Giters

experiencor / keras-yolo2

Easy training on custom dataset. Various backends (MobileNet and SqueezeNet) supported. A YOLO demo to detect raccoon run entirely in brower is accessible at https://git.io/vF7vI (not on Windows).

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Inaccurate Result and stepsize

Zumbalamambo opened this issue · comments

commented

Why is that the step size never crosses 32 while it has to be 320? Also the resultant model is so inaccurate

> 
> __________________________________________________________________________________________________
> Layer (type)                    Output Shape         Param #     Connected to                     
> ==================================================================================================
> input_7 (InputLayer)            (None, 416, 416, 3)  0                                            
> __________________________________________________________________________________________________
> model_4 (Model)                 (None, 13, 13, 1024) 50547936    input_7[0][0]                    
> __________________________________________________________________________________________________
> DetectionLayer (Conv2D)         (None, 13, 13, 40)   41000       model_4[1][0]                    
> __________________________________________________________________________________________________
> reshape_2 (Reshape)             (None, 13, 13, 5, 8) 0           DetectionLayer[0][0]             
> __________________________________________________________________________________________________
> input_8 (InputLayer)            (None, 1, 1, 1, 10,  0                                            
> __________________________________________________________________________________________________
> lambda_5 (Lambda)               (None, 13, 13, 5, 8) 0           reshape_2[0][0]                  
>                                                                  input_8[0][0]                    
> ==================================================================================================
> Total params: 50,588,936
> Trainable params: 50,568,264
> Non-trainable params: 20,672
> __________________________________________________________________________________________________
> Epoch 1/53
>  32/320 [==>...........................] - ETA: 6:49 - loss: 10.2203Epoch 00001: val_loss improved from inf to 10.10496, saving model to model.h5
>  32/320 [==>...........................] - ETA: 10:32 - loss: 10.2203 - val_loss: 0.0000e+00Epoch 2/53
>  32/320 [==>...........................] - ETA: 4:31 - loss: 10.0364Epoch 00002: val_loss improved from 10.10496 to 10.03809, saving model to model.h5
>  32/320 [==>...........................] - ETA: 7:14 - loss: 10.0364 - val_loss: 0.0000e+00Epoch 3/53
>  32/320 [==>...........................] - ETA: 5:10 - loss: 10.0151Epoch 00003: val_loss improved from 10.03809 to 10.01886, saving model to model.h5
>  32/320 [==>...........................] - ETA: 7:54 - loss: 10.0151 - val_loss: 0.0000e+00