Train a model: RCNN accuracy decreases

[r-isr]

Hi, Could you please provide an insight,

In training I get the following phenomena , RPN acuracy is improved but RCN accuracy is decreased. Also note what happens to L1 and Log losses of RCNN part, the first increases and the last decreased.

Do you have any idea which parameter may need to be tuned?

Thanks,

Epoch[0] Batch [100] Speed: 0.95 samples/sec Train-RPNAcc=0.808787, RPNLogLoss=0.439269, RPNL1Loss=0.095613, RCNNAcc=0.874149, RCNNLogLoss=0.446150, RCNNL1Loss=0.086455,
Epoch[0] Batch [2000] Speed: 0.93 samples/sec Train-RPNAcc=0.915812, RPNLogLoss=0.214620, RPNL1Loss=0.061609, RCNNAcc=0.875118, RCNNLogLoss=0.347196, RCNNL1Loss=0.104136,
Epoch[0] Batch [4000] Speed: 0.96 samples/sec Train-RPNAcc=0.933159, RPNLogLoss=0.173305, RPNL1Loss=0.057984, RCNNAcc=0.871077, RCNNLogLoss=0.339611, RCNNL1Loss=0.107748,
Epoch[0] Batch [5000] Speed: 0.98 samples/sec Train-RPNAcc=0.938542, RPNLogLoss=0.160328, RPNL1Loss=0.054555, RCNNAcc=0.869417, RCNNLogLoss=0.333381, RCNNL1Loss=0.109924,
Epoch[0] Batch [5500] Speed: 0.98 samples/sec Train-RPNAcc=0.940943, RPNLogLoss=0.154480, RPNL1Loss=0.053179, RCNNAcc=0.868047, RCNNLogLoss=0.332574, RCNNL1Loss=0.111403,

[jessemelpolio]

Let it train... Don't be impatient. As you know, we've trained for a long time and got the baseline results. It's quite normal in the beginning.