Bill Ross comments

Results 31 comments of


                                            Bill Ross

expected dense_10 shape (None, <dense_size>) but got array shape (<batch_size>, 1)

Batch size 2048->512: OOM at same Iteration 155, slightly different stats: ``` Limit: 10916963943 InUse: 10899258624 MaxInUse: 10916435968 NumAllocs: 26261 MaxAllocSize: 55796736 ```

expected dense_10 shape (None, <dense_size>) but got array shape (<batch_size>, 1)

Keeping batch 512, going back to just greyscale+kwds, input=2980, dense_size=315: ``` Model parameter count: 1127264 Additional unused parameters: -126234 ``` http://phobrain.com/pr/home/gallery/t2_w_5e-05_1e-08.png

expected dense_10 shape (None, <dense_size>) but got array shape (<batch_size>, 1)

Going back to my GInfty-removed debug version (grey+keywords), batch_size 2048, it runs 10 epochs per Iteration, and winds up after 100 Iterations at loss: 0.0018 - binary_accuracy: 1.0000 - val_loss:...

expected dense_10 shape (None, <dense_size>) but got array shape (<batch_size>, 1)

I noticed the overfit.. it looks like validation accuracy just bounces around after the first few Iterations, like in the graphs above, but I haven't graphed yet. Adding BatchNorm+Dropout(.1) after...

expected dense_10 shape (None, <dense_size>) but got array shape (<batch_size>, 1)

Dropping the BatchNorm, top accuracy is 82%, so I'd run 100 times and use the top 20 models, and without playing with the model much, I've boosted top accuracy from...

expected dense_10 shape (None, <dense_size>) but got array shape (<batch_size>, 1)

I think the loss was changing more before I added the 6 Dense's at the end, or maybe it was the Dropout; from above: ``` loss: 2.3588e-04 - binary_accuracy: 1.0000...

expected dense_10 shape (None, <dense_size>) but got array shape (<batch_size>, 1)

88% is the highest yet. I'm grinding out predictions from my top 100 models, which will take 2 days. ``` Written: vert_v/m_model_28_2048_5_69_86.pairs (1482800) in 0:06:55.677205 ``` I might be getting...

expected dense_10 shape (None, <dense_size>) but got array shape (<batch_size>, 1)

You can try keywords and histograms directly, with green and yellow '+' options in search Mode: AI, and pure histogram nets in different groupings in the Sigma-[0|theta|0theta] options, act now...

expected dense_10 shape (None, <dense_size>) but got array shape (<batch_size>, 1)

I call it a crime against nature that I can't run my 95MB models on a 1030 Gtx or two, and let them chug away on predictions full-tilt, vs. wasting...

How to retrieve bounding boxes/generate bounding boxes

Maybe if you build a temporary global tree, and then descend it while building the final subtree arrays. The global tree could even be done with a different package in...