Why it is better to go with existing pre-trained model already in Github rather than training one's own model from scratch?

In today's world these are normal hurdles while we are building a system in AI.

• Not every one has access to the right dataset.
• Not every one has access to the right machine.

Training from scratch you might end up taking more number of days costing so much more than when using already trained models.

Already trained models are always best way to proceed because take for example you have a requirement where your client wants you to detect the Hyundai cars from a series of images. How would you go about it ?

One way is to take 1000s of images of Hyundai car and train the model on it and its ready to detect Hyundai cars on new images. Question- how will you get 1000s of images of Hyundai car and is the dataset enough to converge the network and increase the accuracy? - Answer is, it is not the best way. Second and preferred way would be to use inception model already trained to detect cars and then use a method named - transfer learning to create a model and transfer the weights from inception model and train the model with specific Hyundai images - now what will happen is that the model will converge faster and will be more accurate to identify Hyundai images.