when i was training my data,there is a issue exist as follow:> d:\orgin\faster-rcnn-tensorflow-python3.5-master\faster-rcnn-tensorflow-python3.5-master\lib\layer_utils\proposal_target_layer.py(139)_sample_rois() -> keep_inds = np.append(fg_inds, bg_inds) (Pdb)

[tdf1995]

This problem appears after completing several rounds of training

[xupinjie]

i met the same problem.
did you solve it?

[tdf1995]

@Jasonxu033 there is a problem in my dataset

[xupinjie]

@tdf1995 what problem?

[niuniu111]

I met the same problem. keep_inds = np.append(fg_inds, bg_inds) (Pdb), did you solve it?

[xupinjie]

@niuniu111 Some probles in your dateset. There are some pictures couldnot fit the annotation.

[niuniu111]

I use VGG16 to run normally. When running RESNET, there is a similar error. Does this mean that the data is OK?

[xupinjie]

@niuniu111 sorry i didn't try resnet

[lccate]

I changed the learning rate from 0.0001 to 0.1,then the (Pdb) error occurs, I hope it will help you

[LarryJiang134]

For me, this error occurs randomly. I have trained entire pascal voc before yet the second day when I tried to train the network, I saw this error message. I don't think it is code which is problematic. This might because my GPU or RAM are not stable. Try to run it on another device.
Just my observation. Hope it is helpful.

[niuniu111]

It's hopeful,Thanks

[kmust-why]

imdb.append_flipped_images() Will this code affect?

[yhs95]

f:\pycode\faster-rcnn-tensorflow-python3.5-master\lib\layer_utils\proposal_tar
get_layer.py(138)_sample_rois()
-> keep_inds = np.append(fg_inds, bg_inds)
(Pdb)

[kmust-why]

Is training your own data set?

[kmust-why]

You can try to reduce the value of SCALES.

[yhs95]

I tried it, is it modified in config.py? How much is it appropriate to change the value of SCALES?

[kmust-why]

I use SCALES = [1,2,4,8,16], RATIOS = [0.5,1,2], you can try it.

[yhs95]

The picture I used for training is a grayscale image.According to what you said, using SCALES = [1,2,4,8,16], RATIOS = [0.5,1,2] will not work.As shown.After entering c, this result appears
InvalidArgumentError (see above for traceback): Incompatible shapes: [0,24] vs.
[256,24]。

[kmust-why]

Have you solved this problem later? I have encountered this problem again today.

[Tju-AI]

This problem appears after completing several rounds of training

same question,have you solved it??

[niuniu111]

The picture I used for training is a grayscale image.According to what you said, using SCALES = [1,2,4,8,16], RATIOS = [0.5,1,2] will not work.As shown.After entering c, this result appears
InvalidArgumentError (see above for traceback): Incompatible shapes: [0,24] vs.
[256,24]。

same question, have you solved this problem? What is the reason?

[niuniu111]

Some probles in your dateset. There are some pictures couldnot fit the annotation.

I also meet the question,I want to know what mean "there are some pictures could not fit the annotation."can you answer the question specifically?

[morpheusthewhite]

Same question

[morpheusthewhite]

It seems to be caused by the following lines

    if fg_inds.size > 0 and bg_inds.size > 0:
        fg_rois_per_image = min(fg_rois_per_image, fg_inds.size)
        fg_inds = npr.choice(fg_inds, size=int(fg_rois_per_image), replace=False)
        bg_rois_per_image = rois_per_image - fg_rois_per_image
        to_replace = bg_inds.size < bg_rois_per_image
        bg_inds = npr.choice(bg_inds, size=int(bg_rois_per_image), replace=to_replace)
    elif fg_inds.size > 0:
        to_replace = fg_inds.size < rois_per_image
        fg_inds = npr.choice(fg_inds, size=int(rois_per_image), replace=to_replace)
        fg_rois_per_image = rois_per_image
    elif bg_inds.size > 0:
        to_replace = bg_inds.size < rois_per_image
        bg_inds = npr.choice(bg_inds, size=int(rois_per_image), replace=to_replace)
        fg_rois_per_image = 0
    else:
        import pdb
        pdb.set_trace()  # <----- this command launches python debugger and so it stops execution

    # The indices that we're selecting (both fg and bg)
    keep_inds = np.append(fg_inds, bg_inds)

Even if the error refers to the append the stop is caused by pdb.set_trace(). That line of code was put there probably to debug some unwanted situation regarding overlapping buonding boxes

[morpheusthewhite]

I solved by raising an exception instead of opening pdb

(in lib/layer_util/proposal_target_layer.py)

    if fg_inds.size > 0 and bg_inds.size > 0:
        fg_rois_per_image = min(fg_rois_per_image, fg_inds.size)
        fg_inds = npr.choice(fg_inds, size=int(fg_rois_per_image), replace=False)
        bg_rois_per_image = rois_per_image - fg_rois_per_image
        to_replace = bg_inds.size < bg_rois_per_image
        bg_inds = npr.choice(bg_inds, size=int(bg_rois_per_image), replace=to_replace)
    elif fg_inds.size > 0:
        to_replace = fg_inds.size < rois_per_image
        fg_inds = npr.choice(fg_inds, size=int(rois_per_image), replace=to_replace)
        fg_rois_per_image = rois_per_image
    elif bg_inds.size > 0:
        to_replace = bg_inds.size < rois_per_image
        bg_inds = npr.choice(bg_inds, size=int(rois_per_image), replace=to_replace)
        fg_rois_per_image = 0
    else:
        raise Exception()

and catching it in train.py

     # Compute the graph with summary
     try:
           rpn_loss_cls, rpn_loss_box, loss_cls, loss_box, total_loss = self.net.train_step(sess, blobs, train_op)
     except Exception:
           print('image invalid, skipping')
           continue

[Monti03]

if you are using pascal_voc.py to deal with your dataset, you can try to remove -1 from _load_pascal_annotation in those rows:
x1 = float(bbox.find('xmin').text) - 1
y1 = float(bbox.find('ymin').text) - 1
x2 = float(bbox.find('xmax').text) - 1
y2 = float(bbox.find('ymax').text) - 1

[niuniu111]

when I use your approach, an exception is rainsing,which is proved image is wrong?@morpheusthewhite

[morpheusthewhite]

@niuniu111 yes, it is raised every time something in your dataset is not correct.
I met that problem when the annotation was associated to the wrong image; you should check it out

[gwspotex]

that's mean your fg & bg all return 0..
step1:
u can print the fg & bg then u can find it in dataset
step2:
if the dataset is OK u probably change RPN to small anchor & ratio
step3:
u can e-Maill to me :gwspotex@gmail.coM

[gwspotex]

that's mean your fg & bg all return 0..
step1:
u can print the fg & bg then u can find it in dataset
step2:
if the dataset is OK u probably change RPN to small anchor & ratio
step3:
u can e-Maill to me :gwspotex@gmail.com

[PangJian123]

i solve it.
First,find two lines in config.py:
tf.app.flags.DEFINE_float('roi_bg_threshold_high', 0.5, "Overlap threshold for a ROI to be considered background (class = 0 if overlap in [LO, HI))")
tf.app.flags.DEFINE_float('roi_bg_threshold_low', 0.1, "Overlap threshold for a ROI to be considered background (class = 0 if overlap in [LO, HI))")
Second,I modify two values :0.5 changed to 0.3 and 0.1 changed to 0.0.
I hope it can help you.

[dBeker]

@morpheusthewhite can you send your changes as PR?