def predict(row):
df = pd.DataFrame.from_dict([row], orient='columns')
#open model RF
with open('website/model.pkl', 'rb') as f:
model = pickle.load(f)
df1 = transform_test(df)
#predict probability using model
prediction = model.predict_proba(df1.values.reshape(1, -1))[0][1]
row['prediction'] = prediction
#determine contribution of features to prediction using treeinterpreter
prediction, bias, contributions = ti.predict(model,
df1.values.reshape(1, -1))
#empty list to hold important features which contributed to prediction
important_features = []
#names for features
column_features = [
'name_length', 'num_payouts', 'user_age', 'org_facebook',
'org_twitter', 'body_length', 'gts', 'sale_duration', 'tickets_sold'
]
#take the top 3 features which had the highest contribution
for feature, key in sorted(
zip(abs(contributions[0][:, 1]), column_features))[::-1][:3]:
important_features.append(key)
row['contributions'] = important_features
#returns the prob of fraud which we can later classify using the threshold chosen
return row
def predict(row):
df = pd.DataFrame.from_dict([row], orient='columns')
#open model RF
with open('website/model.pkl', 'rb') as f:
model = pickle.load(f)
df1 = transform_test(df)
#predict probability using model
prediction = model.predict_proba(df1.values.reshape(1, -1))[0][1]
row['prediction'] = prediction
#returns the prob of fraud which we can later classify using the threshold chosen
return row
return src, tgt, office, visda, noe
src, tgt, office, visda, noe = get_datasetname(args)
batch_size = {"train": 36, "val": 36, "test": 4}
for i in range(10):
batch_size["val" + str(i)] = 4
if visda == False:
data_transforms = {
'train': tran.transform_train(resize_size=28, crop_size=28),
'val': tran.transform_train(resize_size=28, crop_size=28),
}
data_transforms = tran.transform_test(data_transforms=data_transforms,
resize_size=28,
crop_size=28)
dsets = {
"train":
ImageList(open(src).readlines(), transform=data_transforms["train"]),
"val":
ImageList(open(tgt).readlines(), transform=data_transforms["val"]),
"test":
ImageList(open(tgt).readlines(), transform=data_transforms["val"])
}
dset_loaders = {
x: torch.utils.data.DataLoader(dsets[x],
batch_size=batch_size[x],
shuffle=True,
num_workers=4)
for x in ['train', 'val']
metavar='S',
help='method:l2 or l2+bss')
parser.add_argument('--lr',
type=float,
default=0.01,
help='init learning rate')
args = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
data_transforms = {
'train': trans.transform_train(resize_size=256, crop_size=224),
'val': trans.transform_train(resize_size=256, crop_size=224),
}
data_transforms = trans.transform_test(data_transforms=data_transforms,
resize_size=256,
crop_size=224)
# set dataset
batch_size = {"train": 48, "val": 100, "test": 100}
for i in range(10):
batch_size["val" + str(i)] = 4
trainpath = args.trainpath
testpath = args.testpath
dsets = {
"train":
datasets.ImageFolder(root=trainpath, transform=data_transforms["train"]),
"val":
datasets.ImageFolder(root=testpath, transform=data_transforms["val"]),
global precisions
precision = getPrecision(self.max_depth, self.min_impurity_decrease)
ax.scatter(self.max_depth, self.min_impurity_decrease, precision)
precisions[self.i] = precision
print(self.max_depth, "; ", self.min_impurity_decrease, " done")
input, output, input_names, output_names, column_ignored = csv_parsing.parse(
"census-income.names", "census-income.data", {
"ignore": [24],
"ignoreColumnThresold": 0.1
})
transformer, formated_input = transform.transform(input, input_names)
test, test_output = csv_test.parse_test("census-income.test", output_names,
column_ignored)
formated_test = transform.transform_test(test, input_names, transformer)
print("Initialization done")
fig = plt.figure()
ax = fig.add_subplot(111, projection='3d')
i = 0
for max_depth in np.arange(1, 10):
for min_impurity_decrease in np.arange(0, 0.0001, 0.00001):
precisions.append(0)
threads.append(Computation(i, max_depth, min_impurity_decrease))
threads[-1].start()
i += 1
if (i % 4 == 0):
for n in range(i - 4, i):