From ef56ff861c51d76ed0c280f8963e3754908f39a0 Mon Sep 17 00:00:00 2001 From: xinyingxinying <63766413+xinyingxinying@users.noreply.github.com> Date: Fri, 17 Apr 2020 20:32:06 +0800 Subject: [PATCH] # add Deform Conv --- ppdet/modeling/ops.py | 298 ++++++++++++++++++++---------------------- 1 file changed, 141 insertions(+), 157 deletions(-) diff --git a/ppdet/modeling/ops.py b/ppdet/modeling/ops.py index 28abd477c..7597e250a 100644 --- a/ppdet/modeling/ops.py +++ b/ppdet/modeling/ops.py @@ -14,8 +14,6 @@ import numpy as np from numbers import Integral -import math -import six from paddle import fluid from paddle.fluid.param_attr import ParamAttr @@ -26,12 +24,140 @@ from ppdet.utils.bbox_utils import bbox_overlaps, box_to_delta __all__ = [ 'AnchorGenerator', 'DropBlock', 'RPNTargetAssign', 'GenerateProposals', 'MultiClassNMS', 'BBoxAssigner', 'MaskAssigner', 'RoIAlign', 'RoIPool', - 'MultiBoxHead', 'SSDLiteMultiBoxHead', 'SSDOutputDecoder', - 'RetinaTargetAssign', 'RetinaOutputDecoder', 'ConvNorm', - 'MultiClassSoftNMS', 'LibraBBoxAssigner' + 'MultiBoxHead', 'SSDOutputDecoder', 'RetinaTargetAssign', + 'RetinaOutputDecoder', 'ConvNorm', 'DeformConvNorm', 'MultiClassSoftNMS', + 'LibraBBoxAssigner' ] +def _conv_offset(input, filter_size, stride, padding, act=None, name=None): + out_channel = filter_size * filter_size * 3 + out = fluid.layers.conv2d( + input, + num_filters=out_channel, + filter_size=filter_size, + stride=stride, + padding=padding, + param_attr=ParamAttr( + initializer=fluid.initializer.Constant(value=0), + name=name + ".w_0"), + bias_attr=ParamAttr( + initializer=fluid.initializer.Constant(value=0), + name=name + ".b_0"), + act=act, + name=name) + return out + + +def DeformConvNorm(input, + num_filters, + filter_size, + stride=1, + groups=1, + norm_decay=0., + norm_type='affine_channel', + norm_groups=32, + dilation=1, + lr_scale=1, + freeze_norm=False, + act=None, + norm_name=None, + initializer=None, + bias_attr=False, + name=None): + if bias_attr: + bias_para = ParamAttr( + name=name + "_bias", + initializer=fluid.initializer.Constant(value=0), + learning_rate=lr_scale * 2) + else: + bias_para = False + offset_mask = _conv_offset( + input=input, + filter_size=filter_size, + stride=stride, + padding=(filter_size - 1) // 2, + act=None, + name=name + "_conv_offset") + offset_channel = filter_size**2 * 2 + mask_channel = filter_size**2 + offset, mask = fluid.layers.split( + input=offset_mask, + num_or_sections=[offset_channel, mask_channel], + dim=1) + mask = fluid.layers.sigmoid(mask) + conv = fluid.layers.deformable_conv( + input=input, + offset=offset, + mask=mask, + num_filters=num_filters, + filter_size=filter_size, + stride=stride, + padding=(filter_size - 1) // 2 * dilation, + dilation=dilation, + groups=groups, + deformable_groups=1, + im2col_step=1, + param_attr=ParamAttr( + name=name + "_weights", + initializer=initializer, + learning_rate=lr_scale), + bias_attr=bias_para, + name=name + ".conv2d.output.1") + + norm_lr = 0. if freeze_norm else 1. + pattr = ParamAttr( + name=norm_name + '_scale', + learning_rate=norm_lr * lr_scale, + regularizer=L2Decay(norm_decay)) + battr = ParamAttr( + name=norm_name + '_offset', + learning_rate=norm_lr * lr_scale, + regularizer=L2Decay(norm_decay)) + + if norm_type in ['bn', 'sync_bn']: + global_stats = True if freeze_norm else False + out = fluid.layers.batch_norm( + input=conv, + act=act, + name=norm_name + '.output.1', + param_attr=pattr, + bias_attr=battr, + moving_mean_name=norm_name + '_mean', + moving_variance_name=norm_name + '_variance', + use_global_stats=global_stats) + scale = fluid.framework._get_var(pattr.name) + bias = fluid.framework._get_var(battr.name) + elif norm_type == 'gn': + out = fluid.layers.group_norm( + input=conv, + act=act, + name=norm_name + '.output.1', + groups=norm_groups, + param_attr=pattr, + bias_attr=battr) + scale = fluid.framework._get_var(pattr.name) + bias = fluid.framework._get_var(battr.name) + elif norm_type == 'affine_channel': + scale = fluid.layers.create_parameter( + shape=[conv.shape[1]], + dtype=conv.dtype, + attr=pattr, + default_initializer=fluid.initializer.Constant(1.)) + bias = fluid.layers.create_parameter( + shape=[conv.shape[1]], + dtype=conv.dtype, + attr=battr, + default_initializer=fluid.initializer.Constant(0.)) + out = fluid.layers.affine_channel( + x=conv, scale=scale, bias=bias, act=act) + + if freeze_norm: + scale.stop_gradient = True + bias.stop_gradient = True + return out + + def ConvNorm(input, num_filters, filter_size, @@ -178,6 +304,16 @@ def DropBlock(input, block_size, keep_prob, is_test): return output +def CreateTensorFromNumpy(numpy_array): + paddle_array = fluid.layers.create_parameter( + attr=ParamAttr(), + shape=numpy_array.shape, + dtype=numpy_array.dtype, + default_initializer=NumpyArrayInitializer(numpy_array)) + paddle_array.stop_gradient = True + return paddle_array + + @register @serializable class AnchorGenerator(object): @@ -560,8 +696,6 @@ class BBoxAssigner(object): @register class LibraBBoxAssigner(object): - __shared__ = ['num_classes'] - def __init__(self, batch_size_per_im=512, fg_fraction=.25, @@ -807,7 +941,6 @@ class LibraBBoxAssigner(object): hs = boxes[:, 3] - boxes[:, 1] + 1 keep = np.where((ws > 0) & (hs > 0))[0] boxes = boxes[keep] - max_overlaps = max_overlaps[keep] fg_inds = np.where(max_overlaps >= fg_thresh)[0] bg_inds = np.where((max_overlaps < bg_thresh_hi) & ( max_overlaps >= bg_thresh_lo))[0] @@ -1078,155 +1211,6 @@ class MultiBoxHead(object): self.pad = pad -@register -@serializable -class SSDLiteMultiBoxHead(object): - def __init__(self, - min_ratio=20, - max_ratio=90, - base_size=300, - min_sizes=None, - max_sizes=None, - aspect_ratios=[[2.], [2., 3.], [2., 3.], [2., 3.], [2., 3.], - [2., 3.]], - steps=None, - offset=0.5, - flip=True, - clip=False, - pad=0, - conv_decay=0.0): - super(SSDLiteMultiBoxHead, self).__init__() - self.min_ratio = min_ratio - self.max_ratio = max_ratio - self.base_size = base_size - self.min_sizes = min_sizes - self.max_sizes = max_sizes - self.aspect_ratios = aspect_ratios - self.steps = steps - self.offset = offset - self.flip = flip - self.pad = pad - self.clip = clip - self.conv_decay = conv_decay - - def _separable_conv(self, input, num_filters, name): - dwconv_param_attr = ParamAttr( - name=name + 'dw_weights', regularizer=L2Decay(self.conv_decay)) - num_filter1 = input.shape[1] - depthwise_conv = fluid.layers.conv2d( - input=input, - num_filters=num_filter1, - filter_size=3, - stride=1, - padding="SAME", - groups=int(num_filter1), - act=None, - use_cudnn=False, - param_attr=dwconv_param_attr, - bias_attr=False) - bn_name = name + '_bn' - bn_param_attr = ParamAttr( - name=bn_name + "_scale", regularizer=L2Decay(0.0)) - bn_bias_attr = ParamAttr( - name=bn_name + "_offset", regularizer=L2Decay(0.0)) - bn = fluid.layers.batch_norm( - input=depthwise_conv, - param_attr=bn_param_attr, - bias_attr=bn_bias_attr, - moving_mean_name=bn_name + '_mean', - moving_variance_name=bn_name + '_variance') - bn = fluid.layers.relu6(bn) - pwconv_param_attr = ParamAttr( - name=name + 'pw_weights', regularizer=L2Decay(self.conv_decay)) - pointwise_conv = fluid.layers.conv2d( - input=bn, - num_filters=num_filters, - filter_size=1, - stride=1, - act=None, - use_cudnn=True, - param_attr=pwconv_param_attr, - bias_attr=False) - return pointwise_conv - - def __call__(self, inputs, image, num_classes): - def _permute_and_reshape(input, last_dim): - trans = fluid.layers.transpose(input, perm=[0, 2, 3, 1]) - compile_shape = [0, -1, last_dim] - return fluid.layers.reshape(trans, shape=compile_shape) - - def _is_list_or_tuple_(data): - return (isinstance(data, list) or isinstance(data, tuple)) - - if self.min_sizes is None and self.max_sizes is None: - num_layer = len(inputs) - self.min_sizes = [] - self.max_sizes = [] - step = int( - math.floor(((self.max_ratio - self.min_ratio)) / (num_layer - 2 - ))) - for ratio in six.moves.range(self.min_ratio, self.max_ratio + 1, - step): - self.min_sizes.append(self.base_size * ratio / 100.) - self.max_sizes.append(self.base_size * (ratio + step) / 100.) - self.min_sizes = [self.base_size * .10] + self.min_sizes - self.max_sizes = [self.base_size * .20] + self.max_sizes - - locs, confs = [], [] - boxes, mvars = [], [] - - for i, input in enumerate(inputs): - min_size = self.min_sizes[i] - max_size = self.max_sizes[i] - if not _is_list_or_tuple_(min_size): - min_size = [min_size] - if not _is_list_or_tuple_(max_size): - max_size = [max_size] - step = [ - self.steps[i] if self.steps else 0.0, self.steps[i] - if self.steps else 0.0 - ] - box, var = fluid.layers.prior_box( - input, - image, - min_sizes=min_size, - max_sizes=max_size, - steps=step, - aspect_ratios=self.aspect_ratios[i], - variance=[0.1, 0.1, 0.2, 0.2], - clip=self.clip, - flip=self.flip, - offset=0.5) - - num_boxes = box.shape[2] - box = fluid.layers.reshape(box, shape=[-1, 4]) - var = fluid.layers.reshape(var, shape=[-1, 4]) - num_loc_output = num_boxes * 4 - num_conf_output = num_boxes * num_classes - # get loc - mbox_loc = self._separable_conv(input, num_loc_output, - "loc_{}".format(i + 1)) - loc = _permute_and_reshape(mbox_loc, 4) - # get conf - mbox_conf = self._separable_conv(input, num_conf_output, - "conf_{}".format(i + 1)) - conf = _permute_and_reshape(mbox_conf, num_classes) - - locs.append(loc) - confs.append(conf) - boxes.append(box) - mvars.append(var) - - ssd_mbox_loc = fluid.layers.concat(locs, axis=1) - ssd_mbox_conf = fluid.layers.concat(confs, axis=1) - prior_boxes = fluid.layers.concat(boxes) - box_vars = fluid.layers.concat(mvars) - - prior_boxes.stop_gradient = True - box_vars.stop_gradient = True - return ssd_mbox_loc, ssd_mbox_conf, prior_boxes, box_vars - - @register @serializable class SSDOutputDecoder(object): -- GitLab