def get_model(self):
image = paddle.static.data(
name='image', shape=[None, 1, 28, 28], dtype='float32')
label = paddle.static.data(name='label', shape=[None, 1], dtype='int64')
model = MobileNet()
out = model.net(input=image, class_dim=10)
cost = paddle.nn.functional.loss.cross_entropy(input=out, label=label)
avg_cost = paddle.mean(x=cost)
startup_prog = paddle.static.default_startup_program()
train_prog = paddle.static.default_main_program()
return startup_prog, train_prog
def get_model(self):
image = fluid.layers.data(
name='image', shape=[1, 28, 28], dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
model = MobileNet()
out = model.net(input=image, class_dim=10)
cost = fluid.layers.cross_entropy(input=out, label=label)
avg_cost = fluid.layers.mean(x=cost)
startup_prog = fluid.default_startup_program()
train_prog = fluid.default_main_program()
return startup_prog, train_prog
def test_analysis_helper(self):
image = paddle.static.data(
name='image', shape=[None, 1, 28, 28], dtype='float32')
label = paddle.static.data(name='label', shape=[None, 1], dtype='int64')
model = MobileNet()
out = model.net(input=image, class_dim=10)
cost = paddle.nn.functional.loss.cross_entropy(input=out, label=label)
avg_cost = paddle.mean(x=cost)
acc_top1 = paddle.metric.accuracy(input=out, label=label, k=1)
acc_top5 = paddle.metric.accuracy(input=out, label=label, k=5)
optimizer = paddle.optimizer.Momentum(
momentum=0.9,
learning_rate=0.01,
weight_decay=paddle.regularizer.L2Decay(4e-5))
optimizer.minimize(avg_cost)
main_prog = paddle.static.default_main_program()
places = paddle.static.cuda_places() if paddle.is_compiled_with_cuda(
) else paddle.static.cpu_places()
exe = paddle.static.Executor(places[0])
def transform(x):
return np.reshape(x, [1, 28, 28])
train_dataset = paddle.vision.datasets.MNIST(
mode='train', backend='cv2', transform=transform)
train_loader = paddle.io.DataLoader(
train_dataset,
places=places,
feed_list=[image, label],
drop_last=True,
return_list=False,
batch_size=64)
exe.run(paddle.static.default_startup_program())
vars = ['conv2d_0.tmp_0', 'fc_0.tmp_0', 'fc_0.tmp_1', 'fc_0.tmp_2']
var_collector1 = VarCollector(main_prog, vars, use_ema=True)
values = var_collector1.abs_max_run(
train_loader, exe, step=None, loss_name=avg_cost.name)
vars = [v.name for v in main_prog.list_vars() if v.persistable]
var_collector2 = VarCollector(main_prog, vars, use_ema=False)
values = var_collector2.run(train_loader,
exe,
step=None,
loss_name=avg_cost.name)
var_collector2.pdf(values)
def test_analysis_helper(self):
image = fluid.layers.data(name='image',
shape=[1, 28, 28],
dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
model = MobileNet()
out = model.net(input=image, class_dim=10)
cost = fluid.layers.cross_entropy(input=out, label=label)
avg_cost = fluid.layers.mean(x=cost)
acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1)
acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5)
optimizer = fluid.optimizer.Momentum(
momentum=0.9,
learning_rate=0.01,
regularization=fluid.regularizer.L2Decay(4e-5))
optimizer.minimize(avg_cost)
main_prog = fluid.default_main_program()
places = fluid.cuda_places() if fluid.is_compiled_with_cuda(
) else fluid.cpu_places()
exe = fluid.Executor(places[0])
train_reader = paddle.fluid.io.batch(paddle.dataset.mnist.train(),
batch_size=64)
train_loader = fluid.io.DataLoader.from_generator(
feed_list=[image, label],
capacity=512,
use_double_buffer=True,
iterable=True)
train_loader.set_sample_list_generator(train_reader, places)
exe.run(fluid.default_startup_program())
vars = ['conv2d_0.tmp_0', 'fc_0.tmp_0', 'fc_0.tmp_1', 'fc_0.tmp_2']
var_collector1 = VarCollector(main_prog, vars, use_ema=True)
values = var_collector1.abs_max_run(train_loader,
exe,
step=None,
loss_name=avg_cost.name)
vars = [v.name for v in main_prog.list_vars() if v.persistable]
var_collector2 = VarCollector(main_prog, vars, use_ema=False)
values = var_collector2.run(train_loader,
exe,
step=None,
loss_name=avg_cost.name)
var_collector2.pdf(values)
def test_accuracy(self):
float_infer_model_path_prefix = "./mv1_float_inference"
image = paddle.static.data(name='image',
shape=[None, 1, 28, 28],
dtype='float32')
label = paddle.static.data(name='label',
shape=[None, 1],
dtype='int64')
model = MobileNet()
out = model.net(input=image, class_dim=10)
cost = paddle.nn.functional.loss.cross_entropy(input=out, label=label)
avg_cost = paddle.mean(x=cost)
acc_top1 = paddle.metric.accuracy(input=out, label=label, k=1)
acc_top5 = paddle.metric.accuracy(input=out, label=label, k=5)
optimizer = paddle.optimizer.Momentum(
momentum=0.9,
learning_rate=0.01,
weight_decay=paddle.regularizer.L2Decay(4e-5))
optimizer.minimize(avg_cost)
main_prog = paddle.static.default_main_program()
val_prog = main_prog.clone(for_test=True)
#place = paddle.CPUPlace()
place = paddle.CUDAPlace(
0) if paddle.is_compiled_with_cuda() else paddle.CPUPlace()
exe = paddle.static.Executor(place)
exe.run(paddle.static.default_startup_program())
def transform(x):
return np.reshape(x, [1, 28, 28])
train_dataset = paddle.vision.datasets.MNIST(mode='train',
backend='cv2',
transform=transform)
test_dataset = paddle.vision.datasets.MNIST(mode='test',
backend='cv2',
transform=transform)
train_loader = paddle.io.DataLoader(train_dataset,
places=place,
feed_list=[image, label],
drop_last=True,
batch_size=64,
return_list=False)
valid_loader = paddle.io.DataLoader(test_dataset,
places=place,
feed_list=[image, label],
batch_size=64,
return_list=False)
def sample_generator_creator():
def __reader__():
for data in test_dataset:
image, label = data
yield image, label
return __reader__
def train(program):
iter = 0
for data in train_loader():
cost, top1, top5 = exe.run(
program,
feed=data,
fetch_list=[avg_cost, acc_top1, acc_top5])
iter += 1
if iter % 100 == 0:
print(
'train iter={}, avg loss {}, acc_top1 {}, acc_top5 {}'.
format(iter, cost, top1, top5))
def test(program, outputs=[avg_cost, acc_top1, acc_top5]):
iter = 0
result = [[], [], []]
for data in valid_loader():
cost, top1, top5 = exe.run(program,
feed=data,
fetch_list=outputs)
iter += 1
if iter % 100 == 0:
print(
'eval iter={}, avg loss {}, acc_top1 {}, acc_top5 {}'.
format(iter, cost, top1, top5))
result[0].append(cost)
result[1].append(top1)
result[2].append(top5)
print(' avg loss {}, acc_top1 {}, acc_top5 {}'.format(
np.mean(result[0]), np.mean(result[1]), np.mean(result[2])))
return np.mean(result[1]), np.mean(result[2])
train(main_prog)
top1_1, top5_1 = test(val_prog)
paddle.static.save_inference_model(
path_prefix=float_infer_model_path_prefix,
feed_vars=[image, label],
fetch_vars=[avg_cost, acc_top1, acc_top5],
executor=exe,
program=val_prog)
quant_config = {
'weight_quantize_type': 'channel_wise_abs_max',
'activation_quantize_type': 'moving_average_abs_max',
'not_quant_pattern': ['skip_quant'],
'quantize_op_types': ['conv2d', 'depthwise_conv2d', 'mul']
}
train_config = {
"num_epoch":
1, # training epoch num
"max_iter":
20,
"save_iter_step":
10,
"learning_rate":
0.0001,
"weight_decay":
0.0001,
"use_pact":
False,
"quant_model_ckpt_path":
"./quantaware_with_infermodel_checkpoints/",
"teacher_model_path_prefix":
float_infer_model_path_prefix,
"model_path_prefix":
float_infer_model_path_prefix,
"distill_node_pair": [
"teacher_fc_0.tmp_0", "fc_0.tmp_0",
"teacher_batch_norm_24.tmp_4", "batch_norm_24.tmp_4",
"teacher_batch_norm_22.tmp_4", "batch_norm_22.tmp_4",
"teacher_batch_norm_18.tmp_4", "batch_norm_18.tmp_4",
"teacher_batch_norm_13.tmp_4", "batch_norm_13.tmp_4",
"teacher_batch_norm_5.tmp_4", "batch_norm_5.tmp_4"
]
}
def test_callback(compiled_test_program, feed_names, fetch_list,
checkpoint_name):
outputs = fetch_list
iter = 0
result = [[], [], []]
for data in valid_loader():
cost, top1, top5 = exe.run(compiled_test_program,
feed=data,
fetch_list=fetch_list)
iter += 1
if iter % 100 == 0:
print(
'eval iter={}, avg loss {}, acc_top1 {}, acc_top5 {}'.
format(iter, cost, top1, top5))
result[0].append(cost)
result[1].append(top1)
result[2].append(top5)
print("quant model checkpoint: " + checkpoint_name +
' avg loss {}, acc_top1 {}, acc_top5 {}'.format(
np.mean(result[0]), np.mean(result[1]), np.mean(
result[2])))
return np.mean(result[1]), np.mean(result[2])
def test_quant_aware_with_infermodel(exe, place):
quant_aware_with_infermodel(exe,
place,
scope=None,
train_reader=train_loader,
quant_config=quant_config,
train_config=train_config,
test_callback=test_callback)
def test_export_quant_infermodel(exe, place, checkpoint_path,
quant_infermodel_save_path):
export_quant_infermodel(
exe,
place,
scope=None,
quant_config=quant_config,
train_config=train_config,
checkpoint_path=checkpoint_path,
export_inference_model_path_prefix=quant_infermodel_save_path)
#place = paddle.CPUPlace()
place = paddle.CUDAPlace(
0) if paddle.is_compiled_with_cuda() else paddle.CPUPlace()
exe = paddle.static.Executor(place)
test_quant_aware_with_infermodel(exe, place)
checkpoint_path = "./quantaware_with_infermodel_checkpoints/epoch_0_iter_10"
quant_infermodel_save_path = "./quantaware_with_infermodel_export"
test_export_quant_infermodel(exe, place, checkpoint_path,
quant_infermodel_save_path)
train_config["use_pact"] = True
test_quant_aware_with_infermodel(exe, place)
train_config["use_pact"] = False
checkpoint_path = "./quantaware_with_infermodel_checkpoints/epoch_0_iter_10"
quant_infermodel_save_path = "./quantaware_with_infermodel_export"
test_export_quant_infermodel(exe, place, checkpoint_path,
quant_infermodel_save_path)
def test_accuracy(self):
image = paddle.static.data(name='image',
shape=[None, 1, 28, 28],
dtype='float32')
label = paddle.static.data(name='label',
shape=[None, 1],
dtype='int64')
model = MobileNet()
out = model.net(input=image, class_dim=10)
cost = paddle.nn.functional.loss.cross_entropy(input=out, label=label)
avg_cost = paddle.mean(x=cost)
acc_top1 = paddle.metric.accuracy(input=out, label=label, k=1)
acc_top5 = paddle.metric.accuracy(input=out, label=label, k=5)
optimizer = paddle.optimizer.Momentum(
momentum=0.9,
learning_rate=0.01,
weight_decay=paddle.regularizer.L2Decay(4e-5))
optimizer.minimize(avg_cost)
main_prog = paddle.static.default_main_program()
val_prog = main_prog.clone(for_test=True)
place = paddle.CUDAPlace(
0) if paddle.is_compiled_with_cuda() else paddle.CPUPlace()
exe = paddle.static.Executor(place)
exe.run(paddle.static.default_startup_program())
def transform(x):
return np.reshape(x, [1, 28, 28])
train_dataset = paddle.vision.datasets.MNIST(mode='train',
backend='cv2',
transform=transform)
test_dataset = paddle.vision.datasets.MNIST(mode='test',
backend='cv2',
transform=transform)
train_loader = paddle.io.DataLoader(train_dataset,
places=place,
feed_list=[image, label],
drop_last=True,
batch_size=64,
return_list=False)
valid_loader = paddle.io.DataLoader(test_dataset,
places=place,
feed_list=[image, label],
batch_size=64,
return_list=False)
def sample_generator_creator():
def __reader__():
for data in test_dataset:
image, label = data
yield image, label
return __reader__
def train(program):
iter = 0
for data in train_loader():
cost, top1, top5 = exe.run(
program,
feed=data,
fetch_list=[avg_cost, acc_top1, acc_top5])
iter += 1
if iter % 100 == 0:
print(
'train iter={}, avg loss {}, acc_top1 {}, acc_top5 {}'.
format(iter, cost, top1, top5))
def test(program, outputs=[avg_cost, acc_top1, acc_top5]):
iter = 0
result = [[], [], []]
for data in valid_loader():
cost, top1, top5 = exe.run(program,
feed=data,
fetch_list=outputs)
iter += 1
if iter % 100 == 0:
print(
'eval iter={}, avg loss {}, acc_top1 {}, acc_top5 {}'.
format(iter, cost, top1, top5))
result[0].append(cost)
result[1].append(top1)
result[2].append(top5)
print(' avg loss {}, acc_top1 {}, acc_top5 {}'.format(
np.mean(result[0]), np.mean(result[1]), np.mean(result[2])))
return np.mean(result[1]), np.mean(result[2])
train(main_prog)
top1_1, top5_1 = test(val_prog)
paddle.fluid.io.save_inference_model(
dirname='./test_quant_post',
feeded_var_names=[image.name, label.name],
target_vars=[avg_cost, acc_top1, acc_top5],
main_program=val_prog,
executor=exe,
model_filename='model',
params_filename='params')
quant_post_static(exe,
'./test_quant_post',
'./test_quant_post_inference',
sample_generator=sample_generator_creator(),
model_filename='model',
params_filename='params',
batch_nums=10)
quant_post_prog, feed_target_names, fetch_targets = paddle.fluid.io.load_inference_model(
dirname='./test_quant_post_inference',
executor=exe,
model_filename='__model__',
params_filename='__params__')
top1_2, top5_2 = test(quant_post_prog, fetch_targets)
print("before quantization: top1: {}, top5: {}".format(top1_1, top5_1))
print("after quantization: top1: {}, top5: {}".format(top1_2, top5_2))
def test_accuracy(self):
image = paddle.static.data(
name='image', shape=[None, 1, 28, 28], dtype='float32')
image.stop_gradient = False
label = paddle.static.data(name='label', shape=[None, 1], dtype='int64')
model = MobileNet()
out = model.net(input=image, class_dim=10)
cost = paddle.nn.functional.loss.cross_entropy(input=out, label=label)
avg_cost = paddle.mean(x=cost)
acc_top1 = paddle.metric.accuracy(input=out, label=label, k=1)
acc_top5 = paddle.metric.accuracy(input=out, label=label, k=5)
optimizer = paddle.optimizer.Momentum(
momentum=0.9,
learning_rate=0.01,
weight_decay=paddle.regularizer.L2Decay(4e-5))
optimizer.minimize(avg_cost)
main_prog = paddle.static.default_main_program()
val_prog = main_prog.clone(for_test=True)
place = paddle.CUDAPlace(0) if paddle.is_compiled_with_cuda(
) else paddle.CPUPlace()
exe = paddle.static.Executor(place)
exe.run(paddle.static.default_startup_program())
def transform(x):
return np.reshape(x, [1, 28, 28])
train_dataset = paddle.vision.datasets.MNIST(
mode='train', backend='cv2', transform=transform)
test_dataset = paddle.vision.datasets.MNIST(
mode='test', backend='cv2', transform=transform)
train_loader = paddle.io.DataLoader(
train_dataset,
places=place,
feed_list=[image, label],
drop_last=True,
return_list=False,
batch_size=64)
valid_loader = paddle.io.DataLoader(
test_dataset,
places=place,
feed_list=[image, label],
batch_size=64,
return_list=False)
def train(program):
iter = 0
for data in train_loader():
cost, top1, top5 = exe.run(
program,
feed=data,
fetch_list=[avg_cost, acc_top1, acc_top5])
iter += 1
if iter % 100 == 0:
print(
'train iter={}, avg loss {}, acc_top1 {}, acc_top5 {}'.
format(iter, cost, top1, top5))
def test(program):
iter = 0
result = [[], [], []]
for data in valid_loader():
cost, top1, top5 = exe.run(
program,
feed=data,
fetch_list=[avg_cost, acc_top1, acc_top5])
iter += 1
if iter % 100 == 0:
print('eval iter={}, avg loss {}, acc_top1 {}, acc_top5 {}'.
format(iter, cost, top1, top5))
result[0].append(cost)
result[1].append(top1)
result[2].append(top5)
print(' avg loss {}, acc_top1 {}, acc_top5 {}'.format(
np.mean(result[0]), np.mean(result[1]), np.mean(result[2])))
return np.mean(result[1]), np.mean(result[2])
train(main_prog)
top1_1, top5_1 = test(main_prog)
config = {
'weight_quantize_type': 'channel_wise_abs_max',
'activation_quantize_type': 'moving_average_abs_max',
'quantize_op_types': ['depthwise_conv2d', 'mul', 'conv2d'],
}
quant_train_prog_pact = quant_aware(
main_prog,
place,
config,
for_test=False,
act_preprocess_func=pact,
optimizer_func=get_optimizer,
executor=exe)
quant_eval_prog = quant_aware(val_prog, place, config, for_test=True)
train(quant_train_prog_pact)
quant_eval_prog, int8_prog = convert(
quant_eval_prog, place, config, save_int8=True)
top1_2, top5_2 = test(quant_eval_prog)
# values before quantization and after quantization should be close
print("before quantization: top1: {}, top5: {}".format(top1_1, top5_1))
print("after quantization: top1: {}, top5: {}".format(top1_2, top5_2))
def test_accuracy(self):
image = paddle.static.data(
name='image', shape=[None, 1, 28, 28], dtype='float32')
label = paddle.static.data(name='label', shape=[None, 1], dtype='int64')
model = MobileNet()
out = model.net(input=image, class_dim=10)
cost = paddle.nn.functional.loss.cross_entropy(input=out, label=label)
avg_cost = paddle.mean(x=cost)
acc_top1 = paddle.metric.accuracy(input=out, label=label, k=1)
acc_top5 = paddle.metric.accuracy(input=out, label=label, k=5)
optimizer = paddle.optimizer.Momentum(
momentum=0.9,
learning_rate=0.01,
weight_decay=paddle.regularizer.L2Decay(4e-5))
optimizer.minimize(avg_cost)
main_prog = paddle.static.default_main_program()
val_prog = main_prog.clone(for_test=True)
place = paddle.CUDAPlace(0) if paddle.fluid.is_compiled_with_cuda(
) else paddle.CPUPlace()
exe = paddle.static.Executor(place)
exe.run(paddle.static.default_startup_program())
train_loader = paddle.io.DataLoader.from_generator(
feed_list=[image, label],
capacity=512,
use_double_buffer=True,
iterable=True)
valid_loader = paddle.io.DataLoader.from_generator(
feed_list=[image, label],
capacity=512,
use_double_buffer=True,
iterable=True)
train_reader = paddle.batch(paddle.dataset.mnist.train(), batch_size=64)
eval_reader = paddle.batch(paddle.dataset.mnist.test(), batch_size=64)
train_loader.set_sample_list_generator(train_reader, place)
valid_loader.set_sample_list_generator(eval_reader, place)
def train(program):
iter = 0
for data in train_loader():
cost, top1, top5 = exe.run(
program,
feed=data,
fetch_list=[avg_cost, acc_top1, acc_top5])
iter += 1
if iter % 100 == 0:
print(
'train iter={}, avg loss {}, acc_top1 {}, acc_top5 {}'.
format(iter, cost, top1, top5))
def test(program, outputs=[avg_cost, acc_top1, acc_top5]):
iter = 0
result = [[], [], []]
for data in valid_loader():
cost, top1, top5 = exe.run(program,
feed=data,
fetch_list=outputs)
iter += 1
if iter % 100 == 0:
print('eval iter={}, avg loss {}, acc_top1 {}, acc_top5 {}'.
format(iter, cost, top1, top5))
result[0].append(cost)
result[1].append(top1)
result[2].append(top5)
print(' avg loss {}, acc_top1 {}, acc_top5 {}'.format(
np.mean(result[0]), np.mean(result[1]), np.mean(result[2])))
return np.mean(result[1]), np.mean(result[2])
train(main_prog)
top1_1, top5_1 = test(val_prog)
paddle.static.save_inference_model(
dirname='./test_quant_post_dynamic',
feeded_var_names=[image.name, label.name],
target_vars=[avg_cost, acc_top1, acc_top5],
main_program=val_prog,
executor=exe,
model_filename='model',
params_filename='params')
quant_post_dynamic(
model_dir='./test_quant_post_dynamic',
save_model_dir='./test_quant_post_inference',
model_filename='model',
params_filename='params',
generate_test_model=True)
quant_post_prog, feed_target_names, fetch_targets = paddle.static.load_inference_model(
dirname='./test_quant_post_inference/test_model', executor=exe)
top1_2, top5_2 = test(quant_post_prog, fetch_targets)
print("before quantization: top1: {}, top5: {}".format(top1_1, top5_1))
print("after quantization: top1: {}, top5: {}".format(top1_2, top5_2))
def test_accuracy(self):
image = fluid.layers.data(name='image',
shape=[1, 28, 28],
dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
model = MobileNet()
out = model.net(input=image, class_dim=10)
cost = fluid.layers.cross_entropy(input=out, label=label)
avg_cost = fluid.layers.mean(x=cost)
acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1)
acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5)
optimizer = fluid.optimizer.Momentum(
momentum=0.9,
learning_rate=0.01,
regularization=fluid.regularizer.L2Decay(4e-5))
optimizer.minimize(avg_cost)
main_prog = fluid.default_main_program()
val_prog = main_prog.clone(for_test=True)
place = fluid.CUDAPlace(
0) if fluid.is_compiled_with_cuda() else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
feeder = fluid.DataFeeder([image, label], place, program=main_prog)
train_reader = paddle.fluid.io.batch(paddle.dataset.mnist.train(),
batch_size=64)
eval_reader = paddle.fluid.io.batch(paddle.dataset.mnist.test(),
batch_size=64)
def train(program):
iter = 0
for data in train_reader():
cost, top1, top5 = exe.run(
program,
feed=feeder.feed(data),
fetch_list=[avg_cost, acc_top1, acc_top5])
iter += 1
if iter % 100 == 0:
print(
'train iter={}, avg loss {}, acc_top1 {}, acc_top5 {}'.
format(iter, cost, top1, top5))
def test(program, outputs=[avg_cost, acc_top1, acc_top5]):
iter = 0
result = [[], [], []]
for data in train_reader():
cost, top1, top5 = exe.run(program,
feed=feeder.feed(data),
fetch_list=outputs)
iter += 1
if iter % 100 == 0:
print(
'eval iter={}, avg loss {}, acc_top1 {}, acc_top5 {}'.
format(iter, cost, top1, top5))
result[0].append(cost)
result[1].append(top1)
result[2].append(top5)
print(' avg loss {}, acc_top1 {}, acc_top5 {}'.format(
np.mean(result[0]), np.mean(result[1]), np.mean(result[2])))
return np.mean(result[1]), np.mean(result[2])
train(main_prog)
top1_1, top5_1 = test(val_prog)
fluid.io.save_inference_model(
dirname='./test_quant_post',
feeded_var_names=[image.name, label.name],
target_vars=[avg_cost, acc_top1, acc_top5],
main_program=val_prog,
executor=exe,
model_filename='model',
params_filename='params')
quant_post_static(exe,
'./test_quant_post',
'./test_quant_post_inference',
sample_generator=paddle.dataset.mnist.test(),
model_filename='model',
params_filename='params',
batch_nums=10)
quant_post_prog, feed_target_names, fetch_targets = fluid.io.load_inference_model(
dirname='./test_quant_post_inference',
executor=exe,
model_filename='__model__',
params_filename='__params__')
top1_2, top5_2 = test(quant_post_prog, fetch_targets)
print("before quantization: top1: {}, top5: {}".format(top1_1, top5_1))
print("after quantization: top1: {}, top5: {}".format(top1_2, top5_2))
def test_accuracy(self):
image = fluid.layers.data(name='image',
shape=[1, 28, 28],
dtype='float32')
image.stop_gradient = False
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
model = MobileNet()
out = model.net(input=image, class_dim=10)
cost = fluid.layers.cross_entropy(input=out, label=label)
avg_cost = fluid.layers.mean(x=cost)
acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1)
acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5)
optimizer = fluid.optimizer.Momentum(
momentum=0.9,
learning_rate=0.01,
regularization=fluid.regularizer.L2Decay(4e-5))
optimizer.minimize(avg_cost)
main_prog = fluid.default_main_program()
val_prog = main_prog.clone(for_test=True)
place = fluid.CUDAPlace(
0) if fluid.is_compiled_with_cuda() else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
feeder = fluid.DataFeeder([image, label], place, program=main_prog)
train_reader = paddle.fluid.io.batch(paddle.dataset.mnist.train(),
batch_size=64)
eval_reader = paddle.fluid.io.batch(paddle.dataset.mnist.test(),
batch_size=64)
def train(program):
iter = 0
for data in train_reader():
cost, top1, top5 = exe.run(
program,
feed=feeder.feed(data),
fetch_list=[avg_cost, acc_top1, acc_top5])
iter += 1
if iter % 100 == 0:
print(
'train iter={}, avg loss {}, acc_top1 {}, acc_top5 {}'.
format(iter, cost, top1, top5))
def test(program):
iter = 0
result = [[], [], []]
for data in eval_reader():
cost, top1, top5 = exe.run(
program,
feed=feeder.feed(data),
fetch_list=[avg_cost, acc_top1, acc_top5])
iter += 1
if iter % 100 == 0:
print(
'eval iter={}, avg loss {}, acc_top1 {}, acc_top5 {}'.
format(iter, cost, top1, top5))
result[0].append(cost)
result[1].append(top1)
result[2].append(top5)
print(' avg loss {}, acc_top1 {}, acc_top5 {}'.format(
np.mean(result[0]), np.mean(result[1]), np.mean(result[2])))
return np.mean(result[1]), np.mean(result[2])
train(main_prog)
top1_1, top5_1 = test(main_prog)
config = {
'weight_quantize_type': 'channel_wise_abs_max',
'activation_quantize_type': 'moving_average_abs_max',
'quantize_op_types': ['depthwise_conv2d', 'mul', 'conv2d'],
}
quant_train_prog_pact = quant_aware(main_prog,
place,
config,
for_test=False,
act_preprocess_func=pact,
optimizer_func=get_optimizer,
executor=exe)
quant_eval_prog = quant_aware(val_prog, place, config, for_test=True)
train(quant_train_prog_pact)
quant_eval_prog, int8_prog = convert(quant_eval_prog,
place,
config,
save_int8=True)
top1_2, top5_2 = test(quant_eval_prog)
# values before quantization and after quantization should be close
print("before quantization: top1: {}, top5: {}".format(top1_1, top5_1))
print("after quantization: top1: {}, top5: {}".format(top1_2, top5_2))
def test_accuracy(self):
image = paddle.static.data(name='image',
shape=[None, 1, 28, 28],
dtype='float32')
label = paddle.static.data(name='label',
shape=[None, 1],
dtype='int64')
model = MobileNet()
out = model.net(input=image, class_dim=10)
cost = paddle.nn.functional.loss.cross_entropy(input=out, label=label)
avg_cost = paddle.mean(x=cost)
acc_top1 = paddle.metric.accuracy(input=out, label=label, k=1)
acc_top5 = paddle.metric.accuracy(input=out, label=label, k=5)
optimizer = paddle.optimizer.Momentum(
momentum=0.9,
learning_rate=0.01,
weight_decay=paddle.regularizer.L2Decay(4e-5))
optimizer.minimize(avg_cost)
main_prog = paddle.static.default_main_program()
val_prog = main_prog.clone(for_test=True)
place = paddle.CUDAPlace(0) if paddle.fluid.is_compiled_with_cuda(
) else paddle.static.CPUPlace()
exe = paddle.static.Executor(place)
exe.run(paddle.static.default_startup_program())
train_loader = paddle.io.DataLoader.from_generator(
feed_list=[image, label],
capacity=512,
use_double_buffer=True,
iterable=True)
valid_loader = paddle.io.DataLoader.from_generator(
feed_list=[image, label],
capacity=512,
use_double_buffer=True,
iterable=True)
train_reader = paddle.batch(paddle.dataset.mnist.train(),
batch_size=64)
eval_reader = paddle.batch(paddle.dataset.mnist.test(), batch_size=64)
train_loader.set_sample_list_generator(train_reader, place)
valid_loader.set_sample_list_generator(eval_reader, place)
def train(program):
iter = 0
for data in train_loader():
cost, top1, top5 = exe.run(
program,
feed=data,
fetch_list=[avg_cost, acc_top1, acc_top5])
iter += 1
if iter % 100 == 0:
print(
'train iter={}, avg loss {}, acc_top1 {}, acc_top5 {}'.
format(iter, cost, top1, top5))
def test(program):
iter = 0
result = [[], [], []]
for data in valid_loader():
cost, top1, top5 = exe.run(
program,
feed=data,
fetch_list=[avg_cost, acc_top1, acc_top5])
iter += 1
if iter % 100 == 0:
print(
'eval iter={}, avg loss {}, acc_top1 {}, acc_top5 {}'.
format(iter, cost, top1, top5))
result[0].append(cost)
result[1].append(top1)
result[2].append(top5)
print(' avg loss {}, acc_top1 {}, acc_top5 {}'.format(
np.mean(result[0]), np.mean(result[1]), np.mean(result[2])))
return np.mean(result[1]), np.mean(result[2])
train(main_prog)
top1_1, top5_1 = test(main_prog)
config = {
'weight_quantize_type': 'channel_wise_abs_max',
'activation_quantize_type': 'moving_average_abs_max',
'quantize_op_types': ['depthwise_conv2d', 'mul', 'conv2d'],
}
quant_train_prog = quant_aware(main_prog,
place,
config,
for_test=False)
quant_eval_prog = quant_aware(val_prog, place, config, for_test=True)
train(quant_train_prog)
quant_eval_prog, int8_prog = convert(quant_eval_prog,
place,
config,
save_int8=True)
top1_2, top5_2 = test(quant_eval_prog)
# values before quantization and after quantization should be close
print("before quantization: top1: {}, top5: {}".format(top1_1, top5_1))
print("after quantization: top1: {}, top5: {}".format(top1_2, top5_2))