bubbliiiing / faster-rcnn-keras Goto Github PK

大佬，请问下，在roi_helper的calc_iou函数中，第102行有一段代码：
labels[label_pos:4+label_pos] = [1, 1, 1, 1]
这里将labels的这几个位置置为1是干嘛用的呢？
第115行，Y2 = np.concatenate([np.array(y_class_regr_label),np.array(y_class_regr_coords)],axis=1)
这样得到的Y2的形状将是[n_rois, 4*(n_classes-1)+4*(n_classes-1)]，就是[n_rois, 8*(n_classes-1)]。
然后在train.py中，第162行，将Y2直接送入model_classifier中进行训练。
loss_class = model_classifier.train_on_batch([X, X2[:, sel_samples, :]], [Y1[:, sel_samples, :], Y2[:, sel_samples, :]])
然而，model_classifier的输出list的第二个元素来自
out_regr = TimeDistributed(Dense(4 * (nb_classes-1), activation='linear', kernel_initializer='zero'), name='dense_regress_{}'.format(nb_classes))(out)，这里的形状是[1, n_rois, 4*(n_classes-1)]。这样，y_true和y_pred维度就不一样了，这样计算loss有问题吗？
我运行下来是没有问题的，但是不太理解这里是为什么。大佬能指点一二吗？

请问一下utils/config.py中self.model_path = "logs/model.h5"，该文件是voc_weight.h5吗？如果不是，能提供一下吗

1.在作者train.py文件的第164行：
write_log(callback, ['detection_cls_loss', 'detection_reg_loss', 'detection_acc'], loss_class, train_step)。

2.我用tensorboard查看log发现detection_acc的图有点像loss的曲线，
所以我打印了检测名称：
print(model_classifier.metrics_names)，
得到的结果是：
['loss', 'dense_class_2_loss', 'dense_regress_2_loss', 'dense_class_2_acc']，
所以我怀疑这里有点问题，请作者确认一下。

keras版本和tensorflow版本？

不知道该怎么改

8/2000 [..............................] - ETA: 2:14:03 - rpn_cls: 2.5315 - rpn_regr: 3.9138 - detector_cls: 1.7752 - detector_regr: 0.4206Traceback (most recent call last):
File "f:/deeplearning-exp/faster-rcnn-keras-master/train.py", line 119, in
X2, Y1, Y2, IouS = calc_iou(R, config, boxes[0], width, height, NUM_CLASSES)
File "f:\deeplearning-exp\faster-rcnn-keras-master\utils\roi_helpers.py", line 115, in calc_iou
Y2 = np.concatenate([np.array(y_class_regr_label),np.array(y_class_regr_coords)],axis=1)
File "<array_function internals>", line 6, in concatenate
numpy.AxisError: axis 1 is out of bounds for array of dimension 1
这一行似乎应该改成Y2 = np.concatenate([np.array(y_class_regr_label),np.array(y_class_regr_coords)],axis=0)，是么？

File "/mistgpu/miniconda/envs/tensorflow-gpu/lib/python3.6/site-packages/keras/engine/base_layer.py", line 793, in compute_output_shape
'method on your layer (%s).' % self.class.name)
NotImplementedError: Please run in eager mode or implement the compute_output_shape method on your layer (BatchNormalization).

请问在代码的哪一块能获取到box的置信度？

ModuleNotFoundError: No module named 'utils.config'; 'utils' is not a package

classification[mask_neg][val_locs] = -1 这一句代码好像不会改变原 classification 的值。

我做了下面的测试

import numpy as np
x = np.array([1,2,3,4,5,6])
print(x)                   # [1 2 3 4 5 6]
mask = x > 2
print(mask)            # [False False  True  True  True  True]
x[mask] = 100
print(x)                   # [  1   2 100 100 100 100]
x[mask][0] = -1
print(x)                   # [  1   2 100 100 100 100]

输出：
[1 2 3 4 5 6]
[False False True True True True]
[ 1 2 100 100 100 100]
[ 1 2 100 100 100 100]

方便说下你这版的tensorflow和keras的版本吗

博主你好，我想问一下这个代码是否可以在training的过程中validation？

我用c++ tensorflow预测到这一步了 但是后面我看python代码进行了2次预测，请问下c++下面怎么处理这个结果呢

训练自己的数据集1000张 7个类别，loss稳定在0.3左右，定位效果不错，为啥分类都是错的呢，只有一个类别map在70几，其他的 都是0？

直接运行nets/resnet.py出现以下错误：
`Traceback (most recent call last):
File "D:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\framework\ops.py", line 1864, in _create_c_op
c_op = c_api.TF_FinishOperation(op_desc)
tensorflow.python.framework.errors_impl.InvalidArgumentError: Shape must be rank 1 but is rank 0 for 'bn_conv1/Reshape_4' (op: 'Reshape') with input shapes: [1,1,1,64], [].

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "E:/PycharmProjects/faster-rcnn-keras/nets/resnet.py", line 250, in
model = ResNet50(inputs)
File "E:/PycharmProjects/faster-rcnn-keras/nets/resnet.py", line 162, in ResNet50
x = BatchNormalization(name='bn_conv1')(x)
File "D:\Program Files\Anaconda3\lib\site-packages\keras\engine\base_layer.py", line 457, in call
output = self.call(inputs, **kwargs)
File "E:/PycharmProjects/faster-rcnn-keras/nets/resnet.py", line 91, in call
epsilon=self.epsilon)
File "D:\Program Files\Anaconda3\lib\site-packages\keras\backend\tensorflow_backend.py", line 1908, in batch_normalization
mean = tf.reshape(mean, (-1))
File "D:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\ops\gen_array_ops.py", line 9093, in reshape
"Reshape", tensor=tensor, shape=shape, name=name)
File "D:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\framework\op_def_library.py", line 788, in _apply_op_helper
op_def=op_def)
File "D:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\util\deprecation.py", line 507, in new_func
return func(*args, **kwargs)
File "D:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\framework\ops.py", line 3616, in create_op
op_def=op_def)
File "D:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\framework\ops.py", line 2027, in init
control_input_ops)
File "D:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\framework\ops.py", line 1867, in _create_c_op
raise ValueError(str(e))
ValueError: Shape must be rank 1 but is rank 0 for 'bn_conv1/Reshape_4' (op: 'Reshape') with input shapes: [1,1,1,64], [].

Process finished with exit code 1
`
异常是从BatchNormalization抛出的，请问有人遇到这样的问题吗

UP主，啥时候把代码升级到2.1

看了下输出的结果发现recall非常低，即便precession已经降到0了，recall也只有80多，尝试了下调学习率（根本不太可能有用吧。。。。），没有效果，怀疑是RPN的问题，增加了RPN里两个loss的权重，也不行，请问还有别的优化方法吗？另外关于epoch_length参数是要设置成和自己的样本数量一致吗？如果我设置了一个小于样本数量的值a，就随机从我的样本中抽取a个进行训练？望各位不吝赐教/(ㄒoㄒ)/~~

您好，我有几个问题想请教您！
1.请问这个模型的初始权重(voc_weights.h5)文件可以换成其他网络吗？比如换成vgg16或者resnet101什么的。

2.您设定的epoch为100个，每个epoch2000(step?)次，我的电脑(win10.1050ti.4G)一个epoch训练时间大概要半个小时。之前我在琢磨git上的Faster-RCNN-TensorFlow-Python3，他迭代一次所需的时间只有1-2s(我上个厕所就迭代了一千多次。。。但后面他的demo.py文件我运行起来无法生成预测的图片，怎么改都不行，加上训练所需时间实在太短，感觉不真实，所以就放弃研究那个模型了。)请问您有研究过那个模型吗？为什么两个模型的训练时间相差这么大？是我电脑环境的问题吗

我用的
tf-gpu 1.12.0
cudn 9.0.176
cudnn 7.3.0.29

Epoch 1/50: 0%| | 0/114 [00:04<?, ?it/s<class 'dict'>]
Traceback (most recent call last):
File "F:/JetBrains/PycharmProjects/faster-rcnn-keras/train.py", line 208, in
callback)
File "F:/JetBrains/PycharmProjects/faster-rcnn-keras/train.py", line 52, in fit_one_epoch
X2, Y1, Y2 = calc_iou(R, config, boxes[i], NUM_CLASSES)
File "F:\JetBrains\PycharmProjects\faster-rcnn-keras\utils\roi_helpers.py", line 122, in calc_iou
np.int32)] = 1
IndexError: index 1 is out of bounds for axis 1 with size 1

就是这行代码：
y_class_regr_label[
np.arange(np.shape(gt_roi_loc)[0])[:pos_roi_per_this_image], np.array(gt_roi_label[:pos_roi_per_this_image],
np.int32)] = 1

Traceback (most recent call last):
File "f:/dece/faster-rcnn-keras-master/get_map.py", line 706, in
+ " ; Recall=" + "{0:.2f}%".format(rec[score05_idx]*100) + " ; Precision=" + "{0:.2f}%".format(prec[score05_idx]*100)) ore05_idx]*100))
IndexError: index 0 is out of bounds for axis 0 with size 0

报错显示如上

作者您好!

在FasterRCNN的論文裡面，計算tx, ty, tw, th的公式如圖所示
而在encode_boxes有看到
`

    encoded_box[:, :2][assign_mask] = box_center - assigned_priors_center
    encoded_box[:, :2][assign_mask] /= assigned_priors_wh
    encoded_box[:, :2][assign_mask] *= 4
    encoded_box[:, 2:4][assign_mask] = np.log(box_wh / assigned_priors_wh)
    encoded_box[:, 2:4][assign_mask] *= 4

`

和decode_boxes
`

    # 真实框距离先验框中心的xy轴偏移情况
    decode_bbox_center_x = mbox_loc[:, 0] * prior_width / 4
    decode_bbox_center_x += prior_center_x
    decode_bbox_center_y = mbox_loc[:, 1] * prior_height / 4
    decode_bbox_center_y += prior_center_y
    
    # 真实框的宽与高的求取
    decode_bbox_width = np.exp(mbox_loc[:, 2] / 4)
    decode_bbox_width *= prior_width
    decode_bbox_height = np.exp(mbox_loc[:, 3] /4)
    decode_bbox_height *= prior_height

`

想請問為何要針對encode_box*4和對decode_box/4呢?

请问训练自己的数据集是需要从头开始训练，还是可以使用你提供的预训练好的权重进行继续训练啊。

是按照博主的环境配置的

（我的其他项目是可以用gpu训练的。）
同时，我也在博主的train.py中加了
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
可是训练的时候使用的还是cpu，无法用gpu加速，训练得太慢了。

可爱的博主知道是什么原因吗？
求指教！

请问我rtx2060，6g的显存，batch_size=1的时候都会爆是为什么呀？

hello, when we have done the train, the predict.py can plot the predict boxes, but ,how can we evaluate the recall , precission and mAP

这个模型是你自己训练的还是下载的论文原作者的？感觉去检测视频时候，效果很差。

能训练，但是训练几十个step之后就会OOM。减少了classifier_layers函数中的channel能延迟OOM到来，但没法根治问题。

博主的程序是不是一次训练一张图片，可否一次训练多张？

./nets/frcnn.py
classifier = get_resnet50_classifier(feature_map_input, roi_input, 14, num_classes)

./nets/classifier.py

def get_resnet50_classifier(base_layers, input_rois, roi_size, num_classes=21):
    # batch_size, 38, 38, 1024 -> batch_size, num_rois, 14, 14, 1024
    out_roi_pool = RoiPoolingConv(roi_size)([base_layers, input_rois])

    # batch_size, num_rois, 14, 14, 1024 -> num_rois, 1, 1, 2048
    out = resnet50_classifier_layers(out_roi_pool)

    # batch_size, num_rois, 1, 1, 2048 -> batch_size, num_rois, 2048
    out = TimeDistributed(Flatten())(out)

    # batch_size, num_rois, 2048 -> batch_size, num_rois, num_classes
    out_class   = TimeDistributed(Dense(num_classes, activation='softmax', kernel_initializer=random_normal(stddev=0.02)), name='dense_class_{}'.format(num_classes))(out)
    # batch_size, num_rois, 2048 -> batch_size, num_rois, 4 * (num_classes-1)
    out_regr    = TimeDistributed(Dense(4 * (num_classes - 1), activation='linear', kernel_initializer=random_normal(stddev=0.02)), name='dense_regress_{}'.format(num_classes))(out)
    return [out_class, out_regr]

./nets/resnet.py

def conv_block_td(input_tensor, kernel_size, filters, stage, block, strides=(2, 2)):
    nb_filter1, nb_filter2, nb_filter3 = filters
    conv_name_base  = 'res' + str(stage) + block + '_branch'
    bn_name_base    = 'bn' + str(stage) + block + '_branch'

-----> here is the issue--------------------------------------------------------------
    x = TimeDistributed(Conv2D(nb_filter1, (1, 1), strides=strides, kernel_initializer='normal'), name=conv_name_base + '2a')(input_tensor)
--------------------------------------------------------------------------------------
    x = TimeDistributed(BatchNormalization(), name=bn_name_base + '2a')(x)
    x = Activation('relu')(x)
...

def resnet50_classifier_layers(x):
    # batch_size, num_rois, 14, 14, 1024 -> batch_size, num_rois, 7, 7, 2048
    x = conv_block_td(x, 3, [512, 512, 2048], stage=5, block='a', strides=(2, 2))
    # batch_size, num_rois, 7, 7, 2048 -> batch_size, num_rois, 7, 7, 2048
    x = identity_block_td(x, 3, [512, 512, 2048], stage=5, block='b')
    # batch_size, num_rois, 7, 7, 2048 -> batch_size, num_rois, 7, 7, 2048
    x = identity_block_td(x, 3, [512, 512, 2048], stage=5, block='c')
    # batch_size, num_rois, 7, 7, 2048 -> batch_size, num_rois, 1, 1, 2048
    x = TimeDistributed(AveragePooling2D((7, 7)), name='avg_pool')(x)

我在colab上试着运行，但在上面箭头处出现错误，请问是什么原因？
TypeError: Dimension value must be integer or None or have an index method, got value '(None, None, 1024)' with type '<class 'tuple'>'

博主大大，数据集的百度网盘链接失效了，麻烦再发一下数据集，球球了

I calcualated the maP with Cartucho repository (https://github.com/Cartucho/mAP), with weights reported and the value is only 63 , different from 77 reported. Thank you

想问下博主如何让模型从0开始训练

大佬，您好我将您的模型用于voc数据集训练时，发现后两个损失值一直在上升，即detectior_cls和detectior_regr,最近做毕设用到这个模型，希望您能百忙之中尽快回复
1/2000 [..............................] - ETA: 11:45:41 - rpn_cls: 0.9137 - rpn_regr: 2.7919 - detector_cls: 0.1544 - detector_regr: 0.2115
2/2000 [..............................] - ETA: 7:20:18 - rpn_cls: 1.0191 - rpn_regr: 2.2316 - detector_cls: 0.2252 - detector_regr: 0.3208
3/2000 [..............................] - ETA: 5:54:12 - rpn_cls: 1.0463 - rpn_regr: 1.8886 - detector_cls: 0.2790 - detector_regr: 0.4301
4/2000 [..............................] - ETA: 5:11:04 - rpn_cls: 1.0501 - rpn_regr: 1.7989 - detector_cls: 0.3448 - detector_regr: 0.5418
5/2000 [..............................] - ETA: 4:41:25 - rpn_cls: 1.0798 - rpn_regr: 1.6981 - detector_cls: 0.3767 - detector_regr: 0.5938
6/2000 [..............................] - ETA: 4:21:43 - rpn_cls: 1.0718 - rpn_regr: 1.5993 - detector_cls: 0.4244 - detector_regr: 0.6206

Python 3.6

运行resnet的话，会报错这个TypeError。
检查之后发现可能是这行代码的问题
第77行
if sorted(reduction_axes) == range(K.ndim(x))[:-1]:

range部分是不是要变成list？
我把它变成list之后就可以顺畅运行了。

本来只想把之前同事修改的训练代码改回up的代码，没想到这是一条不归路。疯狂复制才回到了正轨。。

Hi,
Does this repo support pre-computed proposals?
BTW， have you tested on Nvidia RTX3090 GPU?

File "train.py", line 207, in <module>
    fit_one_epoch(model_rpn, model_all, epoch, epoch_size, epoch_size_val, gen, gen_val, Interval_Epoch, callback)
  File "train.py", line 55, in fit_one_epoch
    X2, Y1, Y2 = calc_iou(R, config, boxes[i], NUM_CLASSES)
  File "/media/user/date/qianyi/deeplearningwork/faster-rcnn-keras/utils/roi_helpers.py", line 110, in calc_iou
    Y1 = np.eye(num_classes)[np.array(gt_roi_label,np.int32)]
IndexError: index 6 is out of bounds for axis 0 with size 2

NUM_CLASSES=2，只有一个类可以进行训练吗出错如上，求助，谢谢

I am training the model with basic resnet 50 weights from imagenet with voc pascal images. I got this error after a few ages. I think it could come from images without enough positive and negative bounding boxes. Can you solve?

• faster-rcnn-keras-master\faster-rcnn-keras-master\utils\roi_helpers.py", line 122, in calc_iou

  y_class_regr_label = np.reshape(y_class_regr_label, [np.shape(gt_roi_loc)[0], -1])

• File "<array_function internals>", line 6, in reshape

• lib\site-packages\numpy\core\fromnumeric.py", line 301, in reshape

  return _wrapfunc(a, 'reshape', newshape, order=order)

• ib\site-packages\numpy\core\fromnumeric.py", line 61, in _wrapfunc

  return bound(*args, **kwds)

• ValueError: cannot reshape array of size 0 into shape (0,newaxis)