Unable to convert caffe ResNet model about keras HOT 5 CLOSED

I have replicated the first error you mention, but it looks like it is not in the first Scale layer, but as this line of Theano's error suggests:

Apply node that caused the error: GpuReshape{4}(scale5a_branch1_gamma, TensorConstant{[1 1 1 7]})

it is in layer 'scale5a_branch1', which is close to the end of the network.

If we look at the result of model.summary():

bn5a_branch1 (BatchNormalization)(None, 2048, 7, 7) 4096 res5a_branch1[0][0]

---

bn5a_branch2c (BatchNormalization(None, 2048, 7, 7) 4096 res5a_branch2c[0][0]

---

scale5a_branch1 (Scale) (None, 2048, 7, 7) 4096 bn5a_branch1[0][0]

---

scale5a_branch2c (Scale) (None, 2048, 7, 7) 4096 bn5a_branch2c[0][0]

---

res5a (Merge) (None, 2048, 7, 7) 0 scale5a_branch1[0][0]
scale5a_branch2c[0][0]

it seems that for some reason it is trying to reshape the 'gamma' parameter, which has a shape of (2048) to (1,1,1,7), when in fact it should reshape it to (1,2048,1,1) for applying the corresponding operations on the input data.

About your tests with Tensorflow I must say that I am not providing compatibility for it, so several unknown errors might arise.

I will continue taking a look at this, if you have any other idea or clue let me know.

from keras.

Any update on this front? I'm getting the same error..

from keras.

I am sorry, I didn't have time to give direct support to this topic and enable the caffe2keras conversion for residual networks.

Although, an implementation of ResNet50 together with its trained weights was recently added to Keras by @fchollet. You can find the info to use it in this link.

from keras.

Hey Marc!

I understand. However Im not trying to convert a resnet pretrained in Imagenet but in places2 (which we then could add to keras). Any idea on when you might have a look at this?

from keras.

I am pretty busy right now, so I'm afraid can't promise anything. Although, PRs are welcome in case anybody is willing to help with this or any other possible functionality.

from keras.