cool
", "rawMarkdown": "cool", "votes": null }, { "id": "282346", "postDate": "02/13/2018 22:10:48", "content": "Thanks for the link.
", "rawMarkdown": "Thanks for the link.", "votes": null }, { "id": "285045", "postDate": "02/19/2018 06:01:12", "content": "Anybody else tried this out ? I can build tfRecords fairly easily which is often a barrier for using tensorflow +Slim models. so I did that, followed the instructions on setting up the model to be best of my ability and ran some train and evaluations. But I'm getting really lousy results - it trains really slowly - even after 140,000 steps ( 1/2 sec a step for me, the total loss is still quite high - around .5, .6. Then at evaluation ( which runs really slowly at that ) the mAP using 'coco_mask_metrics' for small objects is .003 ! and when I look at the predicted boxes on some sample images in tensorboard, the fit is truly awful . So far I definitely don't get anything worth submitting to the LB. Anybody else doing better ?
", "rawMarkdown": "Anybody else tried this out ? I can build tfRecords fairly easily which is often a barrier for using tensorflow +Slim models. so I did that, followed the instructions on setting up the model to be best of my ability and ran some train and evaluations. But I'm getting really lousy results - it trains really slowly - even after 140,000 steps ( 1/2 sec a step for me, the total loss is still quite high - around .5, .6. Then at evaluation ( which runs really slowly at that ) the mAP using 'coco_mask_metrics' for small objects is .003 ! and when I look at the predicted boxes on some sample images in tensorboard, the fit is truly awful . So far I definitely don't get anything worth submitting to the LB. Anybody else doing better ?", "votes": null }, { "id": "285076", "postDate": "02/19/2018 08:41:01", "content": "I did not manage to run masks yet, but I tried with faster rcnn resnet101. Check results if it helps.
", "rawMarkdown": "I did not manage to run masks yet, but I tried with faster rcnn resnet101. Check [results](https://drive.google.com/drive/folders/1bsP2kUQjwObL0PQd-nrae1_tRpiB1_99?usp=sharing) if it helps.", "votes": null }, { "id": "285078", "postDate": "02/19/2018 08:52:56", "content": "Welcome.
", "rawMarkdown": "Welcome.", "votes": null }, { "id": "285304", "postDate": "02/19/2018 21:36:35", "content": "thanks for sharing.. I got nowhere near that. before you edited your message, I think you were asking for how the masks get encoded.. I use this block of code :
\n\n\n", "rawMarkdown": "thanks for sharing.. I got nowhere near that. before you edited your message, I think you were asking for how the masks get encoded.. I use this block of code : \n\n> mask_images=[]\n> for mask_file in next(os.walk(path + masks_dir))[2]:\n> mask_path = path + masks_dir + mask_file\n> with tf.gfile.GFile(mask_path, 'rb') as fid:\n> encoded_mask_png = fid.read()\n> encoded_png_io = io.BytesIO(encoded_mask_png)\n> mask_ = PIL.Image.open(encoded_png_io)\n> if mask_.format != 'PNG':\n> raise ValueError('Mask format not PNG')\n> mask_np = np.asarray(mask_)\n> # google models want a 0, 1 mask file \n> mask_np=(mask_np/255).astype(np.uint8)\n> img = PIL.Image.fromarray(mask_np)\n> output = io.BytesIO()\n> img.save(output, format='PNG')\n> mask_images.append(output.getvalue())\n> \n> instance_masks.append(mask_np) \n> class_ids.append(1)\n> class_texts.append(\"cells\".encode('utf-8'))\n> masks = np.stack(instance_masks, axis=2)\n> xmins, xmaxs, ymins, ymaxs = get_mask_info(masks)\n> tf_example = tf.train.Example(features=tf.train.Features(feature={\n> 'image/height': dataset_util.int64_feature(height), ... 'image/object/mask' : >dataset_util.bytes_list_feature(mask_images), \n> \n> }))", "votes": null }, { "id": "285323", "postDate": "02/19/2018 23:21:10", "content": "\n \nmask_images=[]\n for mask_file in next(os.walk(path + masks_dir))[2]:\n mask_path = path + masks_dir + mask_file\n with tf.gfile.GFile(mask_path, 'rb') as fid:\n encoded_mask_png = fid.read()\n encoded_png_io = io.BytesIO(encoded_mask_png)\n mask_ = PIL.Image.open(encoded_png_io)\n if mask_.format != 'PNG':\n raise ValueError('Mask format not PNG')\n mask_np = np.asarray(mask_)\ngoogle models want a 0, 1 mask file
\n\n\nmask_np=(mask_np/255).astype(np.uint8)\n img = PIL.Image.fromarray(mask_np)\n output = io.BytesIO()\n img.save(output, format='PNG')\n mask_images.append(output.getvalue())\n\n instance_masks.append(mask_np) \n class_ids.append(1)\n class_texts.append(\"cells\".encode('utf-8'))\n masks = np.stack(instance_masks, axis=2)\n xmins, xmaxs, ymins, ymaxs = get_mask_info(masks)\n tf_example = tf.train.Example(features=tf.train.Features(feature={\n 'image/height': dataset_util.int64_feature(height), ... 'image/object/mask' : >dataset_util.bytes_list_feature(mask_images), \n\n }))\n
Hello all, could you tell me hoe to prepaire dataset? Does it need to follow coco dataset rule ? Thanks
", "rawMarkdown": "Hello all, could you tell me hoe to prepaire dataset? Does it need to follow coco dataset rule ? Thanks", "votes": null }, { "id": "285346", "postDate": "02/20/2018 02:17:02", "content": "There are several examples out there but I found this one easiest to follow:Using your own dataset
\n\nIt doesn't tell you how to add masks to the TFrecord. for that you need to populate another tag called
\n\n\n", "rawMarkdown": "There are several examples out there but I found this one easiest to follow:[Using your own dataset][1]\n\n\n [1]: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/using_your_own_dataset.md\n\nIt doesn't tell you how to add masks to the TFrecord. for that you need to populate another tag called\n> 'image/object/mask' : dataset_util.bytes_list_feature(mask_images),", "votes": null }, { "id": "285379", "postDate": "02/20/2018 04:55:09", "content": "\n'image/object/mask' : dataset_util.bytes_list_feature(mask_images), \n
seems like I've got past my issues with training. After 16K steps here's Tensorboard Images
\n\nand here's the .. mAP results from coco_mask_metrics
", "rawMarkdown": "seems like I've got past my issues with training. After 16K steps here's [Tensorboard Images][1]\n\n\nand here's the .. [mAP results from coco_mask_metrics][2]\n\n\n [1]: https://drive.google.com/file/d/1bwZ0P2AmXB7zVkDBttDNZ32jkmHVz9Ec/view?usp=sharing\n [2]: https://drive.google.com/file/d/1zKHN1eUPbCtjIN7tPq3phtxBSOH1VjIf/view?usp=sharing", "votes": null }, { "id": "285400", "postDate": "02/20/2018 06:20:20", "content": "Great. But it looks too low mAP. Do you use DSB 2018 dataset?
", "rawMarkdown": "Great. But it looks too low mAP. Do you use DSB 2018 dataset?", "votes": null }, { "id": "285420", "postDate": "02/20/2018 07:53:36", "content": "You can also check different conversion scripts for different datasets in models/research/object_detection/dataset_tools.
", "rawMarkdown": "You can also check different conversion scripts for different datasets in [models/research/object_detection/dataset_tools](https://github.com/tensorflow/models/tree/master/research/object_detection/dataset_tools).", "votes": null }, { "id": "285422", "postDate": "02/20/2018 08:00:53", "content": "Nice to hear that skthetwo. What base model did you use, did you use pretrained weights or did you trained it from scratch, did you change max_detections_per_class (by default is 100) under second_stage_post_processing?
", "rawMarkdown": "Nice to hear that skthetwo. What base model did you use, did you use pretrained weights or did you trained it from scratch, did you change max_detections_per_class (by default is 100) under second_stage_post_processing?", "votes": null }, { "id": "285500", "postDate": "02/20/2018 13:01:38", "content": "Thanks. I will try it. Could you tell me how about LB did you achieve with this code? I am wondering about pytorch, keras and tensorflow framework?
", "rawMarkdown": "Thanks. I will try it. Could you tell me how about LB did you achieve with this code? I am wondering about pytorch, keras and tensorflow framework?", "votes": null }, { "id": "285526", "postDate": "02/20/2018 14:03:53", "content": "I used the directives:
\n\n\n\n\nfine_tune_checkpoint: \"/home/shantanu/models/research/output/TFR/model.ckpt\"\n from_detection_checkpoint: true
\n
when model.ckpt comes from faster_rcnn_resnet101_coco_11_06_2017 so presumably I did use pre-trained weights. And I set the maxes to :
\n\n\n\n\n\nmax_detections_per_class: 300\n max_total_detections: 300\n
and since my num_classes is 1, that's 300 detections I reckon. thanks btw for sharing your detection results - it validated for me that the code \"works\" so I had an incentive to debug MY code.
", "rawMarkdown": "I used the directives:\n >fine_tune_checkpoint: \"/home/shantanu/models/research/output/TFR/model.ckpt\"\n >from_detection_checkpoint: true\n\nwhen model.ckpt comes from faster_rcnn_resnet101_coco_11_06_2017 so presumably I did use pre-trained weights. And I set the maxes to :\n\n > max_detections_per_class: 300\n > max_total_detections: 300\n \nand since my num_classes is 1, that's 300 detections I reckon. thanks btw for sharing your detection results - it validated for me that the code \"works\" so I had an incentive to debug MY code.", "votes": null }, { "id": "285532", "postDate": "02/20/2018 14:33:06", "content": "yes I did use the DSB2018 dataset and yes the mAP looks too low - too low for me to even bother submitting it to the LB. But you should have seen the mAP results before I fixed a couple of bugs in my code ! Also I bear in mind though that this is after only 16K steps when the total loss during training has many steps with losses > 1 and some losses > 5 even and config settings more or less at default ! Lets see how well is does after more training and tweaking.
", "rawMarkdown": "yes I did use the DSB2018 dataset and yes the mAP looks too low - too low for me to even bother submitting it to the LB. But you should have seen the mAP results before I fixed a couple of bugs in my code ! Also I bear in mind though that this is after only 16K steps when the total loss during training has many steps with losses > 1 and some losses > 5 even and config settings more or less at default ! Lets see how well is does after more training and tweaking.", "votes": null }, { "id": "287197", "postDate": "02/23/2018 20:55:02", "content": "Hi , do you have any update LB score? I am chosing between tensorflow or keras implementation. For now, keras may give better performance 0.416
", "rawMarkdown": "Hi , do you have any update LB score? I am chosing between tensorflow or keras implementation. For now, keras may give better performance 0.416", "votes": null } ], "comments": [ { "id": 282287, "author_name": "wayne999", "author_url": "", "post_date": "02/13/2018 20:43:20", "content": "cool
", "votes": null, "replies": [] }, { "id": 282346, "author_name": "isathish", "author_url": "", "post_date": "02/13/2018 22:10:48", "content": "Thanks for the link.
", "votes": null, "replies": [ { "id": 285078, "author_name": "maesaer", "author_url": "", "post_date": "02/19/2018 08:52:56", "content": "Welcome.
", "votes": null, "replies": [] } ] }, { "id": 285045, "author_name": "skthetwo", "author_url": "", "post_date": "02/19/2018 06:01:12", "content": "Anybody else tried this out ? I can build tfRecords fairly easily which is often a barrier for using tensorflow +Slim models. so I did that, followed the instructions on setting up the model to be best of my ability and ran some train and evaluations. But I'm getting really lousy results - it trains really slowly - even after 140,000 steps ( 1/2 sec a step for me, the total loss is still quite high - around .5, .6. Then at evaluation ( which runs really slowly at that ) the mAP using 'coco_mask_metrics' for small objects is .003 ! and when I look at the predicted boxes on some sample images in tensorboard, the fit is truly awful . So far I definitely don't get anything worth submitting to the LB. Anybody else doing better ?
", "votes": null, "replies": [ { "id": 285076, "author_name": "maesaer", "author_url": "", "post_date": "02/19/2018 08:41:01", "content": "I did not manage to run masks yet, but I tried with faster rcnn resnet101. Check results if it helps.
", "votes": null, "replies": [] }, { "id": 285304, "author_name": "skthetwo", "author_url": "", "post_date": "02/19/2018 21:36:35", "content": "thanks for sharing.. I got nowhere near that. before you edited your message, I think you were asking for how the masks get encoded.. I use this block of code :
\n\n\n", "votes": null, "replies": [] }, { "id": 285379, "author_name": "skthetwo", "author_url": "", "post_date": "02/20/2018 04:55:09", "content": "\n \nmask_images=[]\n for mask_file in next(os.walk(path + masks_dir))[2]:\n mask_path = path + masks_dir + mask_file\n with tf.gfile.GFile(mask_path, 'rb') as fid:\n encoded_mask_png = fid.read()\n encoded_png_io = io.BytesIO(encoded_mask_png)\n mask_ = PIL.Image.open(encoded_png_io)\n if mask_.format != 'PNG':\n raise ValueError('Mask format not PNG')\n mask_np = np.asarray(mask_)\ngoogle models want a 0, 1 mask file
\n\n\nmask_np=(mask_np/255).astype(np.uint8)\n img = PIL.Image.fromarray(mask_np)\n output = io.BytesIO()\n img.save(output, format='PNG')\n mask_images.append(output.getvalue())\n\n instance_masks.append(mask_np) \n class_ids.append(1)\n class_texts.append(\"cells\".encode('utf-8'))\n masks = np.stack(instance_masks, axis=2)\n xmins, xmaxs, ymins, ymaxs = get_mask_info(masks)\n tf_example = tf.train.Example(features=tf.train.Features(feature={\n 'image/height': dataset_util.int64_feature(height), ... 'image/object/mask' : >dataset_util.bytes_list_feature(mask_images), \n\n }))\n
seems like I've got past my issues with training. After 16K steps here's Tensorboard Images
\n\nand here's the .. mAP results from coco_mask_metrics
", "votes": null, "replies": [] }, { "id": 285400, "author_name": "john1231983", "author_url": "", "post_date": "02/20/2018 06:20:20", "content": "Great. But it looks too low mAP. Do you use DSB 2018 dataset?
", "votes": null, "replies": [] }, { "id": 285422, "author_name": "maesaer", "author_url": "", "post_date": "02/20/2018 08:00:53", "content": "Nice to hear that skthetwo. What base model did you use, did you use pretrained weights or did you trained it from scratch, did you change max_detections_per_class (by default is 100) under second_stage_post_processing?
", "votes": null, "replies": [] }, { "id": 285526, "author_name": "skthetwo", "author_url": "", "post_date": "02/20/2018 14:03:53", "content": "I used the directives:
\n\n\n\n\nfine_tune_checkpoint: \"/home/shantanu/models/research/output/TFR/model.ckpt\"\n from_detection_checkpoint: true
\n
when model.ckpt comes from faster_rcnn_resnet101_coco_11_06_2017 so presumably I did use pre-trained weights. And I set the maxes to :
\n\n\n\n\n\nmax_detections_per_class: 300\n max_total_detections: 300\n
and since my num_classes is 1, that's 300 detections I reckon. thanks btw for sharing your detection results - it validated for me that the code \"works\" so I had an incentive to debug MY code.
", "votes": null, "replies": [] } ] }, { "id": 285323, "author_name": "john1231983", "author_url": "", "post_date": "02/19/2018 23:21:10", "content": "Hello all, could you tell me hoe to prepaire dataset? Does it need to follow coco dataset rule ? Thanks
", "votes": null, "replies": [ { "id": 285346, "author_name": "skthetwo", "author_url": "", "post_date": "02/20/2018 02:17:02", "content": "There are several examples out there but I found this one easiest to follow:Using your own dataset
\n\nIt doesn't tell you how to add masks to the TFrecord. for that you need to populate another tag called
\n\n\n", "votes": null, "replies": [] }, { "id": 285420, "author_name": "maesaer", "author_url": "", "post_date": "02/20/2018 07:53:36", "content": "\n'image/object/mask' : dataset_util.bytes_list_feature(mask_images), \n
You can also check different conversion scripts for different datasets in models/research/object_detection/dataset_tools.
", "votes": null, "replies": [] }, { "id": 285500, "author_name": "john1231983", "author_url": "", "post_date": "02/20/2018 13:01:38", "content": "Thanks. I will try it. Could you tell me how about LB did you achieve with this code? I am wondering about pytorch, keras and tensorflow framework?
", "votes": null, "replies": [] }, { "id": 285532, "author_name": "skthetwo", "author_url": "", "post_date": "02/20/2018 14:33:06", "content": "yes I did use the DSB2018 dataset and yes the mAP looks too low - too low for me to even bother submitting it to the LB. But you should have seen the mAP results before I fixed a couple of bugs in my code ! Also I bear in mind though that this is after only 16K steps when the total loss during training has many steps with losses > 1 and some losses > 5 even and config settings more or less at default ! Lets see how well is does after more training and tweaking.
", "votes": null, "replies": [] }, { "id": 287197, "author_name": "john1231983", "author_url": "", "post_date": "02/23/2018 20:55:02", "content": "Hi , do you have any update LB score? I am chosing between tensorflow or keras implementation. For now, keras may give better performance 0.416
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "282274": "[link](https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/instance_segmentation.md)", "282287": "cool", "282346": "Thanks for the link.", "285045": "Anybody else tried this out ? I can build tfRecords fairly easily which is often a barrier for using tensorflow +Slim models. so I did that, followed the instructions on setting up the model to be best of my ability and ran some train and evaluations. But I'm getting really lousy results - it trains really slowly - even after 140,000 steps ( 1/2 sec a step for me, the total loss is still quite high - around .5, .6. Then at evaluation ( which runs really slowly at that ) the mAP using 'coco_mask_metrics' for small objects is .003 ! and when I look at the predicted boxes on some sample images in tensorboard, the fit is truly awful . So far I definitely don't get anything worth submitting to the LB. Anybody else doing better ?", "285076": "I did not manage to run masks yet, but I tried with faster rcnn resnet101. Check [results](https://drive.google.com/drive/folders/1bsP2kUQjwObL0PQd-nrae1_tRpiB1_99?usp=sharing) if it helps.", "285078": "Welcome.", "285304": "thanks for sharing.. I got nowhere near that. before you edited your message, I think you were asking for how the masks get encoded.. I use this block of code : \n\n> mask_images=[]\n> for mask_file in next(os.walk(path + masks_dir))[2]:\n> mask_path = path + masks_dir + mask_file\n> with tf.gfile.GFile(mask_path, 'rb') as fid:\n> encoded_mask_png = fid.read()\n> encoded_png_io = io.BytesIO(encoded_mask_png)\n> mask_ = PIL.Image.open(encoded_png_io)\n> if mask_.format != 'PNG':\n> raise ValueError('Mask format not PNG')\n> mask_np = np.asarray(mask_)\n> # google models want a 0, 1 mask file \n> mask_np=(mask_np/255).astype(np.uint8)\n> img = PIL.Image.fromarray(mask_np)\n> output = io.BytesIO()\n> img.save(output, format='PNG')\n> mask_images.append(output.getvalue())\n> \n> instance_masks.append(mask_np) \n> class_ids.append(1)\n> class_texts.append(\"cells\".encode('utf-8'))\n> masks = np.stack(instance_masks, axis=2)\n> xmins, xmaxs, ymins, ymaxs = get_mask_info(masks)\n> tf_example = tf.train.Example(features=tf.train.Features(feature={\n> 'image/height': dataset_util.int64_feature(height), ... 'image/object/mask' : >dataset_util.bytes_list_feature(mask_images), \n> \n> }))", "285323": "Hello all, could you tell me hoe to prepaire dataset? Does it need to follow coco dataset rule ? Thanks", "285346": "There are several examples out there but I found this one easiest to follow:[Using your own dataset][1]\n\n\n [1]: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/using_your_own_dataset.md\n\nIt doesn't tell you how to add masks to the TFrecord. for that you need to populate another tag called\n> 'image/object/mask' : dataset_util.bytes_list_feature(mask_images),", "285379": "seems like I've got past my issues with training. After 16K steps here's [Tensorboard Images][1]\n\n\nand here's the .. [mAP results from coco_mask_metrics][2]\n\n\n [1]: https://drive.google.com/file/d/1bwZ0P2AmXB7zVkDBttDNZ32jkmHVz9Ec/view?usp=sharing\n [2]: https://drive.google.com/file/d/1zKHN1eUPbCtjIN7tPq3phtxBSOH1VjIf/view?usp=sharing", "285400": "Great. But it looks too low mAP. Do you use DSB 2018 dataset?", "285420": "You can also check different conversion scripts for different datasets in [models/research/object_detection/dataset_tools](https://github.com/tensorflow/models/tree/master/research/object_detection/dataset_tools).", "285422": "Nice to hear that skthetwo. What base model did you use, did you use pretrained weights or did you trained it from scratch, did you change max_detections_per_class (by default is 100) under second_stage_post_processing?", "285500": "Thanks. I will try it. Could you tell me how about LB did you achieve with this code? I am wondering about pytorch, keras and tensorflow framework?", "285526": "I used the directives:\n >fine_tune_checkpoint: \"/home/shantanu/models/research/output/TFR/model.ckpt\"\n >from_detection_checkpoint: true\n\nwhen model.ckpt comes from faster_rcnn_resnet101_coco_11_06_2017 so presumably I did use pre-trained weights. And I set the maxes to :\n\n > max_detections_per_class: 300\n > max_total_detections: 300\n \nand since my num_classes is 1, that's 300 detections I reckon. thanks btw for sharing your detection results - it validated for me that the code \"works\" so I had an incentive to debug MY code.", "285532": "yes I did use the DSB2018 dataset and yes the mAP looks too low - too low for me to even bother submitting it to the LB. But you should have seen the mAP results before I fixed a couple of bugs in my code ! Also I bear in mind though that this is after only 16K steps when the total loss during training has many steps with losses > 1 and some losses > 5 even and config settings more or less at default ! Lets see how well is does after more training and tweaking.", "287197": "Hi , do you have any update LB score? I am chosing between tensorflow or keras implementation. For now, keras may give better performance 0.416" }, "source": "meta" }