def quantize(data, shift_bits, target_bits=relay.const(7, dtype='int32')):
"""Quantize output of layer, to be consistent with source code @yx
Question: should the shift_bits participating to network control flow?
At mxnet quantization with truman's code, the bits number of max_v
is converted to normal interger using function `asscalar()`. However,
I cannot find the related function in relay.
I am confused with the control flow logic in model network, whether
the condition `shift_bits == -1` should join in model network or just
left it in python code flow. By Longtao.Wang
Parameters
----------
shift_bits: tvm.relay.Expr
The shift_bits parameter is never used according to @yx's source code,
which always be constant Expr(-1).
"""
max_v = relay.max(relay.abs(data))
min_v = relay.min(data)
ln_max_v = relay.log(relay.cast(max_v, 'float32'))
ln_2 = relay.log(relay.const(2.))
total_bits = relay.ceil(relay.divide(ln_max_v, ln_2)) # ceil( ln(max_v) / ln(2) )
shift_bits = relay.subtract(total_bits.astype('int32'), target_bits)
shift_bits = relay.maximum(shift_bits, relay.const(0))
denominator = relay.left_shift(relay.const(1),
relay.cast(shift_bits, 'int32'))
out = relay.divide(data, denominator)
# According to @yx's code, use divide operation instead of shift op for
# possible negative number round.
# out = relay.right_shift(data, shift_bits)
out = relay.cast(relay.clip(out, a_min=-128, a_max=127), 'int8')
return out, max_v, min_v, shift_bits
def expected():
p0 = relay.var("p0", shape=(), dtype="int32")
less = relay.less(p0, relay.const(10, dtype="int32"))
z0 = relay.min(less)
f0 = relay.Function([p0], z0)
f0 = f0.with_attr("Primitive", tvm.tir.IntImm("int32", 1))
x = relay.var("x", shape=(), dtype="int32")
f = relay.Call(f0, [x])
return relay.Function([x], f)
def before():
x = relay.var("x", shape=(), dtype="int32")
less = relay.less(x, relay.const(10, dtype="int32"))
z = relay.min(less)
return relay.Function([x], z)