VGG19-FCN for Image Classification

TensorFlow implementation of Very Deep Convolutional Networks for Large-Scale Image Recognition.
This repository contains the examples of natural image classification using pre-trained model as well as training a VGG19-like network from scratch on CIFAR-10 dataset (91.81% accuracy on testing set). The pre-trained model on CIFAR-10 can be download from here.

Requirements

Implementation Details

The VGG networks are defined in src/nets/vgg.py.
An example of image classification using pre-trained model is in examples/vgg_pretrained.py.
An example of train a network from scratch on CIFAR-10 is in examples/vgg_cifar.py.

For testing the pre-trained model

The last three fully connected layers are converted to convolutional layers making it a fully convolutional network. Then the input images can be arbitrary size.
Global average pooling is used to get fix size of class scores for all the test images.
Images are rescaled so that the smallest side equals 224 before fed into the model to reduce the computational complexity as well as keep the high classification performance, since for too large scale image, responses of relevant parts maybe averaged by responses of non-relevant of the image through global average pooling.

For training from scratch on CIFAR-10

The convolutional layers part is the same as VGG19. The fully connected layers are implemented as convolutional layers with filter size 1.
Since CIFAR-10 contains smaller number of training set and classes than ImageNet, two fully connected layers consist of 1024 units are used before the output linear layer.
Batch normalization for all of the layers except the output layer. All the activation functions are ReLUs.
During training, dropout with keep probability 0.4 is applied to two fully connected layers. I have tried to add dropout after each max pooling layers, but it harms the performance both for training and testing set. 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 50 epochs, and finally decays to 1e-5 after 70 epochs.
Each color channel of the input images are subtracted by the mean value computed from the training set.

Usage

ImageNet Classification

Preparation

Download the pre-trained parameters VGG19 NPY here. This is original downloaded from here.
Setup path in examples/vgg_pretrained.py: VGG_PATH is the path for pre-trained vgg model. DATA_PATH is the path to put testing images.

Run

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

python vgg_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.

Train the network on CIFAR-10

Preparation

Download CIFAR-10 dataset from here
Setup path in examples/vgg_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.
Train the model
Go to examples/ and run the script:

python vgg_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.

Evaluate the model

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

python vgg_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 104, which indicates the one I uploaded.
The output will be the accuracy of training and testing set.

Results

Image classification using pre-trained VGG19 model

Top five predictions (VGG19) are shown. The probabilities are shown keeping two decimal places. Note that the pre-trained model are trained on ImageNet.

Data Source | Image | Result |
|:—|:—:|:—|
COCO || 1: probability: 1.00, label: brown bear, bruin
2: probability: 0.00, label: American black bear, black bear
3: probability: 0.00, label: ice bear, polar bear
4: probability: 0.00, label: sloth bear, Melursus ursinus
5: probability: 0.00, label: chow, chow chow
COCO || 1: probability: 0.61, label: street sign
2: probability: 0.27, label: traffic light, traffic signal, stoplight
3: probability: 0.02, label: mailbox, letter box
4: probability: 0.02, label: parking meter
5: probability: 0.01, label: pay-phone, pay-station
COCO ||1: probability: 0.48, label: passenger car, coach, carriage
2: probability: 0.36, label: trolleybus, trolley coach, trackless trolley
3: probability: 0.10, label: minibus
4: probability: 0.02, label: school bus
5: probability: 0.01, label: streetcar, tram, tramcar, trolley, trolley car
COCO ||1: probability: 0.17, label: burrito
2: probability: 0.13, label: plate
3: probability: 0.10, label: Dungeness crab, Cancer magister
4: probability: 0.06, label: mashed potato
5: probability: 0.06, label: guacamole
ImageNet ||1: probability: 1.00, label: goldfish, Carassius auratus
2: probability: 0.00, label: rock beauty, Holocanthus tricolor
3: probability: 0.00, label: anemone fish
4: probability: 0.00, label: coral reef
5: probability: 0.00, label: puffer, pufferfish, blowfish, globefish
Self Collection | |1: probability: 0.33, label: tabby, tabby cat
2: probability: 0.20, label: Egyptian cat
3: probability: 0.11, label: tiger cat
4: probability: 0.03, label: Cardigan, Cardigan Welsh corgi
5: probability: 0.02, label: bookcase
Self Collection | |1: probability: 1.00, label: streetcar, tram, tramcarr
2: probability: 0.00, label: trolleybus, trolley coach
3: probability: 0.00, label: passenger car, coach, carriage
4: probability: 0.00, label: electric locomotive
5: probability: 0.00, label: minibus