deep-learning-models

deep-learning-models copied to clipboard

• Squeezenet has very small footprint ( 5.1 MB model file ).
• It has AlexNet accuracy.
• Feedforward prediction time is near 5 FPS with CPU (2.4 GHz Intel Core i5).
• This project contains both Tensorflow and Theano weights ported from Caffe weights.
• Weights are placed in model folder at keras-squeezenet repository.
• Code is rewritten and follows the rules of other model codes in keras/applications.

TO-DO

• Links for the model files should be fixed according to release version of this library (look at squeezenet.py).

Reference:

• SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size

Keras Project Reference:

• keras-squeezenet

Original Project Reference:

• Squeezenet

• Docstring placement and indentation are fixed.
• Download links for weights are added.
• Function arguments are changed.

Sorry for the delay. I do intend on merging this, I just have limited bandwidth.

@rcmalli do you think if it would be a good idea to add also head-less versions of the architecture (see, e.g. ResNet50) so to ease the potential usage in transfer learning?