Ashish Kumar

Ashish Kumar

Data Scientist | Traveller |
Trekker | Dog Father

What are the numbers at the end of ResNet Architecture and does it have any significance on model performance and overfitting?

Posted: less than 1 minute read

numbers-in-resnet.jpg

ResNet is one of the standard architectures which is widely used in computer vision problems. It is both fast and accurate for many datasets and problems.

Let’s see what the number 34 in ResNet34 signifies. These numbers tell how many layers are present in this variant of architecture. Other options are 18, 50, 101 and 152.

So how do we select a particular architecture for our problem?

This depends on the size of the data we have for training. Models with a higher number of layers take longer to train and are prone to overfitting. In case we have a small amount of data we should start with a smaller number of layers but with bigger data architecture with more layers can give more accurate results.

You may also enjoy

Turn Jupyter Notebook into a Standalone Web Application using Binder

3 minute read

Being a Data Scientist, there are situations when we feel that our findings can be better explained with the help of a working prototype. A prototype can be very helpful in giving a birds-eye view of the solution to stakeholders. But due to limited or no knowledge of web development, we either end up looking for some help and if this doesn’t wo...

Are error metrics and loss the same during model training?

1 minute read

A “Loss” and “Metric” are the 2 things which we look at the end of each epoch1 to check the performance of a model. These are very closely related but are not exactly the same. Let’s look at their significance. An Epoch is an event when our model has looked at each of the data points in the training set once. ↩

Worry about the concepts and techniques and not about the libraries or software used?

less than 1 minute read

It doesn’t matter what software and libraries we learn because it only takes a few days to switch from one software or library to another. Whatever tools we use the concept remains the same. These libraries will be updated at a much faster pace in the coming time as more people will be getting into this field, with different skillsets and idea...