After training for 10000 iterations, I want to train for another 10000.
Instead of training from 0 - 20000, I was wondering if I could load the params_10000.npy and train 10000 iterations more.

Tried loading the model like vgg16.npy but loss comes out to be huge.

Is there a way to train on top of previous trained model?

where is this class file ?

hi,
I want to know the parameter about this code

self.anchor_min_height = 40 * self.image_resize_factor

the 40 is that:

45 - 5 = 40
45 is the feature map height

am I right ?

thx

hi,
I want to use your model to change my own datasets.
But Know ,I want read your source code in the data_engine.py file.

    def proposal_prepare(self, imdb):
        anchors = self.generate_anchors()
        proposals = np.zeros(
            [self.anchor_size * self.convmap_width * self.convmap_height, 4])

        for i in range(self.convmap_height):
            h = i * 16 + 8
            for j in range(self.convmap_width):
                w = j * 16 + 8
                for k in range(self.anchor_size):
                    index = i * self.convmap_width * self.anchor_size + \
                            j * self.anchor_size + k
                    anchor = anchors[k, :]
                    proposals[index, :] = anchor + np.array([w, h, w, h])

the numbers 16 and 8 is acording to the four times pooling layer to use 16, but what does the 8 means ?

thx

Hello, thank you for this great work.

Is the model already trained in this repo? If I use test.py I see it already detects people, so maybe it is just ready to use?

Another question: does train.py has a stop criteria or loss function threshold? I set it to train with precarious dataset and after 43 hours I have this output:

step : 8400 time : 153299.0062 loss : 0.46541554 l_r : 0.0001

When will it stop?

Thank you

I want to train my data. How to convert voc dataset format to your dataset format or how to use voc format directly?

I have got the following error while running your code.

Traceback (most recent call last):
  File "demo.py", line 10, in <module>
    from image_pylib import IMGLIB
ModuleNotFoundError: No module named 'image_pylib'

@huangshiyu13 Hi, thank you for sharing the code. I have dowloaded the two datasets. I can train the model with the following data path settings:
imageLoadDir = Dataset/precarious_dataset/train/
anoLoadDir = precarious_dataset/trainAno/.
If I also want to use the synthetic_dataset dataset, how should I set the data path?

Hello, I want to try train my own dataset. can you help to guide us how to train with our own dataset. we will really appreciate that. thankyou :)

I want to train my data, but i could not fine tune the model with my own data, so how to fine tune the model?

How to convert a trained npy file to tensorflow type model?

Hi, I am interested in the RPNplus, the train code must train a new model, how to fine tune the model?

在duke-MTMC-reID数据集中，出现的人占图片的比例都很小，抠出来的框有时候比人大一圈，所以算出来的score很低。模型好像没有办法框出很小的人，我们看了半天没看懂代码，能不能告诉一下怎么修改bbox的最小尺寸使得能够框出很小的人？谢谢！

hi,
I want to use RPNPlus to detect the object?
and I want to use this model to extract the proposal area.
And I changed the Input Image size to (224 ,224 ,1). But I don't know how to change the 'wandhG'?

wandhG = [[100.0, 100.0], [300.0, 300.0], [500.0, 500.0],
          [200.0, 100.0], [370.0, 185.0], [440.0, 220.0],
          [100.0, 200.0], [185.0, 370.0], [220.0, 440.0]]

thx

I train the model in synthetic_dataset. When I run demo.py with params_10000.npy (instead of model.npy), the result is very bad, even can't detect person. How do you train your model @huangshiyu13 ?

can not download model files(RPN_model & VGG16_model)？ anyone can give me, ([email protected]), thanks very much.

Inside proposal_prepare, why did you increment h and w to be incremented *16 + 8 from i,j?

I want to use the architecture to detect other objects, such as birds, or other kinds of objects. Is it possbile ? Or any modifications ?

And I use the pre-trained model on my data, there are always several boxes overlapped on one body, how can I fix this , is it related to the threshold of nms? Or I need to retrain by myself?

Hi @huangshiyu13,

When the generator data is available ?

Thanks

This model is useful in some data sets. But in some data sets such as duke-MTMC-reID, the proportion of people in the picture is very small. And in this case, the bbox min size is bigger than the people's size, so some people can't detection successfully. How to deal with this problem? Thank u very much!

首先，感谢你分享的代码，我现在遇到一个问题：计算机RAM 8GB，GPU 显存12GB，现在想用的数据集大概有4万多张，但是一训练直接就会因为内存不足而崩溃，请问不增加内存的情况下能解决这个问题吗？

i used the following code in the terminal
python train.py 0
And I got the following errors

Traceback (most recent call last):' File "train.py", line 240, in
with tf.device(gpuNow):'

`File "/home/farshid/anaconda3/lib/python3.5/contextlib.py", line 59, in enter
return next(self.gen)'

File "/home/farshid/anaconda3/lib/python3.5/site-packages/tensorflow/python/framework/ops.py", line 3045, in device' device_function = pydev.merge_device(device_name_or_function)'

File "/home/farshid/anaconda3/lib/python3.5/site-packages/tensorflow/python/framework/device.py", line 282, in merge_device' spec = DeviceSpec.from_string(spec or "")'

File "/home/farshid/anaconda3/lib/python3.5/site-packages/tensorflow/python/framework/device.py", line 228, in from_string' return DeviceSpec().parse_from_string(spec)'

`File "/home/farshid/anaconda3/lib/python3.5/site-packages/tensorflow/python/framework/device.py", line 146, in parse_from_string'

splits = [x.split(":") for x in spec.split("/")]'AttributeError: 'int' object has no attribute 'split'
`

Thank you for your sharing code and dataset. I am writing to you about your synthetic_dataset. I wanted to train a model to deal with the non-straight pedestrian detection and wondered if I can use the synthetic_dataset mentioned in your paper - "Expecting the Unexpected : Training Detectors for Unusual Pedestrians with Adversarial Imposters CVPR. 2017“ to train it. I found that the unreasonable datasets in the synthetic dataset didn't be deleted and I didn't know how to get only imposter images as you did in the paper. Then, I wish you can give me some advice.
Than you!

Here is the link.

https://github.com/YunYang1994/TensorFlow2.0-Examples/tree/master/4-Object_Detection/RPN

After nn starts the training, it outputs a type error in data_engine.py in prepare_data,
line: fg_inx = fg_idx[:fg_num]

TypeError: slice indices must be integers or None or have an index method

I just downloaded the precarious_dataset and updated the folders on train.py