def run(
N,
IC,
OC,
IH,
IW,
KH,
KW,
PH,
PW,
SH,
SW,
has_bias=True,
):
inp_v = np.random.normal(size=(N, IC, IH, IW))
w_v = np.random.normal(size=(N, OC, IC, KH, KW))
b_v = np.random.normal(size=(1, OC, 1, 1))
inp_scale = dtype.get_scale(inp_dtype)
w_scale = dtype.get_scale(w_dtype)
b_scale = dtype.get_scale(b_dtype)
inpv = dtype.convert_to_qint8(inp_v * inp_scale, inp_dtype)
wv = dtype.convert_to_qint8(w_v * w_scale, w_dtype)
bv = dtype.convert_to_qint32(b_v * b_scale, b_dtype)
inp_int8 = tensor(inpv, dtype=inp_dtype)
w_int8 = Parameter(wv, dtype=w_dtype)
b_int32 = Parameter(bv, dtype=b_dtype)
inp_fp32 = inp_int8.astype("float32")
w_fp32 = w_int8.astype("float32")
b_fp32 = b_int32.astype("float32")
def run_batch_conv_bias(inp, w, b):
b = b if has_bias else Parameter(np.zeros_like(b.numpy()))
result = F.quantized.batch_conv_bias_activation(
inp,
w,
b,
stride=(SH, SW),
padding=(PH, PW),
dtype=out_dtype,
)
return result.astype("float32")
expected = F.conv2d(inp_fp32, w_fp32[0],
b_fp32 if has_bias else None)[0]
expected = expected.astype(out_dtype).astype("float32")
expected = F.flatten(expected)
result = run_batch_conv_bias(inp_int8, w_int8, b_int32)
result = F.flatten(result)
np.testing.assert_allclose(result.numpy(),
expected.numpy(),
atol=outp_scale)
def run(
N,
IC,
OC,
IH,
IW,
KH,
KW,
PH,
PW,
SH,
SW,
has_bias=True,
nonlinear_mode="identity",
):
inp_v = np.random.normal(size=(N, IC, IH, IW))
w_v = np.random.normal(size=(OC, IC, KH, KW))
b_v = np.random.normal(size=(1, OC, 1, 1))
inp_scale = dtype.get_scale(inp_dtype)
w_scale = dtype.get_scale(w_dtype)
b_scale = dtype.get_scale(b_dtype)
inpv = dtype.convert_to_qint8(inp_v * inp_scale, inp_dtype)
wv = dtype.convert_to_qint8(w_v * w_scale, w_dtype)
bv = dtype.convert_to_qint32(b_v * b_scale, b_dtype)
inp_int8 = tensor(inpv, dtype=inp_dtype)
w_int8 = Parameter(wv, dtype=w_dtype)
b_int32 = Parameter(bv, dtype=b_dtype)
inp_fp32 = inp_int8.astype("float32")
w_fp32 = w_int8.astype("float32")
b_fp32 = b_int32.astype("float32")
def convert_to_nchw4(var):
var = F.reshape(var, (var.shape[0], var.shape[1] // 4, 4,
var.shape[2], var.shape[3]))
var = F.transpose(var, (0, 1, 3, 4, 2))
return var
def run_conv2d(inp, w, b):
O = F.conv2d(
inp,
w,
b if has_bias else None,
stride=(SH, SW),
padding=(PH, PW),
)
if nonlinear_mode == "relu":
return F.relu(O)
else:
return O
def run_conv_bias(inp, w, b, format="NCHW"):
b = b if has_bias else Parameter(np.zeros_like(b.numpy()))
if format == "NCHW4":
inp = convert_to_nchw4(inp)
w = convert_to_nchw4(w)
b = convert_to_nchw4(b)
return F.quantized.conv_bias_activation(
inp,
w,
b,
stride=(SH, SW),
padding=(PH, PW),
dtype=out_dtype,
nonlinear_mode=nonlinear_mode,
)
format = "NCHW4" if is_cuda_available() else "NCHW"
expected = run_conv2d(inp_fp32, w_fp32, b_fp32)
expected = expected.astype(out_dtype).astype("float32")
result = run_conv_bias(inp_int8, w_int8, b_int32,
format=format).astype("float32")
if format == "NCHW4":
result = F.transpose(result, (0, 1, 4, 2, 3))
expected = F.flatten(expected)
result = F.flatten(result)
np.testing.assert_allclose(result.numpy(),
expected.numpy(),
atol=outp_scale)
def run(
N,
IC,
OC,
IH,
IW,
KH,
KW,
PH,
PW,
SH,
SW,
has_bias=True,
nonlinear_mode="IDENTITY",
):
inp_v = np.random.normal(size=(N, IC, IH, IW))
w_v = np.random.normal(size=(OC, IC, KW, KW))
b_v = np.random.normal(size=(1, OC, 1, 1))
inp_scale = mgb.dtype.get_scale(inp_dtype)
w_scale = mgb.dtype.get_scale(w_dtype)
b_scale = mgb.dtype.get_scale(b_dtype)
inpv = mgb.dtype.convert_to_qint8(inp_v * inp_scale, inp_dtype)
wv = mgb.dtype.convert_to_qint8(w_v * w_scale, w_dtype)
bv = mgb.dtype.convert_to_qint32(b_v * b_scale, b_dtype)
inp_int8 = tensor(inpv, dtype=inp_dtype)
w_int8 = Parameter(wv, dtype=w_dtype)
b_int32 = Parameter(bv, dtype=b_dtype)
inp_fp32 = inp_int8.astype("float32")
w_fp32 = w_int8.astype("float32")
b_fp32 = b_int32.astype("float32")
jit.trace.enabled = True
b_symbolic = True
def convert_to_nchw4(var):
return var.reshape(var.shapeof(0),
var.shapeof(1) // 4, 4, var.shapeof(2),
var.shapeof(3)).dimshuffle(0, 1, 3, 4, 2)
@jit.trace(symbolic=b_symbolic)
def run_conv2d(inp, w, b):
O = F.conv2d(
inp,
w,
b if has_bias else None,
stride=(SH, SW),
padding=(PH, PW),
)
if nonlinear_mode == "RELU":
return F.relu(O)
else:
return O
@jit.trace(symbolic=b_symbolic)
def run_conv_bias(inp, w, b, format="NCHW"):
b = b if has_bias else np.zeros_like(b)
if format == "NCHW4":
inp = convert_to_nchw4(inp)
w = convert_to_nchw4(w)
b = F.flatten(b)
return F.conv_bias_activation(
inp,
w,
b,
stride=(SH, SW),
padding=(PH, PW),
dtype=out_dtype,
nonlinear_mode=nonlinear_mode,
)
format = "NCHW4" if is_cuda_available() else "NCHW"
expected = run_conv2d(inp_fp32, w_fp32, b_fp32)
expected = expected.astype(out_dtype).astype("float32")
result = run_conv_bias(inp_int8, w_int8, b_int32,
format=format).astype("float32")
if format == "NCHW4":
result = result.dimshuffle(0, 1, 4, 2, 3)
expected = F.flatten(expected)
result = F.flatten(result)
assertTensorClose(result.numpy(), expected.numpy())
