Untitled
unknown
python
4 years ago
9.5 kB
6
Indexable
import torch
import numpy as np
import glob
from sklearn.manifold import TSNE
import plotly.express as px
import plotly.graph_objects as go
from sklearn.manifold import TSNE
from tqdm import tqdm
from PIL import Image
import matplotlib.pyplot as plt
from sklearn.metrics import roc_auc_score
import time
import plotly.figure_factory as ff
models =["customnet", "resnet18_scratch", "resnet50_scratch", "resnet18_pretrained", "resnet50_pretrained"]
models =["customnet"]
dataset = "lungs"
V = "18"
logs_folder = "logs_V"+V
colors = ["darkorange","yellow","green","blue","black","red"]
# 2D/3D Proejctions ( SIMCLR, SHIFT)
D2 = True # if False 3D projections are done
PNH = True # if False SHIFT features are enabled
Nomalized = False
#groups = {"150":"Emphysema 5-10", "300":"Emphysema 10-15", "450":"Emphysema 15-20", "600":"Emphysema 20-100" }
groups = {"150":"Emphysema 0.1-0.5","300":"Emphysema 0.5-1","450":"Emphysema 1-2","600":"Emphysema 2-3",
"750":"Emphysema 3-5", "900":"Emphysema 5-10", "1050":"Emphysema 10-15", "1200":"Emphysema 15-20",
"1350":"Emphysema 20-100"}
tsne_imgs = {}
for model in models:
f_name_train = r"C:\Users\Leonard\Desktop\ConLea\CSI_AISSAM\{}\{}_{}_unsup_simclr_CSI_shift_{}_one_class_{}\feats_1_{}_train_f_name.pth".format(logs_folder,dataset,model, mode, str(class_idx), dataset)
f_name_one_class = r"C:\Users\Leonard\Desktop\ConLea\CSI_AISSAM\{}\{}_{}_unsup_simclr_CSI_shift_{}_one_class_{}\feats_1_{}_f_name.pth".format(logs_folder,dataset,model, mode, str(class_idx), dataset)
if PNH and model not in ["resnet50_pretrained"]:
f_size = 512
path_mode_train = r"C:\Users\Leonard\Desktop\ConLea\CSI_AISSAM\{}\{}_{}_unsup_simclr_CSI_shift_{}_one_class_{}\feats_1_{}_train_penultimate.pth".format(logs_folder,dataset,model, mode, str(class_idx), dataset)
path_mode_one_class = r"C:\Users\Leonard\Desktop\ConLea\CSI_AISSAM\{}\{}_{}_unsup_simclr_CSI_shift_{}_one_class_{}\feats_1_{}_penultimate.pth".format(logs_folder,dataset,model, mode,str(class_idx), dataset)
else:
print("we are here")
f_size = 2048
path_mode_train = r"C:\Users\Leonard\Desktop\ConLea\CSI_AISSAM\{}\{}_{}_unsup_simclr_CSI_shift_{}_one_class_{}\feats_1_{}_train_penultimate.pth".format(logs_folder,dataset,model, mode, str(class_idx), dataset)
path_mode_one_class = r"C:\Users\Leonard\Desktop\ConLea\CSI_AISSAM\{}\{}_{}_unsup_simclr_CSI_shift_{}_one_class_{}\feats_1_{}_penultimate.pth".format(logs_folder,dataset,model, mode,str(class_idx), dataset)
for i in range(0,1):
F = torch.zeros((100, f_size))
L = [99]*100
f_names_ = [99]*100
for idx in range(number_of_classes):
if idx!=class_idx:
f_name_path = r"C:\Users\Leonard\Desktop\ConLea\CSI_AISSAM\{}\{}_{}_unsup_simclr_CSI_shift_{}_one_class_{}\feats_1_one_class_{}_f_name.pth".format(logs_folder,dataset,model,mode,str(class_idx),str(idx))
if PNH:
path_mode = r"C:\Users\Leonard\Desktop\ConLea\CSI_AISSAM\{}\{}_{}_unsup_simclr_CSI_shift_{}_one_class_{}\feats_1_one_class_{}_penultimate.pth".format(logs_folder,dataset,model,mode,str(class_idx),str(idx))
feat = torch.load(path_mode)[:,i,:]
print(feat.shape)
F = torch.cat((F, feat), dim = 0)
L += ["class "+str(item) for item in [idx]*torch.load(path_mode)[:,i,:].shape[0]]
f_names_ += list(np.unique(torch.load(f_name_path), axis=0))
F , L, f_names_ = F[100:], L[100:], f_names_[100:]
L = [item for sublist in [[groups[str(idx)]]*150 for idx in np.arange(150,1500,150)] for item in sublist]
F = torch.cat((F,torch.load(path_mode_train)[:,i,:]), dim=0)
L += ["Anchor train"]*torch.load(path_mode_train)[:,i,:].shape[0]
f_names_ += list(np.unique(torch.load(f_name_train), axis=0))
# Add some samples from test:
F = torch.cat((F, torch.load(path_mode_one_class)[:,i,:]), dim=0)
L += ["Anchor test"]*torch.load(path_mode_one_class)[:,i,:].shape[0]
f_names_ += list(np.unique(torch.load(f_name_one_class), axis=0))
if D2:
print("2D T-Sne Projections for model {} with Features size : {},{}".format(model, F.shape[0],F.shape[1]))
tsne = TSNE(n_components=2, random_state=0 ,
perplexity=110.0,
learning_rate=200.0,
n_iter=1000,
min_grad_norm=1e-07,
verbose=0,
method='barnes_hut',
square_distances='legacy')
projections = tsne.fit_transform(F)
if Nomalized:
projections = projections / np.linalg.norm(projections, keepdims=True, axis=-1)
if i==0:
tsne_imgs[model] = projections
fig = px.scatter(
projections, x=0, y=1,
color=L, labels=L, symbol=L,
color_discrete_sequence = colors
)
else:
print("3D T-Sne Projections for {} with Features size : {},{}".format(model, F.shape[0],F.shape[1]))
tsne = TSNE(n_components=3, random_state=0 ,
perplexity=100.0,
learning_rate=200.0,
n_iter=1000,
min_grad_norm=1e-07,
verbose=0,
method='barnes_hut',
square_distances='legacy')
projections = tsne.fit_transform(F)
projections = projections / np.linalg.norm(projections, keepdims=True, axis=-1)
fig = px.scatter_3d(
projections, x=0, y=1, z=2, color = L,
color_discrete_sequence = px.colors.qualitative.Dark24)
fig.update_layout(
margin = dict(l=10, r=40, t=30, b=10),
paper_bgcolor = "LightSteelBlue",
plot_bgcolor = "LightSteelBlue",
width = 1000,
height = 500,
legend = dict(yanchor="top", y=1, xanchor="left", x=1),
bargroupgap = 0.0,
title = "Representation of {} Class VS Abnormal Classes".format("Normal"),
yaxis_title = "Shifted samples with None : {} ".format(rot[str(i)]))
fig.update_traces(marker_size=4)
fig.show()
model = "resnet18_scratch"
projections = tsne_imgs[model]
no_of_images = len(projections) # number of images. It is recommended to use a square of 2 number
ellipside = True # elipsoid or rectangular visualization
image_width = 64 # width and height of each visualized images
image_names = f_names_
# use tsne to cluster images in 2 dimensions
reduced = projections
reduced_transformed = reduced - np.min(reduced, axis=0)
reduced_transformed /= np.max(reduced_transformed, axis=0)
image_xindex_sorted = np.argsort(np.sum(reduced_transformed, axis=1))
# draw all images in a merged image
merged_width = int(np.ceil(np.sqrt(no_of_images))*image_width) + 8000
merged_image = np.ones((merged_width, merged_width, 3), dtype='uint8')*200
ind, ood = [],[]
ind_f, ood_f = [],[]
for counter, index in tqdm(enumerate(image_xindex_sorted)):
if ellipside:
a = np.ceil(reduced_transformed[counter, 0] * (merged_width-image_width-1)+1)
b = np.ceil(reduced_transformed[counter, 1] * (merged_width-image_width-1)+1)
a = int(a - np.mod(a-1,image_width) + 1)
b = int(b - np.mod(b-1,image_width) + 1)
image_address = image_names[counter]
img = np.asarray(Image.open(image_address.replace(".npz",".png")).resize((image_width, image_width)))
img = np.expand_dims(img, axis=-1)
if "one_class_train" in image_address:
# for the anchor class:
ind_f.append(image_address)
ind.append((a, a+image_width, b, b+image_width))
else:
ood_f.append(image_address)
ood.append((a, a+image_width, b, b+image_width))
merged_image[a:a+image_width, b:b+image_width,:] = img[:,:,:]
fig = px.imshow(merged_image)
save = False
colors = ["_","darkorange","yellow","green", "blue", "lightpink"]
groups = {"150":"Emphysema 0.1-0.5","300":"Emphysema 0.5-1","450":"Emphysema 1-2","600":"Emphysema 2-3",
"750":"Emphysema 3-5", "900":"Emphysema 5-10", "1050":"Emphysema 10-15", "1200":"Emphysema 15-20",
"1350":"Emphysema 20-100"}
class_ = 0
for g in np.arange(150, 750, 150):
class_ += 1
fig.add_trace( go.Scatter(
x=[(ood_[2]+ood_[3])//2 for ood_ in ood[g-150:g]],
y=[(ood_[0]+ood_[1])//2 for ood_ in ood[g-150:g]],
mode="markers",
marker = dict(color=colors[class_],),
opacity = 0.3,
name = groupes[str(g)]))
class_ = 0
for idx,(ind_,ood_) in tqdm(enumerate(zip(ind[:1000], ood[:1000]))):
if idx%150==0 and idx<600:
class_ += 1
if idx==600:
class_ += 1
c = colors[class_]
fig.add_shape(
type='rect',
x0=ood_[2], x1=ood_[3], y0=ood_[0], y1=ood_[1],
xref='x', yref='y',
line_color=c)
fig.add_shape(
type='rect',
x0=ind_[2], x1=ind_[3], y0=ind_[0], y1=ind_[1],
xref='x', yref='y',
line_color='cyan')
fig.update_layout(height=1000, width=1000, newshape=dict(line_color='cyan'))
fig.show()
if save:
merged_image = Image.fromarray(merged_image)
merged_image.save("test.png")Editor is loading...