An implementation of SqueezeNet in chainer.
Shamelessly forked from ejlb's implementation, with slight changes to layer sizes and support for pretrained weights added. (Weights included.)
Below are some benchmarks on a ImageNet-like dataset (1 million 255x255 images with 128 batch size)
| Model | Size (mb) | Parameters (million) | Accuracy |
|---|---|---|---|
| ZFNet | 117 | 16.42 | 0.5835 |
| SqueezeNet | 4.6 | 1.288 | 0.5207 |
- ZFNet instead of AlexNet (I already had a trained ZFNet model).
- This is the basic squeezenet without deep compression. I had to add a dense layer to the end of SqueezeNet to get correct shape for my labels. I also had to add batch normalisation to the fire modules so it would fit.
- The paper says images are 224x224 but the code and parameters suggests they use 227x227. I also added some padding, relu and initialisation that was used in the squeezenet code but not mentioned in the paper.
- Accuracy of squeezenet is lower in this test but there is a 25x size reduction over ZFNet (SqueezeNet size is the same as in the paper) and a 12x reduction in params.
-
Weights are contained in the file
weights.npz -
These are converted from the caffemodel file available here
-
This required a small change in weight sizes from the original code (a single layer, fire4 I think, went from 16 filters to 32)
-
Load weights with the following:
from chainer import serializers from model import SqueezeNet model = SqueezeNet() serializer.load_npz('weights.npz', model)
- Inputs should be of shape (batch_size, 3, 227, 227)
- It takes about 5.5 seconds to do a batch of 20 on my Macbook Pro