import torch
import numpy as np

# initialize a tensor directly as data
data = [[1,2],[3,4]]
x_data = torch.tensor(data).to('xpu')

print(x_data)

# from a numpy array
arr = np.array(data)
x_np = torch.from_numpy(arr).to('xpu')

print(x_np)

# from another tensor (could override the shape, datatype)
x_ones = torch.ones_like(x_data) # retains properties of x_data
print(f"Ones Tensor: \n {x_ones} \n")

x_rand = torch.rand_like(x_data, dtype=torch.float) # overrides datatype
print(f"Random Tensor: \n {x_rand} \n")

# use `shape` to define dimentions
shape = (3,5)
rand_tensor = torch.rand(shape)
ones_tensor = torch.ones(shape)
zeros_tensor = torch.zeros(shape)

print(f"Random: \n {rand_tensor} \n")
print(f"Ones: \n {ones_tensor} \n")
print(f"Zeros: \n {zeros_tensor} \n")

# tensors have attributes

tensor_attr = torch.rand(3,4).to('xpu')

# print(f"Shape: {tensor_attr.shape}")
# print(f"Datatype: {tensor_attr.dtype}")
# print(f"Device: {tensor_attr.device}")

# indexing & slicing

tensor = torch.ones(4,4).to('xpu')
print(f"First row: {tensor[0]}")
print(f"First column: {tensor[:,0]}")
print(f"Last column: {tensor[...,-1]}")
print(tensor)

# joining tensors using torch.cat
t1 = torch.cat([tensor, tensor, tensor], dim=1).to('xpu')
print(t1)



# different ways to do matrix multiplication (y1, y2, y3 will have same values)
y1 = tensor @ tensor.T
y2 = tensor.matmul(tensor.T)

y3 = torch.rand_like(y1) # create a new tensor with same shape as y1
torch.matmul(tensor, tensor.T, out=y3)


# comput the element-wise product (z1, z2, z3 will have same values)
z1 = tensor * tensor
z2 = tensor.mul(tensor)

z3 = torch.rand_like(z1)
torch.mul(tensor, tensor, out=z3)

# convert a one-item tensor into a number with `item()`
agg = tensor.sum()
agg_item = agg.item()
print(agg, type(agg))
print(agg_item, type(agg_item))

# in-place operations, will assign the result into the operand
tensor.add_(5) # will assign the result to tensor (ie. override the original tensor)