Visualize Object detection model SSD mobilenet v2

General Discussion
,
#1

Can anyone please help me to visualize the inner features of the SSD model at the time of inference, like the feature map of each layer?

2 Likes
#2

BTW I have no clue on how to get access to the last layer of the model. You know class prediction, bbox prediction, etc.

1 Like
#3

You can have access to the inner layers such as the backbone in Keras fashion. First you must run the object detection model similar as the notebook eager_few_shot_od_training_tf2_colab.ipynb, to get the backbone should detection_model.feature_extractor.classifaction_backbone, with that you can extract features.

2 Likes
#4

Thanks for the help, but I did know how to do feature extraction from backbone layers like VGG16, Mobilenet, etc. But I want to do the extraction from the object detection layers.

1 Like
#5

You’re welcome, and I’m also looking at how to extract the object detector layers. If you get an answer, please share it. :smiley:

1 Like
#6

I was tracing back the objectdetection object in the repository of the API. Maybe it was luck, but I found out how to get the object detection layers at least the Prediction head:
detection_model._box_predictor._prediction_heads[‘class_predictions_with_background’]._class_predictor_layers[0]
Hopefully helps

1 Like
#7

my bad my last approach get the layer but no as tensor, I mean the layer doesnt response to layer.outputs[0] because is not a keras tensor to get the feature map. Well I’m still as the last time

1 Like
#8

Considering the feature maps in an SSD model:

What specific feature map do you want take here?

2 Likes
#9

The prediction class layers, but I need at least the last prediction layer as a tensor. I found the prediction_head; this one lacks the output feature. Hence is not compatible to use to get the feature map of the model in that layer.

1 Like
#10

Once you are able to load the detection checkpoints into their respective model class I think it should be possible to access the layers just by their indices.

OD API models may often seem to be a bit opaque. I would also suggest taking a look at the Model Garden object detection models since they seem to be more transparent and accessible with regards to the interpretability aspects:

https://github.com/tensorflow/models/tree/master/official/vision/detection

2 Likes
#11

Thanks, I use to think the same that once a checkpoint is loaded was possible to access the layers by their indices or similar. But was not the case with the object detection API. I will take a look at your proposal nevertheless is frustrating the current state and documentation of the API.

1 Like