Upload decisions_users.py

utils/decisions_users.py

@@ -0,0 +1,315 @@
+import os
+import sys 
+import json
+import shutil
+import urllib.request
+from pathlib import Path
+import pathlib
+import time
+import urllib
+from ast import literal_eval
+import albumentations as A
+import tensorflow as tf
+
+import cv2
+import numpy as np
+import pandas as pd
+import plotly.express as px
+import matplotlib.pyplot as plt
+from PIL import Image
+
+import streamlit as st
+
+import seaborn as sns
+
+sys.path.append(f'{os.getcwd()}/utils')
+from utils.eval_users import get_product_dev_page_layout
+
+# print(os.getcwd())
+
+# Hide GPU from visible devices
+tf.config.set_visible_devices([], 'GPU')
+# Enable GPU memory growth - avoid allocating all memory at start
+# gpus = tf.config.experimental.list_physical_devices(device_type='GPU')
+# for gpu in gpus:
+#     tf.config.experimental.set_memory_growth(device=gpu, enable=True)
+
+from utils.control import show_tsne_vis,show_random_samples
+
+from utils.annoy_sampling import load_annoy_tree
+
+
+from utils.model_utils import load_model
+from utils.model_utils import get_feature_vector, get_feature_extractor_model, get_predictions_and_roi
+
+sns.set_style('darkgrid')
+plt.rcParams['axes.grid'] = False
+
+# st.set_page_config(layout="wide")
+
+#https://github.com/IliaLarchenko/albumentations-demo/blob/3cb6528a513fe3b35dbb2c2a63cdcfbb9bb2a932/src/utils.py#L149
+
+GRAD_CAM_IMAGE_DIR = f'{os.getcwd()}/data/gradcam_vis_data/'
+TEST_CSV_FILE =   f'{os.getcwd()}/data/test_set_pred_prop.csv'
+
+annoy_tree_save_path =f'{os.getcwd()}/data/filtered_train_embedding/representative_samples_emb.annoy'
+test_emb_path = f'{os.getcwd()}/data/filtered_train_embedding/test_embeddings.npy'
+test_emb_id_path =f'{os.getcwd()}/data/filtered_train_embedding/test_ids.npy'
+
+train_emb_id_path =f'{os.getcwd()}/data/filtered_train_embedding/representative_train_ids.npy'
+
+repr_id_path =f'{os.getcwd()}/data/filtered_train_embedding/representative_train_ids.npy'
+borderline_id_path =f'{os.getcwd()}/data/filtered_train_embedding/borderline_train_ids.npy'
+
+MODEL_PATH = f'{os.getcwd()}/model/keras_model_0422/'
+
+ROOT_FIG_DIR = f'{os.getcwd()}/figures/'
+
+test_emb = np.load(test_emb_path)
+test_ids = np.load(test_emb_id_path)
+test_id_list = list(test_ids)
+test_labels = [_id.split("\\")[1] for _id in test_id_list]
+print(" NUmber of test samples: ",len(test_labels))
+test_features = test_emb.reshape(-1,1792)
+
+
+# train embedding list
+train_ids = np.load(train_emb_id_path)
+train_id_list = list(train_ids)
+train_labels = [_id.split("\\")[1] for _id in train_id_list]
+print(" NUmber of training samples: ",len(train_labels))
+
+annoy_tree = load_annoy_tree(test_features.shape[1],annoy_tree_save_path)
+def annoy_matching(annoy_f,query_item, query_index, n=10):
+            return annoy_f.get_nns_by_vector(query_item, n)
+
+def get_img(fn ,thumbnail=False):
+    img = Image.open(fn)
+    if thumbnail:
+        img.thumbnail((100,100))
+    return img
+
+
+def plot_n_similar(seed_id,similar_ids, test_path,n=10, scale=5):
+    f,ax = plt.subplots(1,n+1,figsize=((n+1)*scale,scale))
+    # print(os.path.basename(test_labels[seed_id])[:-4])
+    title = "SEED ID:{0}\nLabel:{1}".format(seed_id,os.path.basename(test_labels[seed_id]))
+    # print("path:", test_labels[seed_id].replace("F:/","E:/"))
+    ax[0].imshow(get_img(test_path.replace("F:/","E:/")),cmap='gray')
+    ax[0].set_title(title,fontsize=12)
+    for i in range(len(similar_ids)):
+        ax[i+1].imshow(get_img(similar_ids[i].replace("F:/","E:/")),cmap='gray')
+        title = "ID:{0}\nDistance: {1:.3f}\nLabel:{2}".format(i,0.1223,os.path.basename(similar_ids[i])[:-4])  
+        ax[i+1].set_title(title,fontsize=10)
+    f.suptitle("Images similar to seed_id {0}".format(seed_id),fontsize=18)
+    plt.subplots_adjust(top=0.5)
+    return f
+
+def load_image(filename,change_url=True):
+    # if change_url:
+    print(filename)
+    print(os.path.exists(filename))
+    img = cv2.imread(filename)
+    return img
+
+@st.cache(allow_output_mutation=True)
+def get_model(model_path):
+    new_model = tf.keras.models.load_model(model_path)
+    # keras_model = load_model(model_path)
+    return new_model
+
+@st.cache(allow_output_mutation=True)
+def get_feature_vector_model(model_path):
+    keras_model = tf.keras.models.load_model(model_path)
+    feature_extractor =  tf.keras.Model(keras_model.inputs,keras_model.layers[-3].output)
+    return feature_extractor
+
+def load_pd_data_frame(df_csv_path):
+    return pd.read_csv(df_csv_path)
+
+def get_path_list_from_df(df_data):
+    return list(df_data['path'])
+
+def get_class_probs_from_df(df_data):
+    return list(df_data['class_probs'])
+
+
+def visualize_bar_plot(df_data):
+    fig = px.bar(df_data, x="probability", y="class", orientation='h')
+    return fig
+
+def run_instance_exp(img_path, img_path_list,prob_list,grad_vis_path_list):
+
+    st.subheader('Instance Exploration')
+    # st.columns((1,1,1))  with row4_2:
+    LABELS = ['CNV', 'DRUSEN', 'DME', 'NORMAL']
+
+    left_column, middle_column, right_column = st.columns((1,1,1))
+    display_image = load_image(img_path)
+    # fig = px.imshow(display_image)
+    # left_column.plotly_chart(fig, use_container_width=True)
+    left_column.image(cv2.resize(display_image, (180,180)),caption = "Selected Input")
+
+    # get class probabilities
+    indx = img_path_list.index(img_path)
+    print(img_path)
+    prb_tmp = prob_list[indx]
+    print(f"{prb_tmp[1:-1]}")
+    clss_probs = literal_eval('"'+prb_tmp[1:-1]+'"')
+    print(clss_probs[1:-1].split(' '))
+    prob_cls = [float(p) for p in clss_probs[1:-1].split(' ')]
+    tmp_df = pd.DataFrame.from_dict({'class':LABELS,'probability':prob_cls})
+    print(tmp_df.head())
+    fig = plt.figure(figsize=(15, 13))
+    sns.barplot(x='probability', y='class', data=tmp_df)
+    middle_column.pyplot(fig)
+    # st.caption('Predictions')
+
+    tmp_grad_img = GRAD_CAM_IMAGE_DIR + img_path.split("\\")[-2] +'/'+img_path.split("\\") [-1] 
+    
+    display_image = load_image(tmp_grad_img,replace=False)
+    # left_column.plotly_chart(fig, use_container_width=True)
+    right_column.image(display_image,caption = "ROI")
+
+    # seed_id = 900
+    seed_id = test_id_list.index(img_path)
+    query_item = test_features[seed_id]
+    print(query_item.shape)
+    closest_idxs = annoy_matching(annoy_tree,query_item, seed_id, 10)
+    closest_fns = [train_ids[close_i] for close_i in closest_idxs]
+    st.subheader('Top-10 Similar Samples from Gallery Set')
+    st.pyplot(plot_n_similar(seed_id,closest_fns, img_path,n=10, scale=4))
+
+
+
+def run_instance_exp_keras_model(img_path, new_model, feature_extractor_model):
+
+    st.subheader('Instance Exploration')
+    # st.columns((1,1,1))  with row4_2:
+    LABELS = ['CNV', 'DRUSEN', 'DME', 'NORMAL']
+
+    left_column, middle_column, right_column = st.columns((1,1,1))
+    print(img_path)
+    org_img_path = img_path
+
+    img_path = f'{os.getcwd()}/data/oct2017/test/' + img_path.split("\\")[-2] +'/'+img_path.split("\\") [-1]
+    # img_path.replace("F:/XAI/data/OCT2017/","/home/hodor/dev/Learning/XAI/streamlit_demo/multipage-app/data/xai_framework_data/")
+    display_image = load_image(img_path)
+    # fig = px.imshow(display_image)
+    # left_column.plotly_chart(fig, use_container_width=True)
+    left_column.image(cv2.resize(display_image, (180,180)),caption = "Selected Input")
+   
+
+    roi_img, probs = get_predictions_and_roi(img_path, new_model)
+
+    ## probs
+    # print(np.asarray(probs))
+    # print(probs.shape)
+    prob_cls =np.asarray(probs)[0]
+    # print(prob_cls)
+    tmp_df = pd.DataFrame.from_dict({'class':LABELS,'probability':prob_cls})
+    fig = plt.figure(figsize=(5, 4))
+    sns.barplot(x='probability', y='class', data=tmp_df)
+    middle_column.pyplot(fig)
+    # middle_column.write("Probabilities")
+
+    # grad img 
+    right_column.image(roi_img, caption = "Decision ROI")
+
+    # seed_id = 900
+    seed_id = test_id_list.index(org_img_path)
+    query_item = get_feature_vector(img_path,feature_extractor_model)
+    query_item = query_item.reshape(-1,1792)
+    # print(query_item.shape)
+    closest_idxs = annoy_matching(annoy_tree,query_item[0,:], seed_id, 10)
+    closest_fns = [train_ids[close_i] for close_i in closest_idxs]
+
+    closest_fns_tmp = [f'{os.getcwd()}/data/oct2017/train/' + each_fn.split("\\")[-2] +'/'+each_fn.split("\\") [-1] 
+                for each_fn in closest_fns]
+    # print(closest_fns)
+    st.subheader('Top-10 Similar Samples from Gallery Set')
+    st.pyplot(plot_n_similar(seed_id,closest_fns_tmp, img_path,n=10,scale=3))
+
+def main():
+
+    new_model =  get_model(MODEL_PATH)
+    feature_extractor_model = get_feature_vector_model(MODEL_PATH)
+
+    row4_1, row4_2 = st.tabs(["Global Level Explanations", "Instance Level Explanations"])
+
+    with row4_1:
+        borderline_cases = np.load(borderline_id_path)
+        representative_cases = np.load(repr_id_path)
+        borderline_id_list = list(borderline_cases)
+        # print(borderline_id_list)
+
+        borderline_id_list = [f'{os.getcwd()}/data/oct2017/train/' + each_fn.split("\\")[-2] +'/'+each_fn.split("\\") [-1] 
+                for each_fn in borderline_id_list]
+        representative_id_list = list(representative_cases)
+        representative_id_list = [f'{os.getcwd()}/data/oct2017/train/' + each_fn.split("\\")[-2] +'/'+each_fn.split("\\") [-1] 
+                for each_fn in representative_id_list]
+        # st.info('GLOABAL EXPLANATION!! ')
+        option = st.selectbox('Please select to explore Representative or Borderline Samples', ["Representative Samples","Borderline Cases"])
+        if option:
+            clss = st.selectbox('Select a category(class)', ["CNV","DME", "NORMAL", "DRUSEN"])
+            side_1, side_2 = st.columns(2)
+
+        with side_1:
+            check_emb = st.checkbox('Embdedding Space Visuzalization')
+
+        with side_2:
+            check_samp = st.checkbox('Random Sample Visuzalization')
+
+        if check_emb and check_samp:
+            st.write("Emb and vis")
+            if option.startswith("Rep"):
+                filter_lst = list(filter(lambda k: clss in k, representative_id_list))
+                show_random_samples(filter_lst,clss)
+                show_tsne_vis(f"{ROOT_FIG_DIR}/tsne_representative.png", title="Representative")
+            else:
+                filter_lst = list(filter(lambda k: clss in k, borderline_id_list))
+                show_random_samples(filter_lst,clss)
+                show_tsne_vis(f"{ROOT_FIG_DIR}/tsne_borderline.png", title="Borderline")
+        elif check_emb:
+            st.write("embedding vis")
+            if option.startswith("Rep"):
+                show_tsne_vis(f"{ROOT_FIG_DIR}/tsne_representative.png", title="Representative")
+            else:
+                show_tsne_vis(f"{ROOT_FIG_DIR}/tsne_borderline.png", title="Borderline")
+        elif check_samp:
+            st.write("rand vis")
+            if option.startswith("Rep"):
+                filter_lst = list(filter(lambda k: clss in k, representative_id_list))
+                show_random_samples(filter_lst,clss)
+            else:
+                filter_lst = list(filter(lambda k: clss in k, borderline_id_list))
+                show_random_samples(filter_lst,clss)
+    with row4_2:
+        DF_TEST_PROP = load_pd_data_frame(TEST_CSV_FILE)
+        IMG_PATH_LISTS = get_path_list_from_df(DF_TEST_PROP)
+        IMG_CLSS_PROBS_LIST = get_class_probs_from_df(DF_TEST_PROP)
+        grad_vis_path_list = None
+        row2_col1, row2_col2 = st.columns(2)
+        with row2_col1:
+            option = st.selectbox('Please select a sample image, then click Explain Me button', IMG_PATH_LISTS)
+        with row2_col2:
+            st.info("Press the button")
+            pressed = st.button('Explain ME')
+        
+        if pressed:
+            st.empty()
+            st.sidebar.write('Please wait for the magic to happen! This may take up to a minute.')
+            run_instance_exp_keras_model(option, new_model,feature_extractor_model)
+
+
+        # # new_model = load_model(MODEL_PATH)
+        # option = st.sidebar.selectbox('Please select a sample image, then click Explain Me button', IMG_PATH_LISTS)
+        # pressed = st.sidebar.button('Explain ME')
+
+# main()
+# expander_faq = st.expander("More About Our Project")
+# expander_faq.write("Hi there! If you have any questions about our project, or simply want to check out the source code, please visit our github repo: https://github.com/kaplansinan/MLOPS")
+
+
+def get_product_dev_page_layout():
+    return main()