main_plots.py

# -*- coding: utf-8 -*-
"""
Created on Sun Jan  1 23:33:56 2023

@author: Moha-Cate
"""

import numpy as np

import os

import matplotlib.pyplot as plt
import seaborn as sns
import pandas as pd

sns.set_style("whitegrid")


# print('################# Notice:')
# print('################# Notice:')
# print('################# Notice:')
# print('seaborn version: ', sns.__version__)
# print('if the colors in legends are mismatched, make sure to have seaborn verion 0.11.0')
# print('################# Notice end')


        
df_summary = pd.read_csv('csv_results_summary.csv', low_memory=False)        

df_summary = df_summary.drop(columns=['severity', 'perturb_type'])

try:
    os.mkdir('./figures')
except:
    s=1 
        
#%%        
        
df_here = df_summary.copy()

df_here=df_here.loc[df_here['class']==1]

# df_here_indomain = df_here.loc[df_here['InDomain']==1]
# df_here_outdomain = df_here.loc[df_here['InDomain']==0]
# df_here_indomain_orig = df_here.loc[df_here['Severity']==0]
# df_here_outdomain_orig = df_here.loc[df_here['Severity']==0]

df_here_orig = df_here.loc[df_here['Severity']==0].loc[pd.isna(df_here['Perturbation Type'])]
df_here_d1 = df_here.loc[df_here['Severity']==0].loc[pd.notna(df_here['Perturbation Type'])]
df_here_d2 = df_here.loc[df_here['Severity'] > 0]

######################
x="Segmentation Model"
y="jacc"
# hue="Method"
# hue_order = ['MWen','no-change'] #hue_order=hue_order,  hue=hue,
#fig, axes = plt.subplots(1, 2, figsize=(15, 5), sharey=True)
plt.figure(1, figsize=(4, 5), dpi=300)
ax_orig = sns.boxplot(x=x, y=y, data=df_here_orig, 
                 showfliers = False,
                 showmeans=True,
                 meanprops = {"marker": "x", "markerfacecolor": "white", 
                              "markeredgecolor": "white"},
                palette=sns.color_palette("husl")) #  flierprops = dict(markerfacecolor = '0.50', markersize = 1)
ax_orig.set(ylim=(-.05, 1))
# plt.legend(bbox_to_anchor=(0.16, 1), loc="lower left", borderaxespad=0. , ncol=3)
#ax_in_orig.set(xticklabels=[])
ax_orig.set(ylabel= 'IOU')
ax_orig.text(0.45, 1.1, "Without Perturbation", color='black', fontsize=12)
# ax_orig.set(title="Without Perturbation")
plt.savefig('./figures/IOU_orig.png', dpi=300, bbox_inches = 'tight')


######################
x="Segmentation Model"
y="jacc"
col_category = 'Perturbation Type'

df_here_d1.loc[df_here_d1['Perturbation Type']=='11', 'Perturbation Type'] = 'De-illumination + De-spot'
df_here_d1.loc[df_here_d1['Perturbation Type']=='10', 'Perturbation Type'] = 'De-spot'
df_here_d1.loc[df_here_d1['Perturbation Type']=='1', 'Perturbation Type'] = 'De-illumination'

plt.figure(3, figsize=(4, 5), dpi=300)
ax3 = sns.catplot(x=x, y=y,
                col = col_category, showmeans=True,
                data=df_here_d1, kind="box",
                showfliers = False, height=4, aspect=1 ,           
                 meanprops = {"marker": "x", "markerfacecolor": "white", 
                              "markeredgecolor": "white"},
                palette=sns.color_palette("husl"))
#ax3.set(ylim=(-.5, 10))
ax3.set_titles("{col_name}")
ax3.set(ylabel= 'IOU')
# ax3.fig.suptitle("Perturbation Types", fontweight='bold', fontsize=11)
ax3.fig.text(0.45, 1, "Perturbation Types", color='black', fontsize=12)
# plt.tight_layout()
plt.savefig('./figures/IOU_d1.png', dpi=300, bbox_inches = 'tight')


######################
x="Segmentation Model"
y="jacc"
hue="Severity"
hue_order = [1,2,3,4,5]
col_category = 'Perturbation Type'

for label in df_here_d2['Perturbation Type'].unique().tolist():
    df_here_d2.loc[df_here_d2['Perturbation Type']==label, 'Perturbation Type'] = label.replace('_', ' ').title()

plt.figure(3, figsize=(4, 5), dpi=300)
ax3 = sns.catplot(x=x, y=y,
                hue=hue, col = col_category, showmeans=True, col_wrap=4,
                data=df_here_d2, kind="box", hue_order=hue_order, 
                showfliers = False, height=4, aspect=1 ,           
                 meanprops = {"marker": "x", "markerfacecolor": "white", 
                              "markeredgecolor": "white"},
                palette=sns.color_palette("husl")
                )
#ax3.set(ylim=(-.5, 10))
ax3.set_titles("{col_name}")
ax3.set(ylabel= 'IOU')
# ax3.fig.suptitle("Perturbation Types", fontweight='bold', fontsize=11)
ax3.fig.text(0.45, 1, "Perturbation Types", color='black', fontsize=12)
# plt.tight_layout()
plt.savefig('./figures/IOU_d2.png', dpi=300, bbox_inches = 'tight')

CREATE_INDIV = False

if CREATE_INDIV:
    for label in df_here_d2['Perturbation Type'].unique().tolist():
        plt.figure(3, figsize=(4, 5), dpi=300)
        ax3 = sns.catplot(x=x, y=y,
                        hue=hue, col = col_category, showmeans=True,
                        data=df_here_d2.loc[df_here_d2['Perturbation Type']==label], kind="box", hue_order=hue_order, 
                        showfliers = False, height=4, aspect=1 ,           
                        meanprops = {"marker": "x", "markerfacecolor": "white", 
                                    "markeredgecolor": "white"},
                        palette=sns.color_palette("husl")
                        )
        #ax3.set(ylim=(-.5, 10))
        ax3.set_titles("{col_name}")
        ax3.set(ylabel= 'IOU')
        # ax3.fig.suptitle("Perturbation Types", fontweight='bold', fontsize=11)
        ax3.fig.text(0.45, 1, "Perturbation Types", color='black', fontsize=12)
        # plt.tight_layout()
        flabel = label.replace(' ', '_')
        plt.savefig(f'./figures/d2_figs/IOU_{flabel}_d2.png', dpi=300, bbox_inches = 'tight')


    
################


plt.close('all')
#%% find the images to show

img_size = 256
binary = False

import torch
from models.cenet import CENet
from models.deeplabv3plus import DeepLabV3Plus
from models.unetpp import UnetPlusPlus

from utils.data_utils import * 

img_size = (img_size, img_size)

torch_models = ['cenet']
polar_models = ['mnet']

list_models = ['UNet++', 'DeepLabV3+', 'CE-Net']
list_model_names = ['unetpp', 'deeplabv3plus', 'cenet']
list_saved_models = ['../models/unetpp/unetpp_model.h5', 
                     '../models/deeplabv3plus/deeplabv3plus_model.h5',
                     '../models/cenet/cenet_model.pth']
thresholds_jacc = [0.35, 0.35, 0]
prefer = ['contrast', 'Hue_minus', 'Hue_minus']

fig_jacc, axs_jacc = plt.subplots(len(list_models), 5, figsize=(7.5, 7.5), dpi=300)

repeat = ()

for counter_model in range(len(list_models)):
    model_name=list_model_names[counter_model]
    df_here_method = df_summary.loc[df_summary['Segmentation Model'] == list_models[counter_model]]
    flag_done_thisMethod_jacc = False
    
    model = {
        'cenet': CENet,
        'deeplabv3plus': DeepLabV3Plus,
        'unetpp': UnetPlusPlus
    }[model_name]((img_size[0],img_size[1],3), 2 if binary else 3) # only important for unet models, SOTA models have their own size/n_channels and this will be disregarded
    
    if model_name not in torch_models:
        model.load(list_saved_models[counter_model])
    else:
        model.load(list_saved_models[counter_model], cpu=True)

    maskIDs = df_here_method.maskID.unique() if counter_model != 2 else [repeat[0]]
    for maskID in maskIDs:
        if flag_done_thisMethod_jacc:
            continue
        else:
            rec = df_here_method.loc[df_here_method.maskID==maskID]
            # print(rec[['maskID', 'Perturbation Type', 'jacc', 'class']])
            rec_noPert = rec.loc[rec['Perturbation Type'].isnull()]
            # print(rec_noPert[['maskID', 'Perturbation Type', 'jacc', 'class']])
            jacc_origin = rec_noPert.jacc.to_numpy()
            # rec_Pert = rec.loc[(rec['Perturbation Type'].notnull())] #  & (rec['Severity'] == 5)
            rec_Pert = rec.loc[(rec['Perturbation Type'] == prefer[counter_model])] if counter_model != 2 else repeat[1]
            # if counter_model > 0: print(rec_Pert)
            # print(rec_Pert[['maskID', 'Perturbation Type', 'jacc', 'class']])
            # idx = rec_Pert.index if counter_model != 2 else rec.loc[(rec['Perturbation Type'] == repeat[1])].loc[(rec['Severity'] == repeat[2])].index
            for IND in rec_Pert.index:
                # print(jacc_origin.shape, rec_Pert.jacc[IND])
                if counter_model==2 or jacc_origin[IND%3]-rec_Pert.jacc[IND] > thresholds_jacc[counter_model]:
                    img_orig, mask_gt_cat = process_pair(rec_noPert.FullFileName.iloc[0], rec_noPert.MaskFileName.iloc[0], img_size, binary=binary, polar=(model_name in polar_models), channelsFirst=(model_name in torch_models))
                    img_perturbed, _ = process_pair(rec_Pert.FullFileName[IND], rec_noPert.MaskFileName.iloc[0], img_size, binary=binary, polar=(model_name in polar_models), channelsFirst=(model_name in torch_models))
                    tensor = np.expand_dims(img_orig, axis=0)
                    tensor_perturbed = np.expand_dims(img_perturbed, axis=0)
                    
                    pred_orig = model.predict(tensor)
                    pred_orig = np.squeeze(pred_orig)
                    
                    pred_perturebed = model.predict(tensor_perturbed)
                    pred_perturebed = np.squeeze(pred_perturebed)
                    
                    if model_name in torch_models:
                      pred_orig = np.moveaxis(pred_orig, 0, -1)
                      pred_perturebed = np.moveaxis(pred_perturebed, 0, -1)
                      img_orig = np.moveaxis(img_orig, 0, -1)
                      img_perturbed = np.moveaxis(img_perturbed, 0, -1)
                      mask_gt_cat = np.moveaxis(mask_gt_cat, 0, -1)
                               
                    mask_gt = np.argmax(mask_gt_cat, axis=-1)
        
                    im_pred = np.argmax(pred_orig, axis=-1) 
                    im_pred_perturebed = np.argmax(pred_perturebed, axis=-1)        
        
                    axs_jacc[counter_model, 0].imshow(img_orig)
                    axs_jacc[counter_model, 0].set_title('Orig. Image')
                    axs_jacc[counter_model, 0].set_xticks([])
                    axs_jacc[counter_model, 0].set_yticks([])
                    axs_jacc[counter_model, 1].imshow(mask_gt)
                    axs_jacc[counter_model, 1].set_title('Ground Truth')
                    axs_jacc[counter_model, 1].set_xticks([])
                    axs_jacc[counter_model, 1].set_yticks([])
                    axs_jacc[counter_model, 2].imshow(im_pred)
                    axs_jacc[counter_model, 2].set_title('Pred. Orig.')
                    axs_jacc[counter_model, 2].set_xticks([])
                    axs_jacc[counter_model, 2].set_yticks([])
                    axs_jacc[counter_model, 2].set_xlabel( list_models[counter_model] ) 
                    axs_jacc[counter_model, 3].imshow(img_perturbed)
                    axs_jacc[counter_model, 3].set_title('Pert. Image')
                    axs_jacc[counter_model, 3].set_xticks([])
                    axs_jacc[counter_model, 3].set_yticks([])
                    #axs[counter_model, 3].yaxis.set_label_position("right")
                    axs_jacc[counter_model, 3].set_ylabel( 'Severity: ' + str(int(rec_Pert.loc[IND, 'Severity'])) ) 
                    axs_jacc[counter_model, 3].set_xlabel(rec_Pert.loc[IND, 'Perturbation Type'].replace('_', ' ').title())
                    axs_jacc[counter_model, 4].imshow(im_pred_perturebed)            
                    axs_jacc[counter_model, 4].set_title('Pred. Pert.')
                    axs_jacc[counter_model, 4].set_xticks([])
                    axs_jacc[counter_model, 4].set_yticks([])
                    axs_jacc[counter_model, 4].set_xlabel( list_models[counter_model] ) 
        
                    flag_done_thisMethod_jacc=True
                    if counter_model==1: repeat = (maskID, rec_Pert.loc[IND:IND+1, rec_Pert.columns])
                    break


fig_jacc.tight_layout() 
# plt.show()

fig_jacc.savefig('./figures/fundus_byJACC_models.png', dpi=300, bbox_inches = 'tight')