crowdnet基於paddlepaddle

 做人羣密度估計的demo，基本復現了 crowdnet，效果一般，具體細節看論文實現

 
from matplotlib import pyplot as plt
from paddle import fluid
import paddle
import json
import random, sys, time
# from scipy import misc
import numpy as np
import cv2
import scipy.spatial, scipy.ndimage

json_stage1_path = '/home/aistudio/data/cus_data/annotation_train_stage1.json'
json_stage2_path = '/home/aistudio/data/cus_data/annotation_train_stage2.json'

def create_density_map(annos, imshape):
    gt = np.zeros(imshape, dtype='uint8')

    for dot in annos:
        try:
            gt[int(dot[1]), int(dot[0])] = 1
        except IndexError:
            print (dot[1], dot[0], sys.exc_info())

    density = np.zeros(gt.shape, dtype=np.float32)
    gt_count = np.count_nonzero(gt)
    if gt_count == 0:
        return density
    pts = np.array(zip(np.nonzero(gt)[1], np.nonzero(gt)[0]))
    # leafsize = 2048
    # tree = scipy.spatial.KDTree(pts.copy(), leafsize=leafsize)
    # k = gt_count
    # if gt_count>=5:
    #     k = 5
    # distances, locations = tree.query(pts, k=k, eps=20.)

    for i, pt in enumerate(pts):
        pt2d = np.zeros(gt.shape, dtype=np.float32)
        pt2d[pt[1], pt[0]] = 1.
        if gt_count > 1:
            # d_avg = distances[i].sum() / float(k-1)
            # sigma = d_avg * 0.18 /
            # # sigma = distances[i][1] * 1
            # print(sigma)
            sigma=10
        else:
            sigma = np.average(np.array(gt.shape)) / 2. /2.
        density += scipy.ndimage.filters.gaussian_filter(pt2d, sigma, mode='constant')
    return density

with open(json_stage1_path, 'r') as f:
    dict1 = json.loads(f.read())
with open(json_stage2_path, 'r') as f:
    dict2 = json.loads(f.read())

if dict1 is not None:
    l1 = [(key['name'], key['type'], key['annotation'], key['num'], key['id'], key['ignore_region']) for key in
          dict1['annotations']]  # 2859
    print(len(l1))
if dict2 is not None:
    l2 = [(key['name'], key['type'], key['annotation'], key['num'], key['id'], key['ignore_region']) for key in
          dict2['annotations']]  # 2859
    print(len(l2))
l1.extend(l2)  # 6477
print(len(l1))
train_data, test_data = l1[:6000], l1[6000:]  # [xxxx][yyyy]
index_np = np.empty(shape=(len(l1), 2), dtype='|S41')

resize_shape = (256,256)

for i, d in enumerate(l1):
#         print(d[0])
    img = cv2.imread('/home/aistudio/data/cus_data/train/' + d[0].split('/')[-1])
#         print(img)
    if d[5]:
        pts = np.array([[anno['x'], anno['y']] for anno in d[5][0]], dtype=np.int32)
        cv2.fillPoly(img, [pts], (0, 0, 0))

    original_shape = img.shape[:2]
    img = cv2.resize(img, resize_shape).astype('float32') / 255.0
    annos = []
    if d[1] == 'bbox':
        annos = [((anno['x'] + anno['w'] / 2) * resize_shape[1] / original_shape[1],
                  (anno['y'] + anno['h'] / 9) * resize_shape[0] / original_shape[0]) for anno in d[2]]
        # 這裏的x axis=1,y axis=0 , 但是下面的createmap的時候會換過來用，不必擔心，
    elif d[1] == 'dot':
        annos = [(anno['x'] * resize_shape[1] / original_shape[1],
                  anno['y'] * resize_shape[0] / original_shape[0])
                 for anno in d[2]]

    map_img = create_density_map(annos=annos, imshape=resize_shape)
#         ori_img, map_img = random_transform(img, map_img, **random_transform_args)
#         ori_img = np.transpose(img, [2, 0, 1])
#         map_img = map_img[np.newaxis, ...]
    name = d[0].split('/')[-1].split('.')[0]
    index_np[i] = [name, str(d[3])]

    np.savez('/home/aistudio/data/cus_data/all_img/' + name + '.npz',
             ori=img, map=map_img)
    print(i)
print(index_np[0])
np.save('/home/aistudio/data/cus_data/index_np.npy', index_np)




from matplotlib import pyplot as plt
from paddle import fluid
import paddle
import json
import random, sys
# from scipy import misc
import numpy as np
import cv2
import scipy.spatial, scipy.ndimage

json_path = '/home/linhai/PycharmProjects/baidu_star_2018 _train/annotation/annotation_train_stage1.json'
json_stage2_path = '/home/aistudio/data/data675/baidu_star_2018/annotation/annotation_train_stage2.json'
# 2859

import paddle.fluid as fluid
import numpy as np
# from crowdnet import shallow_deconv_net, crowd_deconv_net, shallow_deconv_without_bn, crowd_deconv_without_bn
# from image_reader_annotation import img_reader_generator_with_ignore

config = fluid.CheckpointConfig(
    checkpoint_dir="/tmp/ckpt", max_num_checkpoints=2,
    epoch_interval=2, step_interval=50)

BATCH = 16
use_cuda = 1
resize_shape = (256, 256)




def crowd_deconv_without_bn(img):
    x = img
    # x = fluid.nets.img_conv_group(
    #     input=img,
    #     conv_num_filter=[64] * 2,
    #     conv_filter_size=3,
    #     pool_size=2,
    #     pool_stride=[2, 1],
    #     conv_act='relu',
    #     conv_with_batchnorm=True,
    #     conv_batchnorm_drop_rate=[0.0, 0.25],
    #     pool_type='max',
    # )
    x = fluid.layers.conv2d(input=x, num_filters=64, filter_size=3, padding=1, act='relu')  # padding='same')
    # print(x.shape)
    # x = fluid.layers.batch_norm(input=x, act='relu')
    x = fluid.layers.conv2d(input=x, num_filters=64, filter_size=3, padding=1, act='relu')  # padding='same')
    # print(x.shape)
    # x = fluid.layers.batch_norm(input=x, act='relu')
    x = fluid.layers.pool2d(input=x, pool_size=2, pool_stride=2)  # padding='same')
    print(x.shape)
    x = fluid.layers.dropout(x=x, dropout_prob=0.25)


    x = fluid.layers.conv2d(input=x, num_filters=128, filter_size=3, padding=1, act='relu')  # padding='same')
    # print(x.shape)
    # x = fluid.layers.batch_norm(input=x, act='relu')
    x = fluid.layers.conv2d(input=x, num_filters=128, filter_size=3, padding=1, act='relu')  # padding='same')
    # print(x.shape)
    # x = fluid.layers.batch_norm(input=x, act='relu')
    x = fluid.layers.pool2d(input=x, pool_size=2, pool_stride=2)  # padding='same')
    print(x.shape)
    x = fluid.layers.dropout(x=x, dropout_prob=0.25)

    x = fluid.layers.conv2d(input=x, num_filters=256, filter_size=3, padding=1, act='relu')  # padding='same')
    # x = fluid.layers.batch_norm(input=x, act='relu')
    x = fluid.layers.conv2d(input=x, num_filters=256, filter_size=3, padding=1, act='relu')  # padding='same')
    # x = fluid.layers.batch_norm(input=x, act='relu')
    x = fluid.layers.conv2d(input=x, num_filters=256, filter_size=3, padding=1, act='relu')  # padding='same')
    # x = fluid.layers.batch_norm(input=x, act='relu')
    print(x.shape)
    x = fluid.layers.pool2d(input=x, pool_size=2, pool_stride=2)  # padding='same')
    x = fluid.layers.dropout(x=x, dropout_prob=0.5)
    print(x.shape)
    x = fluid.layers.conv2d(input=x, num_filters=512, filter_size=3, padding=1, act='relu')  # padding='same')
    # x = fluid.layers.batch_norm(input=x, act='relu')
    x = fluid.layers.conv2d(input=x, num_filters=512, filter_size=3, padding=1, act='relu')  # padding='same')
    # x = fluid.layers.batch_norm(input=x, act='relu')
    x = fluid.layers.conv2d(input=x, num_filters=512, filter_size=3, padding=1,act='relu' )  # padding='same')
#     x = fluid.layers.batch_norm(input=x, act='relu')
    x = fluid.layers.pool2d(input=x, pool_size=3, pool_stride=1, pool_padding=1)  # padding='same')
    x = fluid.layers.dropout(x=x, dropout_prob=0.5)

    x = fluid.layers.conv2d(input=x, num_filters=512, filter_size=3, padding=1, act='relu')  # padding='same')
    # x = fluid.layers.batch_norm(input=x, act='relu')
    x = fluid.layers.conv2d(input=x, num_filters=512, filter_size=3, padding=1, act='relu')  # padding='same')
    # x = fluid.layers.batch_norm(input=x, act='relu')
    x = fluid.layers.conv2d(input=x, num_filters=512, filter_size=3, padding=1)  # padding='same')
    x = fluid.layers.batch_norm(input=x, act='relu')
    # x = fluid.layers.pool2d(input=x, pool_size=3, pool_stride=2, pool_padding=1)  # padding='same')
    x = fluid.layers.dropout(x=x, dropout_prob=0.5)
    print('clowd_net output shape:',x.shape)

    return x


def shallow_deconv_without_bn(img, kernel_size=(5, 5)):
    x = img
    x = fluid.layers.conv2d(input=x, num_filters=24, filter_size=kernel_size, padding=2, act='relu')
    # print(x.shape )
    # x = fluid.layers.batch_norm(input=x, act='relu')
    x = fluid.layers.pool2d(input=x, pool_type='avg', pool_size=5, pool_stride=2, pool_padding=2)  # padding='same')
    # print(x.shape )

    x = fluid.layers.conv2d(input=x, num_filters=24, filter_size=kernel_size, padding=2, act='relu')
    # x = fluid.layers.batch_norm(input=x, act='relu')
    x = fluid.layers.pool2d(input=x, pool_type='avg', pool_size=5, pool_stride=2, pool_padding=2)  # padding='same')
    # print(x.shape )

    x = fluid.layers.conv2d(input=x, num_filters=24, filter_size=kernel_size, padding=2, )
    x = fluid.layers.batch_norm(input=x, act='relu')
    x = fluid.layers.pool2d(input=x, pool_type='avg', pool_size=5, pool_stride=2, pool_padding=2)  # padding='same')
    print('shallow_net output shape:',x.shape)
    return x
def img_reader_generator(batch=BATCH, resize_shape=resize_shape, is_train=True):
    index = np.load('/home/aistudio/data/cus_data/index_np.npy')
    l1, l2 = index[:6000], index[6000:]

    def train():
        np.random.shuffle(l1)
        for i in range(6000 / batch):
            # t1 = time.clock()
            # ran = np.random.randint(0, 4, shape=batch)
            data = []
            for x in l1[i * batch:(i + 1) * batch]:  # x[0]=name x[1] =num
                x1 = np.load('/home/aistudio/data/cus_data/all_img/' + x[0] + '.npz')
                ori_img, map_img = x1['ori'], x1['map']
                if random.random() > .5:
                    ori_img = ori_img[:,::-1]
                    map_img = map_img[:,::-1]
                ori_img =np.transpose(ori_img, [2, 0, 1])
                map_img = map_img[np.newaxis, ...]
                data.append((ori_img, map_img, np.float(x[1])))
            # print('time', time.clock() - t1)
            yield data

    def test():
        for i in range(477 / batch):
            data = []
            temp_list = l2[i * batch:(i + 1) * batch]
            for x in temp_list:  # x[0]=name x[1] =num
                x1 = np.load('/home/aistudio/data/cus_data/all_img/' + x[0] + '.npz')
                ori_img, map_img = x1['ori'], x1['map']
                data.append((ori_img, map_img, np.float(x[1])))

            if (i + 2) * batch > 477:
                for x in l2[(i + 1) * batch:]:
                    x1 = np.load('/home/aistudio/data/cus_data/all_img/' + x[0] + '.npz')
                    ori_img, map_img = x1['ori'], x1['map']
                    data.append((ori_img, map_img, np.float(x[1])))

            yield data

    if is_train:
        return train
    else:
        return test
config = fluid.CheckpointConfig(
    checkpoint_dir="/tmp/ckpt", max_num_checkpoints=2,
    epoch_interval=2, step_interval=50)

BATCH = 16
use_cuda = 1
resize_shape = (256, 256)


json_path2 = '/home/linhai/PycharmProjects/baidu_star_2018 _train/annotation/annotation_train_stage1.json'

save_dirname = 'model/paddle.deconv_beta01_600e.model'
check_point_dirname = 'model/check_point_dirname'

optimizer_func = lambda: fluid.optimizer.Adam(learning_rate=0.0001)


def train_program():
    img_data = fluid.layers.data(name='img_data', shape=[3, 256, 256], dtype='float32')
    img_label = fluid.layers.data(name='img_label', shape=[1, 256, 256], dtype='float32')
    img_num = fluid.layers.data(name='img_num', shape=[1], dtype='float32')
    print(img_label)

    # net_out1 = crowd_deconv_net(img_data)  # !!!!!!!!!!!!!!!!!!!!!!!!!
    # net_out2 = shallow_deconv_net(img_data)
    net_out1 = crowd_deconv_without_bn(img_data)  # !!!!!!!!!!!!!!!!!!!!!!!!!
    net_out2 = shallow_deconv_without_bn(img_data)
    concat_out = fluid.layers.concat([net_out1, net_out2], axis=1)
    concat_out = fluid.layers.batch_norm(input=concat_out, act='relu')
    conv_end = fluid.layers.conv2d(input=concat_out, num_filters=1, filter_size=1)
    print(conv_end.shape)
    map_out = fluid.layers.resize_bilinear(conv_end, out_shape=(256, 256))
    print(map_out.shape)
    cost = fluid.layers.square_error_cost(input=map_out, label=img_label)
    avg_cost1 = fluid.layers.mean(cost)

    sum_ = fluid.layers.reduce_sum(map_out, dim=[1, 2, 3])
    sum_ = fluid.layers.reshape(sum_, [-1, 1])
    loss_ = fluid.layers.abs(fluid.layers.elementwise_sub(sum_, img_num))
    loss = fluid.layers.mean(fluid.layers.elementwise_div(loss_, img_num))

    sum_loss = loss + avg_cost1 * 6e5

    # acc = fluid.layers.accuracy()

    return [sum_loss, loss]


place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()

train_reader = img_reader_generator(batch=BATCH, resize_shape=resize_shape, is_train=True)
test_reader = img_reader_generator(batch=BATCH, resize_shape=resize_shape, is_train=False)

trainer = fluid.Trainer(
    train_func=train_program,
    optimizer_func=optimizer_func,
    place=place,
    checkpoint_config=config)

feed_order = ['img_data', 'img_label', 'img_num']


def event_handler(event):
    if isinstance(event, fluid.EndStepEvent):
        # if event.step %   == 0:
        print("Step:{0}, Epoch:{1} sum_loss:{2} loss:{3}".format(
            event.step, event.epoch, map(np.array, event.metrics)[0][0], event.metrics[1][0]))
    elif isinstance(event, fluid.EndEpochEvent):
        print('Test :')
        avg_cost = trainer.test(
            reader=test_reader,
            feed_order=feed_order,
        )
        # print(avg_cost)
        print('Epoch{0}, Test Loss {1}, \n'.format(event.epoch, avg_cost))
        trainer.save_params(save_dirname)




# fluid.io.load_params(save_dirname)
trainer.train(num_epochs=400,
              event_handler=event_handler,
              reader=train_reader,
              feed_order=feed_order
              )