I have been using the following image to test the accuracy of the model. For the model in Lab C1_W4_Lab1, this picture was identified as a horse but for the model in C1_W4_Lab2 this picture was accurately identified as a human.
Now my question is the model in both the labs were absolutely the same, the only difference being the addition of the validation set. since validation set does not contribute directly in model training other than the feedback. what caused the model to identify the picture correctly
You have found a strange photo to test humans and horses :D, there are features here that could be mistaken for a horse, no wonder why the model is once predicting a horse and then a human.
Now to the results, well, if the model is not trained on identical data which cannot be, because the preprocessing might include splits at random and shuffling, then it is possible that the model on the second trial, trained on such data that it learned to distinguish this photo as a human. I am wondering what the class probability outputs for each trials was for human and horses classes!
1 Like
for horse the probablity is 0.96
for Human probablity is 0.99
I have found this photo on pixabay 
.
If the pre processing split caused the change in results or in this case detect more accurately, is there a way we can split in a certain way rather than being it completely random
2 Likes
Interesting stuff, by using seeds in python environment you may get reproducible results.