ImageNet Classification

AlexNet, ZFNet, ResNet, ResNetV2, SENet were implemented and some models trained on ILSVRC2012 and mini ImageNet dataset are available.

Accuracy

model name	dataset	top-1 val err	top-5 val err	top-1 test err	top-5 test err
ResNetSE-50	mini imageNet	16.36%	4.80%	16.78%	4.78%
ocd-ResNetV2-50	ILSVRC 2012	21.95%	5.81%	---	---
Average	ILSVRC 2012	21.14%	5.50%	---	---

Note: Only the best models for each dataset are listed here, "Average" is predicted by the average score of two ResNetV2-50 models trained independently on ILSVRC 2012 dataset. For error rates of all 13 provided models, please check test_single_image.ipynb.

The weights files are provided here: To Be Update

中国大陆地区可从这里下载 For friends in China Mainland: 网盘下载

It is strongly recommended to save the weights files in ./h5 directory, since test function will automatically load models there.

Requirements

For training:

• tensorflow >= 2.0.0
• opencv-python
• tqdm >= 4.42.0
• numpy

For testing:

• tensorflow >= 2.0.0
• opencv-python
• numpy
• tqdm >= 4.42.0 (if applying seam carving)

Models

All the architectures are from corresponding references, and may have been slightly adjusted.

For example, "AlexNet-BN" replaced all Local Response Normalization layers to Batch Normalization. This scheme significantly reduces both the top-1 and top-5 error rates.

model	references
AlexNet, AlexNet-BN	ImageNet Classification with Deep Convolutional Neural Networks
ZFNet-BN	Visualizing and Understanding Convolutional Networks
ResNetV2	Identity Mappings in Deep Residual Networks
ResNetSE	Squeeze-and-Excitation Networks

All the models above were trained on mini imageNet dataset and ResNetV2 was trained on ILSVRC 2012 as well. These models were implemented in models directory, and can used for your own training task.

Dataset

• mini ImageNet - A subset of ImageNet, with 60,000 images in total contains 100 categories listed in ./data/miniImageNet/labels_to_content.txt.
• ILSVRC 2012 - The dataset used in ILSVRC 2012, contained 1,000 categories listed in ./data/ILSVRC2012/labels_to_content.txt and a training set of 128,1167 images.

Usage

For testing single image

You can directly use the provided weights files to predict your own image. ./test_single_image.ipynb is the user interface with detailed instructions and intuitive visualization. Check it for more information.

Meanwhile, ./test_single_image.py is another predictor and able to compute average scores on two different models. You may need to modify the code before running, deprecated if you don't know what the code is doing.

For training

The following is where the code needs to be modified:

• ./constants.py - Modify the input size, batch size, epoch number, weight decay, number of classes, size of training set & validation set, numerical characteristics of dataset as your intensions.
• ./utils/data.py - A file including training/validation images and its labels will be loaded during the training process (such as train_labels.txt and valid_labels.txt in data's subdirectory). The paths of these files and image root directory need to be defined there. Moreover, you can modify the data augmentation algorithm as you like.
• ./train.py - Choose the training model implemented in models directory or your own. Set checkpoint file path, learning rate and its decay strategy.

Then, run these command to start your training:

$ python3 train.py

The default code will train ILSVRC2012 dataset on ResNetV2-50 model. For GTX 1080 Ti, it takes about 4 days to run 50 epochs.

For testing 10-crop on test set

The following is where the code needs to be modified:

• ./constants.py - Modify the input size, number of classes and numerical characteristics of trained dataset.
• ./utils/data.py - Modify the paths of image root directory, andval_labels.txt or test_labels.txt mentioned above.
• ./test.py - Choose the model and set the path of weights file.

Then, run these command to start your testing, top-1 and top-5 accuracy will be reported:

$ python3 test.py

The default code will test ILSVRC2012 validation set of ./h5/ocd-ResNetV2-50.h5 model.

Limitation

All trainable variables were saved on GPU rather than CPU, and it will only use a single GPU to compute even if there are multiple GPUs available. This is because my computer has only one GPU, not able to debug or run multi-gpu training.

As the consequence, it's easy to have the problem that CUDA out of memory if the model is too deep and batch size is too large. Experiments shows that this will happen when using ResNetSE-50 model with batch size of 128. Try to decrease the batch size, simplify your model or use tf.float16 to train if you encounter this problem.

References

• The numerical characteristics of ILSVRC 2012 dataset and some implementation details of data input & data augmentation come from ImageNet_ResNet_Tensorflow2.0 by Apm5.

Note: Numerical characteristics of mini ImageNet can be extracted by ./data/utils/image_cal.py, it can also extract them from ILSVRC2012 dataset as well but costs plenty of time since the training set is too large.