def test_qadd_relu_same_qparams(self):
add_relu = torch.ops.quantized.add_relu
add = torch.ops.quantized.add
A = torch.arange(-25, 25, dtype=torch.float)
B = torch.arange(-25, 25, dtype=torch.float)
scale = 2.0
zero_point = 127
qA = torch.quantize_linear(A, scale=scale, zero_point=zero_point,
dtype=torch.quint8)
qB = torch.quantize_linear(B, scale=scale, zero_point=zero_point,
dtype=torch.quint8)
# Add ReLU ground truth
C = (qA.dequantize() + qB.dequantize()).numpy()
qC = _quantize(C, scale, zero_point)
qC_hat = add(qA, qB, scale=scale, zero_point=zero_point)
np.testing.assert_equal(qC, qC_hat.int_repr(),
"Quantized addition failed.")
# Add + ReLU ground truth
Crelu = C.copy()
Crelu[C < 0] = 0
qCrelu = _quantize(Crelu, scale, zero_point)
qCrelu_hat = add_relu(qA, qB, scale=scale, zero_point=zero_point)
np.testing.assert_equal(qCrelu, qCrelu_hat.int_repr(),
"Quantized addition with ReLU failed.")
def qlinear_ref(X_q, X_scale, X_zp, W_q, W_scale, W_zp, b_q, Y_scale, Y_zp):
row_offsets_ref = X_q.sum(axis=1).astype(np.int32).reshape((-1, 1))
col_offsets_ref = W_q.sum(axis=1).astype(np.int32).reshape((1, -1))
assert X_q.ndim == 2
batch_size, input_channels = X_q.shape
Prod_XqWq_ref = (np.matmul(X_q.astype(np.int32),
W_q.astype(np.int32).T) -
W_zp * row_offsets_ref - X_zp * col_offsets_ref +
input_channels * X_zp * W_zp)
Y_q_ref = _quantize(Prod_XqWq_ref + b_q, Y_scale / (X_scale * W_scale),
Y_zp)
return Y_q_ref
def test_qnnpack_relu(self, Q):
X, (scale, zero_point), (qmin, qmax), (torch_type, np_type) = Q
relu = torch.ops.quantized.qnnpack_relu
Y = X.copy()
X = torch.from_numpy(X)
qX = torch.quantize_linear(X, scale=scale, zero_point=zero_point, dtype=torch_type)
qY_hat = relu(qX)
Y[Y < 0] = 0
qY = _quantize(Y, scale, zero_point, dtype=np_type)
np.testing.assert_equal(qY, qY_hat.int_repr())