for cascade_lv in xrange(1,etc.cascade_level)

[tengshaofeng]

dear mks0601: i am so appreciated with your great work. I have read your code these days. The following code comes my question:

code start

for cascade_lv in xrange(1,etc.cascade_level): #question

if cascade_lv > 1:
    print "Training start!"
    train_start = time.time()
    fp_loss = open("./result/loss_" + str(cascade_lv) + "_.txt", "w")

    for e in xrange(etc.epoch_num[cascade_lv]):

        fp_conf_mat = open("./result/conf_mat_" + str(e) + "_.txt","w")  
        epoch_start = time.time()
        average_loss = 0

        for b_iter in xrange(etc.batch_iter):

            pos_id = random.sample(xrange(len(pos_db)),etc.pos_batch)
            neg_id = random.sample(xrange(len(neg_db)),etc.neg_batch)

            if cascade_lv == 0:            #question
                db_12[:etc.pos_batch,:] = pos_db[pos_id,:etc.dim_12]
                db_12[etc.pos_batch:,:] = neg_db[neg_id,:etc.dim_12]
                lb[:etc.pos_batch,:] = np.ones((etc.pos_batch,1),np.float32)
                lb[etc.pos_batch:,:] = np.zeros((etc.neg_batch,1),np.float32)
                zipped = zip(db_12, lb)
                np.random.shuffle(zipped)

code ending

So, "for cascade_lv in xrange(1,etc.cascade_level)", cascade_lv is in [1,2,3].

However "if cascade_lv == 0: ", i want to know xrange or range??? Look forward to your reply

[tengshaofeng]

Another problem, when i run your code with all the input data ready, comes the following error : Traceback (most recent call last): File "train.py", in saver_detect.restore(sess,etc.save_dir+"24-net.ckpt") raise ValueError("Restore called with invalid save path %s" % save_path) ValueError: Restore called with invalid save path

I thought that calling restore function before saver.save? So dose the code is the latest code ?

[mks0601]

Hi tengshaofeng

First of all, i ran some experiments so that i modified some conditions. For the first issue, in train.py 193 line,

for cascade_lv in xrange(1,etc.cascade_level):

if cascade_lv > 1:

the xrange of cascade_lv should be xrange(0,etc.cascade_level) and the if statement should be removed.

Maybe i trained the model whose cascade_lv is 2(48-net). So, hard neg mining from cascade_lv 1(24-net) and train the 48-net.

Also, for the second issue, you may understand why there is restore function before save. I loaded 24-net and hard neg mining and train 48-net.

[tengshaofeng]

I have understand what you said. Thanks very much. restore error from following code:

code start

if cascade_lv == 0: saver_detect.restore(sess,etc.save_dir+"12-net.ckpt") elif cascade_lv == 1: saver_detect.restore(sess,etc.save_dir+"24-net.ckpt")

saver_calib_12.restore(sess,etc.save_dir+"12-net_calib.ckpt")
saver_calib_24.restore(sess,etc.save_dir+"24-net_calib.ckpt")

code end

So shoud i run train_cali.py first?

[mks0601]

training order is as follow: (12-net, 12-calib-net), (24-net, 24-calib-net), (48-net, 48-calib-net) the order in parenthesis does not matter

that restore is to collect hard negative from 12-net and 12-calib net

[tengshaofeng]

ok, it is so clear. Thank you very much. Here is the last question. It seems that your "normalization layer" keeps blank. Tensorflow does not have the function to acheive it ? I am the beginner to tensorflow. Sorry to bother you all the time.

[mks0601]

I do not know whether there is normalization layer or not and that is why i omitted the normalization layer.

[tengshaofeng]

I think there is a normalization layer (LRN).

[96imranahmed]

Hi @tengshaofeng, did you implement the LRN in the end? Did it give you any training speed boosts?

Also can I confirm, to train, do you guys run train.py and then afterwards run train_calib.py? I'm confused as to how you train both the 12-net and 12-calib-net in parallel?

Also, thank you so much for this implementation @mks0601!

[mks0601]

As I said, 12-net and 12-calib net is different network. You do not need to train these networks simultaneously.

[96imranahmed]

@mks0601 , Right ok - that's what I was thinking (having read through the paper)! So I was just a bit confused!

Thanks for the response!

[tengshaofeng]

Hi,@96imranahmed, for the 1st question, I have implemented the LRN, but I didn't investigate whether it can boost the training speed. I just do it like the paper said.
for the 2nd question, You can train the all the 12-24-48-calib-net first then train the 12-24-48-net. if you have several machines gpu, you can train the 12-net on one machine, and train 24-cali-net on another machine.

[ShuoYe2020]

Dear @mks0601 @tengshaofeng ，Sorry to bother you again, I'm running 'test.py' now, This should be the last step for the CNN_cascade, however when the code execute at 'load_db_test()', I found all the 'FDDB-fold-0X-ellipseList.txt' is empty, I am confused, could you please tell me when did you write data in this file?