Inception_V1

Implementation of the Inception V1 convolutional neural network using TensorFlow 2.
Reference Article Inception V1: https://arxiv.org/pdf/1409.4842v1.pdf
Batch Normalization reference Article: https://arxiv.org/pdf/1502.03167v3.pdf
Reference code: https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_v1.py
Article more code examples: https://paperswithcode.com/paper/going-deeper-with-convolutions

Training Results

Two models were trained: one without Batch Normalization (BN) and another with Batch Normalization (BN).

Inception V1	Accuracy without BN	Accuracy with BN
softmax 0	89.02%	87.08%
softmax 1	89.84%	97.54%
softmax 2	98.76%	96.58%

Inception V1	Loss without BN	Loss with BN
softmax 0	0.5322	0.7147
softmax 1	0.3873	0.6366
softmax 2	0.0662	0.1775

Inception V1	Error Rate without BN	Error Rate with BN
softmax 0	10.98%	12.92%
softmax 1	10.16%	2.46%
softmax 2	1.24%	3.42%

The Mnist dataset with 10,000 images was used to train the network.


(Training graph for the model without Batch Normalization.)

(Training graph for the model with Batch Normalization.)

Test Results

5,000 images were used to test the model.

Model without Batch Normalization

Accuracy of softmax functions:
softmax 0: 0.8902000188827515
softmax 1: 0.8984000086784363
softmax 2: 0.9876000285148621

(Confusion matrix for one of the softmax functions. Error rate of 1.24%)

(Confusion matrix for one of the softmax functions. Error rate of 10.16%)

(Confusion matrix for one of the softmax functions. Error rate of 10.98%)

Model with Batch Normalization

Accuracy of softmax functions:
softmax 0: 0.8708000183105469
softmax 1: 0.9753999710083008
softmax 2: 0.9657999873161316

(Confusion matrix for one of the softmax functions. Error rate of 3.42%)

(Confusion matrix for one of the softmax functions. Error rate of 2.46%)

(Confusion matrix for one of the softmax functions. Error rate of 12.92%)