def set_vars(self):
self.qat = ImperativeQuantAware()
self.train_batch_num = 30
self.train_batch_size = 32
self.test_batch_num = 100
self.test_batch_size = 32
self.eval_acc_top1 = 0.99
def _remove_preprocess(self, model):
state_dict = model.state_dict()
self.imperative_qat = ImperativeQuantAware(
weight_bits=self.config['weight_bits'],
activation_bits=self.config['activation_bits'],
weight_quantize_type=self.config['weight_quantize_type'],
activation_quantize_type=self.config['activation_quantize_type'],
moving_rate=self.config['moving_rate'],
quantizable_layer_type=self.config['quantizable_layer_type'])
with paddle.utils.unique_name.guard():
model.__init__()
self.imperative_qat.quantize(model)
state_dict = model.state_dict()
model.set_state_dict(state_dict)
return model
def test_warning(self):
path = "./dynamic_outscale_infer_model_with_warnings/lenet"
imperative_out_scale = ImperativeQuantAware()
with fluid.dygraph.guard():
lenet = ImperativeLenet()
with warnings.catch_warnings(record=True) as w:
warnings.simplefilter("always")
imperative_out_scale.save_quantized_model(
layer=lenet,
path=path,
input_spec=[
paddle.static.InputSpec(shape=[None, 1, 28, 28],
dtype='float32')
])
warning_message = "Warning: No Layer of the model while to be " \
"saved contains the out_threshold attribute, so the " \
"generated inference model would not contain the " \
"out_threshold."
num = get_vaild_warning_num(warning_message, w)
assert num == 1
def func_out_scale_acc(self):
seed = 1000
lr = 0.001
weight_quantize_type = 'abs_max'
activation_quantize_type = 'moving_average_abs_max'
imperative_out_scale = ImperativeQuantAware(
weight_quantize_type=weight_quantize_type,
activation_quantize_type=activation_quantize_type)
with fluid.dygraph.guard():
np.random.seed(seed)
fluid.default_main_program().random_seed = seed
fluid.default_startup_program().random_seed = seed
lenet = ImperativeLenet()
lenet = fix_model_dict(lenet)
imperative_out_scale.quantize(lenet)
reader = paddle.batch(paddle.dataset.mnist.test(),
batch_size=32,
drop_last=True)
adam = AdamOptimizer(learning_rate=lr,
parameter_list=lenet.parameters())
loss_list = train_lenet(lenet, reader, adam)
lenet.eval()
param_save_path = "test_save_quantized_model/lenet.pdparams"
save_dict = lenet.state_dict()
paddle.save(save_dict, param_save_path)
save_path = "./dynamic_outscale_infer_model/lenet"
imperative_out_scale.save_quantized_model(
layer=lenet,
path=save_path,
input_spec=[
paddle.static.InputSpec(shape=[None, 1, 28, 28],
dtype='float32')
])
for i in range(len(loss_list) - 1):
self.assertTrue(loss_list[i] > loss_list[i + 1],
msg='Failed to do the imperative qat.')
def test_save_quantized_model(self):
lr = 0.001
load_param_path = "test_save_quantized_model/lenet.pdparams"
save_path = "./dynamic_outscale_infer_model_from_checkpoint/lenet"
weight_quantize_type = 'abs_max'
activation_quantize_type = 'moving_average_abs_max'
imperative_out_scale = ImperativeQuantAware(
weight_quantize_type=weight_quantize_type,
activation_quantize_type=activation_quantize_type)
with fluid.dygraph.guard():
lenet = ImperativeLenet()
load_dict = paddle.load(load_param_path)
imperative_out_scale.quantize(lenet)
lenet.set_dict(load_dict)
reader = paddle.batch(
paddle.dataset.mnist.test(), batch_size=32, drop_last=True)
adam = AdamOptimizer(
learning_rate=lr, parameter_list=lenet.parameters())
loss_list = train_lenet(lenet, reader, adam)
lenet.eval()
imperative_out_scale.save_quantized_model(
layer=lenet,
path=save_path,
input_spec=[
paddle.static.InputSpec(
shape=[None, 1, 28, 28], dtype='float32')
])
for i in range(len(loss_list) - 1):
self.assertTrue(
loss_list[i] > loss_list[i + 1],
msg='Failed to do the imperative qat.')
class TestImperativeQatAmp(unittest.TestCase):
"""
Test the combination of qat and amp.
"""
@classmethod
def setUpClass(cls):
timestamp = time.strftime('%Y-%m-%d-%H-%M-%S', time.localtime())
cls.root_path = os.path.join(os.getcwd(),
"imperative_qat_amp_" + timestamp)
cls.save_path = os.path.join(cls.root_path, "model")
cls.download_path = 'dygraph_int8/download'
cls.cache_folder = os.path.expanduser('~/.cache/paddle/dataset/' +
cls.download_path)
cls.lenet_url = "https://paddle-inference-dist.cdn.bcebos.com/int8/unittest_model_data/lenet_pretrained.tar.gz"
cls.lenet_md5 = "953b802fb73b52fae42896e3c24f0afb"
seed = 1
np.random.seed(seed)
paddle.static.default_main_program().random_seed = seed
paddle.static.default_startup_program().random_seed = seed
@classmethod
def tearDownClass(cls):
try:
shutil.rmtree(cls.root_path)
except Exception as e:
print("Failed to delete {} due to {}".format(
cls.root_path, str(e)))
def cache_unzipping(self, target_folder, zip_path):
if not os.path.exists(target_folder):
cmd = 'mkdir {0} && tar xf {1} -C {0}'.format(
target_folder, zip_path)
os.system(cmd)
def download_model(self, data_url, data_md5, folder_name):
download(data_url, self.download_path, data_md5)
file_name = data_url.split('/')[-1]
zip_path = os.path.join(self.cache_folder, file_name)
print('Data is downloaded at {0}'.format(zip_path))
data_cache_folder = os.path.join(self.cache_folder, folder_name)
self.cache_unzipping(data_cache_folder, zip_path)
return data_cache_folder
def set_vars(self):
self.qat = ImperativeQuantAware()
self.train_batch_num = 30
self.train_batch_size = 32
self.test_batch_num = 100
self.test_batch_size = 32
self.eval_acc_top1 = 0.99
def model_train(self, model, batch_num=-1, batch_size=32, use_amp=False):
model.train()
train_reader = paddle.batch(paddle.dataset.mnist.train(),
batch_size=batch_size)
adam = paddle.optimizer.Adam(learning_rate=0.001,
parameters=model.parameters())
scaler = paddle.amp.GradScaler(init_loss_scaling=500)
for batch_id, data in enumerate(train_reader()):
x_data = np.array([x[0].reshape(1, 28, 28)
for x in data]).astype('float32')
y_data = np.array([x[1]
for x in data]).astype('int64').reshape(-1, 1)
img = paddle.to_tensor(x_data)
label = paddle.to_tensor(y_data)
if use_amp:
with paddle.amp.auto_cast():
out = model(img)
acc = fluid.layers.accuracy(out, label)
loss = fluid.layers.cross_entropy(out, label)
avg_loss = fluid.layers.mean(loss)
scaled_loss = scaler.scale(avg_loss)
scaled_loss.backward()
scaler.minimize(adam, scaled_loss)
adam.clear_gradients()
else:
out = model(img)
acc = fluid.layers.accuracy(out, label)
loss = fluid.layers.cross_entropy(out, label)
avg_loss = fluid.layers.mean(loss)
avg_loss.backward()
adam.minimize(avg_loss)
model.clear_gradients()
if batch_id % 100 == 0:
_logger.info("Train | step {}: loss = {:}, acc= {:}".format(
batch_id, avg_loss.numpy(), acc.numpy()))
if batch_num > 0 and batch_id + 1 >= batch_num:
break
def model_test(self, model, batch_num=-1, batch_size=32, use_amp=False):
model.eval()
test_reader = paddle.batch(paddle.dataset.mnist.test(),
batch_size=batch_size)
acc_top1_list = []
for batch_id, data in enumerate(test_reader()):
x_data = np.array([x[0].reshape(1, 28, 28)
for x in data]).astype('float32')
y_data = np.array([x[1]
for x in data]).astype('int64').reshape(-1, 1)
img = paddle.to_tensor(x_data)
label = paddle.to_tensor(y_data)
with paddle.amp.auto_cast(use_amp):
out = model(img)
acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1)
acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5)
acc_top1_list.append(float(acc_top1.numpy()))
if batch_id % 100 == 0:
_logger.info(
"Test | At step {}: acc1 = {:}, acc5 = {:}".format(
batch_id, acc_top1.numpy(), acc_top5.numpy()))
if batch_num > 0 and batch_id + 1 >= batch_num:
break
acc_top1 = sum(acc_top1_list) / len(acc_top1_list)
return acc_top1
def ptq(self):
start_time = time.time()
self.set_vars()
params_path = self.download_model(self.lenet_url, self.lenet_md5,
"lenet")
params_path += "/lenet_pretrained/lenet.pdparams"
with fluid.dygraph.guard():
model = ImperativeLenet()
model_state_dict = paddle.load(params_path)
model.set_state_dict(model_state_dict)
_logger.info("Test fp32 model")
fp32_acc_top1 = self.model_test(model, self.test_batch_num,
self.test_batch_size)
self.qat.quantize(model)
use_amp = True
self.model_train(model, self.train_batch_num,
self.train_batch_size, use_amp)
_logger.info("Test int8 model")
int8_acc_top1 = self.model_test(model, self.test_batch_num,
self.test_batch_size, use_amp)
_logger.info('fp32_acc_top1: %f, int8_acc_top1: %f' %
(fp32_acc_top1, int8_acc_top1))
self.assertTrue(int8_acc_top1 > fp32_acc_top1 - 0.01,
msg='fp32_acc_top1: %f, int8_acc_top1: %f' %
(fp32_acc_top1, int8_acc_top1))
input_spec = [
paddle.static.InputSpec(shape=[None, 1, 28, 28], dtype='float32')
]
paddle.jit.save(layer=model,
path=self.save_path,
input_spec=input_spec)
print('Quantized model saved in {%s}' % self.save_path)
end_time = time.time()
print("total time: %ss" % (end_time - start_time))
def test_ptq(self):
self.ptq()
with _test_eager_guard():
self.ptq()
def setUp(self):
_logger.info("test weight_quantize")
self.imperative_qat = ImperativeQuantAware(
weight_quantize_layer=CustomQAT)
def setUp(self):
_logger.info("test weight_preprocess")
self.imperative_qat = ImperativeQuantAware(
weight_preprocess_layer=PACT)
def func_out_scale_acc(self):
paddle.disable_static()
seed = 1000
lr = 0.1
qat = ImperativeQuantAware()
np.random.seed(seed)
reader = paddle.batch(paddle.dataset.mnist.test(),
batch_size=512,
drop_last=True)
lenet = ImperativeLenetWithSkipQuant()
lenet = fix_model_dict(lenet)
qat.quantize(lenet)
adam = AdamOptimizer(learning_rate=lr,
parameter_list=lenet.parameters())
dynamic_loss_rec = []
lenet.train()
loss_list = train_lenet(lenet, reader, adam)
lenet.eval()
path = "./save_dynamic_quant_infer_model/lenet"
save_dir = "./save_dynamic_quant_infer_model"
qat.save_quantized_model(layer=lenet,
path=path,
input_spec=[
paddle.static.InputSpec(
shape=[None, 1, 28, 28],
dtype='float32')
])
paddle.enable_static()
if core.is_compiled_with_cuda():
place = core.CUDAPlace(0)
else:
place = core.CPUPlace()
exe = fluid.Executor(place)
[inference_program, feed_target_names,
fetch_targets] = (fluid.io.load_inference_model(
dirname=save_dir,
executor=exe,
model_filename="lenet" + INFER_MODEL_SUFFIX,
params_filename="lenet" + INFER_PARAMS_SUFFIX))
model_ops = inference_program.global_block().ops
conv2d_count, matmul_count = 0, 0
conv2d_skip_count, matmul_skip_count = 0, 0
find_conv2d = False
find_matmul = False
for i, op in enumerate(model_ops):
if op.type == 'conv2d':
find_conv2d = True
if op.has_attr("skip_quant"):
conv2d_skip_count += 1
if conv2d_count > 0:
self.assertTrue(
'fake_quantize_dequantize' in model_ops[i - 1].type)
else:
self.assertTrue(
'fake_quantize_dequantize' not in model_ops[i -
1].type)
conv2d_count += 1
if op.type == 'matmul':
find_matmul = True
if op.has_attr("skip_quant"):
matmul_skip_count += 1
if matmul_count > 0:
self.assertTrue(
'fake_quantize_dequantize' in model_ops[i - 1].type)
else:
self.assertTrue(
'fake_quantize_dequantize' not in model_ops[i -
1].type)
matmul_count += 1
if find_conv2d:
self.assertTrue(conv2d_skip_count == 1)
if find_matmul:
self.assertTrue(matmul_skip_count == 1)
def test_qat_save(self):
imperative_qat = ImperativeQuantAware(
weight_quantize_type='abs_max',
activation_quantize_type='moving_average_abs_max',
quantizable_layer_type=[
'Conv2D', 'Linear', 'ReLU', 'LeakyReLU', 'ReLU6', 'Tanh',
'Swish'
])
with fluid.dygraph.guard():
lenet = ImperativeLenet()
imperative_qat.quantize(lenet)
adam = AdamOptimizer(
learning_rate=0.001, parameter_list=lenet.parameters())
train_reader = paddle.batch(
paddle.dataset.mnist.train(), batch_size=32, drop_last=True)
test_reader = paddle.batch(
paddle.dataset.mnist.test(), batch_size=32)
epoch_num = 1
for epoch in range(epoch_num):
lenet.train()
for batch_id, data in enumerate(train_reader()):
x_data = np.array([x[0].reshape(1, 28, 28)
for x in data]).astype('float32')
y_data = np.array(
[x[1] for x in data]).astype('int64').reshape(-1, 1)
img = fluid.dygraph.to_variable(x_data)
label = fluid.dygraph.to_variable(y_data)
out = lenet(img)
acc = fluid.layers.accuracy(out, label)
loss = fluid.layers.cross_entropy(out, label)
avg_loss = fluid.layers.mean(loss)
avg_loss.backward()
adam.minimize(avg_loss)
lenet.clear_gradients()
if batch_id % 100 == 0:
_logger.info(
"Train | At epoch {} step {}: loss = {:}, acc= {:}".
format(epoch, batch_id,
avg_loss.numpy(), acc.numpy()))
lenet.eval()
for batch_id, data in enumerate(test_reader()):
x_data = np.array([x[0].reshape(1, 28, 28)
for x in data]).astype('float32')
y_data = np.array(
[x[1] for x in data]).astype('int64').reshape(-1, 1)
img = fluid.dygraph.to_variable(x_data)
label = fluid.dygraph.to_variable(y_data)
out = lenet(img)
acc_top1 = fluid.layers.accuracy(
input=out, label=label, k=1)
acc_top5 = fluid.layers.accuracy(
input=out, label=label, k=5)
if batch_id % 100 == 0:
_logger.info(
"Test | At epoch {} step {}: acc1 = {:}, acc5 = {:}".
format(epoch, batch_id,
acc_top1.numpy(), acc_top5.numpy()))
# save weights
model_dict = lenet.state_dict()
fluid.save_dygraph(model_dict, "save_temp")
# test the correctness of `paddle.jit.save`
data = next(test_reader())
test_data = np.array([x[0].reshape(1, 28, 28)
for x in data]).astype('float32')
test_img = fluid.dygraph.to_variable(test_data)
lenet.eval()
before_save = lenet(test_img)
# save inference quantized model
path = "./qat_infer_model/lenet"
save_dir = "./qat_infer_model"
paddle.jit.save(
layer=lenet,
path=path,
input_spec=[
paddle.static.InputSpec(
shape=[None, 1, 28, 28], dtype='float32')
])
if core.is_compiled_with_cuda():
place = core.CUDAPlace(0)
else:
place = core.CPUPlace()
exe = fluid.Executor(place)
[inference_program, feed_target_names,
fetch_targets] = fluid.io.load_inference_model(
dirname=save_dir,
executor=exe,
model_filename="lenet" + INFER_MODEL_SUFFIX,
params_filename="lenet" + INFER_PARAMS_SUFFIX)
after_save, = exe.run(inference_program,
feed={feed_target_names[0]: test_data},
fetch_list=fetch_targets)
self.assertTrue(
np.allclose(after_save, before_save.numpy()),
msg='Failed to save the inference quantized model.')
class QAT(object):
"""
Quant Aware Training(QAT): Add the fake quant logic for given quantizable layers, namely add the quant_dequant computational logic both for activation inputs and weight inputs.
"""
def __init__(self,
config=None,
weight_preprocess=None,
act_preprocess=None,
weight_quantize=None,
act_quantize=None):
"""
Args:
model(nn.Layer)
config(dict, optional): configs for quantization. if None, will use default config.
Default: None.
weight_quantize(class, optional): Defines how to quantize weight. Using this
can quickly test if user's quantization method works or not. In this method, user should
both define quantization function and dequantization function, that is, the function's input
is non-quantized weight and function returns dequantized weight. If None, will use
quantization op defined by 'weight_quantize_type'.
Default is None.
act_quantize(class, optional): Defines how to quantize activation. Using this
can quickly test if user's quantization method works or not. In this function, user should
both define quantization and dequantization process, that is, the function's input
is non-quantized activation and function returns dequantized activation. If None, will use
quantization op defined by 'activation_quantize_type'.
Default is None.
weight_preprocess(class, optional): Defines how to preprocess weight before quantization. Using this
can quickly test if user's preprocess method works or not. The function's input
is non-quantized weight and function returns processed weight to be quantized. If None, will
use preprocess method defined by 'weight_preprocess_type'.
Default is None.
act_preprocess(class, optional): Defines how to preprocess activation before quantization. Using this
can quickly test if user's preprocess method works or not. The function's input
is non-quantized activation and function returns processed activation to be quantized. If None,
will use preprocess method defined by 'activation_preprocess_type'.
Default is None.
"""
if config is None:
config = _quant_config_default
else:
assert isinstance(config, dict), "config must be dict"
config = _parse_configs(config)
self.config = config
self.weight_preprocess = PACT if self.config[
'weight_preprocess_type'] == 'PACT' else None
self.act_preprocess = PACT if self.config[
'activation_preprocess_type'] == 'PACT' else None
self.weight_preprocess = weight_preprocess if weight_preprocess is not None else self.weight_preprocess
self.act_preprocess = act_preprocess if act_preprocess is not None else self.act_preprocess
self.weight_quantize = weight_quantize
self.act_quantize = act_quantize
self.imperative_qat = ImperativeQuantAware(
weight_bits=self.config['weight_bits'],
activation_bits=self.config['activation_bits'],
weight_quantize_type=self.config['weight_quantize_type'],
activation_quantize_type=self.config['activation_quantize_type'],
moving_rate=self.config['moving_rate'],
quantizable_layer_type=self.config['quantizable_layer_type'],
weight_preprocess_layer=self.weight_preprocess,
act_preprocess_layer=self.act_preprocess,
weight_quantize_layer=self.weight_quantize,
act_quantize_layer=self.act_quantize)
def quantize(self, model):
self.imperative_qat.quantize(model)
def save_quantized_model(self, model, path, input_spec=None):
if self.weight_preprocess is not None or self.act_preprocess is not None:
model = self._remove_preprocess(model)
self.imperative_qat.save_quantized_model(
layer=model, path=path, input_spec=input_spec)
def _remove_preprocess(self, model):
state_dict = model.state_dict()
self.imperative_qat = ImperativeQuantAware(
weight_bits=self.config['weight_bits'],
activation_bits=self.config['activation_bits'],
weight_quantize_type=self.config['weight_quantize_type'],
activation_quantize_type=self.config['activation_quantize_type'],
moving_rate=self.config['moving_rate'],
quantizable_layer_type=self.config['quantizable_layer_type'])
with paddle.utils.unique_name.guard():
if hasattr(model, "_layers"):
model = model._layers
model.__init__()
self.imperative_qat.quantize(model)
state_dict = model.state_dict()
model.set_state_dict(state_dict)
return model
def main():
# create model
model_list = [x for x in models.__dict__["__all__"]]
assert FLAGS.arch in model_list, \
"Expected FLAGS.arch in {}, but received {}".format(
model_list, FLAGS.arch)
model = models.__dict__[FLAGS.arch](pretrained=not FLAGS.resume)
# quantize model
if FLAGS.enable_quant:
if not FLAGS.use_naive_api:
print("use slim api")
quant_config = {
'weight_quantize_type': FLAGS.weight_quantize_type,
}
dygraph_qat = QAT(quant_config)
else:
print("use navie api")
dygraph_qat = ImperativeQuantAware(
weight_quantize_type=FLAGS.weight_quantize_type, )
dygraph_qat.quantize(model)
# prepare
model = paddle.Model(model)
if FLAGS.resume is not None:
print("Resume from " + FLAGS.resume)
model.load(FLAGS.resume)
train_dataset = ImageNetDataset(FLAGS.data, mode='train')
val_dataset = ImageNetDataset(FLAGS.data, mode='val')
optim = make_optimizer(
np.ceil(
float(len(train_dataset)) / FLAGS.batch_size /
ParallelEnv().nranks),
parameter_list=model.parameters())
model.prepare(optim, paddle.nn.CrossEntropyLoss(), Accuracy(topk=(1, 5)))
# test
if FLAGS.eval_only:
model.evaluate(
val_dataset,
batch_size=FLAGS.batch_size,
num_workers=FLAGS.num_workers)
return
# train
output_dir = os.path.join(FLAGS.output_dir, "checkpoint",
FLAGS.arch + "_checkpoint",
time.strftime('%Y-%m-%d-%H-%M', time.localtime()))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
model.fit(train_dataset,
val_dataset,
batch_size=FLAGS.batch_size,
epochs=FLAGS.epoch,
save_dir=output_dir,
num_workers=FLAGS.num_workers)
# save
if FLAGS.enable_quant:
quant_output_dir = os.path.join(FLAGS.output_dir, FLAGS.arch, "model")
input_spec = paddle.static.InputSpec(
shape=[None, 3, 224, 224], dtype='float32')
dygraph_qat.save_quantized_model(model.network, quant_output_dir,
[input_spec])
print("save all checkpoints in " + output_dir)
print("save quantized inference model in " + quant_output_dir)
def test_out_scale_acc(self):
seed = 1000
lr = 0.1
imperative_out_scale = ImperativeQuantAware()
np.random.seed(seed)
reader = paddle.batch(
paddle.dataset.mnist.test(), batch_size=32, drop_last=True)
lenet = ImperativeLenet()
fixed_state = {}
for name, param in lenet.named_parameters():
p_shape = param.numpy().shape
p_value = param.numpy()
if name.endswith("bias"):
value = np.zeros_like(p_value).astype('float32')
else:
value = np.random.normal(
loc=0.0, scale=0.01,
size=np.product(p_shape)).reshape(p_shape).astype('float32')
fixed_state[name] = value
lenet.set_dict(fixed_state)
imperative_out_scale.quantize(lenet)
adam = AdamOptimizer(
learning_rate=lr, parameter_list=lenet.parameters())
dynamic_loss_rec = []
lenet.train()
for batch_id, data in enumerate(reader()):
x_data = np.array([x[0].reshape(1, 28, 28)
for x in data]).astype('float32')
y_data = np.array(
[x[1] for x in data]).astype('int64').reshape(-1, 1)
img = fluid.dygraph.to_variable(x_data)
label = fluid.dygraph.to_variable(y_data)
out = lenet(img)
loss = fluid.layers.cross_entropy(out, label)
avg_loss = fluid.layers.mean(loss)
avg_loss.backward()
adam.minimize(avg_loss)
lenet.clear_gradients()
dynamic_loss_rec.append(avg_loss.numpy()[0])
if batch_id % 100 == 0:
_logger.info('{}: {}'.format('loss', avg_loss.numpy()))
lenet.eval()
path = "./save_dynamic_quant_infer_model/lenet"
save_dir = "./save_dynamic_quant_infer_model"
imperative_out_scale.save_quantized_model(
layer=lenet,
path=path,
input_spec=[
paddle.static.InputSpec(
shape=[None, 1, 28, 28], dtype='float32')
])
paddle.enable_static()
if core.is_compiled_with_cuda():
place = core.CUDAPlace(0)
else:
place = core.CPUPlace()
exe = fluid.Executor(place)
[inference_program, feed_target_names, fetch_targets] = (
fluid.io.load_inference_model(
dirname=save_dir,
executor=exe,
model_filename="lenet" + INFER_MODEL_SUFFIX,
params_filename="lenet" + INFER_PARAMS_SUFFIX))
model_ops = inference_program.global_block().ops
conv2d_count, mul_count = 0, 0
for i, op in enumerate(model_ops):
if op.type == 'conv2d':
if conv2d_count > 0:
self.assertTrue(
'fake_quantize_dequantize' in model_ops[i - 1].type)
else:
self.assertTrue(
'fake_quantize_dequantize' not in model_ops[i - 1].type)
conv2d_count += 1
if op.type == 'mul':
if mul_count > 0:
self.assertTrue(
'fake_quantize_dequantize' in model_ops[i - 1].type)
else:
self.assertTrue(
'fake_quantize_dequantize' not in model_ops[i - 1].type)
mul_count += 1
def test_out_scale_acc(self):
def _build_static_lenet(main, startup, is_test=False, seed=1000):
with fluid.unique_name.guard():
with fluid.program_guard(main, startup):
main.random_seed = seed
startup.random_seed = seed
img = fluid.layers.data(name='image',
shape=[1, 28, 28],
dtype='float32')
label = fluid.layers.data(name='label',
shape=[1],
dtype='int64')
prediction = StaticLenet(img)
if not is_test:
loss = fluid.layers.cross_entropy(input=prediction,
label=label)
avg_loss = fluid.layers.mean(loss)
else:
avg_loss = prediction
return img, label, avg_loss
reader = paddle.batch(paddle.dataset.mnist.test(),
batch_size=32,
drop_last=True)
weight_quantize_type = 'abs_max'
activation_quant_type = 'moving_average_abs_max'
param_init_map = {}
seed = 1000
lr = 0.1
dynamic_out_scale_list = []
static_out_scale_list = []
# imperative train
_logger.info(
"--------------------------dynamic graph qat--------------------------"
)
imperative_out_scale = ImperativeQuantAware()
with fluid.dygraph.guard():
np.random.seed(seed)
fluid.default_main_program().random_seed = seed
fluid.default_startup_program().random_seed = seed
lenet = ImperativeLenet()
fixed_state = {}
for name, param in lenet.named_parameters():
p_shape = param.numpy().shape
p_value = param.numpy()
if name.endswith("bias"):
value = np.zeros_like(p_value).astype('float32')
else:
value = np.random.normal(loc=0.0,
scale=0.01,
size=np.product(p_shape)).reshape(
p_shape).astype('float32')
fixed_state[name] = value
param_init_map[param.name] = value
lenet.set_dict(fixed_state)
imperative_out_scale.quantize(lenet)
adam = AdamOptimizer(learning_rate=lr,
parameter_list=lenet.parameters())
dynamic_loss_rec = []
lenet.train()
for batch_id, data in enumerate(reader()):
x_data = np.array([x[0].reshape(1, 28, 28)
for x in data]).astype('float32')
y_data = np.array([x[1] for x in data
]).astype('int64').reshape(-1, 1)
img = fluid.dygraph.to_variable(x_data)
label = fluid.dygraph.to_variable(y_data)
out = lenet(img)
loss = fluid.layers.cross_entropy(out, label)
avg_loss = fluid.layers.mean(loss)
avg_loss.backward()
adam.minimize(avg_loss)
lenet.clear_gradients()
dynamic_loss_rec.append(avg_loss.numpy()[0])
if batch_id % 100 == 0:
_logger.info('{}: {}'.format('loss', avg_loss.numpy()))
lenet.eval()
path = "./dynamic_outscale_infer_model/lenet"
dynamic_save_dir = "./dynamic_outscale_infer_model"
imperative_out_scale.save_quantized_model(
layer=lenet,
path=path,
input_spec=[
paddle.static.InputSpec(shape=[None, 1, 28, 28],
dtype='float32')
])
_logger.info(
"--------------------------static graph qat--------------------------"
)
static_loss_rec = []
if core.is_compiled_with_cuda():
place = core.CUDAPlace(0)
else:
place = core.CPUPlace()
exe = fluid.Executor(place)
main = fluid.Program()
infer = fluid.Program()
startup = fluid.Program()
static_img, static_label, static_loss = _build_static_lenet(
main, startup, False, seed)
infer_img, _, infer_pre = _build_static_lenet(infer, startup, True,
seed)
with fluid.unique_name.guard():
with fluid.program_guard(main, startup):
opt = AdamOptimizer(learning_rate=lr)
opt.minimize(static_loss)
scope = core.Scope()
with fluid.scope_guard(scope):
exe.run(startup)
for param in main.all_parameters():
param_tensor = scope.var(param.name).get_tensor()
param_tensor.set(param_init_map[param.name], place)
main_graph = IrGraph(core.Graph(main.desc), for_test=False)
infer_graph = IrGraph(core.Graph(infer.desc), for_test=True)
transform_pass = QuantizationTransformPass(
scope=scope,
place=place,
activation_quantize_type=activation_quant_type,
weight_quantize_type=weight_quantize_type,
quantizable_op_type=['conv2d', 'depthwise_conv2d', 'mul'])
transform_pass.apply(main_graph)
transform_pass.apply(infer_graph)
outscale_pass = OutScaleForTrainingPass(scope=scope, place=place)
outscale_pass.apply(main_graph)
build_strategy = fluid.BuildStrategy()
build_strategy.fuse_all_reduce_ops = False
binary = fluid.CompiledProgram(main_graph.graph).with_data_parallel(
loss_name=static_loss.name, build_strategy=build_strategy)
feeder = fluid.DataFeeder(feed_list=[static_img, static_label],
place=place)
with fluid.scope_guard(scope):
for batch_id, data in enumerate(reader()):
loss_v, = exe.run(binary,
feed=feeder.feed(data),
fetch_list=[static_loss])
static_loss_rec.append(loss_v[0])
if batch_id % 100 == 0:
_logger.info('{}: {}'.format('loss', loss_v))
scale_inference_pass = OutScaleForInferencePass(scope=scope)
scale_inference_pass.apply(infer_graph)
save_program = infer_graph.to_program()
static_save_dir = "./static_outscale_infer_model"
with fluid.scope_guard(scope):
fluid.io.save_inference_model(
dirname=static_save_dir,
feeded_var_names=[infer_img.name],
target_vars=[infer_pre],
executor=exe,
main_program=save_program,
model_filename="lenet" + INFER_MODEL_SUFFIX,
params_filename="lenet" + INFER_PARAMS_SUFFIX)
rtol = 1e-05
atol = 1e-08
for i, (loss_d,
loss_s) in enumerate(zip(dynamic_loss_rec, static_loss_rec)):
diff = np.abs(loss_d - loss_s)
if diff > (atol + rtol * np.abs(loss_s)):
_logger.info(
"diff({}) at {}, dynamic loss = {}, static loss = {}".
format(diff, i, loss_d, loss_s))
break
self.assertTrue(np.allclose(np.array(dynamic_loss_rec),
np.array(static_loss_rec),
rtol=rtol,
atol=atol,
equal_nan=True),
msg='Failed to do the imperative qat.')
# load dynamic model
[dynamic_inference_program, feed_target_names,
fetch_targets] = (fluid.io.load_inference_model(
dirname=dynamic_save_dir,
executor=exe,
model_filename="lenet" + INFER_MODEL_SUFFIX,
params_filename="lenet" + INFER_PARAMS_SUFFIX))
# load static model
[static_inference_program, feed_target_names,
fetch_targets] = (fluid.io.load_inference_model(
dirname=static_save_dir,
executor=exe,
model_filename="lenet" + INFER_MODEL_SUFFIX,
params_filename="lenet" + INFER_PARAMS_SUFFIX))
dynamic_ops = dynamic_inference_program.global_block().ops
static_ops = static_inference_program.global_block().ops
for op in dynamic_ops[:]:
if op.type == "flatten2" or 'fake' in op.type:
dynamic_ops.remove(op)
for op in static_ops[:]:
if 'fake' in op.type:
static_ops.remove(op)
for i in range(len(dynamic_ops)):
if dynamic_ops[i].has_attr("out_threshold"):
self.assertTrue(dynamic_ops[i].type == static_ops[i].type)
self.assertTrue(dynamic_ops[i].attr("out_threshold") ==
static_ops[i].attr("out_threshold"))
class QAT(object):
"""
Quant Aware Training(QAT): Add the fake quant logic for given quantizable layers, namely add the quant_dequant computational logic both for activation inputs and weight inputs.
"""
def __init__(self,
config=None,
weight_preprocess=None,
act_preprocess=None,
weight_quantize=None,
act_quantize=None):
"""
Args:
model(nn.Layer)
config(dict, optional): configs for quantization. if None, will use default config.
Default: None.
weight_quantize(class, optional): Defines how to quantize weight. Using this
can quickly test if user's quantization method works or not. In this method, user should
both define quantization function and dequantization function, that is, the function's input
is non-quantized weight and function returns dequantized weight. If None, will use
quantization op defined by 'weight_quantize_type'.
Default is None.
act_quantize(class, optional): Defines how to quantize activation. Using this
can quickly test if user's quantization method works or not. In this function, user should
both define quantization and dequantization process, that is, the function's input
is non-quantized activation and function returns dequantized activation. If None, will use
quantization op defined by 'activation_quantize_type'.
Default is None.
weight_preprocess(class, optional): Defines how to preprocess weight before quantization. Using this
can quickly test if user's preprocess method works or not. The function's input
is non-quantized weight and function returns processed weight to be quantized. If None, will
use preprocess method defined by 'weight_preprocess_type'.
Default is None.
act_preprocess(class, optional): Defines how to preprocess activation before quantization. Using this
can quickly test if user's preprocess method works or not. The function's input
is non-quantized activation and function returns processed activation to be quantized. If None,
will use preprocess method defined by 'activation_preprocess_type'.
Default is None.
"""
if config is None:
config = _quant_config_default
else:
assert isinstance(config, dict), "config must be dict"
config = _parse_configs(config)
self.config = config
self.weight_preprocess = PACT if self.config[
'weight_preprocess_type'] == 'PACT' else None
self.act_preprocess = PACT if self.config[
'activation_preprocess_type'] == 'PACT' else None
self.weight_preprocess = weight_preprocess if weight_preprocess is not None else self.weight_preprocess
self.act_preprocess = act_preprocess if act_preprocess is not None else self.act_preprocess
self.weight_quantize = weight_quantize
self.act_quantize = act_quantize
self.imperative_qat = ImperativeQuantAware(
weight_bits=self.config['weight_bits'],
activation_bits=self.config['activation_bits'],
weight_quantize_type=self.config['weight_quantize_type'],
activation_quantize_type=self.config['activation_quantize_type'],
moving_rate=self.config['moving_rate'],
quantizable_layer_type=self.config['quantizable_layer_type'],
weight_preprocess_layer=self.weight_preprocess,
act_preprocess_layer=self.act_preprocess,
weight_quantize_layer=self.weight_quantize,
act_quantize_layer=self.act_quantize)
def quantize(self, model, inplace=True):
"""
Quantize the input model.
Args:
model(paddle.nn.Layer): The model to be quantized.
inplace(bool): Whether apply quantization to the input model.
Default: False.
Returns:
quantized_model(paddle.nn.Layer): The quantized model.
"""
assert isinstance(model, paddle.nn.Layer), \
"The model must be the instance of paddle.nn.Layer."
self._model = copy.deepcopy(model)
if inplace:
self.imperative_qat.quantize(model)
quant_model = model
else:
quant_model = copy.deepcopy(model)
self.imperative_qat.quantize(quant_model)
return quant_model
def save_quantized_model(self,
model,
path,
input_spec=None,
onnx_format=False):
"""
Save the quantized inference model.
Args:
model (Layer): The model to be saved.
path (str): The path prefix to save model. The format is
``dirname/file_prefix`` or ``file_prefix``.
input_spec (list[InputSpec|Tensor], optional): Describes the input
of the saved model's forward method, which can be described by
InputSpec or example Tensor. If None, all input variables of
the original Layer's forward method would be the inputs of
the saved model. Default: None.
Returns:
None
"""
if self.weight_preprocess is not None or self.act_preprocess is not None:
training = model.training
model = self._remove_preprocess(model)
if training:
model.train()
else:
model.eval()
self.imperative_qat.save_quantized_model(layer=model,
path=path,
input_spec=input_spec,
onnx_format=onnx_format)
def _remove_preprocess(self, model):
state_dict = model.state_dict()
self.imperative_qat = ImperativeQuantAware(
weight_bits=self.config['weight_bits'],
activation_bits=self.config['activation_bits'],
weight_quantize_type=self.config['weight_quantize_type'],
activation_quantize_type=self.config['activation_quantize_type'],
moving_rate=self.config['moving_rate'],
quantizable_layer_type=self.config['quantizable_layer_type'])
with paddle.utils.unique_name.guard():
if hasattr(model, "_layers"):
model = model._layers
model = self._model
self.imperative_qat.quantize(model)
model.set_state_dict(state_dict)
return model
def __init__(self,
config=None,
weight_preprocess=None,
act_preprocess=None,
weight_quantize=None,
act_quantize=None):
"""
Args:
model(nn.Layer)
config(dict, optional): configs for quantization. if None, will use default config.
Default: None.
weight_quantize(class, optional): Defines how to quantize weight. Using this
can quickly test if user's quantization method works or not. In this method, user should
both define quantization function and dequantization function, that is, the function's input
is non-quantized weight and function returns dequantized weight. If None, will use
quantization op defined by 'weight_quantize_type'.
Default is None.
act_quantize(class, optional): Defines how to quantize activation. Using this
can quickly test if user's quantization method works or not. In this function, user should
both define quantization and dequantization process, that is, the function's input
is non-quantized activation and function returns dequantized activation. If None, will use
quantization op defined by 'activation_quantize_type'.
Default is None.
weight_preprocess(class, optional): Defines how to preprocess weight before quantization. Using this
can quickly test if user's preprocess method works or not. The function's input
is non-quantized weight and function returns processed weight to be quantized. If None, will
use preprocess method defined by 'weight_preprocess_type'.
Default is None.
act_preprocess(class, optional): Defines how to preprocess activation before quantization. Using this
can quickly test if user's preprocess method works or not. The function's input
is non-quantized activation and function returns processed activation to be quantized. If None,
will use preprocess method defined by 'activation_preprocess_type'.
Default is None.
"""
if config is None:
config = _quant_config_default
else:
assert isinstance(config, dict), "config must be dict"
config = _parse_configs(config)
self.config = config
self.weight_preprocess = PACT if self.config[
'weight_preprocess_type'] == 'PACT' else None
self.act_preprocess = PACT if self.config[
'activation_preprocess_type'] == 'PACT' else None
self.weight_preprocess = weight_preprocess if weight_preprocess is not None else self.weight_preprocess
self.act_preprocess = act_preprocess if act_preprocess is not None else self.act_preprocess
self.weight_quantize = weight_quantize
self.act_quantize = act_quantize
self.imperative_qat = ImperativeQuantAware(
weight_bits=self.config['weight_bits'],
activation_bits=self.config['activation_bits'],
weight_quantize_type=self.config['weight_quantize_type'],
activation_quantize_type=self.config['activation_quantize_type'],
moving_rate=self.config['moving_rate'],
quantizable_layer_type=self.config['quantizable_layer_type'],
weight_preprocess_layer=self.weight_preprocess,
act_preprocess_layer=self.act_preprocess,
weight_quantize_layer=self.weight_quantize,
act_quantize_layer=self.act_quantize)
def func_qat(self):
self.set_vars()
imperative_qat = ImperativeQuantAware(
weight_quantize_type=self.weight_quantize_type,
activation_quantize_type=self.activation_quantize_type,
fuse_conv_bn=self.fuse_conv_bn)
with fluid.dygraph.guard():
# For CI coverage
conv1 = Conv2D(
in_channels=3,
out_channels=2,
kernel_size=3,
stride=1,
padding=1,
padding_mode='replicate')
quant_conv1 = QuantizedConv2D(conv1)
data = np.random.uniform(-1, 1, [10, 3, 32, 32]).astype('float32')
quant_conv1(fluid.dygraph.to_variable(data))
conv_transpose = Conv2DTranspose(4, 6, (3, 3))
quant_conv_transpose = QuantizedConv2DTranspose(conv_transpose)
x_var = paddle.uniform(
(2, 4, 8, 8), dtype='float32', min=-1.0, max=1.0)
quant_conv_transpose(x_var)
seed = 1
np.random.seed(seed)
fluid.default_main_program().random_seed = seed
fluid.default_startup_program().random_seed = seed
lenet = ImperativeLenet()
lenet = fix_model_dict(lenet)
imperative_qat.quantize(lenet)
adam = AdamOptimizer(
learning_rate=0.001, parameter_list=lenet.parameters())
train_reader = paddle.batch(
paddle.dataset.mnist.train(), batch_size=32, drop_last=True)
test_reader = paddle.batch(
paddle.dataset.mnist.test(), batch_size=32)
epoch_num = 1
for epoch in range(epoch_num):
lenet.train()
for batch_id, data in enumerate(train_reader()):
x_data = np.array([x[0].reshape(1, 28, 28)
for x in data]).astype('float32')
y_data = np.array(
[x[1] for x in data]).astype('int64').reshape(-1, 1)
img = fluid.dygraph.to_variable(x_data)
label = fluid.dygraph.to_variable(y_data)
out = lenet(img)
acc = fluid.layers.accuracy(out, label)
loss = fluid.layers.cross_entropy(out, label)
avg_loss = fluid.layers.mean(loss)
avg_loss.backward()
adam.minimize(avg_loss)
lenet.clear_gradients()
if batch_id % 100 == 0:
_logger.info(
"Train | At epoch {} step {}: loss = {:}, acc= {:}".
format(epoch, batch_id,
avg_loss.numpy(), acc.numpy()))
if batch_id == 500: # For shortening CI time
break
lenet.eval()
eval_acc_top1_list = []
for batch_id, data in enumerate(test_reader()):
x_data = np.array([x[0].reshape(1, 28, 28)
for x in data]).astype('float32')
y_data = np.array(
[x[1] for x in data]).astype('int64').reshape(-1, 1)
img = fluid.dygraph.to_variable(x_data)
label = fluid.dygraph.to_variable(y_data)
out = lenet(img)
acc_top1 = fluid.layers.accuracy(
input=out, label=label, k=1)
acc_top5 = fluid.layers.accuracy(
input=out, label=label, k=5)
if batch_id % 100 == 0:
eval_acc_top1_list.append(float(acc_top1.numpy()))
_logger.info(
"Test | At epoch {} step {}: acc1 = {:}, acc5 = {:}".
format(epoch, batch_id,
acc_top1.numpy(), acc_top5.numpy()))
# check eval acc
eval_acc_top1 = sum(eval_acc_top1_list) / len(
eval_acc_top1_list)
print('eval_acc_top1', eval_acc_top1)
self.assertTrue(
eval_acc_top1 > 0.9,
msg="The test acc {%f} is less than 0.9." % eval_acc_top1)
# test the correctness of `paddle.jit.save`
data = next(test_reader())
test_data = np.array([x[0].reshape(1, 28, 28)
for x in data]).astype('float32')
y_data = np.array(
[x[1] for x in data]).astype('int64').reshape(-1, 1)
test_img = fluid.dygraph.to_variable(test_data)
label = fluid.dygraph.to_variable(y_data)
lenet.eval()
fp32_out = lenet(test_img)
fp32_acc = fluid.layers.accuracy(fp32_out, label).numpy()
with tempfile.TemporaryDirectory(prefix="qat_save_path_") as tmpdir:
# save inference quantized model
imperative_qat.save_quantized_model(
layer=lenet,
path=os.path.join(tmpdir, "lenet"),
input_spec=[
paddle.static.InputSpec(
shape=[None, 1, 28, 28], dtype='float32')
],
onnx_format=self.onnx_format)
print('Quantized model saved in %s' % tmpdir)
if core.is_compiled_with_cuda():
place = core.CUDAPlace(0)
else:
place = core.CPUPlace()
exe = fluid.Executor(place)
[inference_program, feed_target_names,
fetch_targets] = fluid.io.load_inference_model(
dirname=tmpdir,
executor=exe,
model_filename="lenet" + INFER_MODEL_SUFFIX,
params_filename="lenet" + INFER_PARAMS_SUFFIX)
quant_out, = exe.run(inference_program,
feed={feed_target_names[0]: test_data},
fetch_list=fetch_targets)
paddle.disable_static()
quant_out = fluid.dygraph.to_variable(quant_out)
quant_acc = fluid.layers.accuracy(quant_out, label).numpy()
paddle.enable_static()
delta_value = fp32_acc - quant_acc
self.assertLess(delta_value, self.diff_threshold)
def test_qat_acc(self):
def _build_static_lenet(main, startup, is_test=False, seed=1000):
with fluid.unique_name.guard():
with fluid.program_guard(main, startup):
main.random_seed = seed
startup.random_seed = seed
img = fluid.layers.data(
name='image', shape=[1, 28, 28], dtype='float32')
label = fluid.layers.data(
name='label', shape=[1], dtype='int64')
prediction = StaticLenet(img)
if not is_test:
loss = fluid.layers.cross_entropy(
input=prediction, label=label)
avg_loss = fluid.layers.mean(loss)
else:
avg_loss = prediction
return img, label, avg_loss
reader = paddle.batch(
paddle.dataset.mnist.test(), batch_size=32, drop_last=True)
weight_quantize_type = 'abs_max'
activation_quant_type = 'moving_average_abs_max'
param_init_map = {}
seed = 1000
lr = 0.001
# imperative train
_logger.info(
"--------------------------dynamic graph qat--------------------------"
)
imperative_qat = ImperativeQuantAware(
weight_quantize_type=weight_quantize_type,
activation_quantize_type=activation_quant_type,
quantizable_layer_type=[
'Conv2D', 'Linear', 'ReLU', 'LeakyReLU', 'ReLU6', 'Tanh',
'Swish'
])
with fluid.dygraph.guard():
np.random.seed(seed)
fluid.default_main_program().random_seed = seed
fluid.default_startup_program().random_seed = seed
lenet = ImperativeLenet()
fixed_state = {}
for name, param in lenet.named_parameters():
p_shape = param.numpy().shape
p_value = param.numpy()
if name.endswith("bias"):
value = np.zeros_like(p_value).astype('float32')
else:
value = np.random.normal(
loc=0.0, scale=0.01, size=np.product(p_shape)).reshape(
p_shape).astype('float32')
fixed_state[name] = value
param_init_map[param.name] = value
lenet.set_dict(fixed_state)
imperative_qat.quantize(lenet)
adam = AdamOptimizer(
learning_rate=lr, parameter_list=lenet.parameters())
dynamic_loss_rec = []
lenet.train()
for batch_id, data in enumerate(reader()):
x_data = np.array([x[0].reshape(1, 28, 28)
for x in data]).astype('float32')
y_data = np.array(
[x[1] for x in data]).astype('int64').reshape(-1, 1)
img = fluid.dygraph.to_variable(x_data)
label = fluid.dygraph.to_variable(y_data)
out = lenet(img)
loss = fluid.layers.cross_entropy(out, label)
avg_loss = fluid.layers.mean(loss)
avg_loss.backward()
adam.minimize(avg_loss)
lenet.clear_gradients()
dynamic_loss_rec.append(avg_loss.numpy()[0])
if batch_id % 100 == 0:
_logger.info('{}: {}'.format('loss', avg_loss.numpy()))
if batch_id > 500:
break
lenet.eval()
paddle.jit.save(
layer=lenet,
path="./dynamic_mnist/model",
input_spec=[
paddle.static.InputSpec(
shape=[None, 1, 28, 28], dtype='float32')
])
# static graph train
_logger.info(
"--------------------------static graph qat--------------------------"
)
static_loss_rec = []
if core.is_compiled_with_cuda():
place = core.CUDAPlace(0)
else:
place = core.CPUPlace()
exe = fluid.Executor(place)
main = fluid.Program()
infer = fluid.Program()
startup = fluid.Program()
static_img, static_label, static_loss = _build_static_lenet(
main, startup, False, seed)
infer_img, _, infer_pre = _build_static_lenet(infer, startup, True,
seed)
with fluid.unique_name.guard():
with fluid.program_guard(main, startup):
opt = AdamOptimizer(learning_rate=lr)
opt.minimize(static_loss)
scope = core.Scope()
with fluid.scope_guard(scope):
exe.run(startup)
for param in main.all_parameters():
param_tensor = scope.var(param.name).get_tensor()
param_tensor.set(param_init_map[param.name], place)
main_graph = IrGraph(core.Graph(main.desc), for_test=False)
infer_graph = IrGraph(core.Graph(infer.desc), for_test=True)
transform_pass = QuantizationTransformPass(
scope=scope,
place=place,
activation_quantize_type=activation_quant_type,
weight_quantize_type=weight_quantize_type,
quantizable_op_type=['conv2d', 'depthwise_conv2d', 'mul'])
add_quant_dequant_pass = AddQuantDequantPass(
scope=scope,
place=place,
quantizable_op_type=[
'relu', 'leaky_relu', 'relu6', 'tanh', 'swish'
])
transform_pass.apply(main_graph)
transform_pass.apply(infer_graph)
add_quant_dequant_pass.apply(main_graph)
add_quant_dequant_pass.apply(infer_graph)
build_strategy = fluid.BuildStrategy()
build_strategy.fuse_all_reduce_ops = False
binary = fluid.CompiledProgram(main_graph.graph).with_data_parallel(
loss_name=static_loss.name, build_strategy=build_strategy)
feeder = fluid.DataFeeder(
feed_list=[static_img, static_label], place=place)
with fluid.scope_guard(scope):
for batch_id, data in enumerate(reader()):
loss_v, = exe.run(binary,
feed=feeder.feed(data),
fetch_list=[static_loss])
static_loss_rec.append(loss_v[0])
if batch_id % 100 == 0:
_logger.info('{}: {}'.format('loss', loss_v))
save_program = infer_graph.to_program()
with fluid.scope_guard(scope):
fluid.io.save_inference_model("./static_mnist", [infer_img.name],
[infer_pre], exe, save_program)
rtol = 1e-08
atol = 1e-10
for i, (loss_d,
loss_s) in enumerate(zip(dynamic_loss_rec, static_loss_rec)):
diff = np.abs(loss_d - loss_s)
if diff > (atol + rtol * np.abs(loss_s)):
_logger.info(
"diff({}) at {}, dynamic loss = {}, static loss = {}".
format(diff, i, loss_d, loss_s))
break
self.assertTrue(
np.allclose(
np.array(dynamic_loss_rec),
np.array(static_loss_rec),
rtol=rtol,
atol=atol,
equal_nan=True),
msg='Failed to do the imperative qat.')
def test_save_quantized_model(self):
weight_quantize_type = 'abs_max'
activation_quantize_type = 'moving_average_abs_max'
load_param_path = "test_save_quantized_model/lenet.pdparams"
path = "./dynamic_outscale_infer_model_from_checkpoint/lenet"
dynamic_model_save_dir = "./dynamic_outscale_infer_model_from_checkpoint"
static_model_save_dir = "./static_outscale_infer_model"
imperative_out_scale = ImperativeQuantAware(
weight_quantize_type=weight_quantize_type,
activation_quantize_type=activation_quantize_type)
with fluid.dygraph.guard():
lenet = ImperativeLenet()
load_dict = paddle.load(load_param_path)
imperative_out_scale.quantize(lenet)
lenet.set_dict(load_dict)
imperative_out_scale.save_quantized_model(
layer=lenet,
path=path,
input_spec=[
paddle.static.InputSpec(shape=[None, 1, 28, 28],
dtype='float32')
])
if core.is_compiled_with_cuda():
place = core.CUDAPlace(0)
else:
place = core.CPUPlace()
exe = fluid.Executor(place)
# load dynamic model
[dynamic_inference_program, feed_target_names,
fetch_targets] = (fluid.io.load_inference_model(
dirname=dynamic_model_save_dir,
executor=exe,
model_filename="lenet" + INFER_MODEL_SUFFIX,
params_filename="lenet" + INFER_PARAMS_SUFFIX))
# load static model
[static_inference_program, feed_target_names,
fetch_targets] = (fluid.io.load_inference_model(
dirname=static_model_save_dir,
executor=exe,
model_filename="lenet" + INFER_MODEL_SUFFIX,
params_filename="lenet" + INFER_PARAMS_SUFFIX))
dynamic_ops = dynamic_inference_program.global_block().ops
static_ops = static_inference_program.global_block().ops
for op in dynamic_ops[:]:
if op.type == "flatten2" or 'fake' in op.type:
dynamic_ops.remove(op)
for op in static_ops[:]:
if 'fake' in op.type:
static_ops.remove(op)
op_count = 0
for i in range(len(dynamic_ops)):
if dynamic_ops[i].has_attr("out_threshold"):
op_count += 1
self.assertTrue(dynamic_ops[i].type == static_ops[i].type)
self.assertTrue(dynamic_ops[i].attr("out_threshold") ==
static_ops[i].attr("out_threshold"))
_logger.info("op_cout: {}".format(op_count))
self.assertTrue(op_count == 14)
