C1_W2_Lab_1_mnist.html strange behavior

  1. When I browse to C1_W2_Lab_1_mnist.html in my windows 10 PC, I got following snapshot.

  2. When I browse to C1_W2_Lab_1_mnist.html in my andro mobile, I got following snapshot.

  3. Why the loss(about 0.2) in my mobile is much lower than that on my desktop(about 0.9)? Mobile’s figure is much like that on the lecture video, What is the problem of my PC ?

Dear @hjyang ,

There are few reasons for different results on different devices:

Optimization Libraries:

The underlying optimization libraries or frameworks used for training and inference might be optimized differently for different hardware. Different versions of the libraries or hardware-specific optimizations could lead to numerical discrepancies.

Floating-Point Precision

Different devices may have variations in floating-point precision, leading to small numerical differences in computations. This can accumulate over the course of training and affect the final loss values.

Parallelization and Multithreading

If the training or inference process involves parallelization or multithreading, the order of operations may vary between devices. This can result in slightly different numerical outcomes, particularly in cases where floating-point precision matters.

Random Initialization or Operations

If your model involves random initialization or stochastic operations (e.g., dropout during training), these may result in different random values on different devices, leading to variations in the loss.