Skip to content

Commit 0cdf27a

Browse files
authored
Merge pull request #1253 from apache/dev-postgresql
Merge Dev branch into master
2 parents 892425d + eed23ab commit 0cdf27a

File tree

15 files changed

+1158
-133
lines changed

15 files changed

+1158
-133
lines changed
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,45 @@
1+
<!--
2+
Licensed to the Apache Software Foundation (ASF) under one
3+
or more contributor license agreements. See the NOTICE file
4+
distributed with this work for additional information
5+
regarding copyright ownership. The ASF licenses this file
6+
to you under the Apache License, Version 2.0 (the
7+
"License"); you may not use this file except in compliance
8+
with the License. You may obtain a copy of the License at
9+
10+
http://www.apache.org/licenses/LICENSE-2.0
11+
12+
Unless required by applicable law or agreed to in writing,
13+
software distributed under the License is distributed on an
14+
"AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
15+
KIND, either express or implied. See the License for the
16+
specific language governing permissions and limitations
17+
under the License.
18+
-->
19+
20+
# Singa for Diabetic Readmission Prediction task
21+
22+
## Diabetic Readmission
23+
24+
Diabetic readmission is a significant concern in healthcare, with a substantial number of patients being readmitted to the hospital within a short period after discharge. This not only leads to increased healthcare costs but also poses a risk to patient well-being.
25+
26+
Although diabetes is a manageable condition, early identification of patients at high risk of readmission remains a challenge. A reliable and efficient predictive model can help identify these patients, enabling healthcare providers to intervene early and prevent unnecessary readmissions.
27+
28+
To address this issue, we use Singa to implement a machine learning model for predicting diabetic readmission. The dataset is from [BMC Medical Informatics and Decision-Making](https://bmcmedinformdecismak.biomedcentral.com/articles/10.1186/s12911-021-01423-y). Please download the dataset before running the scripts.
29+
30+
31+
## Structure
32+
33+
* `data` includes the scripts for preprocessing Diabetic Readmission datasets.
34+
35+
* `model` includes the MLP model construction codes by creating
36+
a subclass of `Module` to wrap the neural network operations
37+
of each model.
38+
39+
* `train_mlp.py` is the training script, which controls the training flow by
40+
doing BackPropagation and SGD update.
41+
42+
## Command
43+
```bash
44+
python train.py mlp diabetic
45+
```
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,51 @@
1+
<!--
2+
Licensed to the Apache Software Foundation (ASF) under one
3+
or more contributor license agreements. See the NOTICE file
4+
distributed with this work for additional information
5+
regarding copyright ownership. The ASF licenses this file
6+
to you under the Apache License, Version 2.0 (the
7+
"License"); you may not use this file except in compliance
8+
with the License. You may obtain a copy of the License at
9+
10+
http://www.apache.org/licenses/LICENSE-2.0
11+
12+
Unless required by applicable law or agreed to in writing,
13+
software distributed under the License is distributed on an
14+
"AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
15+
KIND, either express or implied. See the License for the
16+
specific language governing permissions and limitations
17+
under the License.
18+
-->
19+
20+
# Singa for Diabetic Retinopathy Classification
21+
22+
## Diabetic Retinopathy
23+
24+
Diabetic Retinopathy (DR) is a progressive eye disease caused by long-term diabetes, which damages the blood vessels in the retina, the light-sensitive tissue at the back of the eye. It typically develops in stages, starting with non-proliferative diabetic retinopathy (NPDR), where weakened blood vessels leak fluid or blood, causing swelling or the formation of deposits. If untreated, it can progress to proliferative diabetic retinopathy (PDR), characterized by the growth of abnormal blood vessels that can lead to severe vision loss or blindness. Symptoms may include blurred vision, dark spots, or difficulty seeing at night, although it is often asymptomatic in the early stages. Early diagnosis through regular eye exams and timely treatment, such as laser therapy or anti-VEGF injections, can help manage the condition and prevent vision impairment.
25+
26+
The dataset has 5 groups characterized by the severity of Diabetic Retinopathy (DR).
27+
28+
- 0: No DR
29+
- 1: Mild Non-Proliferative DR
30+
- 2: Moderate Non-Proliferative DR
31+
- 3: Severe Non-Proliferative DR
32+
- 4: Proliferative DR
33+
34+
35+
To mitigate the problem, we use Singa to implement a machine learning model to help with Diabetic Retinopathy diagnosis. The dataset is from Kaggle https://www.kaggle.com/datasets/mohammadasimbluemoon/diabeticretinopathy-messidor-eyepac-preprocessed. Please download the dataset before running the scripts.
36+
37+
## Structure
38+
39+
* `data` includes the scripts for preprocessing DR image datasets.
40+
41+
* `model` includes the CNN model construction codes by creating
42+
a subclass of `Module` to wrap the neural network operations
43+
of each model.
44+
45+
* `train_cnn.py` is the training script, which controls the training flow by
46+
doing BackPropagation and SGD update.
47+
48+
## Command
49+
```bash
50+
python train_cnn.py cnn diaret -dir pathToDataset
51+
```
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,297 @@
1+
from singa import singa_wrap as singa
2+
from singa import device
3+
from singa import tensor
4+
from singa import opt
5+
import numpy as np
6+
import time
7+
import argparse
8+
import sys
9+
sys.path.append("../../..")
10+
11+
from PIL import Image
12+
13+
from healthcare.data import diaret
14+
from healthcare.models import diabetic_retinopthy_net
15+
16+
np_dtype = {"float16": np.float16, "float32": np.float32}
17+
18+
singa_dtype = {"float16": tensor.float16, "float32": tensor.float32}
19+
20+
21+
# Data augmentation
22+
def augmentation(x, batch_size):
23+
xpad = np.pad(x, [[0, 0], [0, 0], [4, 4], [4, 4]], 'symmetric')
24+
for data_num in range(0, batch_size):
25+
offset = np.random.randint(8, size=2)
26+
x[data_num, :, :, :] = xpad[data_num, :,
27+
offset[0]:offset[0] + x.shape[2],
28+
offset[1]:offset[1] + x.shape[2]]
29+
if_flip = np.random.randint(2)
30+
if (if_flip):
31+
x[data_num, :, :, :] = x[data_num, :, :, ::-1]
32+
return x
33+
34+
35+
# Calculate accuracy
36+
def accuracy(pred, target):
37+
# y is network output to be compared with ground truth (int)
38+
y = np.argmax(pred, axis=1)
39+
a = y == target
40+
correct = np.array(a, "int").sum()
41+
return correct
42+
43+
44+
# Data partition according to the rank
45+
def partition(global_rank, world_size, train_x, train_y, val_x, val_y):
46+
# Partition training data
47+
data_per_rank = train_x.shape[0] // world_size
48+
idx_start = global_rank * data_per_rank
49+
idx_end = (global_rank + 1) * data_per_rank
50+
train_x = train_x[idx_start:idx_end]
51+
train_y = train_y[idx_start:idx_end]
52+
53+
# Partition evaluation data
54+
data_per_rank = val_x.shape[0] // world_size
55+
idx_start = global_rank * data_per_rank
56+
idx_end = (global_rank + 1) * data_per_rank
57+
val_x = val_x[idx_start:idx_end]
58+
val_y = val_y[idx_start:idx_end]
59+
return train_x, train_y, val_x, val_y
60+
61+
62+
# Function to all reduce NUMPY accuracy and loss from multiple devices
63+
def reduce_variable(variable, dist_opt, reducer):
64+
reducer.copy_from_numpy(variable)
65+
dist_opt.all_reduce(reducer.data)
66+
dist_opt.wait()
67+
output = tensor.to_numpy(reducer)
68+
return output
69+
70+
71+
def resize_dataset(x, image_size):
72+
num_data = x.shape[0]
73+
dim = x.shape[1]
74+
X = np.zeros(shape=(num_data, dim, image_size, image_size),
75+
dtype=np.float32)
76+
for n in range(0, num_data):
77+
for d in range(0, dim):
78+
X[n, d, :, :] = np.array(Image.fromarray(x[n, d, :, :]).resize(
79+
(image_size, image_size), Image.BILINEAR),
80+
dtype=np.float32)
81+
return X
82+
83+
84+
def run(global_rank,
85+
world_size,
86+
dir_path,
87+
max_epoch,
88+
batch_size,
89+
model,
90+
data,
91+
sgd,
92+
graph,
93+
verbosity,
94+
dist_option='plain',
95+
spars=None,
96+
precision='float32'):
97+
# now CPU version only, could change to GPU device for GPU-support machines
98+
dev = device.get_default_device()
99+
dev.SetRandSeed(0)
100+
np.random.seed(0)
101+
if data == 'diaret':
102+
train_x, train_y, val_x, val_y = diaret.load(dir_path=dir_path)
103+
else:
104+
print(
105+
'Wrong dataset!'
106+
)
107+
sys.exit(0)
108+
109+
num_channels = train_x.shape[1]
110+
image_size = train_x.shape[2]
111+
data_size = np.prod(train_x.shape[1:train_x.ndim]).item()
112+
num_classes = (np.max(train_y) + 1).item()
113+
114+
if model == 'cnn':
115+
model = diabetic_retinopthy_net.create_model(num_channels=num_channels,
116+
num_classes=num_classes)
117+
else:
118+
print(
119+
'Wrong model!'
120+
)
121+
sys.exit(0)
122+
123+
# For distributed training, sequential has better performance
124+
if hasattr(sgd, "communicator"):
125+
DIST = True
126+
sequential = True
127+
else:
128+
DIST = False
129+
sequential = False
130+
131+
if DIST:
132+
train_x, train_y, val_x, val_y = partition(global_rank, world_size,
133+
train_x, train_y, val_x,
134+
val_y)
135+
136+
if model.dimension == 4:
137+
tx = tensor.Tensor(
138+
(batch_size, num_channels, model.input_size, model.input_size), dev,
139+
singa_dtype[precision])
140+
elif model.dimension == 2:
141+
tx = tensor.Tensor((batch_size, data_size),
142+
dev, singa_dtype[precision])
143+
np.reshape(train_x, (train_x.shape[0], -1))
144+
np.reshape(val_x, (val_x.shape[0], -1))
145+
146+
ty = tensor.Tensor((batch_size,), dev, tensor.int32)
147+
num_train_batch = train_x.shape[0] // batch_size
148+
num_val_batch = val_x.shape[0] // batch_size
149+
idx = np.arange(train_x.shape[0], dtype=np.int32)
150+
151+
# Attach model to graph
152+
model.set_optimizer(sgd)
153+
model.compile([tx], is_train=True, use_graph=graph, sequential=sequential)
154+
dev.SetVerbosity(verbosity)
155+
156+
# Training and evaluation loop
157+
for epoch in range(max_epoch):
158+
start_time = time.time()
159+
np.random.shuffle(idx)
160+
161+
if global_rank == 0:
162+
print('Starting Epoch %d:' % (epoch))
163+
164+
# Training phase
165+
train_correct = np.zeros(shape=[1], dtype=np.float32)
166+
test_correct = np.zeros(shape=[1], dtype=np.float32)
167+
train_loss = np.zeros(shape=[1], dtype=np.float32)
168+
169+
model.train()
170+
for b in range(num_train_batch):
171+
# if b % 100 == 0:
172+
# print ("b: \n", b)
173+
# Generate the patch data in this iteration
174+
x = train_x[idx[b * batch_size:(b + 1) * batch_size]]
175+
if model.dimension == 4:
176+
x = augmentation(x, batch_size)
177+
if (image_size != model.input_size):
178+
x = resize_dataset(x, model.input_size)
179+
x = x.astype(np_dtype[precision])
180+
y = train_y[idx[b * batch_size:(b + 1) * batch_size]]
181+
182+
# Copy the patch data into input tensors
183+
tx.copy_from_numpy(x)
184+
ty.copy_from_numpy(y)
185+
186+
# Train the model
187+
out, loss = model(tx, ty, dist_option, spars)
188+
train_correct += accuracy(tensor.to_numpy(out), y)
189+
train_loss += tensor.to_numpy(loss)[0]
190+
191+
if DIST:
192+
# Reduce the evaluation accuracy and loss from multiple devices
193+
reducer = tensor.Tensor((1,), dev, tensor.float32)
194+
train_correct = reduce_variable(train_correct, sgd, reducer)
195+
train_loss = reduce_variable(train_loss, sgd, reducer)
196+
197+
if global_rank == 0:
198+
print('Training loss = %f, training accuracy = %f' %
199+
(train_loss, train_correct /
200+
(num_train_batch * batch_size * world_size)),
201+
flush=True)
202+
203+
# Evaluation phase
204+
model.eval()
205+
for b in range(num_val_batch):
206+
x = val_x[b * batch_size:(b + 1) * batch_size]
207+
if model.dimension == 4:
208+
if (image_size != model.input_size):
209+
x = resize_dataset(x, model.input_size)
210+
x = x.astype(np_dtype[precision])
211+
y = val_y[b * batch_size:(b + 1) * batch_size]
212+
tx.copy_from_numpy(x)
213+
ty.copy_from_numpy(y)
214+
out_test = model(tx)
215+
test_correct += accuracy(tensor.to_numpy(out_test), y)
216+
217+
if DIST:
218+
# Reduce the evaulation accuracy from multiple devices
219+
test_correct = reduce_variable(test_correct, sgd, reducer)
220+
221+
# Output the evaluation accuracy
222+
if global_rank == 0:
223+
print('Evaluation accuracy = %f, Elapsed Time = %fs' %
224+
(test_correct / (num_val_batch * batch_size * world_size),
225+
time.time() - start_time),
226+
flush=True)
227+
228+
dev.PrintTimeProfiling()
229+
230+
231+
if __name__ == '__main__':
232+
# Use argparse to get command config: max_epoch, model, data, etc., for single gpu training
233+
parser = argparse.ArgumentParser(
234+
description='Training using the autograd and graph.')
235+
parser.add_argument(
236+
'model',
237+
choices=['cnn'],
238+
default='cnn')
239+
parser.add_argument('data',
240+
choices=['diaret'],
241+
default='diaret')
242+
parser.add_argument('-p',
243+
choices=['float32', 'float16'],
244+
default='float32',
245+
dest='precision')
246+
parser.add_argument('-dir',
247+
'--dir-path',
248+
default="/tmp/diaret",
249+
type=str,
250+
help='the directory to store the Diabetic Retinopathy dataset',
251+
dest='dir_path')
252+
parser.add_argument('-m',
253+
'--max-epoch',
254+
default=300,
255+
type=int,
256+
help='maximum epochs',
257+
dest='max_epoch')
258+
parser.add_argument('-b',
259+
'--batch-size',
260+
default=64,
261+
type=int,
262+
help='batch size',
263+
dest='batch_size')
264+
parser.add_argument('-l',
265+
'--learning-rate',
266+
default=0.005,
267+
type=float,
268+
help='initial learning rate',
269+
dest='lr')
270+
parser.add_argument('-g',
271+
'--disable-graph',
272+
default='True',
273+
action='store_false',
274+
help='disable graph',
275+
dest='graph')
276+
parser.add_argument('-v',
277+
'--log-verbosity',
278+
default=0,
279+
type=int,
280+
help='logging verbosity',
281+
dest='verbosity')
282+
283+
args = parser.parse_args()
284+
285+
sgd = opt.SGD(lr=args.lr, momentum=0.9, weight_decay=1e-5,
286+
dtype=singa_dtype[args.precision])
287+
run(0,
288+
1,
289+
args.dir_path,
290+
args.max_epoch,
291+
args.batch_size,
292+
args.model,
293+
args.data,
294+
sgd,
295+
args.graph,
296+
args.verbosity,
297+
precision=args.precision)

0 commit comments

Comments
 (0)