Intership at PanoAI, Chengdu

PanoAI is a company dedicated to the core development of artificial intelligence vision technology, combining the exploration of application scenarios with technological innovation.

In the first two weeks of internship, I was asked to implement a task of classifying 5 difference types of flower. This was a training task for all the interns.

My spot

Data

And the dataset is stored as .tfrecord, so first I tried to get the images out of the .tfrecord file and plot the images of flowers*. I did this to see the feature of the data. Then I found that the dataset is used at Tensorflow tutorials, there are 3,670 total images.

Visualize the dataset, which has 5 classes.

While in this task, the images were superimposed with Gaussian white noise.

Model

Classifying 5 kinds of flowers is simple enough when you know how to use the Network on TensorflowHub. Most of the time you could just treat it as a layer like this, with other layers of your Tensorflow layers: hub.KerasLayer("https://hub.tensorflow.google.cn/google/imagenet/inception_v3/feature_vector/4", trainable=True)

TensorflowHub is a powerful tool when you want to build a project real quick. It has many trained, ready-to-use models, including resnet_50, inceptionV3, mobilenet.

Loss

I used cross entropy loss in this project, here is what should be notice: When using the loss in tf.keras.losses,

• tf.keras.losses.SparseCategoricalCrossentropy uses a one-hot array to calculate the probability.
• tf.keras.losses.Categorical_Crossentropy uses a category index.

superimposed with Gaussian white noise

*This could also be done by using the Tensorflow Generator, no need to put the data in a list and plot it.

Image Registration Project

Working with a team of NTU Masters.

My colleagues.