仅作为记录,大佬请跳过
矩阵相乘(x*w1)
# numpy
h = x.dot(w1)
# torch
h = x.mm(w1)
大于0的保留,小于0的令为0——达到激活函数ReLU的效果
# numpy
h_relu = np.maximum(h, 0)
# torch
h_relu = h.clamp(min=0)
两个数组相减,相减后各个元素平方的和
# numpy
loss = np.square(y_pred - y).sum()
# torch
loss = (y_pred - y).pow(2).sum().item()
注意torch中还需要取.item()
取矩阵的转置
# numpy
grad_w2 = h_relu.T.dot(grad_y_pred)
# torch
grad_w2 = h_relu.t().mm(grad_y_pred)
torch里是用.t()
复制数组
# numpy
grad_h = grad_h_relu.copy()
# torch
grad_h = grad_h_relu.clone()