paddlex 2.0.0rc4__py3-none-any.whl → 3.0.0b2__py3-none-any.whl

paddlex/ppdet/modeling/architectures/fcos.py

@@ -1,105 +0,0 @@
-# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import paddle
-from paddlex.ppdet.core.workspace import register, create
-from .meta_arch import BaseArch
-
-__all__ = ['FCOS']
-
-
-@register
-class FCOS(BaseArch):
-    """
-    FCOS network, see https://arxiv.org/abs/1904.01355
-
-    Args:
-        backbone (object): backbone instance
-        neck (object): 'FPN' instance
-        fcos_head (object): 'FCOSHead' instance
-        post_process (object): 'FCOSPostProcess' instance
-    """
-
-    __category__ = 'architecture'
-    __inject__ = ['fcos_post_process']
-
-    def __init__(self,
-                 backbone,
-                 neck,
-                 fcos_head='FCOSHead',
-                 fcos_post_process='FCOSPostProcess'):
-        super(FCOS, self).__init__()
-        self.backbone = backbone
-        self.neck = neck
-        self.fcos_head = fcos_head
-        self.fcos_post_process = fcos_post_process
-
-    @classmethod
-    def from_config(cls, cfg, *args, **kwargs):
-        backbone = create(cfg['backbone'])
-
-        kwargs = {'input_shape': backbone.out_shape}
-        neck = create(cfg['neck'], **kwargs)
-
-        kwargs = {'input_shape': neck.out_shape}
-        fcos_head = create(cfg['fcos_head'], **kwargs)
-
-        return {
-            'backbone': backbone,
-            'neck': neck,
-            "fcos_head": fcos_head,
-        }
-
-    def _forward(self):
-        body_feats = self.backbone(self.inputs)
-        fpn_feats = self.neck(body_feats)
-        fcos_head_outs = self.fcos_head(fpn_feats, self.training)
-        if not self.training:
-            scale_factor = self.inputs['scale_factor']
-            bboxes = self.fcos_post_process(fcos_head_outs, scale_factor)
-            return bboxes
-        else:
-            return fcos_head_outs
-
-    def get_loss(self, ):
-        loss = {}
-        tag_labels, tag_bboxes, tag_centerness = [], [], []
-        for i in range(len(self.fcos_head.fpn_stride)):
-            # labels, reg_target, centerness
-            k_lbl = 'labels{}'.format(i)
-            if k_lbl in self.inputs:
-                tag_labels.append(self.inputs[k_lbl])
-            k_box = 'reg_target{}'.format(i)
-            if k_box in self.inputs:
-                tag_bboxes.append(self.inputs[k_box])
-            k_ctn = 'centerness{}'.format(i)
-            if k_ctn in self.inputs:
-                tag_centerness.append(self.inputs[k_ctn])
-
-        fcos_head_outs = self._forward()
-        loss_fcos = self.fcos_head.get_loss(fcos_head_outs, tag_labels,
-                                            tag_bboxes, tag_centerness)
-        loss.update(loss_fcos)
-        total_loss = paddle.add_n(list(loss.values()))
-        loss.update({'loss': total_loss})
-        return loss
-
-    def get_pred(self):
-        bbox_pred, bbox_num = self._forward()
-        output = {'bbox': bbox_pred, 'bbox_num': bbox_num}
-        return output

paddlex/ppdet/modeling/architectures/jde.py

@@ -1,125 +0,0 @@
-# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import paddle
-from paddlex.ppdet.modeling.mot.utils import scale_coords
-from paddlex.ppdet.core.workspace import register, create
-from .meta_arch import BaseArch
-
-__all__ = ['JDE']
-
-
-@register
-class JDE(BaseArch):
-    __category__ = 'architecture'
-    __shared__ = ['metric']
-    """
-    JDE network, see https://arxiv.org/abs/1909.12605v1
-
-    Args:
-        detector (object): detector model instance
-        reid (object): reid model instance
-        tracker (object): tracker instance
-        metric (str): 'MOTDet' for training and detection evaluation, 'ReID'
-            for ReID embedding evaluation, or 'MOT' for multi object tracking
-            evaluation。
-    """
-
-    def __init__(self,
-                 detector='YOLOv3',
-                 reid='JDEEmbeddingHead',
-                 tracker='JDETracker',
-                 metric='MOT'):
-        super(JDE, self).__init__()
-        self.detector = detector
-        self.reid = reid
-        self.tracker = tracker
-        self.metric = metric
-
-    @classmethod
-    def from_config(cls, cfg, *args, **kwargs):
-        detector = create(cfg['detector'])
-        kwargs = {'input_shape': detector.neck.out_shape}
-
-        reid = create(cfg['reid'], **kwargs)
-
-        tracker = create(cfg['tracker'])
-
-        return {
-            "detector": detector,
-            "reid": reid,
-            "tracker": tracker,
-        }
-
-    def _forward(self):
-        det_outs = self.detector(self.inputs)
-
-        if self.training:
-            emb_feats = det_outs['emb_feats']
-            loss_confs = det_outs['det_losses']['loss_confs']
-            loss_boxes = det_outs['det_losses']['loss_boxes']
-            jde_losses = self.reid(emb_feats, self.inputs, loss_confs,
-                                   loss_boxes)
-            return jde_losses
-        else:
-            if self.metric == 'MOTDet':
-                det_results = {
-                    'bbox': det_outs['bbox'],
-                    'bbox_num': det_outs['bbox_num'],
-                }
-                return det_results
-
-            elif self.metric == 'ReID':
-                emb_feats = det_outs['emb_feats']
-                embs_and_gts = self.reid(emb_feats, self.inputs, test_emb=True)
-                return embs_and_gts
-
-            elif self.metric == 'MOT':
-                emb_feats = det_outs['emb_feats']
-                emb_outs = self.reid(emb_feats, self.inputs)
-
-                boxes_idx = det_outs['boxes_idx']
-                bbox = det_outs['bbox']
-
-                input_shape = self.inputs['image'].shape[2:]
-                im_shape = self.inputs['im_shape']
-                scale_factor = self.inputs['scale_factor']
-
-                bbox[:, 2:] = scale_coords(bbox[:, 2:], input_shape, im_shape,
-                                           scale_factor)
-
-                nms_keep_idx = det_outs['nms_keep_idx']
-
-                pred_dets = paddle.concat((bbox[:, 2:], bbox[:, 1:2]), axis=1)
-
-                emb_valid = paddle.gather_nd(emb_outs, boxes_idx)
-                pred_embs = paddle.gather_nd(emb_valid, nms_keep_idx)
-
-                online_targets = self.tracker.update(pred_dets, pred_embs)
-                return online_targets
-
-            else:
-                raise ValueError(
-                    "Unknown metric {} for multi object tracking.".format(
-                        self.metric))
-
-    def get_loss(self):
-        return self._forward()
-
-    def get_pred(self):
-        return self._forward()

paddlex/ppdet/modeling/architectures/keypoint_hrhrnet.py

@@ -1,286 +0,0 @@
-# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-from scipy.optimize import linear_sum_assignment
-from collections import abc, defaultdict
-import numpy as np
-import paddle
-
-from paddlex.ppdet.core.workspace import register, create, serializable
-from .meta_arch import BaseArch
-from .. import layers as L
-from ..keypoint_utils import transpred
-
-__all__ = ['HigherHRNet']
-
-
-@register
-class HigherHRNet(BaseArch):
-    __category__ = 'architecture'
-
-    def __init__(self,
-                 backbone='HRNet',
-                 hrhrnet_head='HigherHRNetHead',
-                 post_process='HrHRNetPostProcess',
-                 eval_flip=True,
-                 flip_perm=None,
-                 max_num_people=30):
-        """
-        HigherHRNet network, see https://arxiv.org/abs/1908.10357；
-        HigherHRNet+swahr, see https://arxiv.org/abs/2012.15175
-
-        Args:
-            backbone (nn.Layer): backbone instance
-            hrhrnet_head (nn.Layer): keypoint_head instance
-            bbox_post_process (object): `BBoxPostProcess` instance
-        """
-        super(HigherHRNet, self).__init__()
-        self.backbone = backbone
-        self.hrhrnet_head = hrhrnet_head
-        self.post_process = post_process
-        self.flip = eval_flip
-        self.flip_perm = paddle.to_tensor(flip_perm)
-        self.deploy = False
-        self.interpolate = L.Upsample(2, mode='bilinear')
-        self.pool = L.MaxPool(5, 1, 2)
-        self.max_num_people = max_num_people
-
-    @classmethod
-    def from_config(cls, cfg, *args, **kwargs):
-        # backbone
-        backbone = create(cfg['backbone'])
-        # head
-        kwargs = {'input_shape': backbone.out_shape}
-        hrhrnet_head = create(cfg['hrhrnet_head'], **kwargs)
-        post_process = create(cfg['post_process'])
-
-        return {
-            'backbone': backbone,
-            "hrhrnet_head": hrhrnet_head,
-            "post_process": post_process,
-        }
-
-    def _forward(self):
-        if self.flip and not self.training and not self.deploy:
-            self.inputs['image'] = paddle.concat(
-                (self.inputs['image'], paddle.flip(self.inputs['image'], [3])))
-        body_feats = self.backbone(self.inputs)
-
-        if self.training:
-            return self.hrhrnet_head(body_feats, self.inputs)
-        else:
-            outputs = self.hrhrnet_head(body_feats)
-
-            if self.flip and not self.deploy:
-                outputs = [paddle.split(o, 2) for o in outputs]
-                output_rflip = [
-                    paddle.flip(paddle.gather(o[1], self.flip_perm, 1), [3])
-                    for o in outputs
-                ]
-                output1 = [o[0] for o in outputs]
-                heatmap = (output1[0] + output_rflip[0]) / 2.
-                tagmaps = [output1[1], output_rflip[1]]
-                outputs = [heatmap] + tagmaps
-            outputs = self.get_topk(outputs)
-
-            if self.deploy:
-                return outputs
-
-            res_lst = []
-            h = self.inputs['im_shape'][0, 0].numpy().item()
-            w = self.inputs['im_shape'][0, 1].numpy().item()
-            kpts, scores = self.post_process(*outputs, h, w)
-            res_lst.append([kpts, scores])
-            return res_lst
-
-    def get_loss(self):
-        return self._forward()
-
-    def get_pred(self):
-        outputs = {}
-        res_lst = self._forward()
-        outputs['keypoint'] = res_lst
-        return outputs
-
-    def get_topk(self, outputs):
-        # resize to image size
-        outputs = [self.interpolate(x) for x in outputs]
-        if len(outputs) == 3:
-            tagmap = paddle.concat(
-                (outputs[1].unsqueeze(4), outputs[2].unsqueeze(4)), axis=4)
-        else:
-            tagmap = outputs[1].unsqueeze(4)
-
-        heatmap = outputs[0]
-        N, J = 1, self.hrhrnet_head.num_joints
-        heatmap_maxpool = self.pool(heatmap)
-        # topk
-        maxmap = heatmap * (heatmap == heatmap_maxpool)
-        maxmap = maxmap.reshape([N, J, -1])
-        heat_k, inds_k = maxmap.topk(self.max_num_people, axis=2)
-
-        outputs = [heatmap, tagmap, heat_k, inds_k]
-        return outputs
-
-
-@register
-@serializable
-class HrHRNetPostProcess(object):
-    '''
-    HrHRNet postprocess contain:
-        1) get topk keypoints in the output heatmap
-        2) sample the tagmap's value corresponding to each of the topk coordinate
-        3) match different joints to combine to some people with Hungary algorithm
-        4) adjust the coordinate by +-0.25 to decrease error std
-        5) salvage missing joints by check positivity of heatmap - tagdiff_norm
-    Args:
-        max_num_people (int): max number of people support in postprocess
-        heat_thresh (float): value of topk below this threshhold will be ignored
-        tag_thresh (float): coord's value sampled in tagmap below this threshold belong to same people for init
-
-        inputs(list[heatmap]): the output list of modle, [heatmap, heatmap_maxpool, tagmap], heatmap_maxpool used to get topk
-        original_height, original_width (float): the original image size
-    '''
-
-    def __init__(self, max_num_people=30, heat_thresh=0.1, tag_thresh=1.):
-        self.max_num_people = max_num_people
-        self.heat_thresh = heat_thresh
-        self.tag_thresh = tag_thresh
-
-    def lerp(self, j, y, x, heatmap):
-        H, W = heatmap.shape[-2:]
-        left = np.clip(x - 1, 0, W - 1)
-        right = np.clip(x + 1, 0, W - 1)
-        up = np.clip(y - 1, 0, H - 1)
-        down = np.clip(y + 1, 0, H - 1)
-        offset_y = np.where(heatmap[j, down, x] > heatmap[j, up, x], 0.25,
-                            -0.25)
-        offset_x = np.where(heatmap[j, y, right] > heatmap[j, y, left], 0.25,
-                            -0.25)
-        return offset_y + 0.5, offset_x + 0.5
-
-    def __call__(self, heatmap, tagmap, heat_k, inds_k, original_height,
-                 original_width):
-
-        N, J, H, W = heatmap.shape
-        assert N == 1, "only support batch size 1"
-        heatmap = heatmap[0].cpu().detach().numpy()
-        tagmap = tagmap[0].cpu().detach().numpy()
-        heats = heat_k[0].cpu().detach().numpy()
-        inds_np = inds_k[0].cpu().detach().numpy()
-        y = inds_np // W
-        x = inds_np % W
-        tags = tagmap[np.arange(J)[None, :].repeat(self.max_num_people), y.
-                      flatten(), x.flatten()].reshape(J, -1, tagmap.shape[-1])
-        coords = np.stack((y, x), axis=2)
-        # threshold
-        mask = heats > self.heat_thresh
-        # cluster
-        cluster = defaultdict(lambda: {
-            'coords': np.zeros((J, 2), dtype=np.float32),
-            'scores': np.zeros(J, dtype=np.float32),
-            'tags': []
-        })
-        for jid, m in enumerate(mask):
-            num_valid = m.sum()
-            if num_valid == 0:
-                continue
-            valid_inds = np.where(m)[0]
-            valid_tags = tags[jid, m, :]
-            if len(cluster) == 0:  # initialize
-                for i in valid_inds:
-                    tag = tags[jid, i]
-                    key = tag[0]
-                    cluster[key]['tags'].append(tag)
-                    cluster[key]['scores'][jid] = heats[jid, i]
-                    cluster[key]['coords'][jid] = coords[jid, i]
-                continue
-            candidates = list(cluster.keys())[:self.max_num_people]
-            centroids = [
-                np.mean(
-                    cluster[k]['tags'], axis=0) for k in candidates
-            ]
-            num_clusters = len(centroids)
-            # shape is (num_valid, num_clusters, tag_dim)
-            dist = valid_tags[:, None, :] - np.array(centroids)[None, ...]
-            l2_dist = np.linalg.norm(dist, ord=2, axis=2)
-            # modulate dist with heat value, see `use_detection_val`
-            cost = np.round(l2_dist) * 100 - heats[jid, m, None]
-            # pad the cost matrix, otherwise new pose are ignored
-            if num_valid > num_clusters:
-                cost = np.pad(cost, ((0, 0), (0, num_valid - num_clusters)),
-                              constant_values=((0, 0), (0, 1e-10)))
-            rows, cols = linear_sum_assignment(cost)
-            for y, x in zip(rows, cols):
-                tag = tags[jid, y]
-                if y < num_valid and x < num_clusters and \
-                   l2_dist[y, x] < self.tag_thresh:
-                    key = candidates[x]  # merge to cluster
-                else:
-                    key = tag[0]  # initialize new cluster
-                cluster[key]['tags'].append(tag)
-                cluster[key]['scores'][jid] = heats[jid, y]
-                cluster[key]['coords'][jid] = coords[jid, y]
-
-        # shape is [k, J, 2] and [k, J]
-        pose_tags = np.array([cluster[k]['tags'] for k in cluster])
-        pose_coords = np.array([cluster[k]['coords'] for k in cluster])
-        pose_scores = np.array([cluster[k]['scores'] for k in cluster])
-        valid = pose_scores > 0
-
-        pose_kpts = np.zeros((pose_scores.shape[0], J, 3), dtype=np.float32)
-        if valid.sum() == 0:
-            return pose_kpts, pose_kpts
-
-        # refine coords
-        valid_coords = pose_coords[valid].astype(np.int32)
-        y = valid_coords[..., 0].flatten()
-        x = valid_coords[..., 1].flatten()
-        _, j = np.nonzero(valid)
-        offsets = self.lerp(j, y, x, heatmap)
-        pose_coords[valid, 0] += offsets[0]
-        pose_coords[valid, 1] += offsets[1]
-
-        # mean score before salvage
-        mean_score = pose_scores.mean(axis=1)
-        pose_kpts[valid, 2] = pose_scores[valid]
-
-        # salvage missing joints
-        if True:
-            for pid, coords in enumerate(pose_coords):
-                tag_mean = np.array(pose_tags[pid]).mean(axis=0)
-                norm = np.sum((tagmap - tag_mean)**2, axis=3)**0.5
-                score = heatmap - np.round(norm)  # (J, H, W)
-                flat_score = score.reshape(J, -1)
-                max_inds = np.argmax(flat_score, axis=1)
-                max_scores = np.max(flat_score, axis=1)
-                salvage_joints = (pose_scores[pid] == 0) & (max_scores > 0)
-                if salvage_joints.sum() == 0:
-                    continue
-                y = max_inds[salvage_joints] // W
-                x = max_inds[salvage_joints] % W
-                offsets = self.lerp(salvage_joints.nonzero()[0], y, x, heatmap)
-                y = y.astype(np.float32) + offsets[0]
-                x = x.astype(np.float32) + offsets[1]
-                pose_coords[pid][salvage_joints, 0] = y
-                pose_coords[pid][salvage_joints, 1] = x
-                pose_kpts[pid][salvage_joints, 2] = max_scores[salvage_joints]
-        pose_kpts[..., :2] = transpred(pose_coords[..., :2][..., ::-1],
-                                       original_height, original_width,
-                                       min(H, W))
-        return pose_kpts, mean_score