20+20+70 for me
", "rawMarkdown": "20+20+70 for me", "votes": null }, { "id": "305701", "postDate": "03/29/2018 09:42:27", "content": "60 (heads with 0.0001), 40 (all with 0.0001), 40 (all with 0.00001) for me. @xiapidan: do you use sgd or adam? The second \"20\" is all or 4+?
", "rawMarkdown": "60 (heads with 0.0001), 40 (all with 0.0001), 40 (all with 0.00001) for me. @xiapidan: do you use sgd or adam? The second \"20\" is all or 4+?", "votes": null }, { "id": "306074", "postDate": "03/29/2018 20:25:41", "content": "Thanks
", "rawMarkdown": "Thanks", "votes": null }, { "id": "306166", "postDate": "03/30/2018 01:34:42", "content": "sdg,4+
", "rawMarkdown": "sdg,4+", "votes": null }, { "id": "306183", "postDate": "03/30/2018 02:06:01", "content": "thanks xianpidan and @john1231983! I wanted to check if you used any image normalization strategy ? I used this \n mean = np.mean(img)\n std = np.std(img)\n img = (img-mean)/std\nI did this since I use MEAN_PIXEL = [0., 0., 0.,] the original mean_pixel values are for ImageNet; thus i think we need some image normalization for generalization to our domain. \nAlso tried just img =* 255 ; another idea is to convert to gray and use CLAHE method \nhttp://scikit-image.org/docs/dev/auto_examples/color_exposure/plot_equalize.html\ngray = exposure.equalize_adapthist(gray, nbins=256)
", "rawMarkdown": "thanks xianpidan and @john1231983! I wanted to check if you used any image normalization strategy ? I used this \n mean = np.mean(img)\n std = np.std(img)\n img = (img-mean)/std\nI did this since I use MEAN_PIXEL = [0., 0., 0.,] the original mean_pixel values are for ImageNet; thus i think we need some image normalization for generalization to our domain. \nAlso tried just img =* 255 ; another idea is to convert to gray and use CLAHE method \nhttp://scikit-image.org/docs/dev/auto_examples/color_exposure/plot_equalize.html\ngray = exposure.equalize_adapthist(gray, nbins=256)", "votes": null }, { "id": "306246", "postDate": "03/30/2018 05:22:25", "content": "same with img = (img-mean)/std
", "rawMarkdown": "same with img = (img-mean)/std", "votes": null }, { "id": "306477", "postDate": "03/30/2018 13:21:07", "content": "Sorry for reply late dueto server down. I not normalize image because the pretrain weight of keras is converted from caffe. Hence, i have to follow the way caffe did as matterport. Do you find improvement using the normalization?\n@xiapian: amazing. You can got 0.52+ with this schedule, right?
", "rawMarkdown": "Sorry for reply late dueto server down. I not normalize image because the pretrain weight of keras is converted from caffe. Hence, i have to follow the way caffe did as matterport. Do you find improvement using the normalization?\n@xiapian: amazing. You can got 0.52+ with this schedule, right?", "votes": null }, { "id": "306925", "postDate": "03/31/2018 10:43:38", "content": "what do you mean by 4+ ?
", "rawMarkdown": "what do you mean by 4+ ?", "votes": null }, { "id": "307336", "postDate": "04/01/2018 12:01:51", "content": "In model.train() function there are 5 kind of layers for layers:
layers: Allows selecting wich layers to train. It can be:\n - A regular expression to match layer names to train\n - One of these predefined values:\n heaads: The RPN, classifier and mask heads of the network\n all: All the layers\n 3+: Train Resnet stage 3 and up\n 4+: Train Resnet stage 4 and up\n 5+: Train Resnet stage 5 and up\nHi @xiapidan , You mean you train the model through: 20 epochs ”heads” + 20 epochs ”4+” + 70 epochs \"all\" ,right?
", "rawMarkdown": "Hi @xiapidan , You mean you train the model through: 20 epochs ”heads” + 20 epochs ”4+” + 70 epochs \"all\" ,right?", "votes": null }, { "id": "307453", "postDate": "04/01/2018 16:23:38", "content": "@john1231983, which image size did you choose 512*512?\nI trained 20+50+70 and use resnet50 and Adam in 512*512 but I got around 0.38
", "rawMarkdown": "john1231983, which image size did you choose 512*512?\nI trained 20+50+70 and use resnet50 and Adam in 512*512 but I got around 0.38", "votes": null }, { "id": "307512", "postDate": "04/01/2018 22:21:46", "content": "@Mehul Sampat, I actually tried CLAHE preprocessing and get worse results, not sure exactly why. Did you have any success with it?
", "rawMarkdown": "Mehul Sampat, I actually tried CLAHE preprocessing and get worse results, not sure exactly why. Did you have any success with it?", "votes": null }, { "id": "308269", "postDate": "04/03/2018 08:37:01", "content": "yes @Caibimiao
", "rawMarkdown": "yes @Caibimiao", "votes": null }, { "id": "309296", "postDate": "04/05/2018 01:44:23", "content": "@RLstat sorry fro the late reply. Nope no sucess with CLAHE pre-processing either. Also adding rotations makes performance worse for me. Did image augmentation with rotations help you ?
", "rawMarkdown": "RLstat sorry fro the late reply. Nope no sucess with CLAHE pre-processing either. Also adding rotations makes performance worse for me. Did image augmentation with rotations help you ?", "votes": null }, { "id": "309303", "postDate": "04/05/2018 02:07:40", "content": "@Mehul Sampat, I am now using random flip (left right, up down) and random crop (only for image larger than 512x512), seems doing OK.
", "rawMarkdown": "Mehul Sampat, I am now using random flip (left right, up down) and random crop (only for image larger than 512x512), seems doing OK.", "votes": null }, { "id": "309732", "postDate": "04/05/2018 21:50:53", "content": "@RLstat, thanks for the information. Btw did you use imgaug for augmentation. \nhttps://github.com/aleju/imgaug\nI did this with imgaug:
\n\naugmentation = imgaug.augmenters.OneOf([\nimgaug.augmenters.Fliplr(1.0),\nimgaug.augmenters.Affine(rotate=(-10,10)),\nimgaug.augmenters.Flipud(1.0)\n])
\n\nI will remove the rotate replace it with a crop function now.
", "rawMarkdown": "RLstat, thanks for the information. Btw did you use imgaug for augmentation. \nhttps://github.com/aleju/imgaug\nI did this with imgaug: \n\naugmentation = imgaug.augmenters.OneOf([\nimgaug.augmenters.Fliplr(1.0),\nimgaug.augmenters.Affine(rotate=(-10,10)),\nimgaug.augmenters.Flipud(1.0)\n])\n\nI will remove the rotate replace it with a crop function now.", "votes": null }, { "id": "309806", "postDate": "04/06/2018 02:20:03", "content": "@Mehul Sampat, no I've not used that repository. I remember I've seen some other post that excessive augmentation won't really help. I am kind of keeping it simple for now. I think post processing is probably more important.
", "rawMarkdown": "Mehul Sampat, no I've not used that repository. I remember I've seen some other post that excessive augmentation won't really help. I am kind of keeping it simple for now. I think post processing is probably more important.", "votes": null }, { "id": "309920", "postDate": "04/06/2018 07:37:35", "content": "@RLstat\nMay I know your LB score without post processing?
", "rawMarkdown": "RLstat\nMay I know your LB score without post processing?", "votes": null }, { "id": "310025", "postDate": "04/06/2018 12:28:57", "content": "Hi, RLstat. I am using flip too. But when I use random_crop, It hurt my performance. Does it work for you? How do you do that? Can you give us some hints about this?
", "rawMarkdown": "Hi, RLstat. I am using flip too. But when I use random_crop, It hurt my performance. Does it work for you? How do you do that? Can you give us some hints about this?", "votes": null }, { "id": "310108", "postDate": "04/06/2018 16:21:44", "content": "@Caibimiao, about 0.43 to 0.45.
", "rawMarkdown": "Caibimiao, about 0.43 to 0.45.", "votes": null }, { "id": "310115", "postDate": "04/06/2018 16:35:42", "content": "Hi, @Zheng Li, giving the credit to @john1231983 at this discussion thread: https://www.kaggle.com/c/data-science-bowl-2018/discussion/52989.
\n\nI pretty much followed his methods for random cropping, i.e, just crop images with both height and width larger than 512. Perform resizing with other images. After that, I will do flipping for the resized/cropped images.
\n\nAlso, I just realized this discussion: https://www.kaggle.com/c/data-science-bowl-2018/discussion/53798. It mentioned the potential problem with the resizing in matterport's implementation and how it might affect the performance if mixing resizing with cropping. I will look into more detail of it.
", "rawMarkdown": "Hi, @Zheng Li, giving the credit to @john1231983 at this discussion thread: https://www.kaggle.com/c/data-science-bowl-2018/discussion/52989. \n\nI pretty much followed his methods for random cropping, i.e, just crop images with both height and width larger than 512. Perform resizing with other images. After that, I will do flipping for the resized/cropped images. \n\nAlso, I just realized this discussion: https://www.kaggle.com/c/data-science-bowl-2018/discussion/53798. It mentioned the potential problem with the resizing in matterport's implementation and how it might affect the performance if mixing resizing with cropping. I will look into more detail of it.", "votes": null }, { "id": "310125", "postDate": "04/06/2018 16:54:52", "content": "@RLstat yup agree with the credit to @john1231983 for the great discussion thread.\nAlso regarding the issue in: https://www.kaggle.com/c/data-science-bowl-2018/discussion/53798\nthis is fixed in the main repo. you can git clone https://github.com/matterport/Mask_RCNN.git to get the latest code.
", "rawMarkdown": "RLstat yup agree with the credit to @john1231983 for the great discussion thread.\nAlso regarding the issue in: https://www.kaggle.com/c/data-science-bowl-2018/discussion/53798\nthis is fixed in the main repo. you can git clone https://github.com/matterport/Mask_RCNN.git to get the latest code.", "votes": null }, { "id": "310455", "postDate": "04/07/2018 15:39:55", "content": "Hi, RLstat. I have changed my code according to the new update but it still work badly. Can you give me some tips about learning schedule and learning rate?
", "rawMarkdown": "Hi, RLstat. I have changed my code according to the new update but it still work badly. Can you give me some tips about learning schedule and learning rate?", "votes": null }, { "id": "310545", "postDate": "04/07/2018 22:38:17", "content": "@Zheng Li, well, I am trying out the scheduling mentioned by @john1231983 and @xiapidan\nearly in this thread , and I am using adam. I've not achieved score better than 0.45 yet with just Mask-RCNN, and rank lower than you on LB. So I am not sure whether my training scheduling will really help you.
", "rawMarkdown": "Zheng Li, well, I am trying out the scheduling mentioned by @john1231983 and @xiapidan\nearly in this thread , and I am using adam. I've not achieved score better than 0.45 yet with just Mask-RCNN, and rank lower than you on LB. So I am not sure whether my training scheduling will really help you.", "votes": null }, { "id": "310604", "postDate": "04/08/2018 04:38:05", "content": "Hi RLstat. I have tried many methods to make this function work, but none of them really improve the result. But this should theoretically improve the accuracy. To be honest, I am very curious about this. I really want to know why this function doesn’t work for me. So can you leave your email address or any contact methods? My email is mdlszhengli@gmail.com. Really want to have a discuss with you and I will share you my experiments. Thanks Zheng
", "rawMarkdown": "Hi RLstat. I have tried many methods to make this function work, but none of them really improve the result. But this should theoretically improve the accuracy. To be honest, I am very curious about this. I really want to know why this function doesn’t work for me. So can you leave your email address or any contact methods? My email is mdlszhengli@gmail.com. Really want to have a discuss with you and I will share you my experiments. Thanks Zheng", "votes": null } ], "comments": [ { "id": 305698, "author_name": "xhqlcwq", "author_url": "", "post_date": "03/29/2018 09:38:24", "content": "20+20+70 for me
", "votes": null, "replies": [ { "id": 305701, "author_name": "john1231983", "author_url": "", "post_date": "03/29/2018 09:42:27", "content": "60 (heads with 0.0001), 40 (all with 0.0001), 40 (all with 0.00001) for me. @xiapidan: do you use sgd or adam? The second \"20\" is all or 4+?
", "votes": null, "replies": [] }, { "id": 306074, "author_name": "mingcong18", "author_url": "", "post_date": "03/29/2018 20:25:41", "content": "Thanks
", "votes": null, "replies": [] }, { "id": 306166, "author_name": "xhqlcwq", "author_url": "", "post_date": "03/30/2018 01:34:42", "content": "sdg,4+
", "votes": null, "replies": [] }, { "id": 306183, "author_name": "mpsampat", "author_url": "", "post_date": "03/30/2018 02:06:01", "content": "thanks xianpidan and @john1231983! I wanted to check if you used any image normalization strategy ? I used this \n mean = np.mean(img)\n std = np.std(img)\n img = (img-mean)/std\nI did this since I use MEAN_PIXEL = [0., 0., 0.,] the original mean_pixel values are for ImageNet; thus i think we need some image normalization for generalization to our domain. \nAlso tried just img =* 255 ; another idea is to convert to gray and use CLAHE method \nhttp://scikit-image.org/docs/dev/auto_examples/color_exposure/plot_equalize.html\ngray = exposure.equalize_adapthist(gray, nbins=256)
", "votes": null, "replies": [] }, { "id": 306246, "author_name": "xhqlcwq", "author_url": "", "post_date": "03/30/2018 05:22:25", "content": "same with img = (img-mean)/std
", "votes": null, "replies": [] }, { "id": 306477, "author_name": "john1231983", "author_url": "", "post_date": "03/30/2018 13:21:07", "content": "Sorry for reply late dueto server down. I not normalize image because the pretrain weight of keras is converted from caffe. Hence, i have to follow the way caffe did as matterport. Do you find improvement using the normalization?\n@xiapian: amazing. You can got 0.52+ with this schedule, right?
", "votes": null, "replies": [] }, { "id": 306925, "author_name": "kesjien", "author_url": "", "post_date": "03/31/2018 10:43:38", "content": "what do you mean by 4+ ?
", "votes": null, "replies": [] }, { "id": 307336, "author_name": "bintan", "author_url": "", "post_date": "04/01/2018 12:01:51", "content": "In model.train() function there are 5 kind of layers for layers:
layers: Allows selecting wich layers to train. It can be:\n - A regular expression to match layer names to train\n - One of these predefined values:\n heaads: The RPN, classifier and mask heads of the network\n all: All the layers\n 3+: Train Resnet stage 3 and up\n 4+: Train Resnet stage 4 and up\n 5+: Train Resnet stage 5 and up\nHi @xiapidan , You mean you train the model through: 20 epochs ”heads” + 20 epochs ”4+” + 70 epochs \"all\" ,right?
", "votes": null, "replies": [] }, { "id": 307453, "author_name": "chiaosun", "author_url": "", "post_date": "04/01/2018 16:23:38", "content": "@john1231983, which image size did you choose 512*512?\nI trained 20+50+70 and use resnet50 and Adam in 512*512 but I got around 0.38
", "votes": null, "replies": [] }, { "id": 307512, "author_name": "ruiforecasting", "author_url": "", "post_date": "04/01/2018 22:21:46", "content": "@Mehul Sampat, I actually tried CLAHE preprocessing and get worse results, not sure exactly why. Did you have any success with it?
", "votes": null, "replies": [] }, { "id": 308269, "author_name": "xhqlcwq", "author_url": "", "post_date": "04/03/2018 08:37:01", "content": "yes @Caibimiao
", "votes": null, "replies": [] }, { "id": 309296, "author_name": "mpsampat", "author_url": "", "post_date": "04/05/2018 01:44:23", "content": "@RLstat sorry fro the late reply. Nope no sucess with CLAHE pre-processing either. Also adding rotations makes performance worse for me. Did image augmentation with rotations help you ?
", "votes": null, "replies": [] }, { "id": 309303, "author_name": "ruiforecasting", "author_url": "", "post_date": "04/05/2018 02:07:40", "content": "@Mehul Sampat, I am now using random flip (left right, up down) and random crop (only for image larger than 512x512), seems doing OK.
", "votes": null, "replies": [] }, { "id": 309732, "author_name": "mpsampat", "author_url": "", "post_date": "04/05/2018 21:50:53", "content": "@RLstat, thanks for the information. Btw did you use imgaug for augmentation. \nhttps://github.com/aleju/imgaug\nI did this with imgaug:
\n\naugmentation = imgaug.augmenters.OneOf([\nimgaug.augmenters.Fliplr(1.0),\nimgaug.augmenters.Affine(rotate=(-10,10)),\nimgaug.augmenters.Flipud(1.0)\n])
\n\nI will remove the rotate replace it with a crop function now.
", "votes": null, "replies": [] }, { "id": 309806, "author_name": "ruiforecasting", "author_url": "", "post_date": "04/06/2018 02:20:03", "content": "@Mehul Sampat, no I've not used that repository. I remember I've seen some other post that excessive augmentation won't really help. I am kind of keeping it simple for now. I think post processing is probably more important.
", "votes": null, "replies": [] }, { "id": 309920, "author_name": "ermiao", "author_url": "", "post_date": "04/06/2018 07:37:35", "content": "@RLstat\nMay I know your LB score without post processing?
", "votes": null, "replies": [] }, { "id": 310025, "author_name": "mdlszhengli", "author_url": "", "post_date": "04/06/2018 12:28:57", "content": "Hi, RLstat. I am using flip too. But when I use random_crop, It hurt my performance. Does it work for you? How do you do that? Can you give us some hints about this?
", "votes": null, "replies": [] }, { "id": 310108, "author_name": "ruiforecasting", "author_url": "", "post_date": "04/06/2018 16:21:44", "content": "@Caibimiao, about 0.43 to 0.45.
", "votes": null, "replies": [] }, { "id": 310115, "author_name": "ruiforecasting", "author_url": "", "post_date": "04/06/2018 16:35:42", "content": "Hi, @Zheng Li, giving the credit to @john1231983 at this discussion thread: https://www.kaggle.com/c/data-science-bowl-2018/discussion/52989.
\n\nI pretty much followed his methods for random cropping, i.e, just crop images with both height and width larger than 512. Perform resizing with other images. After that, I will do flipping for the resized/cropped images.
\n\nAlso, I just realized this discussion: https://www.kaggle.com/c/data-science-bowl-2018/discussion/53798. It mentioned the potential problem with the resizing in matterport's implementation and how it might affect the performance if mixing resizing with cropping. I will look into more detail of it.
", "votes": null, "replies": [] }, { "id": 310125, "author_name": "mpsampat", "author_url": "", "post_date": "04/06/2018 16:54:52", "content": "@RLstat yup agree with the credit to @john1231983 for the great discussion thread.\nAlso regarding the issue in: https://www.kaggle.com/c/data-science-bowl-2018/discussion/53798\nthis is fixed in the main repo. you can git clone https://github.com/matterport/Mask_RCNN.git to get the latest code.
", "votes": null, "replies": [] }, { "id": 310455, "author_name": "mdlszhengli", "author_url": "", "post_date": "04/07/2018 15:39:55", "content": "Hi, RLstat. I have changed my code according to the new update but it still work badly. Can you give me some tips about learning schedule and learning rate?
", "votes": null, "replies": [] }, { "id": 310545, "author_name": "ruiforecasting", "author_url": "", "post_date": "04/07/2018 22:38:17", "content": "@Zheng Li, well, I am trying out the scheduling mentioned by @john1231983 and @xiapidan\nearly in this thread , and I am using adam. I've not achieved score better than 0.45 yet with just Mask-RCNN, and rank lower than you on LB. So I am not sure whether my training scheduling will really help you.
", "votes": null, "replies": [] }, { "id": 310604, "author_name": "mdlszhengli", "author_url": "", "post_date": "04/08/2018 04:38:05", "content": "Hi RLstat. I have tried many methods to make this function work, but none of them really improve the result. But this should theoretically improve the accuracy. To be honest, I am very curious about this. I really want to know why this function doesn’t work for me. So can you leave your email address or any contact methods? My email is mdlszhengli@gmail.com. Really want to have a discuss with you and I will share you my experiments. Thanks Zheng
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "305500": "", "305698": "20+20+70 for me", "305701": "60 (heads with 0.0001), 40 (all with 0.0001), 40 (all with 0.00001) for me. @xiapidan: do you use sgd or adam? The second \"20\" is all or 4+?", "306074": "Thanks", "306166": "sdg,4+", "306183": "thanks xianpidan and @john1231983! I wanted to check if you used any image normalization strategy ? I used this \n mean = np.mean(img)\n std = np.std(img)\n img = (img-mean)/std\nI did this since I use MEAN_PIXEL = [0., 0., 0.,] the original mean_pixel values are for ImageNet; thus i think we need some image normalization for generalization to our domain. \nAlso tried just img =* 255 ; another idea is to convert to gray and use CLAHE method \nhttp://scikit-image.org/docs/dev/auto_examples/color_exposure/plot_equalize.html\ngray = exposure.equalize_adapthist(gray, nbins=256)", "306246": "same with img = (img-mean)/std", "306477": "Sorry for reply late dueto server down. I not normalize image because the pretrain weight of keras is converted from caffe. Hence, i have to follow the way caffe did as matterport. Do you find improvement using the normalization?\n@xiapian: amazing. You can got 0.52+ with this schedule, right?", "306925": "what do you mean by 4+ ?", "307336": "In `model.train()` function there are 5 kind of layers for `layers`:\n\n layers: Allows selecting wich layers to train. It can be:\n - A regular expression to match layer names to train\n - One of these predefined values:\n heaads: The RPN, classifier and mask heads of the network\n all: All the layers\n 3+: Train Resnet stage 3 and up\n 4+: Train Resnet stage 4 and up\n 5+: Train Resnet stage 5 and up", "307409": "Hi @xiapidan , You mean you train the model through: 20 epochs ”heads” + 20 epochs ”4+” + 70 epochs \"all\" ,right?", "307453": "john1231983, which image size did you choose 512*512?\nI trained 20+50+70 and use resnet50 and Adam in 512*512 but I got around 0.38", "307512": "Mehul Sampat, I actually tried CLAHE preprocessing and get worse results, not sure exactly why. Did you have any success with it?", "308269": "yes @Caibimiao", "309296": "RLstat sorry fro the late reply. Nope no sucess with CLAHE pre-processing either. Also adding rotations makes performance worse for me. Did image augmentation with rotations help you ?", "309303": "Mehul Sampat, I am now using random flip (left right, up down) and random crop (only for image larger than 512x512), seems doing OK.", "309732": "RLstat, thanks for the information. Btw did you use imgaug for augmentation. \nhttps://github.com/aleju/imgaug\nI did this with imgaug: \n\naugmentation = imgaug.augmenters.OneOf([\nimgaug.augmenters.Fliplr(1.0),\nimgaug.augmenters.Affine(rotate=(-10,10)),\nimgaug.augmenters.Flipud(1.0)\n])\n\nI will remove the rotate replace it with a crop function now.", "309806": "Mehul Sampat, no I've not used that repository. I remember I've seen some other post that excessive augmentation won't really help. I am kind of keeping it simple for now. I think post processing is probably more important.", "309920": "RLstat\nMay I know your LB score without post processing?", "310025": "Hi, RLstat. I am using flip too. But when I use random_crop, It hurt my performance. Does it work for you? How do you do that? Can you give us some hints about this?", "310108": "Caibimiao, about 0.43 to 0.45.", "310115": "Hi, @Zheng Li, giving the credit to @john1231983 at this discussion thread: https://www.kaggle.com/c/data-science-bowl-2018/discussion/52989. \n\nI pretty much followed his methods for random cropping, i.e, just crop images with both height and width larger than 512. Perform resizing with other images. After that, I will do flipping for the resized/cropped images. \n\nAlso, I just realized this discussion: https://www.kaggle.com/c/data-science-bowl-2018/discussion/53798. It mentioned the potential problem with the resizing in matterport's implementation and how it might affect the performance if mixing resizing with cropping. I will look into more detail of it.", "310125": "RLstat yup agree with the credit to @john1231983 for the great discussion thread.\nAlso regarding the issue in: https://www.kaggle.com/c/data-science-bowl-2018/discussion/53798\nthis is fixed in the main repo. you can git clone https://github.com/matterport/Mask_RCNN.git to get the latest code.", "310455": "Hi, RLstat. I have changed my code according to the new update but it still work badly. Can you give me some tips about learning schedule and learning rate?", "310545": "Zheng Li, well, I am trying out the scheduling mentioned by @john1231983 and @xiapidan\nearly in this thread , and I am using adam. I've not achieved score better than 0.45 yet with just Mask-RCNN, and rank lower than you on LB. So I am not sure whether my training scheduling will really help you.", "310604": "Hi RLstat. I have tried many methods to make this function work, but none of them really improve the result. But this should theoretically improve the accuracy. To be honest, I am very curious about this. I really want to know why this function doesn’t work for me. So can you leave your email address or any contact methods? My email is mdlszhengli@gmail.com. Really want to have a discuss with you and I will share you my experiments. Thanks Zheng" }, "source": "meta" }