Inception Models supported:
- Inception_v1 or GoogLeNet [1]
- Inception_v2 [2]
- Inception_v3 [2]
- Inception_v4 [3]
- Inception_ResNet_v1 [3]
- Inception_ResNet_v2 [3]
- Squeeze and Excite (SE) version of all models
The original Inception_v1 or GoogLeNet architecture had inception blocks of various kernel sizes in parallel branches concatenated together as shown below. The modified inception module is more efficient than the original one in terms of size and performance, as claimed by [1].
GoogLeNet Network Blocks: Original Inception Block (left), Efficient Inception Block (Right) [7]
The Inception_v1 block has two auxillary outputs apart from the final output which has regularization effects. There are total 9 Inception Modules in a single architecture.
GoogLeNet Network (From Left to Right) [1]
Inception_v3 is a more efficient version of Inception_v2 while Inception_v2 first implemented the new Inception Blocks (A, B and C). BatchNormalization (BN) [4] was first implemented in Inception_v2. In Inception_v3, even the auxilliary outputs contain BN and similar blocks as the final output. Inception_v2 architecture is similar to v3 but during the input, a traditional convolutional layer has been replaced by a DepthWise Separable Convolutional layer. The input kernel size of both Incpetion v1 and v2 was 7, but was changed to 3 in later versions.
Inception_v3 architecture is as follows:
Inception-v3 Architecture (Batch Norm and ReLU are used after Conv) [5]
After the Stem blocks, the Inception_v3 architecture contains 5x Inception-A Modules followed by a Grid Size Reduction Block, then 4x Inception-B Modules followed by another Reduction Block. In the end, before the final MLP layer, there are 2x Inception-C Modules. Each type of Incception Module has been depicted in the image below for Inception_v4. The MLP layer contains Global Pooling, Flattening and final Dense layer selective for Classification or Regression tasks.
The primary differences between Inception v3 and v4 are in the Stem structure and the number of Inception (A, B and C) modules in each step. The Inception_v4 architecture along with the three modules types are as follows:
Inception-v4: Whole Network Schema (Leftmost), Stem (2nd Left), Inception-A (Middle), Inception-B (2nd Right), Inception-C (Rightmost) [6]
So, in Inception_v4, Inception Module-A is being used 4 times, Module-B 7 times and Module-C 3 times. As seen in the figure, the modified stem now has branch filter concatenation similar to the Inception Modules.
Inception_ResNet_v1, as shown in the figure below, consists of modfied Inception Modules. The main difference is the skip connections like that of ResNets. Its Stem is similar to Inception_v3. In case of Inception_ResNet_v1, Module-A gets repeated 5 times, Module-B 10 times and Module-C 5 times.
Inception-ResNet-v1: Whole Network Schema (Leftmost), Stem (2nd Left), Inception-A (Middle), Inception-B (2nd Right), Inception-C (Rightmost) [2]
The Stem of the Inception_ResNet_v2 is exactly same as that of Inception_v4, so it is not mentioned again in the figure below. Its Inception Modules are similar in structure that of Inception_ResNet_v1 but heavier. But in this case, Module-A gets repeated 10 times, Module-B 20 times and Module-C 10 times.
Inception-ResNet-v2: Inception-A (Leftmost), Inception-B (Middle), Inception-C (Rightmost) [6]
Keping the future in mind, all the models have been developed in Keras with Tensorflow backend (tf.keras), so they do not support Theano as the backend. But, the speciality about the models is their flexibility. The user has the option for:
- Choosing any of 4 available Inception models or 2 Inception-ResNet models for either 1D or 2D tasks.
- Varying number of input kernel/filter, commonly known as the 'Width' of the model. Default is 32, which is better to use following the paper.
- Varying number of classes for Classification tasks and number of features to be extracted for Regression tasks.
- Varying number of Channels (e.g., 2 channels means 2 sources/datasets will be trained jointly for a common target).
- Auxilliary Outputs: Two optional Auxilliary Outputs are available for each Inception Model, each one is set before the Grid Size Reduction Blocks, respectively. If set 'TRUE', the output will contain 3 columns as predictions. Mentionable that this concept of having two auxilliary outputs, which act similar to 'Deep Supervision' for segmentation models such as UNet, has been adopted from Inception_v1 (GoogLeNet) architecture as shown above.
Details of the 1D implementation process are available in the Jupyter Notebook containing the DEMO provided in the codes section. The datasets used in the DEMO are also available in the 'Documents' folder. The DEMO for 2D version will be added later along with supports for ImageNet or CIFAR-10 weights to be added in the top as an option.
[1] Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., & Anguelov, D. et al. (2021). Going Deeper with Convolutions. arXiv.org. Retrieved 28 August 2021, from https://arxiv.org/abs/1409.4842.
[2] Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2021). Rethinking the Inception Architecture for Computer Vision. arXiv.org. Retrieved 28 August 2021, from http://arxiv.org/abs/1512.00567.
[3] Szegedy, C., Ioffe, S., Vanhoucke, V., & Alemi, A. (2021). Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning. arXiv.org. Retrieved 28 August 2021, from https://arxiv.org/abs/1602.07261.
[4] Ioffe, S., & Szegedy, C. (2021). Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. arXiv.org. Retrieved 28 August 2021, from https://arxiv.org/abs/1502.03167.
[5] Review: Inception-v3 — 1st Runner Up (Image Classification) in ILSVRC 2015. Medium. (2021). Retrieved 22 August 2021, from https://sh-tsang.medium.com/review-inception-v3-1st-runner-up-image-classification-in-ilsvrc-2015-17915421f77c.
[6] Review: Inception-v4 — Evolved From GoogLeNet, Merged with ResNet Idea (Image Classification). Medium. (2021). Retrieved 22 August 2021, from https://towardsdatascience.com/review-inception-v4-evolved-from-googlenet-merged-with-resnet-idea-image-classification-5e8c339d18bc.
[7] Review: GoogLeNet (Inception v1)— Winner of ILSVRC 2014 (Image Classification). Medium. (2021). Retrieved 22 August 2021, from https://medium.com/coinmonks/paper-review-of-googlenet-inception-v1-winner-of-ilsvlc-2014-image-classification-c2b3565a64e7.