search

guoqiangqi / PFLD

Implementation of PFLD A Practical Facial Landmark Detector , reference to https://arxiv.org/pdf/1902.10859.pdf
627 stars 166 forks source link

good model #8

Closed liguiyuan closed 5 years ago

liguiyuan commented 5 years ago

Hi， thanks for your open source, I tried to training model, but is not good to used, can we provide an available model in this project? Thanks @guoqiangqi

guoqiangqi commented 5 years ago

Hi， thanks for your open source, I tried to training model, but is not good to used, can we provide an available model in this project? Thanks @guoqiangqi

What does 'not goog to used' mean ? Have you trainned the moel or just something error happened?

liguiyuan commented 5 years ago

I'm sorry, I trainned the model on WFLW , it can run, but the result is not good, may I get a better model by trainning on 300W dataset? Waiting for your answer.

guoqiangqi commented 5 years ago

The WFLW dataset include massive faces with big angles, it would be harder to detect the face landmark,so the accurary would be lower than 300W,Notice here,the author have pretrainned model with WFLW,then fine-tune the model with 300W,but the accurary is lower than LAB on AFLW . Considerated the fast speed ,proformance of PFLD is also fine.

zoooo0820 commented 5 years ago

@liguiyuan In the PFLD paper, PFLD 1X is only trained by 300-W training data (3148 images), while PFLD 1X+ is pretrained on WFLW. However, when I try this model on 300-W (as PFLD 1X did), the result is bad. Could you please share the result of your re-implement on 300-W or on other public dataset?

liguiyuan commented 5 years ago

I'm not training on 300-W dataset, because I found some problems in training step demo(such as : memory leak, slow convergence etc. )，so I only used author WFLW dataset to training first , after slove this bugs, I will training on 300-W dataset. For your question, you can take this advise:

  1. Could you follow PFLD paper to make dataset?

For 300W, we augment the training data by flipping each sample and rotating them from -30° to +30° with 5° interval. Further, each sample has a region of 20% face size randomly occluded

2.Maybe you can change 'train_model.py' file 'TRACKED_POINTS' lamdmarks : TRACKED_POINTS = [33, 38, 50, 46, 60, 64, 68, 72, 55, 59, 76, 82, 85, 16]

  1. Training longer, such as more than 20 epoch or 30 epoch，because the model slow convergence
Danbinabo commented 5 years ago

用于训练的图片和list.txt怎么存放？具体路径和文件夹存放方式在你们设置？

github-luffy commented 5 years ago

训练一次大概三天左右，训练的最好的模型的mean Error为0.076，不知道还有什么办法可以提升模型精度

guoqiangqi commented 5 years ago

有做数据增强吗？或者在300W上预训练一下，不过提升不一定大。

425418567@qq.com

From: 叶修强 Date: 2019-07-03 17:10 To: guoqiangqi/PFLD CC: Guoqiang QI; Mention Subject: Re: [guoqiangqi/PFLD] good model (#8) 训练一次大概三天左右，训练的最好的模型的mean Error为0.076，不知道还有什么办法可以提升模型精度 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

github-luffy commented 5 years ago

下次用300w训练一下，你们训练的最好情况是怎么样的------------------ 原始邮件 ------------------ 发件人: "Guoqiang QI"notifications@github.com 发送时间: 2019年7月3日(星期三) 下午5:12 收件人: "guoqiangqi/PFLD"PFLD@noreply.github.com; 抄送: "叶修强"luffyxq@qq.com;"Comment"comment@noreply.github.com; 主题: Re: [guoqiangqi/PFLD] good model (#8)

有做数据增强吗？或者在300W上预训练一下，不过提升不一定大。

425418567@qq.com

From: 叶修强 Date: 2019-07-03 17:10 To: guoqiangqi/PFLD CC: Guoqiang QI; Mention Subject: Re: [guoqiangqi/PFLD] good model (#8) 训练一次大概三天左右，训练的最好的模型的mean Error为0.076，不知道还有什么办法可以提升模型精度 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

guoqiangqi commented 5 years ago

我们没有去调参，精度不如你的，注意到PFLD精度上并不是非常好，后面没做了。

425418567@qq.com

From: 叶修强 Date: 2019-07-03 17:34 To: guoqiangqi/PFLD CC: Guoqiang QI; Mention Subject: Re: [guoqiangqi/PFLD] good model (#8) 下次用300w训练一下，你们训练的最好情况是怎么样的------------------ 原始邮件 ------------------ 发件人: "Guoqiang QI"notifications@github.com 发送时间: 2019年7月3日(星期三) 下午5:12 收件人: "guoqiangqi/PFLD"PFLD@noreply.github.com; 抄送: "叶修强"luffyxq@qq.com;"Comment"comment@noreply.github.com; 主题: Re: [guoqiangqi/PFLD] good model (#8)

有做数据增强吗？或者在300W上预训练一下，不过提升不一定大。

425418567@qq.com

From: 叶修强 Date: 2019-07-03 17:10 To: guoqiangqi/PFLD CC: Guoqiang QI; Mention Subject: Re: [guoqiangqi/PFLD] good model (#8) 训练一次大概三天左右，训练的最好的模型的mean Error为0.076，不知道还有什么办法可以提升模型精度 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

liguiyuan commented 5 years ago

训练一次大概三天左右，训练的最好的模型的mean Error为0.076，不知道还有什么办法可以提升模型精度

我训练的效果也差不多，目前最好的mean error: 0.075, failure rate: L1 0.185，还在继续改进中，请问一下你用300W试了吗？

github-luffy commented 5 years ago

300w训练数据你们的属性是怎么设置的，自己设置的还是人脸库本身自带的标签？

guoqiangqi commented 5 years ago

我们没跑。。

425418567@qq.com

From: 叶修强 Date: 2019-07-06 09:05 To: guoqiangqi/PFLD CC: Guoqiang QI; Mention Subject: Re: [guoqiangqi/PFLD] good model (#8) 300w训练数据你们的属性是怎么设置的，自己设置的还是人脸库本身自带的标签？ — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

quliulangle commented 5 years ago

我用V100的显卡跑这个模型，内存会占用到60多个G，而且训练的特变慢，训练了三天，也才6个epoch,然后就停下来了，没有看到效果。

liguiyuan commented 5 years ago

300w训练数据你们的属性是怎么设置的，自己设置的还是人脸库本身自带的标签？

300W的数据集没带有表情、遮挡、模糊等标注，需要自己对这些属性做分类标注

liguiyuan commented 5 years ago

我用V100的显卡跑这个模型，内存会占用到60多个G，而且训练的特变慢，训练了三天，也才6个epoch,然后就停下来了，没有看到效果。

代码有内存溢出的问题，你可以参考issue7（#7）的解决办法，tf的op计算结点过多导致训练特别慢。此外可以参照论文调整下学习率和权重衰减，一般100个epoch就能取得还可以的效果。

guoqiangqi commented 5 years ago

这部分代码放到循坏外。

425418567@qq.com

From: quliulangle Date: 2019-07-07 23:08 To: guoqiangqi/PFLD CC: Guoqiang QI; Mention Subject: Re: [guoqiangqi/PFLD] good model (#8) 我用V100的显卡跑这个模型，内存会占用到60多个G，而且训练的特变慢，训练了三天，也才6个epoch,然后就停下来了，没有看到效果。 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

cheneyshen180 commented 5 years ago

attributes = sess.run(attribute_batch) print("attributes1",attributes) attributes_w_n = tf.to_float(attributes[:, 1:6]) mat_ratio = tf.reduce_mean(attributes_w_n, axis=0) mat_ratio = tf.map_fn(lambda x:1.0/x if not x==0.0 else 0.0,mat_ratio)

mat_ratio = list(map(lambda x: 1.0 / x if not x == 0.0 else float(images.shape[0]), sess.run(mat_ratio)))

attributes_w_n = attributes_w_n * mat_ratio
attributes_w_n = tf.reduce_sum(attributes_w_n, axis=1)
attributes_w_n = sess.run(attributes_w_n)

这个代码需要放到循环外面吗？

guoqiangqi commented 5 years ago

sess.run()不用，中间的op要，然后改调整一下代码就可以

425418567@qq.com

From: cheneyshen180 Date: 2019-07-08 15:05 To: guoqiangqi/PFLD CC: Guoqiang QI; Mention Subject: Re: [guoqiangqi/PFLD] good model (#8) attributes = sess.run(attribute_batch) print("attributes1",attributes) attributes_w_n = tf.to_float(attributes[:, 1:6]) mat_ratio = tf.reduce_mean(attributes_w_n, axis=0) mat_ratio = tf.map_fn(lambda x:1.0/x if not x==0.0 else 0.0,mat_ratio)

mat_ratio = list(map(lambda x: 1.0 / x if not x == 0.0 else float(images.shape[0]), sess.run(mat_ratio)))

attributes_w_n = attributes_w_n * mat_ratio attributes_w_n = tf.reduce_sum(attributes_w_n, axis=1) attributes_w_n = sess.run(attributes_w_n) 这个代码需要放到循环外面吗？ — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

FX-STAR commented 5 years ago

训练一次大概三天左右，训练的最好的模型的mean Error为0.076，不知道还有什么办法可以提升模型精度

是common set上吗？

quliulangle commented 5 years ago

image_batch, landmarks_batch, attribute_batch, euler_batch = list_ops['train_next_element']

attributes_w = tf.to_float(attribute_batch[:, 1:6]) mat_ratio = tf.reduce_mean(attributes_w, axis=0) mat_ratio = list(map(lambda x: 1.0 / x if not x == 0.0 else float(image_batch.shape[0]), sess.run(mat_ratio))) attributes_w = attributes_w * mat_ratio attributes_w = tf.reduce_sum(attributes_w, axis=1) for i in range(epoch_size): images, landmarks, attributes, eulers = sess.run([image_batch, landmarks_batch, attribute_batch, euler_batch]) attributes_w_n = sess.run(attributes_w)

是这么改么？快被这个问题搞疯掉😭😭😭

guoqiangqi commented 5 years ago

我没有数据，大概改了下脚本，你跑一下，看看报什么错，加我qq

425418567@qq.com

From: quliulangle Date: 2019-07-08 18:06 To: guoqiangqi/PFLD CC: Guoqiang QI; Mention Subject: Re: [guoqiangqi/PFLD] good model (#8) image_batch, landmarks_batch, attribute_batch, euler_batch = list_ops['train_next_element'] attributes_w = tf.to_float(attribute_batch[:, 1:6]) mat_ratio = tf.reduce_mean(attributes_w, axis=0) mat_ratio = list(map(lambda x: 1.0 / x if not x == 0.0 else float(image_batch.shape[0]), sess.run(mat_ratio))) attributes_w = attributes_w * mat_ratio attributes_w = tf.reduce_sum(attributes_w, axis=1) for i in range(epoch_size): images, landmarks, attributes, eulers = sess.run([image_batch, landmarks_batch, attribute_batch, euler_batch]) attributes_w_n = sess.run(attributes_w) 是这么改么？快被这个问题搞疯掉???????????? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

keyxuliang commented 5 years ago

interocular_distance使用的是60 and 72两个点的距离吗？

guoqiangqi commented 5 years ago

瞳孔或者眼角

425418567@qq.com

From: keyxuliang Date: 2019-07-09 15:51 To: guoqiangqi/PFLD CC: Guoqiang QI; Mention Subject: Re: [guoqiangqi/PFLD] good model (#8) interocular_distance使用的是60 and 72两个点的距离吗？ — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.