FastMaskRCNN
FastMaskRCNN copied to clipboard
Where We Are
Hi, i'm going to put some effort into making the work over the next semester, but i'm unsure what exactly is working and what is not.
As I understand it, the only difference between this and Faster R-CNN is the use of RoIAlign instead of RoIPool and the mask prediction branch. Are we sure everything execpt those two parts work as expected?
In the Mask R-Cnn paper they train the region proposal network seperately from the rest. This code does not seem to do that? Are we sure it works with joint training?
PS: I've trained the model as it is (with the newest commit) for about 250k iterations, the total_loss converges to about 0.4, is that to be expected?
PPS: i've attached some examples of predictions from the training set
@CharlesShang @amirbar @souryuu
Another different btw mask and faster-rcnn in this implementation is Feature Pyramid Network. This is not trival. Most of the bugs come from this FPN part. I actually think it is better to replicate the results of Mask rcnn without FPN first.
You can train only rpn network by set loss weight of rcnn and mask to 0. I recommend you to do so because joint training does not always converge.
For training loss, the number is dependent on many hyperparameters. It is good that the number is converged but it does not guarantee that you get a good result.
I am currently also working on the getting the RPN part of the project working. Since there has not been a PR merged in a while, what is the expected behavior of the training processing following the README? I am using a 1080Ti and was not able to train without error. I was wondering how much did you guys put in to get the results such as above.
P.S. This is the error, I think it has something to do with parsing the input incorrectly but I am not sure if it's particular to my setup a problem in origin master.
tensorflow.python.framework.errors_impl.InvalidArgumentError: Input to reshape is a tensor with 1 values, but the requested shape has 0
[[Node: Reshape = Reshape[T=DT_FLOAT, Tshape=DT_INT32, _device="/job:localhost/replica:0/task:0/gpu:0"](DecodeRaw_1/_555, Reshape/shape)]]
[[Node: ReverseV2/_599 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/cpu:0", send_device="/job:localhost/replica:0```
You need to use tensorflow 1.1 (I had the exact same problem).. this should be in the README.md
From: William Xie Sent: Thursday, September 7, 21:13 Subject: Re: [CharlesShang/FastMaskRCNN] Where We Are (#141) To: CharlesShang/FastMaskRCNN Cc: stillwalker1234, Author
I am currently also working on the getting the RPN part of the project working. Since there has not been a PR merged in a while, what is the expected behavior of the training processing following the README? I am using a 1080Ti and was not able to train without error. I was wondering how much did you guys put in to get the results such as above. P.S. This is the error, I think it has something to do with parsing the input incorrectly but I am not sure if it's particular to my setup a problem in origin master. tensorflow.python.framework.errors_impl.InvalidArgumentError: Input to reshape is a tensor with 1 values, but the requested shape has 0 [[Node: Reshape = Reshape[T=DT_FLOAT, Tshape=DT_INT32, _device="/job:localhost/replica:0/task:0/gpu:0"](DecodeRaw_1/_555, Reshape/shape)]] [[Node: ReverseV2/_599 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/cpu:0", send_device="/job:localhost/replica:0``` — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHubhttps://github.com/CharlesShang/FastMaskRCNN/issues/141#issuecomment-327896310, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AISpV6eG6t5-k_PghwqeI737YL-FMjHAks5sgEBfgaJpZM4POHBN.
@stillwalker1234 You suggestion was spot on. I was running 1.3 and was in the process of digging deep into the code to see what's wrong. Thanks a lot and I agree that it should be noted in the README until the the code gets updated.
hello @stillwalker1234 , I've trained the model but I'm searching the way to test the trained model and how to get the bounding box and the segmentation . Could you help please ? thnx in advance
can you please share how to test with static images. I tried but im getting this error