C3D has worked for the first time

[carsonDB]

I ran the network with:

• 4 conv layers (each following RELU and Max Pooling)
• 2 fully connected layers

  Learning rate: 0.003

  dataset: UCF-11

Here is the training result after running for 18 hours.

C3D_train.tar.gz

[zbwglory]

It looks fine for me. For the arch, it should be conv->pool->ReLU. You can now try to run UCF-101 split1. You should follow the arch of the girl's paper.

[carsonDB]

@zbwglory Fine, I'll tune my parameters in correspondence with that paper's.

But as for the order of layers. Here is a quote from that paper (at page 4).

Each convolutional layer is followed by a rectified linear unit (ReLU) and a space-time max pooling layer