def test_sgd_gpu(self): np.random.seed(1337) model = TinyBobNet() [x.cuda_() for x in model.parameters()] optimizer = optim.SGD(model.parameters(), lr=0.001) train(model, optimizer, steps=1000, gpu=True) evaluate(model, gpu=True)
def train(model):
optim = optimizer.SGD([model.c1, model.c2, model.l1], lr=0.001)
iters = 5 * classes
BS = 32
log_skip = 5
print(f"training with {classes} emojis over {BS * iters} iters batched by {BS}")
import time
t0 = time.time()
for _ in range(iters):
samp = np.random.randint(0, files.shape[0], size=(BS))
img = to_gpu(Tensor(perturb(files[samp])))
out = model.forward(img)
y = np.zeros((BS, classes), dtype=np.float32)
# cross entropy loss trick
y[range(y.shape[0]), samp] = -classes
y = to_gpu(Tensor(y))
loss = out.mul(y).mean()
optim.zero_grad()
loss.backward()
optim.step()
if _ % log_skip == 0:
t1 = time.time()
iters_sec = log_skip * BS / (t1 - t0)
minutes = (BS * (iters - _) / iters_sec) / 60
sys.stdout.write("\u001b[1000D")
sys.stdout.write("\u001b[0K")
sys.stdout.write(
f"[{100 * _ / iters:.2f}% ~{int(minutes)}min left] loss={loss.cpu().data[0]:.2f}"
)
sys.stdout.write(
f" (running @ {iters_sec:.2f} iters/s, feel free to kill at any time)"
)
sys.stdout.flush()
t0 = t1
def test_sgd(self):
np.random.seed(1337)
model = TinyBobNet()
optimizer = optim.SGD(model.parameters(), lr=0.001)
train(model,
X_train,
Y_train,
optimizer,
steps=1000,
device=self.device)
assert evaluate(model, X_test, Y_test, device=self.device) > 0.95
def test_sgd(self):
np.random.seed(1337)
model = TinyBobNet()
optimizer = optim.SGD(model.parameters(), lr=0.001)
train(model, optimizer, steps=1000, gpu=self.gpu)
evaluate(model, gpu=self.gpu)
def test_sgd(self): np.random.seed(1337) model = TinyBobNet() optimizer = optim.SGD([model.l1, model.l2], lr=0.001) train(model, optimizer, steps=1000) evaluate(model)
