def cache_intermediates(self, G, input_tensor, cache=True):
outputs = []
fusion_outputs = []
for _, _, _, output, fusion_op_name, _, _ in\
execute_iterator(G, [input_tensor], disable_cache=True):
if fusion_outputs:
# handle end of fusion
if fusion_op_name is None:
# python upvalues are not as smart as lua
outputs.append([out.copy() for out in fusion_outputs])
fusion_outputs.clear()
continue
# handle start or continuation of fusion
if fusion_op_name is not None:
fusion_outputs.append([out.copy() for out in output])
continue
outputs.append([out.copy() for out in output])
input_hash = IntermediateCache.__get_hash(G, input_tensor)
cache_path_prefix = self.cache_path_prefix(input_tensor, input_hash)
self.write_entry(cache_path_prefix, input_tensor, outputs)
if cache:
self.add_to_cache(input_hash, (input_tensor, outputs))
return self.cache[input_hash]
return None
def test_graph_calc_iterator_cached(value_cache, mnist_graph, mnist_images):
G = create_graph(mnist_graph, opts={"load_tensors":True})
G.add_dimensions()
input_tensor = import_data(mnist_images[0], height=28, width=28, offset=0, divisor=255)
input_tensor = input_tensor.reshape((28, 28, 1))
normal_steps = 0
fusion_steps = 0
# pylint: disable=unused-variable
for step_idx, step, node, output, fusion_op_name, fusion_params, details in\
execute_iterator(G, [input_tensor], value_cache=value_cache):
if fusion_op_name is not None:
fusion_steps += 1
else:
normal_steps += 1
assert normal_steps == 10 and fusion_steps == 0
def _collect(self, G, input_tensors):
stats = OrderedDict()
for _, _, node, output, _, fusion_node, details in\
execute_iterator(G, input_tensors, disable_cache=True):
key = NodeId(node, fusion_node)
node = (node if fusion_node is None else fusion_node)
stat = gather_stats(
output[0], force_ideal=not isinstance(node, InputParameters))
if isinstance(node, FilterParameters) and details:
stat['min_acc'] = details['min_acc']
stat['max_acc'] = details['max_acc']
stats[key] = stat
return stats
def test_graph_kws(kws_graph, kws_sounds):
G = create_graph(kws_graph, opts={"load_tensors": True})
G.add_dimensions()
input_tensor = import_data(kws_sounds[0],
offset=0,
divisor=128,
nptype='int16')
normal_steps = 0
fusion_steps = 0
# pylint: disable=unused-variable
for step_idx, step, node, output, fusion_op_name, fusion_params, details in\
execute_iterator(G, [input_tensor]):
if fusion_op_name is not None:
fusion_steps += 1
else:
normal_steps += 1
assert normal_steps == 9 and fusion_steps == 0
def _collect_execution(self, G, tensors, qrecs=None, qmode=None):
outputs = []
fusion_outputs = []
for step_idx, step, node, output, fusion_op_name, fusion_node, details in\
execute_iterator(G, tensors, limit=self._limit, qrecs=qrecs, qmode=qmode):
if qrecs:
qrec = qrecs[NodeId(node, fusion_node)]
output = [
qrec.out_qs[i].dequantize(out)
for i, out in enumerate(output)
]
else:
output = output.copy()
del step, fusion_op_name
if fusion_node:
fusion_outputs.append({
"name":
"",
"step_idx":
"{}_{}".format(step_idx, len(fusion_outputs)),
"node":
fusion_node,
"output":
output,
"details":
details
})
else:
stat = {
"name": node.name,
"step_idx": str(step_idx),
"node": node,
"output": output,
"details": details,
"fusion_outputs": []
}
if len(fusion_outputs) > 0:
stat['fusion_outputs'] = fusion_outputs.copy()
fusion_outputs.clear()
outputs.append(stat)
return outputs