Hi, Thanks for your support
I wonder if initialize new networks with pre-trained weights of previous versions (examples: inception v4 - v3) can help lowering training time while expecting good results, tried that with some sample networks that weren’t the same at all and it looks promising.
Want to know if someone else tries this method and can help me find out if this method worked in general or just for my special test, unfortunately, I don’t have access to powerful GPUs so I have to do it more in theory’s but on my test, it was something around 10times faster on a simple network compare to initialize with golorot_uniform method