object detection API can not set batch size > 1 for R-FCN training

[luowei1107]

when I use object detection API for R-FCN training, I set batch size greater than 1. An error occur: "Cannot batch tensors with different shapes in component 1. First element had shape [600,750,3] and element 1 had shape [901,600,3]."

How can I modify code to support batch size greater than 1? Please help!

[varun19299]

That's probably because of a keep_aspect_ratio_resizer. That allows for dynamic sizes, hence no batching is permitted.

Use a fixed_shape_resizer instead.

[c329686183]

@luowei1107 I have the same issue, have u solved it ?

[zik1ng]

Thank you for your sloution. It helps me a lot, I can use batch_size 16 for training now.
@varun19299

[luowei1107]

Thanks. I modify the configs which image_resizer{ keep_aspect_ratio_resizer{min_dimension: 600max_dimension: 1024 }} to image_resizer{ fixed_shape_resizer{height: 600 width: 1024 } to solve it.
@varun19299 @c329686183

[amankhandelia]

@varun19299
I have this query, considering the fact that resizing imaging without respecting the aspect ratio will distort the image, will this impact the performance of the model, especially at the time inference (please consider with respect to Mask RCNN)

[varun19299]

Yes, that’s a drawback, but you could probably use another pre-processing step where you could resize with padding / another strategy. Regardless there will be a distortion.

…

On 06-Sep-2018, at 7:05 PM, Aman Khandelia ***@***.***> wrote: @varun19299 <https://github.com/varun19299> I have this query, considering the fact that resizing imaging without respecting the aspect ratio will distort the image, will this impact the performance of the model, especially at the time inference (please consider with respect to Mask RCNN) — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#4831 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AUME57GPyiN1QuxIXNGsJ0MSZhv0RRZlks5uYSSqgaJpZM4VV7gy>.

[Javadzb]

I kept the image size fixed and increased the batch size for training, however it has not affected evaluation! Is there any way to increase batch size for evaluation/testing?

[veonua]

@varun19299 could you please tell more about pre-processing steps. is it part of the config, or it must be done in step where we create record files.

[andretkj1]

That's probably because of a keep_aspect_ratio_resizer. That allows for dynamic sizes, hence no batching is permitted.

Use a fixed_shape_resizer instead.

my config file already use fixed_shape_resizer {
height: 640
width: 640}
but still can change my batch size other then 1. when i change to other then 1 i got omm error. i am using faster_rcnn_inception_resnet_v2_640x640_coco17_tpu-8. @varun19299