automl icon indicating copy to clipboard operation
automl copied to clipboard

Published 20 hours ago • verified publisher icon google

Why do I get better result with D0 than D7?

Open saeedda opened this issue 4 years ago • 5 comments

I am following section "5. Inference for images" of README and I tried it with two models: D0 and D7. Unexpectedly the result of D0 is much better than D7. See these results:

Result with D0: img_d0_1920x1280

Result with D7: img_d7_1920x1280

Am I doing something wrong? Is this related to --hparams="image_size=1920x1280"?

saeedda avatar Feb 02 '21 16:02 saeedda

try without --hparams="image_size=1920x1280" default D7 is underneath image_size=1536x1536

fitoule avatar Feb 02 '21 16:02 fitoule

and default D0 is 512 So with D0 and image_size=1920x1280 you set a better definition whenever D7 and image_size=1920x1280 is less than default

fitoule avatar Feb 02 '21 16:02 fitoule

try without --hparams="image_size=1920x1280" default D7 is underneath image_size=1536x1536

I tried without --hparams="image_size=1920x1280" and the result is very similar:

img_d7

saeedda avatar Feb 02 '21 17:02 saeedda

and default D0 is 512 So with D0 and image_size=1920x1280 you set a better definition whenever D7 and image_size=1920x1280 is less than default

How this --hparams="image_size=1920x1280" is used by the code? Is input image get resized to default sizes (e.g. 1536x1536 for D7)? Is there any padding or something else?

saeedda avatar Feb 02 '21 17:02 saeedda

you can see the image resized by the code using the script : #Visualize TFRecord python3 dataset/inspect_tfrecords.py

fitoule avatar Feb 02 '21 17:02 fitoule