Question in the config

[jiangzihanict]

Hello Eracah,

Sorry to bother you again.
I am running the code with the TF-Keras version. It occurs Key Error in the config when I use the iclr_supervised detection model. Then I check the configuration files according to the TraceBack, it seems there is no "num_filter_list" attribute. I think it is the number of filters in each layer of the model, but I can't figure out what is the correct number.

Here is the TraceBack:
Traceback (most recent call last):
"dotpy_src/models/base/iclr_supervised.py", line 22, in <module>
encoder_num_filters_list = configs["num_filter_list"]
KeyError: 'num_filter_list'

Looking forward to your reply. Thanks!

@eracah

[jiangzihanict]

Oh, I find the filters_list = [128, 256, 512, 768, 1024, 1280] in Theano version. Is it right?

[eracah]

Hi Zihan,
That is the correct filter_list. Sorry about the keras code. It is not very readable due to the over-modularizing of everything. Some day I should rewrite a pytorch version of the code that is more clear to read

[jiangzihanict]

Thanks Eracah! Actually I can understand your code.
Another question is one more detail. It is a piece of code in the loss function of the supervised_iclr2017 model.

def extract_lbl_channels(y_true):
y_true_xy,y_true_wh,y_true_obj_y_true_cls = tf.split(value=y_true,num_split=3,split_dim=3)

Through tf_fit function I derived the shape of y_true is [4, 15, 5]. It seems that the tf.split function cannot be used this way to get the y_true_xy, y_true_wh, y_true_obj_y_true_cls.
In fact, I'm stuck in this place now because I don't know how to get these tensor.

Thank you for taking the time to answer my question.

[eracah]

Hi Zihan,
So: batch_size, num_boxes, box_dimensionality = [4, 15, 5]
So it's supposed to split the last dimension into 3 parts: (y_true_x, y_true_y), (y_true_w, y_true_h), y_true_obj_y_true_cls

I think potentially tf.split API changed? Which version of TF are you using?

Hope that helps!

[jiangzihanict]

Hi Eracah,
Thanks for your quick reply.
My TF version is 1.14.
There are two ways of splitting a tensor in 1.14[tf.split] (Same as 1.15)(https://www.tensorflow.org/versions/r1.15/api_docs/python/tf/split))

If num_or_size_splits is an integer, then value is split along dimension axis into num_split smaller tensors. This requires that num_split evenly divides value.shape[axis].

If num_or_size_splits is a 1-D Tensor (or list), we call it size_splits and value is split into len(size_splits) elements. The shape of the i-th element has the same size as the value except along dimension axis where the size is size_splits[i].

According to your answer, I split the tensor in the 1-D Tensor way. So I get the three tensors:
y_true_xy.shape = [4,15,2]
y_true_wh.shape = [4,15,2]
y_true_obj_y_true_cls = [4,15,1]

But I encounter new problems. The next line of the code is y_true_obj, y_true_cls = tf.split(value=y_true_obj_y_true_cls, num_split=2, split_dim=3). The y_true_obj_y_true_cls.shape is [4,15,1], the last dim is 1 and cannot be evenly divided by 2. I know the last dimension means which weather this box belongs to. So how can i divide this into 2 splits?
Maybe I don't understand the relation between y_true_obj and y_true_cls.

I have studied the test case of Compute_lose(pred, true). I find the gr_truth is a tensor of shape [5, 24, 24, 6] insdead of [4, 15, 5], so what does each dimension mean?

hur-detect/dotpy_src/losses/iclr2017/iclr_supervised.py

Line 190 in e597729

if __name__ == "__main__":

Thanks for your patience and kind help！

[eracah]

I think compute_loss is supposed to be this one:

hur-detect/dotpy_src/losses/ssd/ssd.py

Line 29 in e597729

def compute_loss(y_true, y_preds):

[eracah]

That one takes in 5 dim tensor not a 6 dim one.

[eracah]

I really just need to write a simple pytorch version of this code. I don't think this tf-keras code is fully tested. The master branch code is tested, but uses netcdf files not hdf5 files and it is theano.

[jiangzihanict]

Thanks, Eracah.
Yes, SSD can run successfully. But can I use this loss function directly in the ICLR2017 model? After reading your paper, I know there is some difference between them. I think the compute_loss of this version is pretty clear as it is consistent with your NIPS paper, is it possible to modify it? I’ll study the SSD and theano version code of compute_loss.