def filter_stats(G, pnode, fnode, anode, channel_details=None, qrec=None):
stats = {}
if isinstance(anode, MultiplicativeBiasParameters):
if anode.has_mul_bias:
stats['mul_biases'] = mul_biases = astats(anode.mul_biases)
mul_biases['qstats'] = calculate_qsnrs(anode.mul_biases,
mul_biases['ibits'],
force_ideal=False)
elif qrec and qrec.ktype.startswith('scaled'):
stats['mul_biases'] = mul_biases = astats(qrec.mul_biases_fps)
mul_biases['qstats'] = calculate_qsnrs(qrec.mul_biases_fps,
mul_biases['ibits'],
force_ideal=False)
in_edges = G.indexed_in_edges(pnode.name)
qweights = in_edges[1].from_node.dqvalue
qbiases = in_edges[2].from_node.dqvalue
stats['biases'] = biases = astats(qbiases)
biases['qstats'] = calculate_qsnrs(qbiases,
biases['ibits'],
force_ideal=False)
stats['weights'] = weights = astats(
qweights,
channel_dim=anode.filter.get_order_idx('out_c'),
channel_details=channel_details)
weights['qstats'] = calculate_qsnrs(qweights,
weights['ibits'],
force_ideal=False)
# store the statistics into the graph for later use
anode.stats = stats
stats['step_idx'] = pnode.step_idx
return anode.stats
def filter_stats(node):
stats = {}
if node.has_bias:
stats['biases'] = biases = astats(node.biases)
biases['qstats'] = calculate_qsnrs(node.biases,
biases['ibits'],
force_ideal=False)
stats['weights'] = weights = astats(node.weights)
weights['qstats'] = calculate_qsnrs(node.weights,
weights['ibits'],
force_ideal=False)
# store the statistics into the graph for later use
node.stats = stats
return node.stats
def balance_filter(pnode, precision_threshold=None, small_value_threshold=0.0000000001, fnode=None, G=None):
node = pnode if fnode is None else fnode
channel_dim = node.filter.get_order_idx('out_c')
stats = astats(node.weights, channel_dim=channel_dim, channel_details=True)
if precision_threshold:
if not any(stat['avg_prec'] < precision_threshold for stat in stats['channel_stats']):
LOG.info("layer %s doesn't have any weights below precision threshold", node.name)
return False
LOG.info("balancing weights of layer %s", node.name)
if node.has_mul_bias and node.mul_biases is not None:
mul_bias = node.mul_biases
else:
mul_bias = np.ones(node.filter.out_c, dtype=np.float32)
scale = np.array([max(abs(stat['max']), abs(stat['min']))
for stat in stats['channel_stats']])
# don't balance channels that are effectively zero
threshold_idx = scale < small_value_threshold
scale[threshold_idx] = 1
weights_shape = [1 if idx != channel_dim else size
for idx, size in enumerate(node.weights.shape)]
node.weights /= scale.reshape(weights_shape)
if node.has_bias:
node.biases /= scale
mul_bias *= scale
if G:
G.changes.modify(pnode, 'has_mul_bias', True, fnode=fnode)
else:
node.has_mul_bias = True
node.mul_biases = mul_bias
return True
def gather_stats(activation,
force_ideal=False,
channel_dim=None,
channel_details=None):
stat = astats(activation,
channel_dim=channel_dim,
channel_details=channel_details)
stat['qstats'] = calculate_qsnrs(activation, stat['ibits'], force_ideal)
return stat
def filter_stats(pnode, fnode, anode, channel_details=None, qrec=None):
stats = {}
if isinstance(anode, MultiplicativeBiasParameters):
if anode.has_mul_bias:
stats['mul_biases'] = mul_biases = astats(anode.mul_biases)
mul_biases['qstats'] = calculate_qsnrs(anode.mul_biases,
mul_biases['ibits'],
force_ideal=False)
elif isinstance(qrec, MultScalableFilterQuantizationRecord):
stats['mul_biases'] = mul_biases = astats(qrec.mul_biases_fps)
mul_biases['qstats'] = calculate_qsnrs(qrec.mul_biases_fps,
mul_biases['ibits'],
force_ideal=False)
if anode.has_bias:
if qrec:
qbiases = qrec.prepare_biases(anode,
anode.biases,
anode.weights,
ktype="float32")
else:
qbiases = anode.biases
stats['biases'] = biases = astats(qbiases)
biases['qstats'] = calculate_qsnrs(qbiases,
biases['ibits'],
force_ideal=False)
if qrec:
qweights = qrec.prepare_weights(anode, anode.weights, ktype="float32")
else:
qweights = anode.weights
stats['weights'] = weights = astats(
qweights,
channel_dim=anode.filter.get_order_idx('out_c'),
channel_details=channel_details)
weights['qstats'] = calculate_qsnrs(qweights,
weights['ibits'],
force_ideal=False)
# store the statistics into the graph for later use
anode.stats = stats
stats['step_idx'] = pnode.step_idx
return anode.stats
def gather_stats(activation, force_ideal=False):
stat = astats(activation)
stat['qstats'] = calculate_qsnrs(activation, stat['ibits'], force_ideal)
return stat
