mole_analyzer / useful_functions.py
jeanpoll
initial commit
eee8e74
raw
history blame
2.02 kB
import mlflow
import pandas as pd
import joblib
import os
from fastai.data.block import MultiCategoryBlock, RandomSplitter, DataBlock, CategoryBlock
from fastai.vision.data import *
from fastai.vision.learner import *
from fastai.vision.all import *
import pathlib
plt = platform.system()
if plt == 'Linux': pathlib.WindowsPath = pathlib.PosixPath
lesion_model = None
pipe_cancer = None
def f_load_cnn_model():
global lesion_model
logged_model = os.path.join("mole_models", "ed52a28a7b504ff7ba851c850221d1dd")
lesion_model = mlflow.fastai.load_model(logged_model)
lesion_model.cbs.remove(lesion_model.cbs[6])
lesion_model.cbs.remove(lesion_model.cbs[4])
lesion_model.cbs.remove(lesion_model.cbs[3])
def get_image_files(df):
# df = df.assign(path=lambda x: working_dataset_path + x["path"])
return df
def get_x(df):
return df["path"]
def get_y(df):
return df["dx"]
def f_create_df_with_files_input(file_path):
list_files = [file_path]
list_labels = [""]
df = pd.DataFrame(data={"path": list_files, "dx": list_labels})
return df
def f_predict_cnn_with_tta(file_path):
df = f_create_df_with_files_input(file_path)
dl = lesion_model.dls.test_dl(df)
dl.after_item = Pipeline([ToTensor, Resize(700, method=ResizeMethod.Crop), RandomResizedCrop(350)])
pred, _targ = lesion_model.tta(dl=dl, n=4, use_max=False)
return pred.tolist()[0]
def f_predict_cnn_simple(file_path):
preds = lesion_model.predict(file_path)
return preds[2].tolist()
def f_load_cancer_classifier():
global pipe_cancer
pipe_cancer = joblib.load(os.path.join("mole_models", "pipeline.rf_classifier_mole"))
def f_predict_cancer(preds, age, sex, localization):
df = pd.DataFrame(data=[preds], columns=lesion_model.dls.vocab)
df["age"] = age
df["sex"] = sex
df["localization"] = localization
df["label"] = ""
label = pipe_cancer.predict(df)[0]
label = "Possibly suspicious" if label=="cancer" else "benign"
return label