tf.keras version that allows any input resolution and doesn't use Lambda layers
#16
Conversation
| import tensorflow as tf | ||
|
|
||
| resnet50_Weights_path='./pretrained_model/resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5' | ||
| IMAGE_ORDERING ='channels_last' |
There was a problem hiding this comment.
This should be passed as a variable, and ideally use tensorflow's default.
TODO: default to image_data_format from keras' config (~/.keras/keras.json)
|
|
||
| resnet50_Weights_path='./pretrained_model/resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5' | ||
| IMAGE_ORDERING ='channels_last' | ||
| MERGE_AXIS=-1 |
There was a problem hiding this comment.
This should follow from data_format, e.g. should always be the channel-axis.
| return x | ||
|
|
||
| def identity_block(input_tensor, kernel_size, filters, stage, block): | ||
| def identity_block(input_tensor, kernel_size, filters, stage, block, data_format='channels_last'): |
There was a problem hiding this comment.
Added data_format as a parameter everywhere, rather than the global constant IMAGE_FORMAT. Also used the same name as tf does for transparency.
|
|
||
| x = ZeroPadding2D((3, 3), data_format=IMAGE_ORDERING)(img_input) | ||
| x = Conv2D(64, (7, 7), data_format=IMAGE_ORDERING, strides=(2, 2),kernel_regularizer=l2(weight_decay), name='conv1')(x) | ||
| class PadMultiple(Layer): |
There was a problem hiding this comment.
This pads to a multiple of 32 or whatever is specified in dims.
| padded_to_multiple = PadMultiple((32,32))(img_input) | ||
|
|
||
| bn_axis = 3 if data_format == 'channels_last' else 1 | ||
| merge_axis = 3 if data_format == 'channels_last' else 1 |
There was a problem hiding this comment.
Is merge_axis ever something else than the channel dimension?
| model=Model( img_input , x ).load_weights(resnet50_Weights_path) | ||
| Model(img_input, x).load_weights(resnet50_Weights_path) | ||
|
|
||
| if light_version: |
There was a problem hiding this comment.
Duplicate code complicates code maintenance, so I merged the light version in here.
|
Hi @prhbrt, thanks a lot for looking into this and for contributing! FYI, we are planning to update and refactor this repo and integrate the model training code with https://github.com/qurator-spk/eynollah for future maintenance, so this comes in very handy. My colleagues @vahidrezanezhad and @michalbubula will be working on this - although due to various reasons, we likely won't be able to get our hands dirty much before March. But we will try our best to review and merge any contributions also beforehand. Btw the two models that you were missing should be available from our HF:
|
|
@cneud Understood! Could you in the meantime provide a list of all model-architectures used for eynollah (specifically python code)? I couldn't find python-code for the column classifier in particular, so that one still has fixed dimensions. Also note that this yolo-version might give slightly different outputs as your patching example, due to boundary conditions. Thank you in advance! |
@prhbrt |
|
I have tried to port this over to eynollah here: qurator-spk/eynollah#202 If I messed it up, this PR is still around but I am closing so we can archive the repository. |
4 participants
Since the UNET architecture only uses layers that can scale with the image dimensions, the fixed dimensions seem artificial. I've added a zero-padding layer that increases the dimensions to the nearest multiple of 32. The padding is cut off in the end.
Moreover, I've removed the lambda-layer, as it creates marshaling warning, and used
ZeroPadding2D's asymmetric padding feature. This removes a warning uponload_model.I converted the eynollah-models to this architecture, and they should load without warnings now and use the
tensorflow.kerasAPI and can be found here.This might allow you to skip the patching as used in Eynollah and speed up the whole process. Please let me know what you think.
Notes and sanity checks:
eynollah-main-regions_20220314andeynollah-textline_light_20210425, so I couldn't convert them.