def test_fake_quant_per_channel_other_prec(self):
kernel_size = 3
quant_desc_input = QuantDescriptor(num_bits=4)
quant_desc_weight = QuantDescriptor(num_bits=3, axis=(0))
quant_conv_object = quant_conv.QuantConv1d(
_NUM_IN_CHANNELS,
_NUM_OUT_CHANNELS,
kernel_size,
bias=False,
quant_desc_input=quant_desc_input,
quant_desc_weight=quant_desc_weight)
test_input = torch.randn(16, _NUM_IN_CHANNELS, 16)
test_input_quantizer = TensorQuantizer(quant_desc_input)
weight_quantizer = TensorQuantizer(quant_desc_weight)
quant_input = test_input_quantizer(test_input)
weight_copy = quant_conv_object.weight.clone()
quant_weight = weight_quantizer(weight_copy)
out1 = F.conv1d(quant_input, quant_weight)
out2 = quant_conv_object(test_input)
np.testing.assert_array_equal(out1.detach().cpu().numpy(), out2.detach().cpu().numpy())
def test_against_unquantized(self):
kernel_size = 3
test_input = torch.randn(16, _NUM_IN_CHANNELS, 24).cuda()
torch.manual_seed(12345)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(12345)
fake_quant_conv1d = quant_conv.QuantConv1d(
_NUM_IN_CHANNELS,
_NUM_OUT_CHANNELS,
kernel_size,
bias=True,
quant_desc_input=QuantDescriptor(num_bits=16),
quant_desc_weight=QuantDescriptor(num_bits=16, axis=(0)))
# Reset seed. Make sure weight and bias are the same
torch.manual_seed(12345)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(12345)
conv1d = nn.Conv1d(_NUM_IN_CHANNELS, _NUM_OUT_CHANNELS, kernel_size, bias=True)
fake_quant_output = fake_quant_conv1d(test_input)
output = conv1d(test_input)
test_utils.compare(fake_quant_output, output, rtol=1e-5, atol=1e-4)
def test_fake_quant_per_channel_bias(self):
kernel_size = 3
quant_conv_object = quant_conv.QuantConv1d(
_NUM_IN_CHANNELS,
_NUM_OUT_CHANNELS,
kernel_size,
bias=True,
quant_desc_weight=QuantDescriptor(axis=(0)))
test_input = torch.randn(16, _NUM_IN_CHANNELS, 16)
quant_input = tensor_quant.fake_tensor_quant(
test_input, torch.max(torch.abs(test_input)))
weight_copy = quant_conv_object.weight.clone()
quant_weight = tensor_quant.fake_tensor_quant(
weight_copy,
torch.max(torch.abs(weight_copy).view(_NUM_OUT_CHANNELS, -1),
dim=1,
keepdim=True)[0].view(_NUM_OUT_CHANNELS, 1, 1))
out1 = F.conv1d(quant_input, quant_weight, bias=quant_conv_object.bias)
out2 = quant_conv_object(test_input)
np.testing.assert_array_equal(out1.detach().cpu().numpy(),
out2.detach().cpu().numpy())
def test_fake_quant_per_tensor(self):
kernel_size = 3
quant_conv_object = quant_conv.QuantConv1d(
_NUM_IN_CHANNELS, _NUM_OUT_CHANNELS, kernel_size, bias=False, quant_desc_weight=QuantDescriptor())
test_input = torch.randn(16, _NUM_IN_CHANNELS, 16)
quant_input = tensor_quant.fake_tensor_quant(test_input, torch.max(torch.abs(test_input)))
weight_copy = quant_conv_object.weight.clone()
quant_weight = tensor_quant.fake_tensor_quant(weight_copy, torch.max(torch.abs(weight_copy)))
out1 = F.conv1d(quant_input, quant_weight)
out2 = quant_conv_object(test_input)
np.testing.assert_array_equal(out1.detach().cpu().numpy(), out2.detach().cpu().numpy())
def test_fake_quant_input(self):
kernel_size = 3
quant_conv_object = quant_conv.QuantConv1d(
_NUM_IN_CHANNELS,
_NUM_OUT_CHANNELS,
kernel_size,
bias=False)
quant_conv_object.weight_quantizer.disable()
test_input = torch.randn(20, _NUM_IN_CHANNELS, 50)
quant_input = tensor_quant.fake_tensor_quant(test_input, torch.max(torch.abs(test_input)))
out1 = F.conv1d(quant_input, quant_conv_object.weight)
out2 = quant_conv_object(test_input)
np.testing.assert_array_equal(out1.detach().cpu().numpy(), out2.detach().cpu().numpy())
def test_no_quant(self):
kernel_size = 8
quant_conv_object = quant_conv.QuantConv1d(
_NUM_IN_CHANNELS,
_NUM_OUT_CHANNELS,
kernel_size,
bias=False)
quant_conv_object.input_quantizer.disable()
quant_conv_object.weight_quantizer.disable()
test_input = torch.randn(16, _NUM_IN_CHANNELS, 256)
weight_copy = quant_conv_object.weight.clone()
quant_weight = weight_copy
out1 = F.conv1d(test_input, quant_weight)
out2 = quant_conv_object(test_input)
np.testing.assert_array_equal(out1.detach().cpu().numpy(), out2.detach().cpu().numpy())
def test_weight_fake_quant_per_channel(self):
kernel_size = 3
quant_conv_object = quant_conv.QuantConv1d(
_NUM_IN_CHANNELS,
_NUM_OUT_CHANNELS,
kernel_size,
bias=False,
quant_desc_weight=QuantDescriptor(axis=(0)))
quant_conv_object.input_quantizer.disable()
test_input = torch.randn(16, _NUM_IN_CHANNELS, 256)
weight_copy = quant_conv_object.weight.clone()
amax = quant_utils.reduce_amax(weight_copy, axis=(1, 2))
quant_weight = tensor_quant.fake_tensor_quant(weight_copy, amax)
out1 = F.conv1d(test_input, quant_weight)
out2 = quant_conv_object(test_input)
np.testing.assert_array_equal(out1.detach().cpu().numpy(), out2.detach().cpu().numpy())