memory load. Recommended when using high-resolution images.
MRCNN.USE_MINI_MASK = True
MRCNN.MINI_MASK_SHAPE = (56, 56) # (height, width) of the mini-mask
# Pooled ROIs
MRCNN.POOL_SIZE = 7 # cls/bbox stream
MRCNN.MASK_POOL_SIZE = 14 # mask stream
MRCNN.MASK_SHAPE = [28, 28]
# ==================================
DATA = AttrDict()
# Input image resize
# Images are resized such that the smallest side is >= IMAGE_MIN_DIM and
# the longest side is <= IMAGE_MAX_DIM. In case both conditions can't
# be satisfied together the IMAGE_MAX_DIM is enforced.
DATA.IMAGE_MIN_DIM = 800
DATA.IMAGE_MAX_DIM = 1024
# If True, pad images with zeros such that they're (max_dim by max_dim)
DATA.IMAGE_PADDING = True # currently, the False option is not supported
# Image mean (RGB)
DATA.MEAN_PIXEL = np.array([123.7, 116.8, 103.9])
# Maximum number of ground truth instances to use in one image
DATA.MAX_GT_INSTANCES = 100
# Bounding box refinement standard deviation for RPN and final detections.
DATA.BBOX_STD_DEV = np.array([0.1, 0.1, 0.2, 0.2])
DATA.IMAGE_SHAPE = []
# Quote from "roytseng-tw/Detectron.pytorch":
# Number of Python threads to use for the data loader (warning: using too many
# threads can cause GIL-based interference with Python Ops leading to *slower*
# training; 4 seems to be the sweet spot in our experience)
DATA.LOADER_WORKER_NUM = 2
# ==================================
ROIS = AttrDict()
# Number of ROIs per image to feed to classifier/mask heads
# The Mask RCNN paper uses 512 but often the RPN doesn't generate
# enough positive proposals to fill this and keep a positive:negative
# ratio of 1:3. You can increase the number of proposals by adjusting the RPN NMS threshold.
ROIS.TRAIN_ROIS_PER_IMAGE = 200
# Percent of positive ROIs used to train classifier/mask heads
ROIS.ROI_POSITIVE_RATIO = 0.33
# Eqn.(1) in FPN paper
# useless when DEV.ASSIGN_BOX_ON_ALL_SCALE is True
ROIS.ASSIGN_ANCHOR_BASE = 224.
ROIS.METHOD = 'roi_align' # or roi_pool
# ==================================
TEST = AttrDict()
TEST.BATCH_SIZE = 0 # set in _set_value()
# Max number of final detections
TEST.DET_MAX_INSTANCES = 100
# Minimum probability value to accept a detected instance
# ROIs below this threshold are skipped
TEST.DET_MIN_CONFIDENCE = 0
# Non-maximum suppression threshold for detection
TEST.DET_NMS_THRESHOLD = 0.3
TEST.SAVE_IM = False
# ==================================
TRAIN = AttrDict()
TRAIN.BATCH_SIZE = 6
# Learning rate and momentum
# The Mask RCNN paper uses lr=0.02, but on TensorFlow it causes
# weights to explode. Likely due to differences in optimzer implementation.
TRAIN.OPTIM_METHOD = 'sgd'
TRAIN.INIT_LR = 0.01
TRAIN.MOMENTUM = 0.9
# Weight decay regularization
TRAIN.WEIGHT_DECAY = 0.0001
TRAIN.GAMMA = 0.1
TRAIN.LR_POLICY = 'steps_with_decay'
TRAIN.END2END = False
# in epoch
TRAIN.SCHEDULE = [6, 4, 3]
TRAIN.LR_WARM_UP = False
TRAIN.LR_WP_ITER = 500
TRAIN.LR_WP_FACTOR = 1. / 3.
TRAIN.CLIP_GRAD = True
TRAIN.MAX_GRAD_NORM = 5.0
# let bn learn and also apply the same weight decay when setting up optimizer
TRAIN.BN_LEARN = False
# evaluate mAP after each stage
TRAIN.DO_VALIDATION = True
TRAIN.SAVE_FREQ_WITHIN_EPOCH = 10
TRAIN.FORCE_START_EPOCH = 0 # when you resume training and change the batch size, this is useful
# apply OT loss in FPN heads
TRAIN.FPN_OT_LOSS = False
TRAIN.FPN_OT_LOSS_FAC = 1.
# ==============================
DEV = AttrDict()
DEV.SWITCH = False
DEV.INIT_BUFFER_WEIGHT = 'scratch' # currently only support this
# set to 1 if use all historic data
DEV.BUFFER_SIZE = 1000
# set to <= 0 if trained from the very first iter
DEV.EFFECT_AFER_EP_PERCENT = 0.
DEV.MULTI_UPSAMPLER = False # does not affect much
# if 1, standard conv
DEV.UPSAMPLE_FAC = 2.
DEV.LOSS_CHOICE = 'l1'
DEV.OT_ONE_DIM_FORM = 'conv' # effective if loss_choice is 'ot'
DEV.LOSS_FAC = 0.5
# compute meta_los of small boxes on an instance or class level
DEV.INST_LOSS = False
DEV.FEAT_BRANCH_POOL_SIZE = 14
# ignore regression loss (only for **DEBUG**);
# doomed if you use it during deployment
DEV.DIS_REG_LOSS = False
# assign anchors on all scales and split anchor based on roi-pooling output size
# if used, then ROIS.ASSIGN_ANCHOR_BASE is inactivated
DEV.ASSIGN_BOX_ON_ALL_SCALE = False
# provide a baseline (no meta_loss) to compare
DEV.BASELINE = False
DEV.BIG_SUPERVISE = False
DEV.BIG_LOSS_CHOICE = 'ce' # default setting (currently only support this)
DEV.BIG_FC_INIT = 'scratch' # or 'coco_pretrain'
DEV.BIG_LOSS_FAC = 1.
DEV.BIG_FC_INIT_LIST = dict()
DEV.STRUCTURE = 'alpha' # 'beta'
DEV.DIS_UPSAMPLER = False
DEV.BIG_FEAT_DETACH = True
# merge compare_feat output into classifier
DEV.CLS_MERGE_FEAT = False
DEV.CLS_MERGE_MANNER = 'simple_add' # 'linear_add'
DEV.CLS_MERGE_FAC = .5
# ==============================
CTRL = AttrDict()
CTRL.CONFIG_NAME = ''
CTRL.PHASE = ''
CTRL.DEBUG = None
# train on minival and test also on minival
CTRL.QUICK_VERIFY = False
CTRL.SHOW_INTERVAL = 50
CTRL.PROFILE_ANALYSIS = False # show time for some pass
# ==============================
TSNE = AttrDict()
TSNE.SKIP_INFERENCE = True # skip the evaluation (compute mAP)
TSNE.A_FEW = False
TSNE.PERPLEXITY = 30
TSNE.METRIC = 'euclidean'
TSNE.N_TOPICS = 2
TSNE.BATCH_SZ = 1024 # 1024 # bigger bs is more sparse
TSNE.TOTAL_EP = 150
TSNE.ELLIPSE = True
TSNE.SAMPLE_CHOICE = 'set1' # for detailed config, see 'def prepare_data()' in tools/tsne/prepare_data.py
TSNE.FIG_FOLDER_SUX = 'debug5' # custom folder name
# ==============================
MISC = AttrDict()
MISC.SEED = 2000
MISC.USE_VISDOM = False
MISC.VIS = AttrDict()
MISC.VIS.PORT = -1 # must be passed from configs on different servers
# the following will be set somewhere else
MISC.LOG_FILE = None
MISC.DET_RESULT_FILE = None
MISC.SAVE_IMAGE_DIR = None
MISC.RESULT_FOLDER = None
MISC.DEVICE_ID = []
MISC.GPU_COUNT = -1
def display(self, log_file, quiet=False):
"""Display *final* configuration values."""
print_log("Configurations:", file=log_file, quiet_termi=quiet)
for a in dir(self):
if not a.startswith("__") and not callable(getattr(self, a)):
value = getattr(self, a)
if isinstance(value, AttrDict):
print_log("{}:".format(a), log_file, quiet_termi=quiet)
for _, key in enumerate(value):
print_log("\t{:30}\t\t{}".format(key, value[key]), log_file, quiet_termi=quiet)
else:
print_log("{}\t{}".format(a, value), log_file, quiet_termi=quiet)
print_log("\n", log_file, quiet_termi=quiet)
def _set_value(self):
"""Set values of computed attributes. Override all previous settings."""
random.seed(self.MISC.SEED)
torch.manual_seed(self.MISC.SEED)
if self.CTRL.QUICK_VERIFY:
self.CTRL.SHOW_INTERVAL = 5
self.TRAIN.SAVE_FREQ_WITHIN_EPOCH = 2
if self.CTRL.DEBUG:
self.CTRL.SHOW_INTERVAL = 1
self.DATA.IMAGE_MIN_DIM = 320
self.DATA.IMAGE_MAX_DIM = 512
self.CTRL.PROFILE_ANALYSIS = False
self.TSNE.A_FEW = True
# set MISC.RESULT_FOLDER, 'results/base_101/train (or inference)/'
self.MISC.RESULT_FOLDER = os.path.join(
'results', self.CTRL.CONFIG_NAME.lower(), self.CTRL.PHASE)
if not os.path.exists(self.MISC.RESULT_FOLDER):
os.makedirs(self.MISC.RESULT_FOLDER)
self.TEST.BATCH_SIZE = 2 * self.TRAIN.BATCH_SIZE
# MUST be left **at the end**
# The strides of each layer of the FPN Pyramid.
if self.MODEL.BACKBONE == 'resnet101':
self.MODEL.BACKBONE_STRIDES = [4, 8, 16, 32, 64]
else:
raise Exception('unknown backbone structure')
# Input image size
self.DATA.IMAGE_SHAPE = np.array(
[self.DATA.IMAGE_MAX_DIM, self.DATA.IMAGE_MAX_DIM, 3])
# Compute backbone size from input image size
self.MODEL.BACKBONE_SHAPES = np.array(
[[int(math.ceil(self.DATA.IMAGE_SHAPE[0] / stride)),
int(math.ceil(self.DATA.IMAGE_SHAPE[1] / stride))]
for stride in self.MODEL.BACKBONE_STRIDES])
if self.MISC.USE_VISDOM:
if self.CTRL.DEBUG:
self.MISC.VIS.PORT = 2042
assert self.MISC.VIS.PORT > 0, 'vis_port not designated!!!'
print('\n[visdom is activated] remember to execute '
'**python -m visdom.server -port={:d}** on server (or pc)!\n'.format(self.MISC.VIS.PORT))
self.MISC.VIS.LINE = 100
self.MISC.VIS.TXT = 200
self.MISC.VIS.IMG = 300
self.MISC.VIS.LOSS_LEGEND = [
'total_loss', 'rpn_cls', 'rpn_bbox',
'mrcnn_cls', 'mrcnn_bbox', 'mrcnn_mask_loss']
if self.DEV.SWITCH and not self.DEV.BASELINE:
self.MISC.VIS.LOSS_LEGEND.append('meta_loss')
if self.DEV.SWITCH and self.DEV.BIG_SUPERVISE:
self.MISC.VIS.LOSS_LEGEND.append('big_loss')
if self.TRAIN.FPN_OT_LOSS:
self.MISC.VIS.LOSS_LEGEND.append('fpn_ot_loss')
if self.MISC.GPU_COUNT == 8:
self.DATA.LOADER_WORKER_NUM = 32
elif self.MISC.GPU_COUNT == 4:
self.DATA.LOADER_WORKER_NUM = 16
if self.DEV.BIG_FC_INIT == 'coco_pretrain':
self.DEV.BIG_FC_INIT_LIST = {
# target network vs pretrain network
'dev_roi.big_fc_layer.weight': 'classifier.linear_class.weight',
'dev_roi.big_fc_layer.bias': 'classifier.linear_class.bias',
}
# TODO (low): add more here; delete some config for brevity
if not self.TRAIN.LR_WARM_UP:
del self.TRAIN['LR_WP_ITER']
del self.TRAIN['LR_WP_FACTOR']
if not self.DEV.BIG_SUPERVISE:
del self.DEV['BIG_LOSS_FAC']
del self.DEV['BIG_FC_INIT']
del self.DEV['BIG_LOSS_CHOICE']
del self.DEV['BIG_FC_INIT_LIST']
if self.DEV.LOSS_CHOICE != 'ot':
del self.DEV['OT_ONE_DIM_FORM']
from tools.collections import AttrDict __C = AttrDict() cfg = __C __C.net_config = """[[16, 16], 'mbconv_k3_t1', [], 0, 1]| [[16, 24], 'mbconv_k5_t3', ['mbconv_k5_t3', 'mbconv_k3_t3'], 2, 2]| [[24, 48], 'mbconv_k5_t6', [], 0, 2]| [[48, 80], 'mbconv_k5_t6', ['mbconv_k7_t3', 'mbconv_k5_t3', 'mbconv_k3_t3'], 3, 2]| [[80, 112], 'mbconv_k3_t3', ['mbconv_k3_t3'], 1, 1]| [[112, 160], 'mbconv_k7_t6', ['mbconv_k7_t3', 'mbconv_k7_t3'], 2, 2]| [[160, 352], 'mbconv_k5_t6', ['mbconv_k3_t3'], 1, 1]| [[352, 416], 'mbconv_k3_t3', [], 0, 1]| [[416, 480], 'mbconv_k3_t3', [], 0, 1]""" __C.train_params = AttrDict() __C.train_params.batch_size = 256 __C.train_params.num_workers = 8 __C.optim = AttrDict() __C.optim.last_dim = 1728 __C.optim.init_dim = 16 __C.optim.bn_momentum = 0.1 __C.optim.bn_eps = 0.001 __C.data = AttrDict() __C.data.dataset = 'imagenet' # cifar10 imagenet
from tools.collections import AttrDict __C = AttrDict() cfg = __C __C.net_config = "" __C.train_params = AttrDict() __C.train_params.batch_size = 256 __C.train_params.num_workers = 8 __C.optim = AttrDict() __C.optim.init_dim = 16 __C.optim.bn_momentum = 0.1 __C.optim.bn_eps = 0.001 __C.data = AttrDict() __C.data.dataset = 'imagenet' __C.data.train_data_type = 'lmdb' __C.data.val_data_type = 'img' __C.data.patch_dataset = False __C.data.num_examples = 1281167 __C.data.input_size = (3, 224, 224)
from tools.collections import AttrDict __C = AttrDict() cfg = __C __C.train_params = AttrDict() __C.train_params.epochs = 300 __C.train_params.use_seed = False __C.train_params.seed = 0 __C.optim = AttrDict() __C.optim.init_lr = 0.025 __C.optim.min_lr = 1e-5 __C.optim.lr_schedule = 'cosine' # cosine poly __C.optim.momentum = 0.9 __C.optim.weight_decay = 4e-5 __C.optim.use_grad_clip = False __C.optim.grad_clip = 10 __C.optim.label_smooth = False # False, True __C.optim.smooth_alpha = 0.25 __C.optim.auxiliary = True __C.optim.auxiliary_weight = 0.4 __C.optim.if_resume = False __C.optim.resume = AttrDict() __C.optim.resume.load_path = '' __C.optim.resume.load_epoch = 0
from tools.collections import AttrDict __C = AttrDict() cfg = __C __C.net_type = 'resnet' # choose resnet or mobilenet __C.train_params = AttrDict() __C.train_params.epochs = 100 __C.train_params.use_seed = False __C.train_params.seed = 0 __C.train_params.print_freq = 50 __C.optim = AttrDict() __C.optim.init_lr = 0.1 __C.optim.min_lr = 0 __C.optim.momentum = 0.9 __C.optim.weight_decay = 1e-4 __C.optim.use_grad_clip = False __C.optim.grad_clip = 10 __C.optim.label_smooth = False __C.optim.smooth_alpha = 0.1 __C.optim.if_resume = False __C.optim.resume_path = '' __C.data = AttrDict() __C.data.data_path = 'PATH/to/DataSet' __C.data.num_workers = 32 __C.data.batch_size = 256
from tools.collections import AttrDict __C = AttrDict() cfg = __C __C.net_config = "" __C.data = AttrDict() __C.data.seed = 0 __C.data.num_workers = 16 __C.data.batch_size = 256 __C.data.dataset = 'imagenet' __C.data.train_data_type = 'img' __C.data.val_data_type = 'img' __C.data.patch_dataset = False __C.data.num_examples = 1281167 __C.data.input_size = (3, 224, 224)
from tools.collections import AttrDict
__C = AttrDict()
search_cfg = __C
__C.search_params = AttrDict()
__C.search_params.arch_update_epoch = 10
__C.search_params.val_start_epoch = 120
__C.search_params.sample_policy = 'prob' # prob uniform
__C.search_params.weight_sample_num = 1
__C.search_params.softmax_temp = 1.
__C.search_params.adjoin_connect_nums = []
__C.search_params.net_scale = AttrDict()
__C.search_params.net_scale.chs = []
__C.search_params.net_scale.fm_sizes = []
__C.search_params.net_scale.stage = []
__C.search_params.net_scale.num_layers = []
__C.search_params.PRIMITIVES_stack = [
'mbconv_k3_t3',
'mbconv_k3_t6',
'mbconv_k5_t3',
'mbconv_k5_t6',
'mbconv_k7_t3',
'mbconv_k7_t6',
'skip_connect',
]
__C.search_params.PRIMITIVES_head = [
'mbconv_k3_t3',
'mbconv_k3_t6',
from tools.collections import AttrDict __C = AttrDict() cfg = __C __C.train_params = AttrDict() __C.train_params.epochs = 250 __C.train_params.use_seed = False __C.train_params.seed = 0 __C.optim = AttrDict() __C.optim.init_lr = 0.5 # 0.5, 0.1 __C.optim.min_lr = 1e-5 __C.optim.lr_schedule = 'linear' # cosine poly, linear __C.optim.momentum = 0.9 __C.optim.weight_decay = 3e-5 __C.optim.use_grad_clip = False __C.optim.grad_clip = 5 __C.optim.label_smooth = True __C.optim.smooth_alpha = 0.1 __C.optim.auxiliary = True __C.optim.auxiliary_weight = 0.4 __C.optim.if_resume = False __C.optim.resume = AttrDict() __C.optim.resume.load_path = '' __C.optim.resume.load_epoch = 0