深度学习中Concat层和Flatten层作用

一些其它层的作用参考:

Keras各种layer的作用及用法--简要总结:https://www.jianshu.com/p/86d667ee3c62

Concat层

  • Layer type: Concat
  • 头文件位置: ./include/caffe/layers/concat_layer.hpp
  • CPU 执行源文件位置:./src/caffe/layers/concat_layer.cpp
  • CUDA GPU 执行源文件位置: ./src/caffe/layers/concat_layer.cu

      Concat层的功能:Concat层实现输入数据的拼接。Concat层是一个实用程序层,它将多个输入blob连接到一个输出blob(按照给定的axis,注意除了规定的axis以外,被concat的输入bolb的其他维度的size必须一致)。

参数定义:

参数(ConcatParameter concat_param) 
定义位置: ./src/caffe/proto/caffe.proto

message ConcatParameter {
  // The axis along which to concatenate -- may be negative to index from the
  // end (e.g., -1 for the last axis).  Other axes must have the
  // same dimension for all the bottom blobs.
  // By default, ConcatLayer concatenates blobs along the "channels" axis (1).
  // 指定拼接的维度,默认为1即以channel通道进行拼接;支持负索引,即-1表示最后一个维度
  optional int32 axis = 2 [default = 1];

  // DEPRECATED: alias for "axis" -- does not support negative indexing.
  // 以后会被弃用,作用同axis一样,但不能指定为负数
  optional uint32 concat_dim = 1 [default = 1];
}

      caffe中数据通常为4个维度,即 num×channels×height×width,因此默认值1表示channels通道进行拼接。而0表示num这个维度,即数量上的叠加。

使用方法:

layer {
  name: "data_all"
  type: "Concat"
  bottom: "data_classfier"
  bottom: "data_boundingbox"
  bottom: "data_facialpoints"
  top: "data_all"
  concat_param {
    axis: 0
  }
}

     除了拼接维度外的其它维度都必须相等。比如上面,输入图像均为 24×24×3,用于分类的有150张图片,用于boundingbox回归的有50张,用于关键点回归的也有50张,则最后拼接的结果就是(150+50+50)×3×24×24

Flatten层

作用:Flatten层用来将输入“压平”,即把多维的输入一维化,常用在从卷积层到全连接层的过渡。Flatten不影响batch的大小。

为了更好的理解Flatten层作用,我把这个神经网络进行可视化如下图:

Caption
  • 4
    点赞
  • 28
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
以下是一个基于深度学习的目标检测代码示例,使用的是 TensorFlow 和 Keras 框架。这个代码示例使用的是 Faster R-CNN 模型,可以在 COCO 数据集上进行训练和测试,同还包括了数据增强和模型评估等功能。 ```python import tensorflow as tf from tensorflow import keras from tensorflow.keras import layers from tensorflow.keras import models from tensorflow.keras import optimizers from tensorflow.keras import backend as K from tensorflow.keras.layers import Input from tensorflow.keras.applications import ResNet50 from tensorflow.keras.layers import Conv2D from tensorflow.keras.layers import MaxPooling2D from tensorflow.keras.layers import Flatten from tensorflow.keras.layers import Dense from tensorflow.keras.layers import Dropout from tensorflow.keras.layers import GlobalAveragePooling2D from tensorflow.keras.layers import GlobalMaxPooling2D from tensorflow.keras.layers import TimeDistributed from tensorflow.keras.layers import AveragePooling2D from tensorflow.keras.layers import BatchNormalization from tensorflow.keras.layers import Activation from tensorflow.keras.layers import Add from tensorflow.keras.layers import ZeroPadding2D from tensorflow.keras.layers import Cropping2D from tensorflow.keras.layers import Lambda from tensorflow.keras.layers import Reshape from tensorflow.keras.layers import Concatenate from tensorflow.keras.layers import Softmax from tensorflow.keras.models import Model from tensorflow.keras.callbacks import TensorBoard, ModelCheckpoint, ReduceLROnPlateau, EarlyStopping from tensorflow.keras.utils import plot_model import numpy as np import os import cv2 import time import argparse from tqdm import tqdm from pycocotools.coco import COCO from pycocotools import mask as maskUtils os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' np.random.seed(42) tf.random.set_seed(42) class Config: NAME = "faster_rcnn" BACKBONE = "resnet50" IMAGE_MIN_DIM = 800 IMAGE_MAX_DIM = 1333 RPN_ANCHOR_SCALES = (32, 64, 128, 256, 512) RPN_ANCHOR_RATIOS = [0.5, 1, 2] RPN_ANCHOR_STRIDE = 16 RPN_NMS_THRESHOLD = 0.7 RPN_TRAIN_ANCHORS_PER_IMAGE = 256 RPN_POSITIVE_RATIO = 0.5 DETECTION_MIN_CONFIDENCE = 0.7 DETECTION_NMS_THRESHOLD = 0.3 DETECTION_MAX_INSTANCES = 100 LEARNING_RATE = 0.001 WEIGHT_DECAY = 0.0001 EPOCHS = 50 BATCH_SIZE = 1 STEPS_PER_EPOCH = 1000 VALIDATION_STEPS = 50 IMAGES_PER_GPU = 1 MEAN_PIXEL = np.array([123.7, 116.8, 103.9]) NUM_CLASSES = 81 # COCO has 80 classes + background class DataGenerator(keras.utils.Sequence): def __init__(self, dataset, config, shuffle=True, augment=True): self.dataset = dataset self.config = config self.shuffle = shuffle self.augment = augment self.image_ids = np.copy(self.dataset.image_ids) self.on_epoch_end() def __len__(self): return int(np.ceil(len(self.dataset.image_ids) / self.config.BATCH_SIZE)) def __getitem__(self, idx): batch_image_ids = self.image_ids[idx * self.config.BATCH_SIZE:(idx + 1) * self.config.BATCH_SIZE] batch_images = [] batch_gt_class_ids = [] batch_gt_boxes = [] for image_id in batch_image_ids: image, gt_class_ids, gt_boxes = load_image_gt(self.dataset, self.config, image_id, augment=self.augment) batch_images.append(image) batch_gt_class_ids.append(gt_class_ids) batch_gt_boxes.append(gt_boxes) batch_images = np.array(batch_images) batch_gt_class_ids = np.array(batch_gt_class_ids) batch_gt_boxes = np.array(batch_gt_boxes) rpn_match, rpn_bbox, rois, roi_gt_class_ids, roi_gt_boxes = build_rpn_targets(batch_images.shape, self.config, batch_gt_class_ids, batch_gt_boxes) inputs = [batch_images, batch_gt_class_ids, batch_gt_boxes, rpn_match, rpn_bbox, rois, roi_gt_class_ids, roi_gt_boxes] outputs = [] return inputs, outputs def on_epoch_end(self): if self.shuffle: np.random.shuffle(self.image_ids) def load_image_gt(dataset, config, image_id, augment=True): image = dataset.load_image(image_id) mask, class_ids = dataset.load_mask(image_id) bbox = maskUtils.toBbox(mask) bbox = np.expand_dims(bbox, axis=-1) class_ids = np.expand_dims(class_ids, axis=-1) gt_boxes = np.concatenate([bbox, class_ids], axis=-1) if augment: image, gt_boxes = augment_image(image, gt_boxes) image, window, scale, padding = resize_image(image, min_dim=config.IMAGE_MIN_DIM, max_dim=config.IMAGE_MAX_DIM, padding=True) gt_boxes[:, :4] = resize_box(gt_boxes[:, :4], scale, padding) gt_class_ids = gt_boxes[:, 4] return image.astype(np.float32) - config.MEAN_PIXEL, gt_class_ids.astype(np.int32), gt_boxes[:, :4].astype(np.float32) def augment_image(image, gt_boxes): if np.random.rand() < 0.5: image = np.fliplr(image) gt_boxes[:, 0] = image.shape[1] - gt_boxes[:, 0] - gt_boxes[:, 2] return image, gt_boxes def resize_image(image, min_dim=None, max_dim=None, padding=False): original_shape = image.shape rows, cols = original_shape[0], original_shape[1] if min_dim: scale = max(1, min_dim / min(rows, cols)) if max_dim: scale = min(scale, max_dim / max(rows, cols)) image = cv2.resize(image, (int(round(cols * scale)), int(round(rows * scale)))) if padding: padded_image = np.zeros((max_dim, max_dim, 3), dtype=np.float32) padded_image[:image.shape[0], :image.shape[1], :] = image window = (0, 0, image.shape[1], image.shape[0]) return padded_image, window, scale, (0, 0, 0, 0) return image, None, scale, None def resize_box(boxes, scale, padding): if padding is not None: boxes[:, 0] += padding[1] # x1 boxes[:, 1] += padding[0] # y1 boxes[:, :4] *= scale return boxes def overlaps(boxes1, boxes2): i_x1 = np.maximum(boxes1[:, 0], boxes2[:, 0]) i_y1 = np.maximum(boxes1[:, 1], boxes2[:, 1]) i_x2 = np.minimum(boxes1[:, 2], boxes2[:, 2]) i_y2 = np.minimum(boxes1[:, 3], boxes2[:, 3]) i_area = np.maximum(i_x2 - i_x1 + 1, 0) * np.maximum(i_y2 - i_y1 + 1, 0) a_area = (boxes1[:, 2] - boxes1[:, 0] + 1) * (boxes1[:, 3] - boxes1[:, 1] + 1) b_area = (boxes2[:, 2] - boxes2[:, 0] + 1) * (boxes2[:, 3] - boxes2[:, 1] + 1) u_area = a_area + b_area - i_area overlaps = i_area / u_area return overlaps def compute_iou(box, boxes, eps=1e-8): iou = overlaps(box[np.newaxis], boxes) return iou def compute_backbone_shapes(config, image_shape): if callable(config.BACKBONE): return config.BACKBONE(image_shape) assert isinstance(config.BACKBONE, str) if config.BACKBONE in ["resnet50", "resnet101"]: if image_shape[0] >= 800: return np.array([[200, 256], [100, 128], [50, 64], [25, 32], [13, 16]]) else: return np.array([[100, 128], [50, 64], [25, 32], [13, 16], [7, 8]]) else: raise ValueError("Invalid backbone name") def generate_anchors(scales, ratios, shape, feature_stride, anchor_stride): scales, ratios = np.meshgrid(np.array(scales), np.array(ratios)) scales, ratios = scales.flatten(), ratios.flatten() heights = scales / np.sqrt(ratios) widths = scales * np.sqrt(ratios) shifts_y = np.arange(0, shape[0], anchor_stride) * feature_stride shifts_x = np.arange(0, shape[1], anchor_stride) * feature_stride shifts_x, shifts_y = np.meshgrid(shifts_x, shifts_y) box_widths, box_centers_x = np.meshgrid(widths, shifts_x) box_heights, box_centers_y = np.meshgrid(heights, shifts_y) box_centers = np.stack([box_centers_y, box_centers_x], axis=2) box_sizes = np.stack([box_heights, box_widths], axis=2) box_centers = np.reshape(box_centers, [-1, 2]) box_sizes = np.reshape(box_sizes, [-1, 2]) boxes = np.concatenate([box_centers - 0.5 * box_sizes, box_centers + 0.5 * box_sizes], axis=1) boxes = np.round(boxes) return boxes def generate_pyramid_anchors(scales, ratios, feature_shapes, feature_strides, anchor_stride): anchors = [] for i in range(len(scales)): anchors.append(generate_anchors(scales[i], ratios, feature_shapes[i], feature_strides[i], anchor_stride)) return np.concatenate(anchors, axis=0) def norm_boxes(boxes, shape): boxes = boxes.astype(np.float32) h, w = shape[:2] scale = np.array([h - 1, w - 1, h - 1, w - 1]) shift = np.array([0, 0, 1, 1]) boxes = np.divide(boxes - shift, scale) boxes = np.maximum(np.minimum(boxes, 1), 0) return boxes def denorm_boxes(boxes, shape): h, w = shape[:2] scale = np.array([h - 1, w - 1, h - 1, w - 1]) shift = np.array([0, 0, 1, 1]) boxes = boxes * scale + shift return boxes.astype(np.int32) def overlaps_graph(boxes1, boxes2): b1 = tf.reshape(tf.tile(tf.expand_dims(boxes1, 1), [1, 1, tf.shape(boxes2)[0]]), [-1, 4]) b2 = tf.tile(boxes2, [tf.shape(boxes1)[0], 1]) b2 = tf.reshape(tf.transpose(b2), [-1, 4]) overlaps = compute_iou(b1, b2) overlaps = tf.reshape(overlaps, [tf.shape(boxes1)[0], tf.shape(boxes2)[0]]) return overlaps def detection_target_graph(proposals, gt_class_ids, gt_boxes, config): proposals = tf.cast(proposals, tf.float32) gt_boxes = tf.cast(gt_boxes, tf.float32) gt_class_ids = tf.cast(gt_class_ids, tf.int64) # Compute overlaps matrix [proposals, gt_boxes] overlaps = overlaps_graph(proposals, gt_boxes) # Compute overlaps with positive anchors roi_iou_max = tf.reduce_max(overlaps, axis=1) positive_roi_bool = (roi_iou_max >= config.RPN_POSITIVE_RATIO) positive_indices = tf.where(positive_roi_bool)[:, 0] # Subsample ROIs. Aim for 33% positive # Positive ROIs positive_count = int(config.RPN_TRAIN_ANCHORS_PER_IMAGE * config.RPN_POSITIVE_RATIO) positive_indices = tf.random.shuffle(positive_indices)[:positive_count] positive_count = tf.shape(positive_indices)[0] # Negative ROIs. Add enough to maintain positive:negative ratio. r = 1.0 / config.RPN_POSITIVE_RATIO negative_count = tf.cast(r * tf.cast(positive_count, tf.float32), tf.int32) - positive_count negative_indices = tf.where(roi_iou_max < config.RPN_POSITIVE_RATIO)[:, 0] negative_count = tf.math.minimum(tf.shape(negative_indices)[0], negative_count) negative_indices = tf.random.shuffle(negative_indices)[:negative_count] # Gather selected ROIs positive_rois = tf.gather(proposals, positive_indices) negative_rois = tf.gather(proposals, negative_indices) # Assign positive ROIs to GT boxes. positive_overlaps = tf.gather(overlaps, positive_indices) roi_gt_box_assignment = tf.cond( tf.greater(tf.shape(positive_overlaps)[1], 0), true_fn=lambda: tf.argmax(positive_overlaps, axis=1), false_fn=lambda: tf.cast(tf.constant([]), tf.int64) ) roi_gt_boxes = tf.gather(gt_boxes, roi_gt_box_assignment) roi_gt_class_ids = tf.gather(gt_class_ids, roi_gt_box_assignment) # Compute bbox refinement for positive ROIs deltas = keras_rcnn.backend.boxutils.bbox_transform(positive_rois, roi_gt_boxes) deltas /= tf.constant(config.BBOX_STD_DEV, dtype=tf.float32) # Append negative ROIs and pad bbox deltas and masks that # are not used for negative ROIs with zeros. rois = tf.concat([positive_rois, negative_rois], axis=0) N = tf.shape(negative_rois)[0] P = tf.math.maximum(config.RPN_TRAIN_ANCHORS_PER_IMAGE - tf.shape(rois)[0], 0) rois = tf.pad(rois, [(0, P), (0, 0)]) roi_gt_boxes = tf.pad(roi_gt_boxes, [(0, N + P), (0, 0)]) roi_gt_class_ids = tf.pad(roi_gt_class_ids, [(0, N + P)]) deltas = tf.pad(deltas, [(0, N + P), (0, 0)]) # Return rois and deltas return rois, roi_gt_class_ids, deltas def build_rpn_targets(image_shape, config, gt_class_ids, gt_boxes): feature_shapes = compute_backbone_shapes(config, image_shape) anchors = generate_pyramid_anchors(config.RPN_ANCHOR_SCALES, config.RPN_ANCHOR_RATIOS, feature_shapes, config.BACKBONE_SHAPES, config.RPN_ANCHOR_STRIDE) rpn_match, rpn_bbox = keras_rcnn.backend.anchor.get_best_anchor(anchors, gt_boxes, config) rpn_match = tf.expand_dims(rpn_match, axis=-1) rpn_bbox = tf.reshape(rpn_bbox, [-1, 4]) rois, roi_gt_class_ids, deltas = tf.py_function(detection_target_graph, [anchors, gt_class_ids, gt_boxes, config], [tf.float32, tf.int64, tf.float32]) rois.set_shape([config.RPN_TRAIN_ANCHORS_PER_IMAGE, 4]) roi_gt_class_ids.set_shape([config.RPN_TRAIN_ANCHORS_PER_IMAGE]) deltas.set_shape([config.RPN_TRAIN_ANCHORS_PER_IMAGE, 4 * config.NUM_CLASSES]) rpn_match.set_shape([None, 1]) rpn_bbox.set_shape([None, 4]) rois = tf.stop_gradient(rois) roi_gt_class_ids = tf.stop_gradient(roi_gt_class_ids) deltas = tf.stop_gradient(deltas) rpn_match = tf.stop_gradient(rpn_match) rpn_bbox = tf.stop_gradient(rpn_bbox) return rpn_match, rpn_bbox, rois, roi_gt_class_ids, deltas def build_rpn_model(config): input_image = Input(shape=[None, None, 3], name="input_image") shared_layers = ResNet50(include_top=False, weights='imagenet', input_tensor=input_image) layer_names = ["conv4_block6_out", "conv5_block3_out", "conv6_relu"] layers = [shared_layers.get_layer(name).output for name in layer_names] output_layers = layers rpn_layers = [] for n, layer in enumerate(output_layers): rpn = Conv2D(512, (3, 3), padding="same", activation="relu", name="rpn_conv%d" % (n + 1))(layer) rpn_class = Conv2D(2 * config.RPN_ANCHOR_SCALES[0], (1, 1), activation="sigmoid", name="rpn_class%d" % (n + 1))(rpn) rpn_bbox = Conv2D(4 * config.RPN_ANCHOR_SCALES[0], (1, 1), activation="linear", name="rpn_bbox%d" % (n + 1))(rpn) rpn_layers.append(rpn_class) rpn_layers.append(rpn_bbox) rpn_class_logits = Concatenate(axis=1, name="rpn_class_logits")(rpn_layers[:len(config.RPN_ANCHOR_SCALES)]) rpn_class = Concatenate(axis=1, name="rpn_class")(rpn_layers[len(config.RPN_ANCHOR_SCALES):]) rpn_bbox = Concatenate(axis=1, name="rpn_bbox")(rpn_layers[len(config.R
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值