Where to find zh_lib?

[ghzhao88]

Can't find some functions used in test.py. Please provide the code for the following modules. from zh_lib.dataset import ActionDataset from zh_lib.load_data import load_info from zh_lib.feed_queue import FeedQueue from zh_lib.label_trans import

[vra]

Hi @ghzhao88 , sorry for the inconvience, we are refactoring the project to make it more cleaner and easier to use. Once finished we will notice you.

[ghzhao88]

Thanks for your quick rely. Can you just send me the few files? I like to verify something.

[ghzhao88]

Another question, to my understanding, the flow.txt should include all files in trainlist01.txt, trainlist02.txt and trainlist03.txt and all files in testlist01.txt should not in flow.txt. But v_YoYo_g01_c01.avi in both flow.txt and testlist01.txt.

[vra]

Hi @ghzhao88 , I just upload zh_lib at here, you can download it, unzip it and put it in the root dir of this repo. For your another question, the flow.txt contains all the samples in both training dataset and testing dataset, you can refer https://github.com/USTC-Video-Understanding/I3D_Finetune/blob/master/rmb_lib/action_dataset.py#L50 to see the details.

[ghzhao88]

The zh_lib.zip you uploaded is empty. Please upload again.

[vra]

Sorry for the mistake, please download again.

[ghzhao88]

I downloaded your pre-trained model and tested using testlist01.txt but the mix result is different. Have you verified the result at latest code? Below is my test result. I generated rgb and flow data using dense-flow. Training Split | RGB | Flow | Fusion Split1 | 91.42% | 95.51% | 60.35%

[vra]

Hi @ghzhao88 , According to this issue, the last fc layer should not use a non-linear layer. So we updated test.py and removed the tf.nn.relu parameter in the the last dense layer at here and here, while the uploaded models were generated using previous version of code. Thus an easy way to verify the pre-trained moedels is adding tf.nn.relu to the above mentioned two lines, like this:

            rgb_fc_out = tf.layers.dense(
                rgb_logits_dropout, _CLASS_NUM[dataset], tf.nn.relu, use_bias=True)
            flow_fc_out = tf.layers.dense(
                flow_logits_dropout, _CLASS_NUM[dataset], tf.nn.relu, use_bias=True)

[ghzhao88]

I added tf.nn.relu in test.py but the mixed result is still the same running testlist01.txt.

test accuracy: 0.6088

[panna19951227]

@ghzhao88 I have met the same problem,Have you addressed it?

[ghzhao88]

Not fixed yet.

[MrCuiHao]

@ghzhao88 你好，请问你解决这个混合预测问题了吗？我现在仅仅是测试了rgb流，Top-1错误率比较正常