How To Use CNN In Tensorflow For Image Recognition?

How to use CNN in TensorFlow for image recognition? For a task like an image classification and object detection.

Learn About What Is CNN?

The CNN or Convolutional Neural Networks is the construction behind image recognition. Also, it is a Deep Learning algorithm consists of:

• Input Layer
• Output Layer
• Multiple Hidden Layer

Learn About What Is TensorFlow?

A TensorFlow let you operate directly with tensors. Also, for you to create a neural network from the ground up.

Moreover, always use the high-level estimators API. Because it enables object detection in TensorFlow. Also, it enables you to establish the objects at a higher-level abstraction.

• Computational graph – it represents the flow of the operations. Occurring during the training of deep learning model.

Furthermore, in the CNN model, it is pretty complex the computational graph. 

• Constant – in TensorFlow, it store constant values. Therefore, these use for model training. Also, this constant does not take inputs.

• Placeholder – for input when running a model. It also takes parameters. Moreover, modified in runtime and computational graph. 

• Variable – modifies the computational graph and also adds parameters and nodes to the graph that is trainable.

Basic TensorFlow CNN: MNIST Dataset Estimators

One way to get started with CNN on TensorFlow is to work with patterns base on standard datasets. Moreover, these datasets are builts for TensorFlow.

Also, give you a predictable running and tunning model.

The general steps for

Architecture:

Convolutional layer – 32 5×5 filters -> Pooling Layer – 2×2 filter -> Convolutional layer – 64 5×5 filters -> Pooling Layer – 2×2 filter -> Dense Layer – 1024 neurons -> DenseLayer – 10 neurons 

Application:

• Firstly, Create the input layer using the reshape function ()
• Create the pooling layers using -> layers.conv2d() max_pooling2d()
• Building the dense layer. Use -> layers.dense()
• Generate predictions by running the -> softmax()
• Calculate loss by running -> losses.sparse_softmax_cross_entropy()
• Configure training operation. Use the -> optimizer.minimize()
• Evaluate metric using -> tf.metrics.accuracy()
• Load data using -> mnist.load_data()
• Define an estimator for the custom object detection model
• Start to train the model by running -> train() on the estimator object.
• Lastly, Evaluate the model on all MNIST images using -> evaluate()

Intermediate TensorFlow CNN: Fashion -MNIST Dataset With Estimators

More advanced examples. Such as, Fashion -MNIST Dataset With Estimators

The convolutional layers.

Convolutional layer – 32-3 and 3 filters -> Max pooling layer – 2×2 filter -> Convolutional layer – 64-4 and 3 filters -> Max pooling layer – 2×2 filter -> Convolutional layer – 128-3 and 3 filters -> Max pooling layer – 2×2 filter -> Flatten-> Dense layer – 128 neurons -> Output layer – 10 neurons

Application:

• Load data – needs one-hot encoding. Also, the dataset is not built into TensorFlow.
• Reshape images -> 28x28x1.
• Fix network parameters and placeholders.
• Create the network architecture using -> conv_net() function, conv2d() and maxpool2d().
• Next loss and optimizer nodes.
• Add an evaluation node.
• Train and test the model.
• Plot accuracy.
• Lastly, training and validation data.