def test_serializable(self):
device_params = test_device_matmul(max_batch=5,
max_hidden=9,
max_classes=17,
repeats=1,
device='CPU:0',
dtype='float32')
json.dumps(device_params)
def test_normal(self):
device_params = test_device_matmul(max_batch=4,
max_hidden=8,
max_classes=128,
repeats=1,
device='CPU:0',
dtype='float32')
approx_cost = interpolate_matmul_cost(device_params)
batch4 = approx_cost(4, 8, 16)
batch3 = approx_cost(3, 8, 16)
self.assertAlmostEqual(batch3, batch4, places=1)
def tes_failure(self):
device_params = test_device_matmul(max_batch=4,
max_hidden=8,
max_classes=128,
repeats=1,
device='GPU:0',
dtype='float32')
approx_cost = interpolate_matmul_cost(device_params)
with self.assertRaises(ValueError):
approx_cost(4, 9, 16)
with self.assertRaises(ValueError):
approx_cost(1, 9, 1)
def test_normal_cpu(self):
device_params = test_device_matmul(max_batch=5,
max_hidden=9,
max_classes=17,
repeats=1,
device='CPU:0',
dtype='float32')
self.assertListEqual([1, 2, 3, 4, 5], device_params['batch_sizes'])
self.assertListEqual([1, 2, 3, 4, 5, 7, 8, 9],
device_params['hidden_sizes'])
self.assertListEqual([1, 2, 3, 4, 5, 7, 8, 9, 15, 16, 17],
device_params['class_sizes'])
self.assertLen(device_params['cost_values'], 5 * 8 * 11)
def test_normalize_to_worst_adaptive(self):
device_params = test_device_matmul(max_batch=8,
max_hidden=25,
max_classes=110,
repeats=1,
device='CPU:0',
dtype='float32')
freq_vocab = build_zipf_vocab(120)
batch_sizes, head_sizes, speed_ups, best_splits = estimate_best_splits(
device_params, freq_vocab, num_tails=2, hidden_size=24, factor=2)
self.assertListEqual([1, 2, 3, 4, 5, 7, 8], batch_sizes)
self.assertListEqual([14, 22, 30], head_sizes)
self.assertLen(speed_ups, 7 * 3)
self.assertLen(best_splits, 7 * 3)
def test_normal_gpu(self):
if not tf.test.is_gpu_available(cuda_only=True):
self.skipTest('No GPU available')
device_params = test_device_matmul(max_batch=32,
max_hidden=8,
max_classes=16,
repeats=1,
device='GPU:0',
dtype='float32')
self.assertListEqual(
[1, 2, 3, 4, 5, 7, 8, 9, 15, 16, 17, 23, 24, 25, 31, 32],
device_params['batch_sizes'])
self.assertListEqual([1, 2, 3, 4, 5, 7, 8],
device_params['hidden_sizes'])
self.assertListEqual([1, 2, 3, 4, 5, 7, 8, 9, 15, 16],
device_params['class_sizes'])
self.assertLen(device_params['cost_values'], 16 * 7 * 10)
def test_normal(self):
device_params = test_device_matmul(max_batch=8,
max_hidden=25,
max_classes=30,
repeats=1,
device='CPU:0',
dtype='float32')
approx_cost = interpolate_matmul_cost(device_params)
freq_vocab = build_zipf_vocab(30)
all_freq = np.array([f for _, f in freq_vocab.most_common()])
all_prob = all_freq / np.sum(all_freq)
prob_accum = np.cumsum(all_prob)
split_size = [2, 9, 19]
adaptive_split_cost(approx_cost,
prob_accum,
split_size,
batch_size=6,
hidden_size=24,
factor=2)