• TensorFlow implementation of Going Deeper with Convolutions (CVPR'15).

• This repository contains the examples of natural image classification using pre-trained model as well as training a Inception network from scratch on CIFAR-10 dataset (93.64% accuracy on testing set). The pre-trained model on CIFAR-10 can be download from here.
• Architecture of GoogLeNet from the paper:

• Python 3.3+
• TensorFlow 1.9+
• Numpy
• Scipy

• The GoogLeNet model is defined in src/nets/googlenet.py.
• Inception module is defined in src/models/inception_module.py.
• An example of image classification using pre-trained model is in examples/inception_pretrained.py.
• An example of train a network from scratch on CIFAR-10 is in examples/inception_cifar.py.

For testing the pre-trained model

• Images are rescaled so that the smallest side equals 224 before fed into the model. This is not the same as the original paper which is an ensemble of 7 similar models using 144 224x224 crops per image for testing. So the performance will not be as good as the original paper.

For training from scratch on CIFAR-10

• All the LRN layers are removed from the convolutional layers.
• Batch normalization and ReLU activation are used in all the convolutional layers including the layers in Inception structure except the output layer.
• Two auxiliary classifiers are used as mentioned in the paper, though 512 instead of 1024 hidden units are used in the two fully connected layers to reduce the computation. However, I found the results are almost the same on CIFAR-10 with and without auxiliary classifiers.
• Since the 32 x 32 images are down-sampled to 1 x 1 before fed into inception_5a, this makes the multi-scale structure of inception layers less useful and harm the performance (around 80% accuracy). To make full use of the multi-scale structures, the stride of the first convolutional layer is reduced to 1 and the first two max pooling layers are removed. The the feature map (32 x 32 x channels) will have almost the same size as described in table 1 (28 x 28 x channel) in the paper before fed into inception_3a. I have also tried only reduce the stride or only remove one max pooling layer. But I found the current setting provides the best performance on the testing set.
• During training, dropout with keep probability 0.4 is applied to two fully connected layers and weight decay with 5e-4 is used as well.
• The network is trained through Adam optimizer. Batch size is 128. The initial learning rate is 1e-3, decays to 1e-4 after 30 epochs, and finally decays to 1e-5 after 50 epochs.
• Each color channel of the input images are subtracted by the mean value computed from the training set.

• Download the pre-trained parameters here. This is original from here.
• Setup path in examples/inception_pretrained.py: PRETRINED_PATH is the path for pre-trained model. DATA_PATH is the path to put testing images.

Go to examples/ and put test image in folder DATA_PATH, then run the script:

python inception_pretrained.py --im_name PART_OF_IMAGE_NAME

• --im_name is the option for image names you want to test. If the testing images are all png files, this can be png. The default setting is .jpg.
• The output will be the top-5 class labels and probabilities.

• Download CIFAR-10 dataset from here
• Setup path in examples/inception_cifar.py: DATA_PATH is the path to put CIFAR-10. SAVE_PATH is the path to save or load summary file and trained model.

Go to examples/ and run the script:

python inception_cifar.py --train \
  --lr LEARNING_RATE \
  --bsize BATCH_SIZE \
  --keep_prob KEEP_PROB_OF_DROPOUT \
  --maxepoch MAX_TRAINING_EPOCH

• Summary and model will be saved in SAVE_PATH. One pre-trained model on CIFAR-10 can be downloaded from here.

Go to examples/ and put the pre-trained model in SAVE_PATH. Then run the script:

python inception_cifar.py --eval \
  --load PRE_TRAINED_MODEL_ID

• The pre-trained ID is epoch ID shown in the save modeled file name. The default value is 99, which indicates the one I uploaded.
• The output will be the accuracy of training and testing set.

• Top five predictions are shown. The probabilities are shown keeping two decimal places. Note that the pre-trained model are trained on ImageNet.
• Result of VGG19 for the same images can be found here. The pre-processing of images for both experiments are the same.

Data Source	Image	Result
COCO		1: probability: 1.00, label: brown bear, bruin, Ursus arctos 2: probability: 0.00, label: ice bear, polar bear 3: probability: 0.00, label: hyena, hyaena 4: probability: 0.00, label: chow, chow chow 5: probability: 0.00, label: American black bear, black bear
COCO		1: probability: 0.79, label: street sign 2: probability: 0.06, label: traffic light, traffic signal, stoplight 3: probability: 0.03, label: parking meter 4: probability: 0.02, label: mailbox, letter box 5: probability: 0.01, label: balloon
COCO		1: probability: 0.94, label: trolleybus, trolley coach 2: probability: 0.05, label: passenger car, coach, carriage 3: probability: 0.00, label: fire engine, fire truck 4: probability: 0.00, label: streetcar, tram, tramcar, trolley 5: probability: 0.00, label: minibus
COCO		1: probability: 0.35, label: burrito 2: probability: 0.17, label: potpie 3: probability: 0.14, label: mashed potato 4: probability: 0.10, label: plate 5: probability: 0.03, label: pizza, pizza pie
ImageNet		1: probability: 1.00, label: goldfish, Carassius auratus 2: probability: 0.00, label: rock beauty, Holocanthus tricolor 3: probability: 0.00, label: puffer, pufferfish, blowfish, globefish 4: probability: 0.00, label: tench, Tinca tinca 5: probability: 0.00, label: anemone fish
Self Collection		1: probability: 0.32, label: Egyptian cat 2: probability: 0.30, label: tabby, tabby cat 3: probability: 0.05, label: tiger cat 4: probability: 0.02, label: mouse, computer mouse 5: probability: 0.02, label: paper towel
Self Collection		1: probability: 1.00, label: streetcar, tram, tramcar, trolley, trolley car 2: probability: 0.00, label: passenger car, coach, carriage 3: probability: 0.00, label: trolleybus, trolley coach, trackless trolley 4: probability: 0.00, label: electric locomotive 5: probability: 0.00, label: freight car

• Here is a similar experiment using VGG19.

learning curve for training set

learning curve for testing set

• The accuracy on testing set is 93.64% around 100 epochs. We can observe the slightly over-fitting behavior at the end of training.

Qian Ge