issues with notebook code

[mzaradzki]

issues with notebook code

[mzaradzki]

Hi

The hard-coded "if False" in the notebook has to be set to True or False depending in the .mat file you downloaded from the previous text cell in the notebook:

Download from console with :
wget http://www.vlfeat.org/matconvnet/models/vgg-face.mat
Alternatively :
wget http://www.robots.ox.ac.uk/~vgg/software/vgg_face/src/vgg_face_matconvnet.tar.gz

I believe as it is (if False) it assumes you downloaded the second link: depending on the link the file is not structured the same way

If this comment does not help (let me know) I will download again the .mat file to clarify this point further

[yfoo]

Hi, Ya, so fooling with settting the first block to False, you get: Traceback (most recent call last): File "keras-vg-demo.py", line 93, in <module> net = data['net'][0,0] KeyError: 'net And setting to true you get: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations. Traceback (most recent call last): File "keras-vg-demo.py", line 143, in <module> copy_mat_to_keras(facemodel) File "keras-vg-demo.py", line 136, in copy_mat_to_keras assert (f_l_weights.shape == kmodel.layers[kindex].get_weights()[0].shape) AssertionError The file name is: vgg-face.mat Thanks! I can send you the code I'm running if that would help, but it's pretty much 100% taken from the notebook.

…

On Sun, Apr 16, 2017 at 10:16 AM, m.zaradzki ***@***.***> wrote: Hi The hard-coded "if False" in the notebook has to be set to True or False depending in the .mat file you downloaded from the previous text cell in the notebook: Download from console with : wget http://www.vlfeat.org/matconvnet/models/vgg-face.mat Alternatively : wget http://www.robots.ox.ac.uk/~vgg/software/vgg_face/src/vgg_ face_matconvnet.tar.gz I believe as it is (if False) it assumes you downloaded the second link: depending on the link the file is not structured the same way If this comment does not help (let me know) I will download again the .mat file to clarify this point further — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#5 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ARYHsOe-F6_g5_B-ZXrnDCcAFNin4tSHks5rwk0HgaJpZM4M-q1T> .

[mzaradzki]

Hi,

The new bug you have is probably related to the settings in the keras.json file that is documented here;
https://keras.io/backend/

One variable that impacts the convolution tensor shapes is: "image_data_format"
What value does it have in your configuration ?

When I wrote this notebook I think I used "th" (legacy values) for it as I also used Theano. Anyway it is easy to translate from one value to another one as this is just about permuting the dimension orders.

This is a point I needs to clarify in the documentation cells of the notebook to facilitates replication !

In the Keras help link above you will find function "set_image_dim_ordering" that should tweak it without changing the Json file.

If you dont manage to fix it with this function I will run it with your settings to see.

[yfoo]

Ok I'll take a look at that setting and report back thanks!

…

On Apr 17, 2017, at 6:02 AM, m.zaradzki ***@***.***> wrote: Hi, The new bug you have is probably related to the settings in the keras.json file that is documented here; https://keras.io/backend/ One variable that impacts the convolution tensor shapes is: "image_data_format" What value does it have in your configuration ? When I wrote this notebook I think I used "th" (legacy values) for it as I also used Theano. Anyway it is easy to translate from one value to another one as this is just about permuting the dimension orders. This is a point I needs to clarify in the documentation cells of the notebook to facilitates replication ! In the Keras help link above you will find function "set_image_dim_ordering" that should tweak it without changing the Json file. If you dont manage to fix it with this function I will run it with your settings to see. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[yfoo]

Hi, this is what I have: { "epsilon": 1e-07, "floatx": "float32", "image_data_format": "channels_last", "backend": "tensorflow"

…

On Mon, Apr 17, 2017 at 6:02 AM, m.zaradzki ***@***.***> wrote: Hi, The new bug you have is probably related to the settings in the keras.json file that is documented here; https://keras.io/backend/ One variable that impacts the convolution tensor shapes is: "image_data_format" What value does it have in your configuration ? When I wrote this notebook I think I used "th" (legacy values) for it as I also used Theano. Anyway it is easy to translate from one value to another one as this is just about permuting the dimension orders. This is a point I needs to clarify in the documentation cells of the notebook to facilitates replication ! In the Keras help link above you will find function "set_image_dim_ordering" that should tweak it without changing the Json file. If you dont manage to fix it with this function I will run it with your settings to see. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#5 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ARYHsNJhWa5GQ3VZUWcnvFufoFx4l45lks5rw2LogaJpZM4M-q1T> .

[mzaradzki]

Ok, great, we are on the right track!

It needs something like the following (A) or (B) at the beginning of the notebook:

A)

from keras import backend as K
K.set_image_data_format( 'channels_first' )

B)

from keras import backend as K
K.set_image_dim_ordering( 'th' )

Reading the doc I think the 2 things are equivalent but that (A) is the most recent version

[yfoo]

Hi, Both of those settings cause the following early on: tensorflow/python/framework/common_shapes.py", line 676, in _call_cpp_shape_fn_impl raise ValueError(err.message) ValueError: Negative dimension size caused by subtracting 2 from 1 for 'max_pooling2d_2/MaxPool' (op: 'MaxPool') with input shapes: [?,112,1,128]. Where as the program gets farther without those with the original error: Traceback (most recent call last): File "keras-vg-demo.py", line 146, in <module> copy_mat_to_keras(facemodel) File "keras-vg-demo.py", line 139, in copy_mat_to_keras assert (f_l_weights.shape == kmodel.layers[kindex].get_weights()[0].shape) AssertionError

…

On Mon, Apr 17, 2017 at 12:34 PM, m.zaradzki ***@***.***> wrote: Ok, great, we are on the right track! It needs something like the following (A) or (B) at the beginning of the notebook: A) from keras import backend as K K.set_image_data_format( 'channels_first' ) B) from keras import backend as K K.set_image_dim_ordering( 'th' ) Reading the doc I think the 2 things are equivalent but that (A) is the most recent version — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#5 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ARYHsHG_gC8WgV9hyyMTSDWNqIfWc7cyks5rw77CgaJpZM4M-q1T> .

[mzaradzki]

I just debugged it running Keras 2, TensorFlow and "image_data_format":"channels_last"

Code has been pushed on Git repo.

The last bug you had came from the fact the Input tensor was potentially not ordered the right way depending on the "image_data_format" value at this line:
mdl.add( Permute((1,2,3), input_shape=(224,224,3)) ) # WARNING : 0 is the sample dim

Similarly in "copy_mat_to_keras" function there was a potential axis ordering issue.
As well as in "pred" function.

So I added the necessary information and code changes.

I used your default value for "image_data_format":"channels_last" as it seems to be more common.

Also if you use Keras 1 in "convblock" and "vgg_face_blank" functions I left as comments the Conv2D syntax.

[yfoo]

Thank you for the update! I'll check it out!

…

On Tue, Apr 18, 2017 at 12:58 AM, m.zaradzki ***@***.***> wrote: I just debugged it running *Keras 2*, *TensorFlow* and *"image_data_format":"channels_last"* Code has been pushed on Git repo. The last bug you had came from the fact the Input tensor was potentially not ordered the right way depending on the "image_data_format" value at this line: mdl.add( Permute((1,2,3), input_shape=(224,224,3)) ) # WARNING : 0 is the sample dim Similarly in "copy_mat_to_keras" function there was a potential axis ordering issue. As well as in "pred" function. So I added the necessary information and code changes. I used your default value for *"image_data_format":"channels_last"* as it seems to be more common. Also if you use Keras 1 in "convblock" and "vgg_face_blank" functions I left as comments the Conv2D syntax. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#5 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ARYHsN7CWuO8RrDfEtayZm-IzyZSJOg2ks5rxG03gaJpZM4M-q1T> .

[yfoo]

Thanks, it works great now! Now to get in and get dirty, and figure this stuff out more deeply :) Can you point me in the direction of how to go about training a network with a new input? For instance, say the vggface network didn't have george clooney included, and you'd like to be able to classify an image of george clooney correctly. Have a good one! Chris

…

On Tue, Apr 18, 2017 at 11:22 AM, Chris Coffey ***@***.***> wrote: Thank you for the update! I'll check it out! On Tue, Apr 18, 2017 at 12:58 AM, m.zaradzki ***@***.***> wrote: > I just debugged it running *Keras 2*, *TensorFlow* and > *"image_data_format":"channels_last"* > > Code has been pushed on Git repo. > > The last bug you had came from the fact the Input tensor was potentially > not ordered the right way depending on the "image_data_format" value at > this line: > mdl.add( Permute((1,2,3), input_shape=(224,224,3)) ) # WARNING : 0 is the > sample dim > > Similarly in "copy_mat_to_keras" function there was a potential axis > ordering issue. > As well as in "pred" function. > > So I added the necessary information and code changes. > > I used your default value for *"image_data_format":"channels_last"* as > it seems to be more common. > > Also if you use Keras 1 in "convblock" and "vgg_face_blank" functions I > left as comments the Conv2D syntax. > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > <#5 (comment)>, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/ARYHsN7CWuO8RrDfEtayZm-IzyZSJOg2ks5rxG03gaJpZM4M-q1T> > . >