def train(self, to_static=False):
np.random.seed(SEED)
paddle.seed(SEED)
paddle.framework.random._manual_program_seed(SEED)
mnist = MNIST()
if to_static:
print("Successfully to apply @to_static.")
build_strategy = paddle.static.BuildStrategy()
# Why set `build_strategy.enable_inplace = False` here?
# Because we find that this PASS strategy of PE makes dy2st training loss unstable.
build_strategy.enable_inplace = False
mnist = paddle.jit.to_static(mnist, build_strategy=build_strategy)
optimizer = paddle.optimizer.Adam(learning_rate=0.001,
parameters=mnist.parameters())
scaler = paddle.amp.GradScaler(init_loss_scaling=1024)
mnist, optimizer = paddle.amp.decorate(models=mnist,
optimizers=optimizer,
level='O2',
save_dtype='float32')
loss_data = []
for epoch in range(self.epoch_num):
start = time()
for batch_id, data in enumerate(self.train_reader()):
dy_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(dy_x_data)
label = paddle.to_tensor(y_data)
label.stop_gradient = True
with paddle.amp.auto_cast(enable=True,
custom_white_list=None,
custom_black_list=None,
level='O2'):
prediction, acc, avg_loss = mnist(img, label=label)
scaled = scaler.scale(avg_loss)
scaled.backward()
scaler.minimize(optimizer, scaled)
loss_data.append(avg_loss.numpy()[0])
# save checkpoint
mnist.clear_gradients()
if batch_id % 2 == 0:
print(
"Loss at epoch {} step {}: loss: {:}, acc: {}, cost: {}"
.format(epoch, batch_id, avg_loss.numpy(), acc.numpy(),
time() - start))
start = time()
if batch_id == 10:
break
return loss_data
def train(self, to_static=False):
paddle.seed(SEED)
mnist = MNIST()
if to_static:
print("Successfully to apply @to_static.")
mnist = paddle.jit.to_static(mnist)
adam = AdamOptimizer(learning_rate=0.001,
parameter_list=mnist.parameters())
scaler = paddle.amp.GradScaler(init_loss_scaling=1024)
loss_data = []
for epoch in range(self.epoch_num):
start = time()
for batch_id, data in enumerate(self.train_reader()):
dy_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(dy_x_data)
label = paddle.to_tensor(y_data)
label.stop_gradient = True
with paddle.amp.auto_cast():
prediction, acc, avg_loss = mnist(img, label=label)
scaled = scaler.scale(avg_loss)
scaled.backward()
scaler.minimize(adam, scaled)
loss_data.append(avg_loss.numpy()[0])
# save checkpoint
mnist.clear_gradients()
if batch_id % 10 == 0:
print(
"Loss at epoch {} step {}: loss: {:}, acc: {}, cost: {}"
.format(epoch, batch_id, avg_loss.numpy(), acc.numpy(),
time() - start))
start = time()
if batch_id == 50:
break
return loss_data
class MnistPredictionServiceSubclass(
mnist_pb2_grpc.MnistPredictionServiceServicer):
def __init__(self):
self.mnist = MNIST()
def MnistPredict(self, request, context):
request_map = request.inputs
inputs = cv2.imdecode(np.frombuffer(request.inputs, dtype='uint8'), 1)
predict_result, predict_probability = self.mnist.interface(inputs)
logging.info("predict_result{}predict_probability{}".format(
predict_result, predict_probability))
response = mnist_pb2.MnistPredictResponse()
response.outputs = str(predict_result[0])
response.probability = predict_probability
return response
import web
import os, sys
import io
import base64
import numpy as np
import json
import cv2
sys.path.append("../train_test_mnist/")
from test_mnist import MNIST
mnist = MNIST()
urls = ('/predict', 'MNIST_SERVER')
class MNIST_SERVER():
def GET(self):
pass
def POST(self):
try:
form = web.input()
except ValueError as e:
return e
data = {"success": False}
image_file_value = base64.urlsafe_b64decode(form['image'])
image = cv2.imdecode(np.frombuffer(image_file_value, dtype='uint8'), 1)
predict_result, predict_probability = mnist.interface(image)
def __init__(self):
self.mnist = MNIST()