VGG is a classical convolutional neural network architecture. It was based on an analysis of how to increase the depth of such networks. The network utilises small 3 x 3 filters. Otherwise, the network is characterized by its simplicity: the only other components being pooling layers and a fully connected layer.
VGG Configurations:
The depth of the configurations increases from the left (A) to the right (E), as more layers are added (the added layers are shown in bold). The convolutional layer parameters are denoted as "conv<receptive field size>-<number of channels>". The ReLU activation function is not shown for brevity.
This directory contains implementation of VGG (Very Deep Convolutional Networks for Large-Scale Image Recognition) by Keras.
Model Summary (VGG11):
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) [(None, 224, 224, 3)] 0
conv2d (Conv2D) (None, 224, 224, 64) 1792
max_pooling2d (MaxPooling2D (None, 112, 112, 64) 0
)
conv2d_1 (Conv2D) (None, 112, 112, 128) 73856
max_pooling2d_1 (MaxPooling (None, 56, 56, 128) 0
2D)
conv2d_2 (Conv2D) (None, 56, 56, 256) 295168
conv2d_3 (Conv2D) (None, 56, 56, 256) 590080
max_pooling2d_2 (MaxPooling (None, 28, 28, 256) 0
2D)
conv2d_4 (Conv2D) (None, 28, 28, 512) 1180160
conv2d_5 (Conv2D) (None, 28, 28, 512) 2359808
max_pooling2d_3 (MaxPooling (None, 14, 14, 512) 0
2D)
conv2d_6 (Conv2D) (None, 14, 14, 512) 2359808
conv2d_7 (Conv2D) (None, 14, 14, 512) 2359808
max_pooling2d_4 (MaxPooling (None, 7, 7, 512) 0
2D)
flatten (Flatten) (None, 25088) 0
dense (Dense) (None, 4096) 102764544
dropout (Dropout) (None, 4096) 0
dense_1 (Dense) (None, 4096) 16781312
dropout_1 (Dropout) (None, 4096) 0
dense_2 (Dense) (None, 1000) 4097000
=================================================================
Total params: 132,863,336
Trainable params: 132,863,336
Non-trainable params: 0
_________________________________________________________________
Model Summary (VGG13):
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) [(None, 224, 224, 3)] 0
conv2d (Conv2D) (None, 224, 224, 64) 1792
conv2d_1 (Conv2D) (None, 224, 224, 64) 36928
max_pooling2d (MaxPooling2D (None, 112, 112, 64) 0
)
conv2d_2 (Conv2D) (None, 112, 112, 128) 73856
conv2d_3 (Conv2D) (None, 112, 112, 128) 147584
max_pooling2d_1 (MaxPooling (None, 56, 56, 128) 0
2D)
conv2d_4 (Conv2D) (None, 56, 56, 256) 295168
conv2d_5 (Conv2D) (None, 56, 56, 256) 590080
max_pooling2d_2 (MaxPooling (None, 28, 28, 256) 0
2D)
conv2d_6 (Conv2D) (None, 28, 28, 512) 1180160
conv2d_7 (Conv2D) (None, 28, 28, 512) 2359808
max_pooling2d_3 (MaxPooling (None, 14, 14, 512) 0
2D)
conv2d_8 (Conv2D) (None, 14, 14, 512) 2359808
conv2d_9 (Conv2D) (None, 14, 14, 512) 2359808
max_pooling2d_4 (MaxPooling (None, 7, 7, 512) 0
2D)
flatten (Flatten) (None, 25088) 0
dense (Dense) (None, 4096) 102764544
dropout (Dropout) (None, 4096) 0
dense_1 (Dense) (None, 4096) 16781312
dropout_1 (Dropout) (None, 4096) 0
dense_2 (Dense) (None, 1000) 4097000
=================================================================
Total params: 133,047,848
Trainable params: 133,047,848
Non-trainable params: 0
_________________________________________________________________
Model Summary (VGG16):
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) [(None, 224, 224, 3)] 0
conv2d (Conv2D) (None, 224, 224, 64) 1792
conv2d_1 (Conv2D) (None, 224, 224, 64) 36928
max_pooling2d (MaxPooling2D (None, 112, 112, 64) 0
)
conv2d_2 (Conv2D) (None, 112, 112, 128) 73856
conv2d_3 (Conv2D) (None, 112, 112, 128) 147584
max_pooling2d_1 (MaxPooling (None, 56, 56, 128) 0
2D)
conv2d_4 (Conv2D) (None, 56, 56, 256) 295168
conv2d_5 (Conv2D) (None, 56, 56, 256) 590080
conv2d_6 (Conv2D) (None, 56, 56, 256) 590080
max_pooling2d_2 (MaxPooling (None, 28, 28, 256) 0
2D)
conv2d_7 (Conv2D) (None, 28, 28, 512) 1180160
conv2d_8 (Conv2D) (None, 28, 28, 512) 2359808
conv2d_9 (Conv2D) (None, 28, 28, 512) 2359808
max_pooling2d_3 (MaxPooling (None, 14, 14, 512) 0
2D)
conv2d_10 (Conv2D) (None, 14, 14, 512) 2359808
conv2d_11 (Conv2D) (None, 14, 14, 512) 2359808
conv2d_12 (Conv2D) (None, 14, 14, 512) 2359808
max_pooling2d_4 (MaxPooling (None, 7, 7, 512) 0
2D)
flatten (Flatten) (None, 25088) 0
dense (Dense) (None, 4096) 102764544
dropout (Dropout) (None, 4096) 0
dense_1 (Dense) (None, 4096) 16781312
dropout_1 (Dropout) (None, 4096) 0
dense_2 (Dense) (None, 1000) 4097000
=================================================================
Total params: 138,357,544
Trainable params: 138,357,544
Non-trainable params: 0
_________________________________________________________________
Model Summary (VGG19):
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) [(None, 224, 224, 3)] 0
conv2d (Conv2D) (None, 224, 224, 64) 1792
conv2d_1 (Conv2D) (None, 224, 224, 64) 36928
max_pooling2d (MaxPooling2D (None, 112, 112, 64) 0
)
conv2d_2 (Conv2D) (None, 112, 112, 128) 73856
conv2d_3 (Conv2D) (None, 112, 112, 128) 147584
max_pooling2d_1 (MaxPooling (None, 56, 56, 128) 0
2D)
conv2d_4 (Conv2D) (None, 56, 56, 256) 295168
conv2d_5 (Conv2D) (None, 56, 56, 256) 590080
conv2d_6 (Conv2D) (None, 56, 56, 256) 590080
conv2d_7 (Conv2D) (None, 56, 56, 256) 590080
max_pooling2d_2 (MaxPooling (None, 28, 28, 256) 0
2D)
conv2d_8 (Conv2D) (None, 28, 28, 512) 1180160
conv2d_9 (Conv2D) (None, 28, 28, 512) 2359808
conv2d_10 (Conv2D) (None, 28, 28, 512) 2359808
conv2d_11 (Conv2D) (None, 28, 28, 512) 2359808
max_pooling2d_3 (MaxPooling (None, 14, 14, 512) 0
2D)
conv2d_12 (Conv2D) (None, 14, 14, 512) 2359808
conv2d_13 (Conv2D) (None, 14, 14, 512) 2359808
conv2d_14 (Conv2D) (None, 14, 14, 512) 2359808
conv2d_15 (Conv2D) (None, 14, 14, 512) 2359808
max_pooling2d_4 (MaxPooling (None, 7, 7, 512) 0
2D)
flatten (Flatten) (None, 25088) 0
dense (Dense) (None, 4096) 102764544
dropout (Dropout) (None, 4096) 0
dense_1 (Dense) (None, 4096) 16781312
dropout_1 (Dropout) (None, 4096) 0
dense_2 (Dense) (None, 1000) 4097000
=================================================================
Total params: 143,667,240
Trainable params: 143,667,240
Non-trainable params: 0
_________________________________________________________________
The network trained and tested with the CIFAR-10 dataset.
The CIFAR-10 dataset consists of 60000 32x32 color images in 10 classes, with 6000 images per class. There are 50000 training images and 10000 test images.
To install prerequisites run the following command:
$ sudo apt install python3 python3-wheel python3-pip python3-venv python3-dev python3-setuptoolsCreate an isolated Python virtual environment using the venv standard
library module. This will keep dependant Python packages from interfering with other Python projects on your system.
$ python3 -m venv venv
$ source venv/bin/activateOnce activated, update core packaging tools (pip, setuptools, and wheel) to the latest versions.
(venv) $ pip install --upgrade pip setuptools wheelTo install requirements, run the following command:
(venv) $ pip install -r requirements.txtTo train the network with CIFAR-10 dataset, run the following command:
(venv) $ python ./train.py \
--configuration vgg19 \
--train-directory ../datasets/CIFAR-10/train \
--batch-size 256 \
--input-image-width 32 \
--input-image-height 32 \
--shuffle True \
--seed 0 \
--dropout-rate 0.5 \
--learning-rate 0.01 \
--momentum 0.9 \
--epochs 50 \
--validation_split 0.20 \
--logs-path ./logs \
--model-path ./checkpointor simply train with default values:
(venv) $ python ./train.pyTrain Script Usage:
(venv) $ python ./train.py --help
usage: train.py [-h] [--configuration CONFIGURATION] [--train-directory TRAIN_DIRECTORY] [--batch-size BATCH_SIZE] [--input-image-width INPUT_IMAGE_WIDTH] [--input-image-height INPUT_IMAGE_HEIGHT] [--shuffle SHUFFLE]
[--seed SEED] [--dropout-rate DROPOUT_RATE] [--learning-rate LEARNING_RATE] [--momentum MOMENTUM] [--epochs EPOCHS] [--validation_split VALIDATION_SPLIT] [--logs-path LOGS_PATH] [--model-path MODEL_PATH]
Trains the model
optional arguments:
-h, --help show this help message and exit
--configuration CONFIGURATION
Model configuration, one of: vgg11, vgg13, vgg16, vgg19
--train-directory TRAIN_DIRECTORY
Directory where the train data is located
--batch-size BATCH_SIZE
Size of the batches of data
--input-image-width INPUT_IMAGE_WIDTH
Input image width
--input-image-height INPUT_IMAGE_HEIGHT
Input image height
--shuffle SHUFFLE Whether to shuffle the data
--seed SEED Random seed for shuffling and transformations
--dropout-rate DROPOUT_RATE
Fraction of the input units to drop
--learning-rate LEARNING_RATE
Learning rate
--momentum MOMENTUM Hyper-parameter that accelerates gradient descent in the relevant direction and dampens oscillations
--epochs EPOCHS Number of epochs
--validation_split VALIDATION_SPLIT
Fraction of the training data to be used as validation data
--logs-path LOGS_PATH
Path of the directory where to save the log files to be parsed by TensorBoard
--model-path MODEL_PATH
Path to save the modelVisualizing Metrics:
TensorBoard is a visualization tool provided with TensorFlow to tracking and visualizing metrics such as loss and accuracy.
To launch TensorBoard, run the following command:
(venv) $ tensorboard --logdir ./logs
Training Results:
| Epoch | Loss | Accuracy | Loss (Val) | Accuracy (Val) |
|---|---|---|---|---|
| 1 | 2.2411 | 0.1463 | 2.2871 | 0.1964 |
| 2 | 1.9682 | 0.2670 | 1.9287 | 0.2781 |
| 3 | 1.6822 | 0.3658 | 1.5490 | 0.4286 |
| ... | ... | ... | ... | ... |
| 50 | 0.0165 | 0.9949 | 1.4107 | 0.7807 |
To evaluate the network with CIFAR-10 dataset, run the following command:
(venv) $ python ./evaluate.py \
--test-directory ../datasets/CIFAR-10/test \
--batch-size 256 \
--model-path ./checkpointor simply evaluate with default values:
(venv) $ python ./evaluate.pyEvaluate Script Usage:
(venv) $ python ./evaluate.py --help
usage: evaluate.py [-h] [--test-directory TEST_DIRECTORY] [--batch-size BATCH_SIZE] [--model-path MODEL_PATH]
Evaluates the model
optional arguments:
-h, --help show this help message and exit
--test-directory TEST_DIRECTORY
Directory where the test data is located
--batch-size BATCH_SIZE
Size of the batches of data
--model-path MODEL_PATH
Path to load the modelEvaluating Results:
| Loss | Accuracy |
|---|---|
| 0.8931 | 0.7247 |
To predict the network with an image, run the following command:
(venv) $ python ./predict.py \
--image-path <image-path>for example to predict this image:
run the following command:
(venv) $ python ./predict.py \
--image-path ../datasets/CIFAR-10/test/airplane/aeroplane_s_000040.pngoutput:
Predicted index: 0
Predicted label (CIFAR-10): airplane
Predict Script Usage:
(venv) $ python ./predict.py --help
usage: predict.py [-h] --image-path IMAGE_PATH [--model-path MODEL_PATH]
Predicts the model
optional arguments:
-h, --help show this help message and exit
--image-path IMAGE_PATH
Image path to predict
--model-path MODEL_PATH
Path to load the modelYou can download pre-trained models from this page or directly download VGG pre-trained model.