amew0 / aihack Goto Github PK

i: input
o: output
k: kernel
s: stride
p: padding

s' = int((s - k + 2*p) / s + 1)   # round down (floor)

Always, check with Random Values to make sure the Conv2d is well setup.

Example:
>>> layer1 = nn.Conv2d (i=1, o=10 , k=3 , s=2 , p=1)

>>> layer1(torch.rand([32,1,512,512])) 
# Interpretation: If my input is 32 images/batch of 1 channel of dimension 512 x 512,
and I pass them through layer1, what do I get?
# You get: 32 x 10 x 256 x 256: 
    what is this?
        For each 32 images, you get 10 features extracted and each of them is of dimension 256 x 256

* experiment with k, s, p

It usually extends nn.Module # So we don't need to build the model scratch Only two functions should be replaced:
    >>> def __init__ ()
        # model architecture is here
    >>> def forward (self,x)
        # glueing the architecture and returning expected output

>>> criterion = nn.CrossEntropyLoss() # EXPLORE other Loss metric also
>>> optimizer = optim.SGD(model.parameters(), lr=0.01) # EXPLORE OTHER optimizers also

Initial parameters of my model are Random, so at first it will guess the output and will get it right at a probability of (1/no.of classes)
So, what do I need to do?

    >>> loss = criterion(outputs, labels)
    >>> optimizer.zero_grad()
    >>> loss.backward()
    >>> optimizer.step()

    1. Calculate how far off it is. (each predicted with each label and count how many of them did it get them right) 
        >>> criterion(outputs, labels)
    2. Change the values a little bit (Learning Rate). 
        >>> loss.backward()

Pytorch by default is designed for training. So we have to explicitly stop training by,
with torch.no_grad(): # What happens if we don't? Gradients / changes are ready to be appended to the parameters. (and we don't want that in this stage)

Why do we need to bother tweaking the weights when it was already done by others partially?

VGG16, RESNET[18/50/etc.], and others Popular ones for our TASK
But we have to adjust them a little to make them compatible, like
    - Input dimenstion expectation
    - Output dimension requirement
But, takes more time to train 

For identification save the filename with the team name and anything after that .pt. e.g. Z_idontknowwhattowritehere.pt

!!IMPORTANT!! To make it compatible with the testing experiment save it ONLY THIS WAY,

x = torch.Tensor(1, 3, 300, 400) # Batch=1 Color-Channel=3 Dimenstion-width=300 Height=400 (update this as required and used by the model)
with torch.no_grad():
    traced_cell = torch.jit.trace(model, (x))
model_name = "./models/Y_model_1.pt"
torch.jit.save(traced_cell, model_name)

Upload it and see how you performed in the testing dataset. 
Good Luck