def relay_lstm_cell(batch_size, input_size, hidden_size):
# based on https://pytorch.org/docs/stable/generated/torch.nn.GRU.html#torch.nn.GRU
state_tensor_type = relay.TensorType((batch_size, hidden_size))
state_tuple_type = relay.TupleType([state_tensor_type, state_tensor_type])
inp = relay.var("input", shape=(batch_size, input_size))
state = relay.Var("state", type_annotation=state_tuple_type)
w_ih = relay.var("w_ih", shape=(4 * hidden_size, input_size))
w_hh = relay.var("w_hh", shape=(4 * hidden_size, hidden_size))
b_ih = relay.var("b_ih", shape=(4 * hidden_size, ))
b_hh = relay.var("b_hh", shape=(4 * hidden_size, ))
hidden = relay.TupleGetItem(state, 0)
cell_state = relay.TupleGetItem(state, 1)
# PyTorch packs the i2h and h2h weights and biases together so we will match that here
w_i_splits = relay.split(w_ih, 4, 0)
w_h_splits = relay.split(w_hh, 4, 0)
b_i_splits = relay.split(b_ih, 4, 0)
b_h_splits = relay.split(b_hh, 4, 0)
w_ii, w_if, w_ig, w_io = w_i_splits[0], w_i_splits[1], w_i_splits[
2], w_i_splits[3]
w_hi, w_hf, w_hg, w_ho = w_h_splits[0], w_h_splits[1], w_h_splits[
2], w_h_splits[3]
b_ii, b_if, b_ig, b_io = b_i_splits[0], b_i_splits[1], b_i_splits[
2], b_i_splits[3]
b_hi, b_hf, b_hg, b_ho = b_h_splits[0], b_h_splits[1], b_h_splits[
2], b_h_splits[3]
def weighted_value(weight, value, bias):
return relay.transpose(
relay.nn.dense(weight, value) +
relay.reshape(bias, (hidden_size, 1)))
i_t = relay.sigmoid(
weighted_value(w_ii, inp, b_ii) + weighted_value(w_hi, hidden, b_hi))
f_t = relay.sigmoid(
weighted_value(w_if, inp, b_if) + weighted_value(w_hf, hidden, b_hf))
g_t = relay.tanh(
weighted_value(w_ig, inp, b_ig) + weighted_value(w_hg, hidden, b_hg))
o_t = relay.sigmoid(
weighted_value(w_io, inp, b_io) + weighted_value(w_ho, hidden, b_ho))
c_t = f_t * cell_state + i_t * g_t
h_t = o_t * relay.tanh(c_t)
h_var = relay.Var("h")
c_var = relay.Var("c")
return relay.Function(
[inp, state, w_ih, w_hh, b_ih, b_hh],
relay.Let(
h_var, h_t,
relay.Let(c_var, c_t,
relay.Tuple([h_var, relay.Tuple([h_var, c_var])]))),
ret_type=relay.TupleType([state_tensor_type, state_tuple_type]))
def test_conv2d_residual_block():
d_shape = (16, 16, 32, 32)
w_shape = (16, 16, 3, 3)
padding = (1, 1)
bias_add, residual_input = get_conv2d_nchw_bias_residual(
d_shape, w_shape, padding)
for func, tol in [
(relay.nn.relu(bias_add + residual_input), 1e-5),
(relay.nn.relu(bias_add) + residual_input, 1e-5),
(relay.sigmoid(bias_add) * residual_input, 1e-5),
(relay.nn.relu(silu(bias_add) * residual_input), 1e-5),
# HardSwish requires higher tolerance since vectoring the residual block epilogue
# in cutlass.
# TODO(masahi): Invesitigate this issue
(relay.nn.relu(hardswish(bias_add) + residual_input), 1e-3),
]:
verify_conv2d(func,
func,
d_shape,
w_shape,
sm=80,
atol=tol,
rtol=tol,
run_benchmark=False)
def after():
data = relay.var("data", shape=(1, 32))
eq1 = relay.var("e1", shape=[], dtype="float32")
eq2 = relay.var("e2", shape=[], dtype="float32")
cb_1 = relay.annotation.compiler_begin(eq1, target)
cb_2 = relay.annotation.compiler_begin(eq2, target)
equality_condition = relay.equal(cb_1, cb_2)
ce_1 = relay.annotation.compiler_end(equality_condition, target)
# if condition
cb_3 = relay.annotation.compiler_begin(data, target)
true_branch = relay.tanh(cb_3)
ce_2 = relay.annotation.compiler_end(true_branch, target)
# else condition
cb_4 = relay.annotation.compiler_begin(data, target)
false_branch = relay.sigmoid(cb_4)
ce_3 = relay.annotation.compiler_end(false_branch, target)
if_condition = relay.If(ce_1, ce_2, ce_3)
cb_5 = relay.annotation.compiler_begin(if_condition, target)
erf_out = relay.erf(cb_5)
ce_4 = relay.annotation.compiler_end(erf_out, target)
func = relay.Function([data, eq1, eq2], ce_4)
mod = tvm.IRModule.from_expr(func)
return mod
def before(dim):
X = relay.var("X", shape=(1, dim))
W = relay.var("W", shape=(3 * dim, dim))
matmul = relay.nn.dense(X, W)
splitted = relay.split(matmul, indices_or_sections=3, axis=1)
out = relay.sigmoid(splitted[0]) + relay.tanh(splitted[1]) * relay.exp(splitted[2])
return relay.Function([X, W], out)
def test_batch_matmul_rewrite():
data = relay.var("data", shape=(1, 4, 16, 16))
data2 = relay.sigmoid(relay.var("data", shape=(4, 16, 64)))
out = relay.nn.conv2d(data,
relay.var("weight"),
kernel_size=(3, 3),
padding=(1, 1),
channels=8)
out = relay.nn.batch_flatten(out)
out = relay.reshape(out, [1, 32, 64])
out = relay.nn.batch_matmul(out, data2)
qmod = quantize_and_build(out)
def _check_batch_matmul(node):
if isinstance(node, Call):
if node.op.name in ["nn.batch_matmul", "nn.conv2d"]:
assert node.checked_type.dtype == "int32"
elif node.op.name == "nn.batch_flatten":
assert node.checked_type.dtype == "int8"
# check if batch_matmul is quantized
relay.analysis.post_order_visit(qmod["main"], _check_batch_matmul)
def get_conv2d_transpose(
x_shape=(1, 32, 8, 8),
k_shape=(32, 16, 3, 3),
groups=1,
padding=(0, 0),
strides=(1, 1),
activation=None,
dtype="float32",
):
x = relay.var("x", shape=(x_shape), dtype=dtype)
kernel = relay.var("kernel", shape=(k_shape), dtype=dtype)
out = relay.nn.conv2d_transpose(
x,
kernel,
channels=k_shape[1] * groups,
kernel_size=k_shape[2:4],
groups=groups,
padding=padding,
strides=strides,
)
dic = {"x": x_shape, "kernel": k_shape}
param_lst = ["kernel"]
if activation == "relu":
return relay.nn.relu(out), dic, param_lst
elif activation == "tanh":
return relay.tanh(out), dic, param_lst
elif activation == "sigmoid":
return relay.sigmoid(out), dic, param_lst
else:
return out, dic, param_lst
def get_blocks(prefix,
data,
in_channel,
out_channel,
include_bn=True,
include_sigmoid=False):
weight = relay.var(prefix + "weight")
bn_gamma = relay.var(prefix + "bn_gamma")
bn_beta = relay.var(prefix + "bn_beta")
bn_mmean = relay.var(prefix + "bn_mean")
bn_mvar = relay.var(prefix + "bn_var")
layer = relay.nn.conv2d(data=data,
weight=weight,
kernel_size=(3, 3),
channels=out_channel,
padding=(1, 1))
if include_bn:
bn_output = relay.nn.batch_norm(layer, bn_gamma, bn_beta, bn_mmean,
bn_mvar)
layer = bn_output[0]
if include_sigmoid:
# dummy layer to prevent pattern detection
layer = relay.sigmoid(layer)
layer = relay.nn.relu(layer)
return layer
def get_conv3d(
x_shape=(1, 32, 8, 8, 8),
k_shape=(16, 32, 3, 3, 3),
groups=1,
padding=(0, 0, 0),
strides=(1, 1, 1),
dilation=(1, 1, 1),
activation=None,
dtype="float32",
):
x = relay.var("x", shape=(x_shape), dtype=dtype)
kernel = relay.const(np.random.randint(0, 1, k_shape).astype(dtype))
out = relay.nn.conv3d(
x,
kernel,
channels=k_shape[0],
kernel_size=k_shape[2:],
groups=groups,
padding=padding,
strides=strides,
dilation=dilation,
)
dic = {"x": x_shape, "kernel": k_shape}
param_lst = ["kernel"]
if activation == "relu":
return relay.nn.relu(out), dic, param_lst
elif activation == "tanh":
return relay.tanh(out), dic, param_lst
elif activation == "sigmoid":
return relay.sigmoid(out), dic, param_lst
else:
return out, dic, param_lst
def do_layer(x, node: Node, params):
if isinstance(node, Conv):
x = conv2d(x, node, params)
elif isinstance(node, Pool):
x = pool2d(x, node)
elif isinstance(node, Identity):
x = x
elif isinstance(node, Activation):
if node.act_type == 'relu':
x = relay.nn.relu(x)
elif node.act_type == 'sigmoid':
x = relay.sigmoid(x)
elif node.act_type == 'tanh':
x = relay.tanh(x)
else:
raise ValueError
elif isinstance(node, Element):
x = x # has been done in terms
elif isinstance(node, Relu):
x = relay.nn.relu(x)
elif isinstance(node, Sequential):
x = sequential(x, node, params)
else:
raise NotImplementedError
return x
def before(dim):
X = relay.var("X", shape=(1, dim))
W = relay.var("W", shape=(3 * dim, dim))
matmul = relay.nn.dense(X, W)
splitted = relay.split(matmul, indices_or_sections=3, axis=1)
out = relay.sigmoid(splitted[0]) + relay.tanh(splitted[1]) * relay.exp(splitted[2])
return relay.Function([X, W], out)
def get_conv1d(
x_shape=((1, 3, 224)),
k_shape=(16, 3, 3),
groups=1,
padding=(1, 1),
strides=(1),
dilation=(1),
channels=None,
activation=None,
dtype="float32",
):
x = relay.var("x", shape=(x_shape), dtype=dtype)
kernel = relay.var("kernel", shape=(k_shape), dtype=dtype)
out = relay.nn.conv1d(
x,
kernel,
kernel_size=k_shape[2:3],
groups=groups,
padding=padding,
strides=strides,
dilation=dilation,
channels=k_shape[0],
)
dic = {"x": x_shape, "kernel": k_shape}
param_lst = ["kernel"]
if activation == "relu":
return relay.nn.relu(out), dic, param_lst
elif activation == "tanh":
return relay.tanh(out), dic, param_lst
elif activation == "sigmoid":
return relay.sigmoid(out), dic, param_lst
else:
return out, dic, param_lst
def test_stop_quantize():
data = relay.var("data", shape=(1, 16, 64, 64))
np_weight0 = np.random.rand(16, 16, 3, 3)
conv0_weight = relay.Constant(tvm.nd.array(np_weight0)).astype("float32")
np_weight1 = np.random.rand(16, 16, 1, 1)
conv1_weight = relay.Constant(tvm.nd.array(np_weight1)).astype("float32")
multiplier = relay.sigmoid(relay.var("data", shape=(1, 16, 1, 1)))
conv0 = relay.nn.conv2d(data,
conv0_weight,
kernel_size=(3, 3),
padding=(1, 1),
channels=16)
act0 = relay.nn.relu(data=conv0)
pool = relay.nn.global_avg_pool2d(data=act0)
conv1 = relay.nn.conv2d(pool,
conv1_weight,
kernel_size=(1, 1),
padding=(0, 0),
channels=16)
act1 = relay.nn.relu(data=conv1)
quantize_and_build(act1 * multiplier)
def get_model(input_shape, var_names):
"""Return a model and any parameters it may have."""
a = relay.var(next(var_names), shape=input_shape, dtype="float32")
out = relay.reshape(a, (1, 1, 1000))
out = relay.sigmoid(out)
out = relay.reshape(out, (1, 1000))
return out
def test_sigmoid(self):
data = relay.var("data", relay.TensorType((-1, 4, 2, 2), "float32"))
net = relay.sigmoid(data)
net = relay.Function(relay.analysis.free_vars(net), net)
mod, params = testing.create_workload(net)
xgraph = xf_relay.from_relay(mod, params)
layers = xgraph.get_layers()
assert layers[0].type[0] == 'Input'
assert layers[1].type[0] == 'Sigmoid'
assert 'relay_id' in layers[1].attrs
def before():
data = relay.var("data", shape=(1, 32))
eq1 = relay.var("e1", shape=[], dtype="float32")
eq2 = relay.var("e2", shape=[], dtype="float32")
eq = relay.equal(eq1, eq2)
true_branch = relay.tanh(data)
false_branch = relay.sigmoid(data)
ife = relay.If(eq, true_branch, false_branch)
out = relay.erf(ife)
func = relay.Function([data, eq1, eq2], out)
mod = tvm.IRModule.from_expr(func)
return mod
def test_mul_rewrite():
"""a test case where rhs of mul is not constant"""
data = relay.var("data", shape=(1, 16, 64, 64))
multiplier = relay.sigmoid(relay.var("data", shape=(1, 16, 1, 1)))
conv = relay.nn.conv2d(
data, relay.var("weight"), kernel_size=(3, 3), padding=(1, 1), channels=16
)
act = relay.nn.relu(data=conv)
quantize_and_build(act * multiplier)
pool = relay.nn.global_avg_pool2d(data=act)
quantize_and_build(act * pool)
def _get_model(shape, input_zp, input_sc, output_zp, output_sc, dtype):
a = relay.var("a", shape=shape, dtype=dtype)
dequantize = relay.qnn.op.dequantize(
a,
input_scale=relay.const(input_sc, "float32"),
input_zero_point=relay.const(input_zp, "int32"),
)
sigmoid = relay.sigmoid(dequantize)
model = relay.qnn.op.quantize(
sigmoid,
output_scale=relay.const(output_sc, "float32"),
output_zero_point=relay.const(output_zp, "int32"),
out_dtype=dtype,
)
return model
def get_conv1d_bias(x_shape=(1, 3, 224), k_shape=(10, 3, 3), activation=None, dtype="float32"):
conv, dic, param_lst = get_conv1d(x_shape=x_shape, k_shape=k_shape, dtype=dtype)
bias = relay.var("bias", shape=(k_shape[0],), dtype=dtype)
out = relay.nn.bias_add(conv, bias)
dic["bias"] = (k_shape[0],)
param_lst += ["bias"]
if activation == "relu":
return relay.nn.relu(out), dic, param_lst
elif activation == "tanh":
return relay.tanh(out), dic, param_lst
elif activation == "sigmoid":
return relay.sigmoid(out), dic, param_lst
else:
return out, dic, param_lst
def expected(dim):
p0 = relay.var("p0", shape=(1, dim))
p1 = relay.var("p1", shape=(3 * dim, dim))
matmul = relay.nn.dense(p0, p1)
f0 = relay.Function([p0, p1], matmul)
p01 = relay.var("p01", shape=(1, 3 * dim))
splitted = relay.split(p01, indices_or_sections=3, axis=1)
out = relay.sigmoid(splitted[0]) + relay.tanh(splitted[1]) * relay.exp(splitted[2])
f1 = relay.Function([p01], out)
X = relay.var("X", shape=(1, dim))
W = relay.var("W", shape=(3 * dim, dim))
y = relay.Call(f0, [X, W])
z = relay.Call(f1, [y])
return relay.Function([X, W], z)
def expected(dim):
p0 = relay.var("p0", shape=(1, dim))
p1 = relay.var("p1", shape=(3 * dim, dim))
matmul = relay.nn.dense(p0, p1)
f0 = relay.Function([p0, p1], matmul)
p01 = relay.var("p01", shape=(1, 3 * dim))
splitted = relay.split(p01, indices_or_sections=3, axis=1)
out = relay.sigmoid(splitted[0]) + relay.tanh(splitted[1]) * relay.exp(splitted[2])
f1 = relay.Function([p01], out)
X = relay.var("X", shape=(1, dim))
W = relay.var("W", shape=(3 * dim, dim))
y = relay.Call(f0, [X, W])
z = relay.Call(f1, [y])
return relay.Function([X, W], z)
def expected(dim):
p0 = relay.var("p0", shape=(1, dim))
p1 = relay.var("p1", shape=(3 * dim, dim))
matmul = relay.nn.dense(p0, p1)
f0 = relay.Function([p0, p1], matmul)
f0 = f0.set_attribute("Primitive", tvm.tir.IntImm("int32", 1))
p01 = relay.var("p01", shape=(1, 3 * dim))
splitted = relay.split(p01, indices_or_sections=3, axis=1)
out = relay.sigmoid(splitted[0]) + relay.tanh(splitted[1]) * relay.exp(splitted[2])
f1 = relay.Function([p01], out)
f1 = f1.set_attribute("Primitive", tvm.tir.IntImm("int32", 1))
X = relay.var("X", shape=(1, dim))
W = relay.var("W", shape=(3 * dim, dim))
y = relay.Call(f0, [X, W])
z = relay.Call(f1, [y])
return relay.Function([X, W], z)
def se_block(data,
name,
input_channels,
ratio=0.25,
layout='NCHW',
dtype='float32'):
pool = relay.nn.global_avg_pool2d(data, layout=layout)
flatten = relay.nn.batch_flatten(data=pool)
dense1 = layers.dense_add_bias(flatten,
units=int(input_channels * ratio),
name=name + '_dense1')
relu = relay.nn.relu(data=dense1)
dense2 = layers.dense_add_bias(relu,
units=input_channels,
name=name + '_dense2')
sigmoid = relay.sigmoid(dense2)
sigmoid = relay.expand_dims(sigmoid,
axis=(-1 if layout == 'NCHW' else 1),
num_newaxis=2)
mul = relay.multiply(data, sigmoid)
return mul
def get_conv3d_transpose(
x_shape=(1, 32, 8, 8, 8),
k_shape=(32, 16, 3, 3, 3),
groups=1,
padding=(0, 0, 0),
strides=(1, 1, 1),
output_padding=(0, 0, 0),
activation=None,
dtype="float32",
data_layout="NCDHW",
kernel_layout="OIDHW",
):
x = relay.var("x", shape=(x_shape), dtype=dtype)
kernel = relay.const(np.random.randint(0, 1, k_shape).astype(dtype))
out = relay.nn.conv3d_transpose(
x,
kernel,
channels=k_shape[1],
kernel_size=k_shape[2:5],
groups=groups,
padding=padding,
strides=strides,
output_padding=output_padding,
data_layout=data_layout,
kernel_layout=kernel_layout,
)
dic = {"x": x_shape, "kernel": k_shape}
param_lst = ["kernel"]
if activation == "relu":
return relay.nn.relu(out), dic, param_lst
elif activation == "tanh":
return relay.tanh(out), dic, param_lst
elif activation == "sigmoid":
return relay.sigmoid(out), dic, param_lst
else:
return out, dic, param_lst
def test_skip_conv():
data = relay.var("data", shape=(1, 16, 64, 64))
np_weight = np.random.rand(16, 16, 3, 3)
conv0_weight = relay.Constant(tvm.nd.array(np_weight)).astype("float32")
conv1_weight = relay.Constant(tvm.nd.array(np_weight)).astype("float32")
multiplier = relay.sigmoid(relay.var("data", shape=(1, 16, 1, 1)))
conv0 = relay.nn.conv2d(data,
conv0_weight,
kernel_size=(3, 3),
padding=(1, 1),
channels=16)
act0 = relay.nn.relu(data=conv0)
conv1 = relay.nn.conv2d(act0,
conv1_weight,
kernel_size=(3, 3),
padding=(1, 1),
channels=16)
act1 = relay.nn.relu(data=conv1)
quantize_and_build(act1 * multiplier)
quantize_and_build(act1 * multiplier, skip_conv_layers=[0])
quantize_and_build(act1 * multiplier, skip_conv_layers=[1])
quantize_and_build(act1 * multiplier, skip_conv_layers=[0, 1])
def get_conv2d_nchw_bias_silu(d_shape, w_shape, padding, out_dtype="float16"):
conv_out = get_conv2d_nchw_bias(d_shape, w_shape, padding, out_dtype=out_dtype)
return conv_out * relay.sigmoid(conv_out)
def lstm_cell(num_hidden, batch_size=1, dtype="float32", name=""):
"""Long-Short Term Memory (LSTM) network cell.
Parameters
----------
num_hidden : int
Number of units in output symbol.
batch_size : int
Batch size (length of states).
Returns
-------
result : tvm.relay.Function
A Relay function that evaluates an LSTM cell.
The function takes in a tensor of input data, a tuple of two
states, and weights and biases for dense operations on the
inputs and on the state. It returns a tuple with two members,
an output tensor and a tuple of two new states.
"""
builder = relay.ScopeBuilder()
input_type = relay.TensorType((batch_size, num_hidden), dtype)
weight_type = relay.TensorType((4*num_hidden, num_hidden), dtype)
bias_type = relay.TensorType((4*num_hidden,), dtype)
dense_type = relay.TensorType((batch_size, 4*num_hidden), dtype)
slice_type = relay.TupleType([input_type, input_type,
input_type, input_type])
ret_type = relay.TupleType([input_type,
relay.TupleType([input_type, input_type])])
inputs = relay.Var("inputs", input_type)
states = relay.Var("states",
relay.TupleType([input_type, input_type]))
i2h_weight = relay.Var("i2h_weight", weight_type)
i2h_bias = relay.Var("i2h_bias", bias_type)
h2h_weight = relay.Var("h2h_weight", weight_type)
h2h_bias = relay.Var("h2h_bias", bias_type)
i2h = builder.let(("i2h", dense_type),
layers.dense_add_bias(
data=inputs,
units=num_hidden * 4,
weight=i2h_weight, bias=i2h_bias,
name="%si2h" % name))
h2h = builder.let(("h2h", dense_type),
layers.dense_add_bias(
data=relay.TupleGetItem(states, 0),
units=num_hidden * 4,
weight=h2h_weight, bias=h2h_bias,
name="%sh2h" % name))
gates = builder.let(("gates", dense_type), relay.add(i2h, h2h))
slice_gates = builder.let(("slice_gates", slice_type),
relay.split(gates,
indices_or_sections=4,
axis=1).astuple())
in_gate = builder.let(("in_gate", input_type),
relay.sigmoid(relay.TupleGetItem(slice_gates, 0)))
forget_gate = builder.let(("forget_gate", input_type),
relay.sigmoid(relay.TupleGetItem(slice_gates, 1)))
in_transform = builder.let(("in_transform", input_type),
relay.tanh(relay.TupleGetItem(slice_gates, 2)))
out_gate = builder.let(("out_gate", input_type),
relay.sigmoid(relay.TupleGetItem(slice_gates, 3)))
next_c = builder.let(("next_c", input_type),
relay.add(relay.multiply(forget_gate,
relay.TupleGetItem(states, 1)),
relay.multiply(in_gate, in_transform)))
next_h = builder.let(("next_h", input_type),
relay.multiply(out_gate, relay.tanh(next_c)))
ret = builder.let(("ret", ret_type),
relay.Tuple([next_h, relay.Tuple([next_h, next_c])]))
builder.ret(ret)
body = builder.get()
return relay.Function([inputs, states, i2h_weight,
i2h_bias, h2h_weight, h2h_bias],
body, ret_type)
def silu(x):
return x * relay.sigmoid(x)
def get_conv2d_nchw_bias_sigmoid(d_shape, w_shape, padding, out_dtype="float16"):
return relay.sigmoid(get_conv2d_nchw_bias(d_shape, w_shape, padding, out_dtype=out_dtype))
def get_net(batch_size, image_shape, num_classes, dtype):
height = image_shape[1]
width = image_shape[2]
data_shape = (batch_size,) + image_shape
net = relay.var("data", shape=data_shape, dtype=dtype)
net = conv_3x3(net, 3, 64, "conv1_1")
net = conv_3x3(net, 64, 64, "conv1_2")
net = relay.nn.max_pool2d(net, pool_size=(2, 2), strides=(2, 2), ceil_mode=True)
net = conv_3x3(net, 64, 128, "conv2_1")
net = conv_3x3(net, 64, 128, "conv2_2")
net = relay.nn.max_pool2d(net, pool_size=(2, 2), strides=(2, 2), ceil_mode=True)
net = conv_3x3(net, 128, 256, "conv3_1")
net = conv_3x3(net, 256, 256, "conv3_2")
net = conv_3x3(net, 256, 256, "conv3_3")
net = relay.nn.max_pool2d(net, pool_size=(2, 2), strides=(2, 2), ceil_mode=True)
net = conv_3x3(net, 256, 512, "conv4_1")
net = conv_3x3(net, 512, 512, "conv4_2")
net = conv_3x3(net, 512, 512, "conv4_3")
net = relay.nn.max_pool2d(net, pool_size=(2, 2), strides=(2, 2), ceil_mode=True)
net = conv_3x3(net, 512, 512, "conv5_1")
net = conv_3x3(net, 512, 512, "conv5_2")
net = conv_3x3(net, 512, 512, "conv5_3")
net = relay.nn.dropout(net, rate=0.5)
net = conv_3x3(net, 512, 72, "conv6", activation=False)
net = relay.transpose(net, (0, 2, 3, 1))
num_class_probs = 9 * 3
num_confidence_scores = 9
#num_box_delta = 9 * 4
pred_class_probs, pred_conf, pred_box_delta = relay.split(net,
(num_class_probs, num_class_probs + num_confidence_scores),
axis=-1)
# Probability
pred_class_probs = relay.reshape(pred_class_probs, (-1, 3))
pred_class_probs = relay.nn.softmax(pred_class_probs)
pred_class_probs = relay.reshape(pred_class_probs, (batch_size, -1, 3))
# Confidence
pred_conf = relay.sigmoid(pred_conf)
pred_conf = relay.reshape(pred_conf, (batch_size, -1, 1))
# Bbox_delta
pred_box_delta = relay.reshape(pred_box_delta, (batch_size, -1, 4))
delta_x, delta_y, delta_w, delta_h = relay.split(pred_box_delta, (1, 2, 3), axis=2)
delta_x = relay.reshape(delta_x, (batch_size, -1))
delta_y = relay.reshape(delta_y, (batch_size, -1))
delta_w = relay.reshape(delta_w, (batch_size, -1))
delta_h = relay.reshape(delta_h, (batch_size, -1))
anchor_box = set_anchors(height, width)
anchor_x = relay.Constant(tvm.nd.array(anchor_box[:, 0]))
anchor_y = relay.Constant(tvm.nd.array(anchor_box[:, 1]))
anchor_w = relay.Constant(tvm.nd.array(anchor_box[:, 2]))
anchor_h = relay.Constant(tvm.nd.array(anchor_box[:, 3]))
box_center_x = anchor_x + delta_x * anchor_w
box_center_y = anchor_y + delta_y * anchor_h
'''
box_width = anchor_w * relay.exp(delta_w)
box_height = anchor_h * relay.exp(delta_h)
'''
box_width = anchor_w + safe_exp(delta_w)
box_height = anchor_h + safe_exp(delta_h)
xmins, ymins, xmaxs, ymaxs = bbox_transform(box_center_x, box_center_y, box_width, box_height)
xmins = relay.minimum(relay.maximum(relay.const(0.0), xmins), relay.const(width - 1.0))
ymins = relay.minimum(relay.maximum(relay.const(0.0), ymins), relay.const(height - 1.0))
xmaxs = relay.maximum(relay.minimum(relay.const(width - 1.0), xmaxs), relay.const(0.0))
ymaxs = relay.maximum(relay.minimum(relay.const(height - 1.0), ymaxs), relay.const(0.0))
det_boxes = relay.stack(bbox_transform_inv(xmins, ymins, xmaxs, ymaxs), axis=-1)
probs = relay.multiply(pred_class_probs, pred_conf)
det_probs = relay.max(probs, axis=2)
det_class = relay.argmax(probs, axis=2)
out = relay.Tuple([det_boxes, det_probs, det_class])
args = relay.analysis.free_vars(out)
return relay.Function(args, out)
def lstm_cell(num_hidden, batch_size=1, dtype="float32", name=""):
"""Long-Short Term Memory (LSTM) network cell.
Parameters
----------
num_hidden : int
Number of units in output symbol.
batch_size : int
Batch size (length of states).
Returns
-------
result : tvm.relay.Function
A Relay function that evaluates an LSTM cell.
The function takes in a tensor of input data, a tuple of two
states, and weights and biases for dense operations on the
inputs and on the state. It returns a tuple with two members,
an output tensor and a tuple of two new states.
"""
builder = relay.ScopeBuilder()
input_type = relay.TensorType((batch_size, num_hidden), dtype)
weight_type = relay.TensorType((4 * num_hidden, num_hidden), dtype)
bias_type = relay.TensorType((4 * num_hidden,), dtype)
dense_type = relay.TensorType((batch_size, 4 * num_hidden), dtype)
slice_type = relay.TupleType([input_type, input_type, input_type, input_type])
ret_type = relay.TupleType([input_type, relay.TupleType([input_type, input_type])])
inputs = relay.Var("inputs", input_type)
states = relay.Var("states", relay.TupleType([input_type, input_type]))
i2h_weight = relay.Var("i2h_weight", weight_type)
i2h_bias = relay.Var("i2h_bias", bias_type)
h2h_weight = relay.Var("h2h_weight", weight_type)
h2h_bias = relay.Var("h2h_bias", bias_type)
i2h = builder.let(
("i2h", dense_type),
layers.dense_add_bias(
data=inputs, units=num_hidden * 4, weight=i2h_weight, bias=i2h_bias, name="%si2h" % name
),
)
h2h = builder.let(
("h2h", dense_type),
layers.dense_add_bias(
data=relay.TupleGetItem(states, 0),
units=num_hidden * 4,
weight=h2h_weight,
bias=h2h_bias,
name="%sh2h" % name,
),
)
gates = builder.let(("gates", dense_type), relay.add(i2h, h2h))
slice_gates = builder.let(
("slice_gates", slice_type), relay.split(gates, indices_or_sections=4, axis=1).astuple()
)
in_gate = builder.let(
("in_gate", input_type), relay.sigmoid(relay.TupleGetItem(slice_gates, 0))
)
forget_gate = builder.let(
("forget_gate", input_type), relay.sigmoid(relay.TupleGetItem(slice_gates, 1))
)
in_transform = builder.let(
("in_transform", input_type), relay.tanh(relay.TupleGetItem(slice_gates, 2))
)
out_gate = builder.let(
("out_gate", input_type), relay.sigmoid(relay.TupleGetItem(slice_gates, 3))
)
next_c = builder.let(
("next_c", input_type),
relay.add(
relay.multiply(forget_gate, relay.TupleGetItem(states, 1)),
relay.multiply(in_gate, in_transform),
),
)
next_h = builder.let(("next_h", input_type), relay.multiply(out_gate, relay.tanh(next_c)))
ret = builder.let(("ret", ret_type), relay.Tuple([next_h, relay.Tuple([next_h, next_c])]))
builder.ret(ret)
body = builder.get()
return relay.Function(
[inputs, states, i2h_weight, i2h_bias, h2h_weight, h2h_bias], body, ret_type
)
def verify_sigmoid(dshape, dtype="float32"):
x = relay.var("x", relay.ty.TensorType(dshape, dtype))
y = relay.sigmoid(x)
func = relay.Function([x], y)
x_data = np.random.uniform(size=dshape).astype(dtype)
verify_results(func, [x_data], "test_sigmoid", rtol=1e-4, atol=1e-4)