PyTorch Notes

gadget functions (torch)

• x.numel() number of elements in x

• broadcast mechanism

• Y = Y + X would allocate a new space for the result of Y + X. Y then is the reference to this new memory. Y += X or Y[:] = Y + X is better. (id(Y) will be unchanged.)

• float() or int() converts a scalar tensor to a standard number, similar to x.item()

• A * B element-wise product.

• A_sum = A.sum(axis=1, keepdims=True) will keep the number of axis. A / A_sum will enable broadcast mechanism.

• A.cumsum(axis=0) accumulated sum.

• linear algebra (matrix, vector) multiplication

  • dot product: torch.dot(x, y) only accept 1-D tensor.

  • matrix-vector multiplication: torch.mv(A, x) matrix and vector. x is 1-D vector. All in column vectors.

  • matrix-matrix multiplication: torch.mm(A, B).

• torch.distributions

  • multinomial.Multinomial(n, probs).sample(N, )

gadget functions (pandas)

• data.fillna(data.mean())

• pd.get_dummies(data, dummy_na=True) Convert categorical variable into dummy/indicator variables. Works for str.

gadget functions (d2l)

• d2l.plot example

  • x = np.arange(0, 3, 0.1) d2l.plot(x, [x ** 2, 2 * x - 1], 'x', 'f(x)', legend=['f(x)', 'Tangent'])

plot

• example

  • plt.plot(y, label=("label"))

  • plt.axhline(y=0, color='black', linestyle='dashed')

  • plt.gca().set_xlabel('x label')

  • plt.gca().set_ylabel('y label')

  • plt.legend()