some question about the capacity of this codes ?

Hi, Q1) is this possible to add or remove a few of layers ? How ?is this possible to freeze the layers ? how ? Q2)how do chose the name of layers : for example , in the ssd512 models , conv2_1 = Conv2D(128, (3, 3), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv2_1')(pool1)

the name of layers must be same name in caffe ? if so , Where do you get the name of the layers?

Q3) my GPU is GTX 1080 (8G) and Ram 16G , Cpu i3 7100k , is this possible to retrain ssd512 on my own dataset?

Q4) what are they fc (6,7) ? is not ssd full convolution ?

fc6 = Conv2D(1024, (3, 3), dilation_rate=(6, 6), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='fc6')(pool5)
 fc7 = Conv2D(1024, (1, 1), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='fc7')(fc6)

Q5) is that possible to retrain this models with TFrecord dataset file ?

Request) please put the ssd512_training.ipynp.

[pierluigiferrari]

1. Adding, removing, and freezing layers works just as with any other Keras model. This is a general Keras question, it doesn’t have anything to do with my SSD implementation in particular, so I would refer you to the Keras API documentation or Stackoverflow to learn how to add or freeze layers.
2. The layer names don’t have to be the same as the Caffe layer names in general, but in my ported weights files I stuck with the original Caffe model layer names, so in order to load these weights by_name, the names of the layers in the model must be the same as the names of the respective layers in the HDF5 weights file. You get the names of a Caffe model from the .prototxt file that defines the Caffe model.
3. Yes.
4. SSD is fully convolutional. fc6 and fc7 are just called that to indicate that they were transformed into convolutional layers from the fc6 and fc7 layers of VGG.
5. This is also a general Keras question, it has nothing to do with my SSD implementation in particular. I’ve recently followed a few Keras and TensorFlow issues that discussed using TFRecord files with Keras models, but I don’t know what the status is there. It is certainly possible by using this SSD implementation inside TensorFlow’s Estimator API, which now fully supports Keras models.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.