tests.utils Namespace Reference

Functions
def	quantize_by_range (x, num_bits)
def	quantize_by_range_fused (x_tuple, num_bits)
def	copy_state_and_quantize (dst, src, num_bits)
def	copy_state_and_quantize_fused (dst, src, num_bits)
def	compare (a, b, rtol=1e-7, atol=1e-6, ctol=1e-6)
def	assert_min_mse (a, b, tol=1e-20)
def	quant_np (x, amax, num_bits=8, fake=False, narrow_range=True)

Function Documentation

quantize_by_range()

def tests.utils.quantize_by_range	(		x,
			num_bits
	)

Quantize torch tensor by range to num_bits with symmetric zero-mean quantizer.

Here is the caller graph for this function:

quantize_by_range_fused()

def tests.utils.quantize_by_range_fused	(		x_tuple,
			num_bits
	)

Quantize multiple torch tensors by combined range to num_bits with symmetric zero-mean quantizer.

copy_state_and_quantize()

def tests.utils.copy_state_and_quantize	(		dst,
			src,
			num_bits
	)

Copy src to dst, quantize all 'weight' entries to num_bits.

Here is the call graph for this function:

copy_state_and_quantize_fused()

def tests.utils.copy_state_and_quantize_fused	(		dst,
			src,
			num_bits
	)

Copy src to dst, quantize all 'weight' entries to num_bits using the aggregate amax.

compare()

def tests.utils.compare	(		a,
			b,
			rtol = 1e-7,
			atol = 1e-6,
			ctol = 1e-6
	)

Compare two tensors and raise AssertionError if their difference is outside of tolerance.

assert_min_mse()

def tests.utils.assert_min_mse	(		a,
			b,
			tol = 1e-20
	)

Assert that the mean squared error between a and b is at least tol.

quant_np()

def tests.utils.quant_np	(		x,
			amax,
			num_bits = 8,
			fake = False,
			narrow_range = True
	)

Quantize x using numpy.