This model is a pretrained model on full imagenet dataset [1] with 14,197,087 images in 21,841 classes. The model is trained by only random crop and mirror augmentation.
The network is based on Inception-BN network [2], and added more capacity. This network runs roughly 2 times slower than standard Inception-BN Network.
We trained this network on a machine with 4 GeForce GTX 980 GPU. Each round costs 23 hours, the released model is the 9 round.
Train Top-1 Accuracy over 21,841 classes: 37.19%
Single image prediction memory requirement: 15MB
ILVRC2012 Validation Performance:
| Over 1,000 classes | Over 21,841 classes | |
|---|---|---|
| Top-1 | 68.3% | 41.9% |
| Top-5 | 89.0% | 69.6% |
| Top=20 | 96.0% | 83.6% |
Note: Directly use 21k prediction may lose diversity in output. You may choose a subset from 21k to make prediction more reasonable.
The compressed file contains:
Inception-symbol.json: symbolic networkInception-0009.params: network parametersynset.txt: prediction label/text mapping
There is no mean image file for this model. We use mean_r=117, mean_g=117 and mean_b=117 to normalize the image.
Models are hosted on http://data.dmlc.ml. You can download it by
wget http://data.dmlc.ml/mxnet/models/imagenet/inception-21k.tar.gz[1] Deng, Jia, et al. "Imagenet: A large-scale hierarchical image database." Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on. IEEE, 2009.
[2] Ioffe, Sergey, and Christian Szegedy. "Batch normalization: Accelerating deep network training by reducing internal covariate shift." arXiv preprint arXiv:1502.03167 (2015).