bruceyang2012 / face-detection-with-mobilenet-ssd Goto Github PK

@bruceyang2012
Hello, it is a great research job.
Can you tell me how to generate the wider_extract.py files. I need to create my own data set.

Thank you very much!

Hi yang shuang , thanks for sharing repo, it is really great work. if you let me , I will get a problem is sharing with you . When I trained wider face data by your face_train.ipynb , I got high val error . Also my epoch is 30 . I got min error 0.10801 .I dont understand why it happen. What can you say doing preprocessing for data , What is your plan to trian data

Hi,
Just wondering why you chose not to include any augmentations in the train_dataset.generate? For example, translation would enable the faces of the people on the ground in the second photo to be detected. I am going to implement something similar and was wondering if you had any rational behind not augmenting. Was it the training time?

Regards,
Moshe

wider_extract.py

if isValid=='0' and (isOccluded!='2') and (isBlur!='2'):
bb.append(bb_each)

Hi @bruceyang2012 !

I'm trying to solve hand detection problem (2 classes: 'background' and 'hand') on custom dataset using your model. I've adjusted data to the desired format, so there is no problem with data loading.

However, after training I get the first 2 coordinates in y_pred [0.5, 0.5] in most of the cases, and [0.5+eps, 0.5-eps] in the rest minority, which means that model cannot classify bounding boxes and doesn't recognize "active" class (hand) at all.

I've tried to play with different parameters of SSDLoss(), like setting beta=2.0 and changing neg_pos_ratio and neg_min but it didn't change the picture.
Can you please give any suggestions on what to do next and what should I change to get reasonable classification results?

Thanks!

Height, Width, Channels : 512 512 3
WARNING:tensorflow:The following Variables were used in a Lambda layer's call (tf.math.multiply_2), but are not present in its tracked objects: <tf.Variable 'detection_conv4_3_norm/detection_conv4_3_norm_gamma:0' shape=(512,) dtype=float32>. This is a strong indication that the Lambda layer should be rewritten as a subclassed Layer.
Freezing classification layers
classification layers freezed
loading classification weights

OSError Traceback (most recent call last)
in
30 print ("loading classification weights")
31 classification_model = './base_models/mobilenet_1_0_224_tf.h5'
---> 32 model.load_weights(classification_model, by_name= True)
33
34 print (colored( ('classification weights %s loaded' % classification_model), 'green'))

2 frames
/usr/local/lib/python3.7/dist-packages/h5py/_hl/files.py in make_fid(name, mode, userblock_size, fapl, fcpl, swmr)
188 if swmr and swmr_support:
189 flags |= h5f.ACC_SWMR_READ
--> 190 fid = h5f.open(name, flags, fapl=fapl)
191 elif mode == 'r+':
192 fid = h5f.open(name, h5f.ACC_RDWR, fapl=fapl)

h5py/_objects.pyx in h5py._objects.with_phil.wrapper()

h5py/_objects.pyx in h5py._objects.with_phil.wrapper()

h5py/h5f.pyx in h5py.h5f.open()

OSError: Unable to open file (file signature not found)

Hey, first of all great project. I'm having great fun working on this.

I trained my own model using my custom data on the architecture mentioned here. After training and quantization I'm getting a model of around 4.7MB. I wanted to decrease this size even further by maybe removing some of the layers. Can you point me on how to remove some layers to get the size lower?

@bruceyang2012 你好，这个有一些精度测试的结果么

Can you give me a tutorial about testing to get the test result you had shown.

classification weights ./base_models/mobilenet_1_0_224_tf.h5 loaded
Traceback (most recent call last):
  File "face_train.py", line 119, in <module>
    box_output_format = ['class_id', 'xmin', 'xmax', 'ymin', 'ymax'])
TypeError: __init__() got an unexpected keyword argument 'image_set_path'

Getting this error while running the face_train code. Any lead will be helpful. thank you.

Hello,
thanks for sharing.
I have two questions:

1. I can do some training to reduce the val_loss under 0.5, however, when I test on single images, the result is wierd. Thus, is there a demo, how can I load the model and test it on single image, is it needed to transform it to the batch-format data?
2. Second, if I do some my own training task, like hand detection task, will the transfer learning with coco or pascal-pretrained model be good?
   Thanks & Regards!

Firstly, thanks for your sharing.
I tried to train the model on Pascalvoc2012 , and i got [email protected] = 0.22. It almost could not work.
I did training on coco, it didn’t converge. So I decided to check the model in using the same group of images as train and test dataset. The result should be 1,if the code was implemented right, but the actual result was closed to 0.

So I wonder whether there is an error in the code or the model could be trained successfully. i want to find where the mistake are.
I have checked the code many times, and please help me.
it is urgent.
many thanks.

In keras_layer_AnchorBoxes.py, there is an error in call() in class AnchorBoxes.
boxes_tensor[:, :, :, 2] = wh_list[:, 0] # Set w
boxes_tensor[:, :, :, 3] = wh_list[:, 1] # Set h
could not broadcast input array from shape (2) into shape (w, h, num_boxes)

Can you add the requirements file with versions of libraries that are to be used