def get_resnet101_train(args):
from symnet.symbol_resnet import get_resnet_train
if not args.pretrained:
args.pretrained = 'model/resnet-101-0000.params'
if not args.save_prefix:
args.save_prefix = 'model/resnet101'
args.img_pixel_means = (0.0, 0.0, 0.0)
args.img_pixel_stds = (1.0, 1.0, 1.0)
args.net_fixed_params = ['conv0', 'stage1', 'gamma', 'beta']
args.rpn_feat_stride = 16
args.rcnn_feat_stride = 16
args.rcnn_pooled_size = (14, 14)
return get_resnet_train(anchor_scales=args.rpn_anchor_scales,
anchor_ratios=args.rpn_anchor_ratios,
rpn_feature_stride=args.rpn_feat_stride,
rpn_pre_topk=args.rpn_pre_nms_topk,
rpn_post_topk=args.rpn_post_nms_topk,
rpn_nms_thresh=args.rpn_nms_thresh,
rpn_min_size=args.rpn_min_size,
rpn_batch_rois=args.rpn_batch_rois,
num_classes=args.rcnn_num_classes,
rcnn_feature_stride=args.rcnn_feat_stride,
rcnn_pooled_size=args.rcnn_pooled_size,
rcnn_batch_size=args.rcnn_batch_size,
rcnn_batch_rois=args.rcnn_batch_rois,
rcnn_fg_fraction=args.rcnn_fg_fraction,
rcnn_fg_overlap=args.rcnn_fg_overlap,
rcnn_bbox_stds=args.rcnn_bbox_stds,
units=(3, 4, 23, 3),
filter_list=(256, 512, 1024, 2048))
def get_resnet50_train(system_dict):
'''
Internal function: Select resnet50 params
Args:
system_dict (dict): Dictionary of all the parameters selected for training
Returns:
mxnet model: Resnet50 model
'''
from symnet.symbol_resnet import get_resnet_train
return get_resnet_train(
anchor_scales=system_dict["rpn_anchor_scales"],
anchor_ratios=system_dict["rpn_anchor_ratios"],
rpn_feature_stride=system_dict["rpn_feat_stride"],
rpn_pre_topk=system_dict["rpn_pre_nms_topk"],
rpn_post_topk=system_dict["rpn_post_nms_topk"],
rpn_nms_thresh=system_dict["rpn_nms_thresh"],
rpn_min_size=system_dict["rpn_min_size"],
rpn_batch_rois=system_dict["rpn_batch_rois"],
num_classes=system_dict["rcnn_num_classes"],
rcnn_feature_stride=system_dict["rcnn_feat_stride"],
rcnn_pooled_size=system_dict["rcnn_pooled_size"],
rcnn_batch_size=system_dict["rcnn_batch_size"],
rcnn_batch_rois=system_dict["rcnn_batch_rois"],
rcnn_fg_fraction=system_dict["rcnn_fg_fraction"],
rcnn_fg_overlap=system_dict["rcnn_fg_overlap"],
rcnn_bbox_stds=system_dict["rcnn_bbox_stds"],
units=(3, 4, 6, 3),
filter_list=(256, 512, 1024, 2048))
def get_resnet101_train(args, config):
from symnet.symbol_resnet import get_resnet_train
if not args.pretrained:
args.pretrained = 'model/resnet-101-0000.params'
if not args.save_prefix:
args.save_prefix = 'model/resnet101'
return get_resnet_train(
anchor_scales=config.rpn['rpn_anchor_scales'],
anchor_ratios=config.rpn['rpn_anchor_ratios'],
rpn_feature_stride=config.rpn['rpn_feat_stride'],
rpn_pre_topk=config.rpn['rpn_pre_nms_topk'],
rpn_post_topk=config.rpn['rpn_post_nms_topk'],
rpn_nms_thresh=config.rpn['rpn_nms_thresh'],
rpn_min_size=config.rpn['rpn_min_size'],
rpn_batch_rois=args.rpn_batch_rois,
num_classes=config.rcnn['rcnn_num_classes'],
rcnn_feature_stride=config.rcnn['rcnn_feat_stride'],
rcnn_pooled_size=config.rcnn['rcnn_pooled_size'],
rcnn_batch_size=args.rcnn_batch_size,
rcnn_batch_rois=args.rcnn_batch_rois,
rcnn_fg_fraction=config.rcnn['rcnn_fg_fraction'],
rcnn_fg_overlap=config.rcnn['rcnn_fg_overlap'],
rcnn_bbox_stds=config.rcnn['rcnn_bbox_stds'],
units=(3, 4, 23, 3),
filter_list=(256, 512, 1024, 2048),
step=args.step)
def get_resnet50_train(system_dict):
from symnet.symbol_resnet import get_resnet_train
return get_resnet_train(anchor_scales=system_dict["rpn_anchor_scales"], anchor_ratios=system_dict["rpn_anchor_ratios"],
rpn_feature_stride=system_dict["rpn_feat_stride"], rpn_pre_topk=system_dict["rpn_pre_nms_topk"],
rpn_post_topk=system_dict["rpn_post_nms_topk"], rpn_nms_thresh=system_dict["rpn_nms_thresh"],
rpn_min_size=system_dict["rpn_min_size"], rpn_batch_rois=system_dict["rpn_batch_rois"],
num_classes=system_dict["rcnn_num_classes"], rcnn_feature_stride=system_dict["rcnn_feat_stride"],
rcnn_pooled_size=system_dict["rcnn_pooled_size"], rcnn_batch_size=system_dict["rcnn_batch_size"],
rcnn_batch_rois=system_dict["rcnn_batch_rois"], rcnn_fg_fraction=system_dict["rcnn_fg_fraction"],
rcnn_fg_overlap=system_dict["rcnn_fg_overlap"], rcnn_bbox_stds=system_dict["rcnn_bbox_stds"],
units=(3, 4, 6, 3), filter_list=(256, 512, 1024, 2048))
def get_resnet101_train(args):
from symnet.symbol_resnet import get_resnet_train
if not args.pretrained:
args.pretrained = 'model/resnet-101-0000.params'
if not args.save_prefix:
args.save_prefix = 'model/resnet101'
args.img_pixel_means = (0.0, 0.0, 0.0)
args.img_pixel_stds = (1.0, 1.0, 1.0)
args.net_fixed_params = ['conv0', 'stage1', 'gamma', 'beta']
args.rpn_feat_stride = 16
args.rcnn_feat_stride = 16
args.rcnn_pooled_size = (14, 14)
return get_resnet_train(anchor_scales=args.rpn_anchor_scales, anchor_ratios=args.rpn_anchor_ratios,
rpn_feature_stride=args.rpn_feat_stride, rpn_pre_topk=args.rpn_pre_nms_topk,
rpn_post_topk=args.rpn_post_nms_topk, rpn_nms_thresh=args.rpn_nms_thresh,
rpn_min_size=args.rpn_min_size, rpn_batch_rois=args.rpn_batch_rois,
num_classes=args.rcnn_num_classes, rcnn_feature_stride=args.rcnn_feat_stride,
rcnn_pooled_size=args.rcnn_pooled_size, rcnn_batch_size=args.rcnn_batch_size,
rcnn_batch_rois=args.rcnn_batch_rois, rcnn_fg_fraction=args.rcnn_fg_fraction,
rcnn_fg_overlap=args.rcnn_fg_overlap, rcnn_bbox_stds=args.rcnn_bbox_stds,
units=(3, 4, 23, 3), filter_list=(256, 512, 1024, 2048))
