search

huoyijie / AdvancedEAST

AdvancedEAST is an algorithm used for Scene image text detect, which is primarily based on EAST, and the significant improvement was also made, which make long text predictions more accurate.https://github.com/huoyijie/raspberrypi-car
https://huoyijie.cn/
MIT License
1.23k stars 380 forks source link

icdar2015上训练效果 #70

Open xxlxx1 opened 5 years ago

xxlxx1 commented 5 years ago

使用icdar2015数据来训练，参数使用作者默认的，为什么loss都很高呢，只训练256大小的 Epoch 1/24 1125/1125 [==============================] - 131s 117ms/step - loss: 1.0463 - val_loss: 1.1207

Epoch 00001: val_loss improved from inf to 1.12069, saving model to model/weights_3T256.001-1.121.h5 Epoch 2/24 1125/1125 [==============================] - 126s 112ms/step - loss: 0.8477 - val_loss: 1.1539

Epoch 00002: val_loss did not improve Epoch 3/24 1125/1125 [==============================] - 126s 112ms/step - loss: 0.7191 - val_loss: 1.1873

Epoch 00003: val_loss did not improve Epoch 4/24 1125/1125 [==============================] - 126s 112ms/step - loss: 0.6319 - val_loss: 1.2241

Epoch 00004: val_loss did not improve Epoch 5/24 1125/1125 [==============================] - 126s 112ms/step - loss: 0.5608 - val_loss: 1.2758

Epoch 00005: val_loss did not improve Epoch 6/24 1125/1125 [==============================] - 126s 112ms/step - loss: 0.5107 - val_loss: 1.2994

Epoch 00006: val_loss did not improve Epoch 00006: early stopping

hcnhatnam commented 5 years ago

I have same the problem @huoyijie

lijian10086 commented 5 years ago

我也遇到相同的问题，求大神解惑@huoyijie

peter-peng-w commented 5 years ago

Maybe you should refer to the issue #27 first and try to follow the way that the author was doing. I am also facing this problem and currently working on it.

lijian10086 commented 5 years ago

@stillarrow @huoyijie 感谢你们的热心回答。我根据the issue #27的作者操作，使用icdar2015的1000个训练样本，其中900个样本用于train，100个样本用于test，使用默认参数，256——》384——》512——》640——736依次训练和加载上一步的最好模型进行初始化。还是出现loss居高不下（0.7左右）的问题。求助，是哪里出了问题？

cfg.py如下： import os

train_task_id = '3T736' initial_epoch = 20 epoch_num = 40 #24 lr = 1e-3 decay = 5e-4

clipvalue = 0.5 # default 0.5, 0 means no clip

patience = 5 load_weights = True lambda_inside_score_loss = 4.0 lambda_side_vertex_code_loss = 1.0 lambda_side_vertex_coord_loss = 1.0

total_img = 1000 validation_split_ratio = 0.1 max_train_img_size = int(train_task_id[-3:]) max_predict_img_size = int(train_task_id[-3:]) # 2400 assert max_train_img_size in [256, 384, 512, 640, 736], \ 'max_train_img_size must in [256, 384, 512, 640, 736]' if max_train_img_size == 256: batch_size = 8 elif max_train_img_size == 384: batch_size = 4 elif max_train_img_size == 512: batch_size = 2 else: batch_size = 1 steps_per_epoch = total_img (1 - validation_split_ratio) // batch_size validation_steps = total_img validation_split_ratio // batch_size

data_dir = 'icpr/' origin_image_dir_name = 'image_10000/' origin_txt_dir_name = 'txt_10000/' train_image_dirname = 'images%s/' % train_task_id train_label_dirname = 'labels%s/' % train_task_id show_gt_image_dir_name = 'show_gtimages%s/' % train_task_id show_act_image_dir_name = 'show_actimages%s/' % train_task_id gen_origin_img = True draw_gt_quad = True draw_act_quad = True valfname = 'val%s.txt' % train_task_id trainfname = 'train%s.txt' % train_task_id

in paper it's 0.3, maybe to large to this problem

shrink_ratio = 0.2

pixels between 0.2 and 0.6 are side pixels

shrink_side_ratio = 0.6 epsilon = 1e-4

num_channels = 3 feature_layers_range = range(5, 1, -1)

feature_layers_range = range(3, 0, -1)

feature_layers_num = len(feature_layers_range)

pixel_size = 4

pixel_size = 2 ** feature_layers_range[-1] locked_layers = False

if not os.path.exists('model'): os.mkdir('model') if not os.path.exists('saved_model'): os.mkdir('saved_model')

model_weightspath = 'model/weights%s.{epoch:03d}-{val_loss:.3f}.h5' \ % train_task_id saved_model_file_path = 'saved_model/eastmodel%s.h5' % train_task_id

saved_model_weights_file_path = 'saved_model/east_modelweights%s.h5'\

% train_task_id

saved_model_weights_file_path = 'saved_model/weights_3T640.020-0.700.h5'

pixel_threshold = 0.9 side_vertex_pixel_threshold = 0.9 trunc_threshold = 0.1 predict_cut_text_line = False predict_write2txt = True

训练过程如下： Epoch 1/40 112/112 [==============================] - 18s 161ms/step - loss: 1.0258 - val_loss: 0.9977

Epoch 00001: val_loss improved from inf to 0.99771, saving model to model/weights_3T256.001-0.998.h5 Epoch 2/40 112/112 [==============================] - 12s 104ms/step - loss: 0.9163 - val_loss: 0.9435

Epoch 00002: val_loss improved from 0.99771 to 0.94348, saving model to model/weights_3T256.002-0.943.h5 Epoch 3/40 112/112 [==============================] - 12s 103ms/step - loss: 0.8685 - val_loss: 0.9589

Epoch 00003: val_loss did not improve from 0.94348 Epoch 4/40 112/112 [==============================] - 12s 103ms/step - loss: 0.8321 - val_loss: 0.9361

Epoch 00004: val_loss improved from 0.94348 to 0.93609, saving model to model/weights_3T256.004-0.936.h5 Epoch 5/40 112/112 [==============================] - 12s 103ms/step - loss: 0.7869 - val_loss: 0.8753

Epoch 00005: val_loss improved from 0.93609 to 0.87533, saving model to model/weights_3T256.005-0.875.h5 Epoch 6/40 112/112 [==============================] - 12s 103ms/step - loss: 0.7634 - val_loss: 0.9075

Epoch 00006: val_loss did not improve from 0.87533 Epoch 7/40 112/112 [==============================] - 11s 102ms/step - loss: 0.7447 - val_loss: 0.9625

Epoch 00007: val_loss did not improve from 0.87533 Epoch 8/40 112/112 [==============================] - 12s 103ms/step - loss: 0.6945 - val_loss: 0.9738

Epoch 00008: val_loss did not improve from 0.87533 Epoch 9/40 112/112 [==============================] - 12s 103ms/step - loss: 0.6495 - val_loss: 0.9505

Epoch 00009: val_loss did not improve from 0.87533 Epoch 10/40 112/112 [==============================] - 12s 103ms/step - loss: 0.6376 - val_loss: 0.9441

Epoch 00010: val_loss did not improve from 0.87533 Epoch 00010: early stopping Epoch 6/40 225/225 [==============================] - 31s 138ms/step - loss: 0.8283 - val_loss: 0.8632

Epoch 00006: val_loss improved from inf to 0.86316, saving model to model/weights_3T384.006-0.863.h5 Epoch 7/40 225/225 [==============================] - 26s 114ms/step - loss: 0.7723 - val_loss: 0.8328

Epoch 00007: val_loss improved from 0.86316 to 0.83276, saving model to model/weights_3T384.007-0.833.h5 Epoch 8/40 225/225 [==============================] - 26s 115ms/step - loss: 0.7441 - val_loss: 0.8024

Epoch 00008: val_loss improved from 0.83276 to 0.80241, saving model to model/weights_3T384.008-0.802.h5 Epoch 9/40 225/225 [==============================] - 26s 114ms/step - loss: 0.6963 - val_loss: 0.8008

Epoch 00009: val_loss improved from 0.80241 to 0.80084, saving model to model/weights_3T384.009-0.801.h5 Epoch 10/40 225/225 [==============================] - 26s 113ms/step - loss: 0.6602 - val_loss: 0.8203

Epoch 00010: val_loss did not improve from 0.80084 Epoch 11/40 225/225 [==============================] - 26s 113ms/step - loss: 0.6240 - val_loss: 0.8538

Epoch 00011: val_loss did not improve from 0.80084 Epoch 12/40 225/225 [==============================] - 26s 114ms/step - loss: 0.5770 - val_loss: 0.8294

Epoch 00012: val_loss did not improve from 0.80084 Epoch 13/40 225/225 [==============================] - 26s 115ms/step - loss: 0.5657 - val_loss: 0.9329

Epoch 00013: val_loss did not improve from 0.80084 Epoch 14/40 225/225 [==============================] - 26s 115ms/step - loss: 0.5337 - val_loss: 0.8109

Epoch 00014: val_loss did not improve from 0.80084 Epoch 00014: early stopping Epoch 6/40 225/225 [==============================] - 31s 138ms/step - loss: 0.8283 - val_loss: 0.8632

Epoch 00006: val_loss improved from inf to 0.86316, saving model to model/weights_3T384.006-0.863.h5 Epoch 7/40 225/225 [==============================] - 26s 114ms/step - loss: 0.7723 - val_loss: 0.8328

Epoch 00007: val_loss improved from 0.86316 to 0.83276, saving model to model/weights_3T384.007-0.833.h5 Epoch 8/40 225/225 [==============================] - 26s 115ms/step - loss: 0.7441 - val_loss: 0.8024

Epoch 00008: val_loss improved from 0.83276 to 0.80241, saving model to model/weights_3T384.008-0.802.h5 Epoch 9/40 225/225 [==============================] - 26s 114ms/step - loss: 0.6963 - val_loss: 0.8008

Epoch 00009: val_loss improved from 0.80241 to 0.80084, saving model to model/weights_3T384.009-0.801.h5 Epoch 10/40 225/225 [==============================] - 26s 113ms/step - loss: 0.6602 - val_loss: 0.8203

Epoch 00010: val_loss did not improve from 0.80084 Epoch 11/40 225/225 [==============================] - 26s 113ms/step - loss: 0.6240 - val_loss: 0.8538

Epoch 00011: val_loss did not improve from 0.80084 Epoch 12/40 225/225 [==============================] - 26s 114ms/step - loss: 0.5770 - val_loss: 0.8294

Epoch 00012: val_loss did not improve from 0.80084 Epoch 13/40 225/225 [==============================] - 26s 115ms/step - loss: 0.5657 - val_loss: 0.9329

Epoch 00013: val_loss did not improve from 0.80084 Epoch 14/40 225/225 [==============================] - 26s 115ms/step - loss: 0.5337 - val_loss: 0.8109

Epoch 00014: val_loss did not improve from 0.80084 Epoch 00014: early stopping Epoch 10/40 450/450 [==============================] - 52s 116ms/step - loss: 0.7323 - val_loss: 0.8008

Epoch 00010: val_loss improved from inf to 0.80084, saving model to model/weights_3T512.010-0.801.h5 Epoch 11/40 450/450 [==============================] - 47s 105ms/step - loss: 0.6656 - val_loss: 0.7684

Epoch 00011: val_loss improved from 0.80084 to 0.76837, saving model to model/weights_3T512.011-0.768.h5 Epoch 12/40 450/450 [==============================] - 47s 105ms/step - loss: 0.6215 - val_loss: 0.7373

Epoch 00012: val_loss improved from 0.76837 to 0.73733, saving model to model/weights_3T512.012-0.737.h5 Epoch 13/40 450/450 [==============================] - 47s 105ms/step - loss: 0.6045 - val_loss: 0.7435

Epoch 00013: val_loss did not improve from 0.73733 Epoch 14/40 450/450 [==============================] - 47s 104ms/step - loss: 0.5627 - val_loss: 0.7855

Epoch 00014: val_loss did not improve from 0.73733 Epoch 15/40 450/450 [==============================] - 47s 104ms/step - loss: 0.5516 - val_loss: 0.7778

Epoch 00015: val_loss did not improve from 0.73733 Epoch 16/40 450/450 [==============================] - 47s 104ms/step - loss: 0.5046 - val_loss: 0.7344

Epoch 00016: val_loss improved from 0.73733 to 0.73437, saving model to model/weights_3T512.016-0.734.h5 Epoch 17/40 450/450 [==============================] - 47s 104ms/step - loss: 0.4785 - val_loss: 0.7575

Epoch 00017: val_loss did not improve from 0.73437 Epoch 18/40 450/450 [==============================] - 47s 104ms/step - loss: 0.4503 - val_loss: 0.8144

Epoch 00018: val_loss did not improve from 0.73437 Epoch 19/40 450/450 [==============================] - 47s 104ms/step - loss: 0.4179 - val_loss: 0.8738

Epoch 00019: val_loss did not improve from 0.73437 Epoch 20/40 450/450 [==============================] - 47s 104ms/step - loss: 0.4001 - val_loss: 0.8318

Epoch 00020: val_loss did not improve from 0.73437 Epoch 21/40 450/450 [==============================] - 47s 104ms/step - loss: 0.3776 - val_loss: 0.8361

Epoch 00021: val_loss did not improve from 0.73437 Epoch 00021: early stopping Epoch 17/40 900/900 [==============================] - 87s 97ms/step - loss: 0.6340 - val_loss: 0.8157

Epoch 00017: val_loss improved from inf to 0.81570, saving model to model/weights_3T640.017-0.816.h5 Epoch 18/40 900/900 [==============================] - 81s 90ms/step - loss: 0.5578 - val_loss: 0.7535

Epoch 00018: val_loss improved from 0.81570 to 0.75353, saving model to model/weights_3T640.018-0.754.h5 Epoch 19/40 900/900 [==============================] - 81s 90ms/step - loss: 0.5180 - val_loss: 0.7031

Epoch 00019: val_loss improved from 0.75353 to 0.70313, saving model to model/weights_3T640.019-0.703.h5 Epoch 20/40 900/900 [==============================] - 81s 91ms/step - loss: 0.4704 - val_loss: 0.6999

Epoch 00020: val_loss improved from 0.70313 to 0.69989, saving model to model/weights_3T640.020-0.700.h5 Epoch 21/40 900/900 [==============================] - 81s 90ms/step - loss: 0.4337 - val_loss: 0.8117

Epoch 00021: val_loss did not improve from 0.69989 Epoch 22/40 900/900 [==============================] - 81s 90ms/step - loss: 0.3875 - val_loss: 0.8635

Epoch 00022: val_loss did not improve from 0.69989 Epoch 23/40 900/900 [==============================] - 81s 90ms/step - loss: 0.3774 - val_loss: 0.8113

Epoch 00023: val_loss did not improve from 0.69989 Epoch 24/40 900/900 [==============================] - 81s 90ms/step - loss: 0.3461 - val_loss: 0.7961

Epoch 00024: val_loss did not improve from 0.69989 Epoch 25/40 900/900 [==============================] - 81s 90ms/step - loss: 0.3265 - val_loss: 0.8327

Epoch 00025: val_loss did not improve from 0.69989 Epoch 00025: early stopping Epoch 21/40 900/900 [==============================] - 107s 119ms/step - loss: 0.5786 - val_loss: 0.6918

Epoch 00021: val_loss improved from inf to 0.69179, saving model to model/weights_3T736.021-0.692.h5 Epoch 22/40 900/900 [==============================] - 100s 111ms/step - loss: 0.5387 - val_loss: 0.7834

Epoch 00022: val_loss did not improve from 0.69179 Epoch 23/40 900/900 [==============================] - 101s 112ms/step - loss: 0.4902 - val_loss: 0.7386

Epoch 00023: val_loss did not improve from 0.69179 Epoch 24/40 900/900 [==============================] - 101s 112ms/step - loss: 0.4416 - val_loss: 0.8202

Epoch 00024: val_loss did not improve from 0.69179 Epoch 25/40 900/900 [==============================] - 101s 112ms/step - loss: 0.4116 - val_loss: 0.7364

Epoch 00025: val_loss did not improve from 0.69179 Epoch 26/40 900/900 [==============================] - 101s 112ms/step - loss: 0.3863 - val_loss: 0.9104

Epoch 00026: val_loss did not improve from 0.69179 Epoch 00026: early stopping 中间模型如下： image

peter-peng-w commented 5 years ago

@lijian10086 It seems that you did not change the parameters in the config file (well, except for the epoch number) but could not reach an eval loss under 0.7. Actually I once did the same experiment on icdar2015 dataset (size 1000) but could reach a eval loss around 0.5, seems a little bit weird..... I think its might be kind of overfitting. image Currently working on OCR using models including this repo, maybe I could share with you the solution in the next few weeks.

lijian10086 commented 5 years ago

@stillarrow 哈哈，终于等到你的回答了。我通过降低lr在736模型下继续训练，最多只能valloss最多只能下到0.64。valloss在0.5左右估计也是远远不够的吧，要达到论文那种0.8F1，估计val要0.1以内吧，你觉得呢？你之前的回答中“ I am also facing this problem and currently working on it.”，是指使用作者的tianchi ICPR dataset 10000样本训练没问题是么？PS：我最近也是在做OCR，能否加我微信，在微信方便即时交流（我的微信号：18024583839）。 image

xiiiiiiii commented 5 years ago

请问你最后valloss训练到多少了？我训练了256，现在在训练387的时候是这样的 image 虽然也不太好 我用的数据集是icdar2017

wsy915 commented 5 years ago

@xiiiiiiii 请问你是怎么训练的呢？我用作者提供的那10000个样本的数据集进行训练，先建了icpr目录，然后建image_10000和txt_10000，其他参数没有改先训练了256，出现了early stop.请问你一开始直接训练256的时候有对代码做什么修改吗

xiiiiiiii commented 5 years ago

我也是这样训练的，early stop是正常现象 防止过拟合，你的训练效果怎么样

---原始邮件--- 发件人: "wsy915"<notifications@github.com> 发送时间: 2019年10月21日(星期一) 上午10:26 收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com>; 抄送: "Mention"<mention@noreply.github.com>;"guopeijun"<494207346@qq.com>; 主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

@xiiiiiiii 请问你是怎么训练的呢？我用作者提供的那10000个样本的数据集进行训练，先建了icpr目录，然后建image_10000和txt_10000，其他参数没有改先训练了256，出现了early stop.请问你一开始直接训练256的时候有对代码做什么修改吗

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

wsy915 commented 5 years ago

我才刚开始训练，才训练到256。然后在第7个epoch就early stop了。你继续训练的时候就是按照作者27的回答改cfg.py里的参数实现的吗？你现在效果理想吗？你如果val_loss一直不降低的话是不是过拟合了？样本数量太少？

------------------ 原始邮件 ------------------ 发件人: "guopeijun"<notifications@github.com>; 发送时间: 2019年10月21日(星期一) 上午10:27 收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com>; 抄送: "王思怡"<785134974@qq.com>; "Comment"<comment@noreply.github.com>; 主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

我也是这样训练的，early stop是正常现象 防止过拟合，你的训练效果怎么样

---原始邮件--- 发件人: "wsy915"<notifications@github.com&gt; 发送时间: 2019年10月21日(星期一) 上午10:26 收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com&gt;; 抄送: "Mention"<mention@noreply.github.com&gt;;"guopeijun"<494207346@qq.com&gt;; 主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

@xiiiiiiii 请问你是怎么训练的呢？我用作者提供的那10000个样本的数据集进行训练，先建了icpr目录，然后建image_10000和txt_10000，其他参数没有改先训练了256，出现了early stop.请问你一开始直接训练256的时候有对代码做什么修改吗

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

xiiiiiiii commented 5 years ago

对，我只用了3k个数据，效果还不错，512是效果最好的时候，之后训练效果也不咋样了，如果1w个数据可能更好吧

---原始邮件--- 发件人: "wsy915"<notifications@github.com> 发送时间: 2019年10月21日(星期一) 上午10:39 收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com>; 抄送: "Mention"<mention@noreply.github.com>;"guopeijun"<494207346@qq.com>; 主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

我才刚开始训练，才训练到256。然后在第7个epoch就early stop了。你继续训练的时候就是按照作者27的回答改cfg.py里的参数实现的吗？你现在效果理想吗？你如果val_loss一直不降低的话是不是过拟合了？样本数量太少？

------------------&nbsp;原始邮件&nbsp;------------------ 发件人: "guopeijun"<notifications@github.com&gt;;
发送时间: 2019年10月21日(星期一) 上午10:27 收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com&gt;;
抄送: "王思怡"<785134974@qq.com&gt;; "Comment"<comment@noreply.github.com&gt;;
主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

我也是这样训练的，early stop是正常现象 防止过拟合，你的训练效果怎么样

---原始邮件---
发件人: "wsy915"<notifications@github.com&amp;gt;
发送时间: 2019年10月21日(星期一) 上午10:26
收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com&amp;gt;;
抄送: "Mention"<mention@noreply.github.com&amp;gt;;"guopeijun"<494207346@qq.com&amp;gt;;
主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

@xiiiiiiii 请问你是怎么训练的呢？我用作者提供的那10000个样本的数据集进行训练，先建了icpr目录，然后建image_10000和txt_10000，其他参数没有改先训练了256，出现了early stop.请问你一开始直接训练256的时候有对代码做什么修改吗

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

wsy915 commented 5 years ago

那请问你在下一次训练的时候的initial_epoch是设置为上次stop的次数之后，还是设置为最后一次保存模型的epoch之后？像我这个，应该设为13吗，还是8？

发自我的iPhone

------------------ 原始邮件 ------------------ 发件人: guopeijun <notifications@github.com> 发送时间: 2019年10月21日 10:41 收件人: huoyijie/AdvancedEAST <AdvancedEAST@noreply.github.com> 抄送: wsy915 <785134974@qq.com>, Comment <comment@noreply.github.com> 主题: 回复：[huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

对，我只用了3k个数据，效果还不错，512是效果最好的时候，之后训练效果也不咋样了，如果1w个数据可能更好吧

---原始邮件--- 发件人: "wsy915"<notifications@github.com&gt; 发送时间: 2019年10月21日(星期一) 上午10:39 收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com&gt;; 抄送: "Mention"<mention@noreply.github.com&gt;;"guopeijun"<494207346@qq.com&gt;; 主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

我才刚开始训练，才训练到256。然后在第7个epoch就early stop了。你继续训练的时候就是按照作者27的回答改cfg.py里的参数实现的吗？你现在效果理想吗？你如果val_loss一直不降低的话是不是过拟合了？样本数量太少？

------------------&amp;nbsp;原始邮件&amp;nbsp;------------------
发件人: "guopeijun"<notifications@github.com&amp;gt;;
发送时间: 2019年10月21日(星期一) 上午10:27
收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com&amp;gt;;
抄送: "王思怡"<785134974@qq.com&amp;gt;; "Comment"<comment@noreply.github.com&amp;gt;;
主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

我也是这样训练的，early stop是正常现象 防止过拟合，你的训练效果怎么样

---原始邮件---
发件人: "wsy915"<notifications@github.com&amp;amp;gt;
发送时间: 2019年10月21日(星期一) 上午10:26
收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com&amp;amp;gt;;
抄送: "Mention"<mention@noreply.github.com&amp;amp;gt;;"guopeijun"<494207346@qq.com&amp;amp;gt;;
主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

@xiiiiiiii 请问你是怎么训练的呢？我用作者提供的那10000个样本的数据集进行训练，先建了icpr目录，然后建image_10000和txt_10000，其他参数没有改先训练了256，出现了early stop.请问你一开始直接训练256的时候有对代码做什么修改吗

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

xiiiiiiii commented 5 years ago

都可以，这个只是保存模型的名称会不一样而已

---原始邮件--- 发件人: "wsy915"<notifications@github.com> 发送时间: 2019年10月21日(星期一) 中午11:06 收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com>; 抄送: "Mention"<mention@noreply.github.com>;"guopeijun"<494207346@qq.com>; 主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

那请问你在下一次训练的时候的initial_epoch是设置为上次stop的次数之后，还是设置为最后一次保存模型的epoch之后？像我这个，应该设为13吗，还是8？

发自我的iPhone

------------------ 原始邮件 ------------------ 发件人: guopeijun <notifications@github.com&gt; 发送时间: 2019年10月21日 10:41 收件人: huoyijie/AdvancedEAST <AdvancedEAST@noreply.github.com&gt; 抄送: wsy915 <785134974@qq.com&gt;, Comment <comment@noreply.github.com&gt; 主题: 回复：[huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

对，我只用了3k个数据，效果还不错，512是效果最好的时候，之后训练效果也不咋样了，如果1w个数据可能更好吧

---原始邮件---
发件人: "wsy915"<notifications@github.com&amp;gt;
发送时间: 2019年10月21日(星期一) 上午10:39
收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com&amp;gt;;
抄送: "Mention"<mention@noreply.github.com&amp;gt;;"guopeijun"<494207346@qq.com&amp;gt;;
主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

我才刚开始训练，才训练到256。然后在第7个epoch就early stop了。你继续训练的时候就是按照作者27的回答改cfg.py里的参数实现的吗？你现在效果理想吗？你如果val_loss一直不降低的话是不是过拟合了？样本数量太少？

------------------&amp;amp;nbsp;原始邮件&amp;amp;nbsp;------------------
发件人: "guopeijun"<notifications@github.com&amp;amp;gt;;
发送时间: 2019年10月21日(星期一) 上午10:27
收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com&amp;amp;gt;;
抄送: "王思怡"<785134974@qq.com&amp;amp;gt;; "Comment"<comment@noreply.github.com&amp;amp;gt;;
主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

我也是这样训练的，early stop是正常现象 防止过拟合，你的训练效果怎么样 

---原始邮件---    
发件人: "wsy915"<notifications@github.com&amp;amp;amp;amp;gt;    
发送时间: 2019年10月21日(星期一) 上午10:26    
收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com&amp;amp;amp;amp;gt;;    
抄送: "Mention"<mention@noreply.github.com&amp;amp;amp;amp;gt;;"guopeijun"<494207346@qq.com&amp;amp;amp;amp;gt;;    
主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)    

@xiiiiiiii  请问你是怎么训练的呢？我用作者提供的那10000个样本的数据集进行训练，先建了icpr目录，然后建image_10000和txt_10000，其他参数没有改先训练了256，出现了early stop.请问你一开始直接训练256的时候有对代码做什么修改吗    

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

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

wsy915 commented 5 years ago

你好，请问你的val_loss最后是多少呢？我训练736的时候，依然只跑了14个epoch，val_loss只到0.19。

发自我的iPhone

------------------ 原始邮件 ------------------ 发件人: guopeijun <notifications@github.com> 发送时间: 2019年10月21日 11:07 收件人: huoyijie/AdvancedEAST <AdvancedEAST@noreply.github.com> 抄送: wsy915 <785134974@qq.com>, Comment <comment@noreply.github.com> 主题: 回复：[huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

都可以，这个只是保存模型的名称会不一样而已

---原始邮件--- 发件人: "wsy915"<notifications@github.com&gt; 发送时间: 2019年10月21日(星期一) 中午11:06 收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com&gt;; 抄送: "Mention"<mention@noreply.github.com&gt;;"guopeijun"<494207346@qq.com&gt;; 主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

那请问你在下一次训练的时候的initial_epoch是设置为上次stop的次数之后，还是设置为最后一次保存模型的epoch之后？像我这个，应该设为13吗，还是8？

发自我的iPhone

------------------ 原始邮件 ------------------
发件人: guopeijun <notifications@github.com&amp;gt;
发送时间: 2019年10月21日 10:41
收件人: huoyijie/AdvancedEAST <AdvancedEAST@noreply.github.com&amp;gt;
抄送: wsy915 <785134974@qq.com&amp;gt;, Comment <comment@noreply.github.com&amp;gt;
主题: 回复：[huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

对，我只用了3k个数据，效果还不错，512是效果最好的时候，之后训练效果也不咋样了，如果1w个数据可能更好吧

---原始邮件---
发件人: "wsy915"<notifications@github.com&amp;amp;gt;
发送时间: 2019年10月21日(星期一) 上午10:39
收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com&amp;amp;gt;;
抄送: "Mention"<mention@noreply.github.com&amp;amp;gt;;"guopeijun"<494207346@qq.com&amp;amp;gt;;
主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

我才刚开始训练，才训练到256。然后在第7个epoch就early stop了。你继续训练的时候就是按照作者27的回答改cfg.py里的参数实现的吗？你现在效果理想吗？你如果val_loss一直不降低的话是不是过拟合了？样本数量太少？ 

------------------&amp;amp;amp;amp;nbsp;原始邮件&amp;amp;amp;amp;nbsp;------------------    
发件人: "guopeijun"<notifications@github.com&amp;amp;amp;amp;gt;;     
发送时间: 2019年10月21日(星期一) 上午10:27    
收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com&amp;amp;amp;amp;gt;;     
抄送: "王思怡"<785134974@qq.com&amp;amp;amp;amp;gt;; "Comment"<comment@noreply.github.com&amp;amp;amp;amp;gt;;     
主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)    

我也是这样训练的，early stop是正常现象 防止过拟合，你的训练效果怎么样     

 ---原始邮件---     
 发件人: "wsy915"<notifications@github.com&amp;amp;amp;amp;amp;gt;     
 发送时间: 2019年10月21日(星期一) 上午10:26     
 收件人: "huoyijie/AdvancedEAST"<AdvancedEAST@noreply.github.com&amp;amp;amp;amp;amp;gt;;     
 抄送: "Mention"<mention@noreply.github.com&amp;amp;amp;amp;amp;gt;;"guopeijun"<494207346@qq.com&amp;amp;amp;amp;amp;gt;;     
 主题: Re: [huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)     

 @xiiiiiiii  请问你是怎么训练的呢？我用作者提供的那10000个样本的数据集进行训练，先建了icpr目录，然后建image_10000和txt_10000，其他参数没有改先训练了256，出现了early stop.请问你一开始直接训练256的时候有对代码做什么修改吗     

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

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

li10141110 commented 4 years ago

@wsy915 你好，你在什么数据上面训练对，最终loss到了多少，我对loss最后知道0.19

wsy915 commented 4 years ago

我是在自己的数据集加上2017rctw上训练的。最后的val_loss大概0.10

发自我的iPhone

------------------ 原始邮件 ------------------ 发件人: Jason Lee <notifications@github.com> 发送时间: 2019年12月16日 12:10 收件人: huoyijie/AdvancedEAST <AdvancedEAST@noreply.github.com> 抄送: wsy915 <785134974@qq.com>, Mention <mention@noreply.github.com> 主题: 回复：[huoyijie/AdvancedEAST] icdar2015上训练效果 (#70)

@wsy915 你好，你在什么数据上面训练对，最终loss到了多少，我对loss最后知道0.19

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