added bettter reset() function to sequential model

[fgolemo]

Hi folks,

after my first attempt I now added a better reset function to the Sequential model.

It works like this:

model = Sequential()

# add layers as usual
model.add(...)

# compile & fit as usual
model.compile(...)
model.fit(...)

# then RESET and fit again. No compilation needed
model.reset()
model.fit(...)

There are two new functions, of better 1 function, 2 modes of operation:

• hard reset (default setting) - will reset the model weights and biases to random values
• soft reset (with parameter use_stored=True) - will reset the model to the stored weights and baises as after compilation, i.e. to the same parameters it started with before.

So it's either

model.reset() # hard reset or
model.reset(use_stored=True) # soft reset

Give me your opinion. The fork was made at commit
8ba647c

[fastai]

Any chance of getting this functionality committed? Not being able to re-initialize weights leads to a lot of compilation overhead... Anything we can do to help?