def grade_Q8():
relu = nn.ReLU()
flatten = nn.Flatten()
f1 = nn.Linear("f1", 28 * 28, 100)
f2 = nn.Linear("f2", 100, 10)
x = nn.Input2d("images", 28, 28, 1)
x = flatten(x)
x = f2(relu(f1(x)))
x.resolve(np.load("msimple_params.npz"))
mnist_test = np.load("mnist_test.npz")
images = mnist_test["images"][:1000]
infer0 = x.compile(golden.Builder())
infer1 = x.compile(cpp.Builder())
label0 = infer0(images=images).argmax(axis=1)
label1 = infer1(images=images).argmax(axis=1)
return np.allclose(label0, label1)
def grade_Q4():
relu = nn.ReLU()
flatten = nn.Flatten()
f1 = nn.Linear("f1", 28 * 28, 32)
f2 = nn.Linear("f2", 32, 10)
x = nn.Input2d("images", 28, 28, 1)
x = flatten(x)
x = f2(relu(f1(x)))
x.resolve(np.load("p5_params.npz"))
mnist_test = np.load("mnist_test.npz")
images = mnist_test["images"]
labels = mnist_test["labels"]
infer = x.compile(golden.Builder())
pred_labels = infer(images=images).argmax(axis=1)
count = sum(labels == pred_labels)
return count > 9500
def grade_Q4():
pool = nn.MaxPool2d(2, 2)
relu = nn.ReLU()
flatten = nn.Flatten()
x = nn.Input2d("images", 28, 28, 1)
c1 = nn.Conv2d("c1", 1, 8, 5) # 28->24
x = pool(relu(c1(x))) # 24->12
c2 = nn.Conv2d("c2", 8, 8, 5) # 12->8
x = pool(relu(c2(x))) # 8->4
f = nn.Linear("f", 8*4*4, 10)
x = f(flatten(x))
x.resolve(np.load("p6_params.npz"))
mnist_test = np.load("mnist_test.npz")
images = mnist_test["images"]
labels = mnist_test["labels"]
infer = x.compile(golden.Builder())
pred_labels = infer(images = images).argmax(axis = 1)
count = sum(labels == pred_labels)
return count > 9500
def grade_Q10():
pool = nn.MaxPool2d(2, 2)
relu = nn.ReLU()
flatten = nn.Flatten()
x = nn.Input2d("images", 28, 28, 1)
c1 = nn.Conv2d("c1", 1, 8, 3) # 28->26
c2 = nn.Conv2d("c2", 8, 8, 3) # 26->24
x = pool(relu(c2(relu(c1(x))))) # 24->12
c3 = nn.Conv2d("c3", 8, 16, 3) # 12->10
c4 = nn.Conv2d("c4", 16, 16, 3) # 10->8
x = pool(relu(c4(relu(c3(x))))) # 8->4
f = nn.Linear("f", 16 * 4 * 4, 10)
x = f(flatten(x))
x.resolve(np.load("mnist_params.npz"))
mnist_test = np.load("mnist_test.npz")
images = mnist_test["images"][:1000]
infer0 = x.compile(golden.Builder())
infer1 = x.compile(cpp.Builder())
label0 = infer0(images=images).argmax(axis=1)
label1 = infer1(images=images).argmax(axis=1)
return np.allclose(label0, label1)
def grade_Q2():
flatten = nn.Flatten()
x = flatten(nn.Input("x"))
return is_same(x, 1, x=(10, 11, 12, 13))