about the train

[Usernamezhx]

first of all. thanks your cool code. about the net architecture,why not use the softmax after the rnn model? thanks in advance

[breadbread1984]

ctc_loss requires input be the logits.

Hi.When I use your code to train in depth image, I met a problem that the loss drops very slowly and oscillatorily. Finally the accuracy of training set is about 90% and the accuracy of testing set is about 50%.Do you know what the problem might be?Thanks in advance

[breadbread1984]

the reported accuracy of the model trained solely on depth image is around 83%

[Usernamezhx]

sorry to bother you again. when I run the train__r3dcnn.py . it will show me that :


I have installed the tensorflow version 1.4.0. I think this may be about the SparseTensor problem. https://github.com/tensorflow/tensorflow/issues/13766
My label sparseTensor looks like this :
SparseTensorValue(indices=array([[0]], dtype=int64), values=array([6]), dense_shape=array([1], dtype=int64))
as you see my label in int six

[Usernamezhx]

I have solved it by update my tensorflow to version 1.5.0...........

[Usernamezhx]

sorry. InvalidArgumentError (see above for traceback): Input to reshape is a tensor with 12042240 values, but the requested shape has 3010560. i change the batch size get the same erro . and I set the single nvidia card get the same erro. when i load my data: data.get_shape().as_list() show that : [?, 10,8,112,112,3]

[Usernamezhx]

yeah, solved . change the code from clips = tf.decode_raw(feature['clips'],out_type = tf.uint8) to clips = tf.decode_raw(feature['clips'],out_type = tf.int32)

[castiel520]

Hi.When I use your code to train in depth image, I met a problem that the loss drops very slowly and oscillatorily. Finally the accuracy of training set is about 90% and the accuracy of testing set is about 50%.Do you know what the problem might be?Thanks in advance

@panxuezm @breadbread1984 Hi guys, I managed to train the network 50000 steps, but i'm confused about how to get the train/test accuracy, here is my eval output:
{'loss': 0.04681665, 'global_step': 290862, 'mean_edit_distance': 0.0}

[breadbread1984]

you can modify the returned estimator in eval mode to print the metric of accuracy.

[castiel520]

@breadbread1984 I manage to get it work and the final accuracy is around 0.28，thanks, I wonder the best accuracy you got on the dataset?

[breadbread1984]

would you try the model I shared at the link on README.md? 在 2018-11-02 18:09:24，"Castiel" <notifications@github.com> 写道： @breadbread1984 I manage to get it work and the final accuracy is around 0.28，thanks, I wonder the best accuracy you got on the test set? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

[castiel520]

@breadbread1984 yes, i tried your C3D model https://github.com/breadbread1984/c3d and download from baidu cloud.

[sainisanjay]

@castiel520 @breadbread1984 May i know how to download pre-trained C3D Model from from baidu cloud. When i tried it asked Chinese phone number for registration. Any idea how to download???

[breadbread1984]

would you suggest a file sharing platform other than dropbox, google drive which I can't access from my country. I will try to share the file with it.