Spaces:
Paused
Paused
Upload 4 files
Browse files- app.py +361 -0
- module.py +275 -0
- requirements.txt +9 -0
- sstvit.py +94 -0
app.py
ADDED
@@ -0,0 +1,361 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
import io
|
3 |
+
import collections
|
4 |
+
from scipy.io import loadmat
|
5 |
+
import matplotlib.pyplot as plt
|
6 |
+
from PIL import Image
|
7 |
+
import numpy as np
|
8 |
+
import torch
|
9 |
+
import argparse
|
10 |
+
import torch.nn as nn
|
11 |
+
import torch.utils.data as Data
|
12 |
+
import torch.backends.cudnn as cudnn
|
13 |
+
from scipy.io import loadmat
|
14 |
+
from scipy.io import savemat
|
15 |
+
from torch import optim
|
16 |
+
from torch.autograd import Variable
|
17 |
+
from sstvit import SSTViT
|
18 |
+
from sklearn.metrics import confusion_matrix
|
19 |
+
|
20 |
+
import matplotlib.pyplot as plt
|
21 |
+
from matplotlib import colors
|
22 |
+
import numpy as np
|
23 |
+
from patchify import patchify, unpatchify
|
24 |
+
import time
|
25 |
+
from matplotlib import colors as mcolors
|
26 |
+
import base64
|
27 |
+
import pandas as pd
|
28 |
+
import st_aggrid
|
29 |
+
import os
|
30 |
+
import json
|
31 |
+
import plotly.express as px
|
32 |
+
|
33 |
+
|
34 |
+
css='''
|
35 |
+
<style>
|
36 |
+
section.main > div {max-width:60rem}
|
37 |
+
</style>
|
38 |
+
'''
|
39 |
+
st.markdown(css, unsafe_allow_html=True)
|
40 |
+
|
41 |
+
class Args(dict):
|
42 |
+
__setattr__ = dict.__setitem__
|
43 |
+
__getattr__ = dict.__getitem__
|
44 |
+
|
45 |
+
args = {
|
46 |
+
'dataset' : 'mg',
|
47 |
+
'flag_test' : 'train',
|
48 |
+
'gpu_id' : 0,
|
49 |
+
'seed' : int(0),
|
50 |
+
'batch_size' : int(64),
|
51 |
+
'test_freq' : int(10),
|
52 |
+
'patches' : int(5),
|
53 |
+
'band_patches' : int(1),
|
54 |
+
'epoches' : int(2000),
|
55 |
+
'learning_rate' : float(5e-4),
|
56 |
+
'gamma' : float(0.9),
|
57 |
+
'weight_decay' : float(0),
|
58 |
+
'train_number' : int(500)
|
59 |
+
}
|
60 |
+
args = Args(args) # dict2object
|
61 |
+
obj = args.copy() # object2dict
|
62 |
+
|
63 |
+
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu_id)
|
64 |
+
|
65 |
+
def test_epoch(model, test_loader):
|
66 |
+
|
67 |
+
pre = np.array([])
|
68 |
+
for batch_idx, (batch_data_t1, batch_data_t2) in enumerate(test_loader):
|
69 |
+
batch_data_t1 = batch_data_t1
|
70 |
+
batch_data_t2 = batch_data_t2
|
71 |
+
|
72 |
+
batch_pred = model(batch_data_t1,batch_data_t2)
|
73 |
+
|
74 |
+
_, pred = batch_pred.topk(1, 1, True, True)
|
75 |
+
pp = pred.squeeze()
|
76 |
+
pre = np.append(pre, pp.data.cpu().numpy())
|
77 |
+
return pre
|
78 |
+
mdic = ['Before','After','Before','After']
|
79 |
+
colors = ['#3b68f8', '#ff0201', '#23fe01'] #-1,0,1,2,3
|
80 |
+
cmap = mcolors.ListedColormap(colors)
|
81 |
+
# Parameter Setting
|
82 |
+
np.random.seed(args.seed)
|
83 |
+
torch.manual_seed(args.seed)
|
84 |
+
torch.cuda.manual_seed(args.seed)
|
85 |
+
cudnn.deterministic = True
|
86 |
+
cudnn.benchmark = False
|
87 |
+
|
88 |
+
def encode_masks_to_rgb(masks):
|
89 |
+
colors = [(0, 0, 255), (255, 0, 0), (0, 255, 0)]
|
90 |
+
# Create an empty RGB image
|
91 |
+
height, width = masks.shape
|
92 |
+
rgb_image = np.zeros((height, width, 3), dtype=np.uint8)
|
93 |
+
|
94 |
+
# Assign colors based on the mask values
|
95 |
+
for i in range(len(colors)):
|
96 |
+
mask_indices = masks == i
|
97 |
+
rgb_image[mask_indices] = colors[i]
|
98 |
+
|
99 |
+
return rgb_image
|
100 |
+
def count_pixel(pred):
|
101 |
+
image = Image.fromarray(pred)
|
102 |
+
|
103 |
+
# Define the colors you want to count in RGB format
|
104 |
+
color2label = {
|
105 |
+
(0, 0, 255): "Non Mangrove",
|
106 |
+
(255, 0, 0): "Mangrove Loss",
|
107 |
+
(0, 255, 0): "Mangrove Before",
|
108 |
+
}
|
109 |
+
|
110 |
+
# Create a flattened list of pixel values
|
111 |
+
pixels = list(image.getdata())
|
112 |
+
# Count the number of pixels for each color
|
113 |
+
color_counts = collections.Counter(pixels)
|
114 |
+
# Calculate the total number of pixels in the image
|
115 |
+
total_pixels = len(pixels)
|
116 |
+
|
117 |
+
# Initialize a dictionary to store the average number of pixels for each class
|
118 |
+
average_counts = {color2label[label]: (count / total_pixels)*100 for label, count in color_counts.items()}
|
119 |
+
|
120 |
+
class_counts = {color2label[label]: count for label, count in color_counts.items()}
|
121 |
+
|
122 |
+
pix_avg = {}
|
123 |
+
pix_count = {}
|
124 |
+
for _, i in color2label.items():
|
125 |
+
try:
|
126 |
+
pix_avg[i] = average_counts[i]
|
127 |
+
pix_count[i] = class_counts[i]
|
128 |
+
except:
|
129 |
+
pix_avg[i] = 0
|
130 |
+
pix_count[i] = 0
|
131 |
+
|
132 |
+
|
133 |
+
x = {
|
134 |
+
"class": list(pix_avg.keys()),
|
135 |
+
"percentage": list(pix_avg.values()),
|
136 |
+
"pixel_count": list(pix_count.values())
|
137 |
+
}
|
138 |
+
# print(x)
|
139 |
+
|
140 |
+
return pd.DataFrame(x)
|
141 |
+
def count_pixel1(pred):
|
142 |
+
image = Image.fromarray(pred)
|
143 |
+
|
144 |
+
# Define the colors you want to count in RGB format
|
145 |
+
color2label = {
|
146 |
+
(0, 0, 255): "Non Mangrove",
|
147 |
+
(255, 0, 0): "Mangrove Loss",
|
148 |
+
(0, 255, 0): "Mangrove After",
|
149 |
+
}
|
150 |
+
|
151 |
+
# Create a flattened list of pixel values
|
152 |
+
pixels = list(image.getdata())
|
153 |
+
# Count the number of pixels for each color
|
154 |
+
color_counts = collections.Counter(pixels)
|
155 |
+
# Calculate the total number of pixels in the image
|
156 |
+
total_pixels = len(pixels)
|
157 |
+
|
158 |
+
# Initialize a dictionary to store the average number of pixels for each class
|
159 |
+
average_counts = {color2label[label]: (count / total_pixels)*100 for label, count in color_counts.items()}
|
160 |
+
|
161 |
+
class_counts = {color2label[label]: count for label, count in color_counts.items()}
|
162 |
+
|
163 |
+
pix_avg = {}
|
164 |
+
pix_count = {}
|
165 |
+
for _, i in color2label.items():
|
166 |
+
try:
|
167 |
+
pix_avg[i] = average_counts[i]
|
168 |
+
pix_count[i] = class_counts[i]
|
169 |
+
except:
|
170 |
+
pix_avg[i] = 0
|
171 |
+
pix_count[i] = 0
|
172 |
+
|
173 |
+
|
174 |
+
x = {
|
175 |
+
"class": list(pix_avg.keys()),
|
176 |
+
"percentage": list(pix_avg.values()),
|
177 |
+
"pixel_count": list(pix_count.values())
|
178 |
+
}
|
179 |
+
# print(x)
|
180 |
+
|
181 |
+
return pd.DataFrame(x)
|
182 |
+
|
183 |
+
file = st.file_uploader("Upload file", type=['mat'])
|
184 |
+
|
185 |
+
if file:
|
186 |
+
|
187 |
+
data_img2 = loadmat(file)['data_img2']
|
188 |
+
data_img1 = loadmat(file)['data_img1']
|
189 |
+
st.subheader("Preview Dataset")
|
190 |
+
col1, col2 = st.columns(2)
|
191 |
+
with col1:
|
192 |
+
fig = plt.figure(figsize=(5, 5))
|
193 |
+
plt.subplot(121)
|
194 |
+
plt.imshow(data_img1)
|
195 |
+
plt.title('Before', fontweight='bold')
|
196 |
+
plt.xticks([])
|
197 |
+
plt.yticks([])
|
198 |
+
plt.subplot(122)
|
199 |
+
plt.imshow(data_img2)
|
200 |
+
plt.title('After', fontweight='bold')
|
201 |
+
plt.xticks([])
|
202 |
+
plt.yticks([])
|
203 |
+
plt.show()
|
204 |
+
st.pyplot(fig)
|
205 |
+
holder = st.empty()
|
206 |
+
if holder.button("Start Prediction"):
|
207 |
+
start = time.time()
|
208 |
+
holder.empty()
|
209 |
+
with st.spinner("Processing, please wait around 7-15 minute"):
|
210 |
+
data_t1 = loadmat(file)['data_t1']
|
211 |
+
data_t2 = loadmat(file)['data_t2']
|
212 |
+
L_post = loadmat(file)['L_post']
|
213 |
+
L_pre = loadmat(file)['L_pre']
|
214 |
+
data_img1 = loadmat(file)['data_img1']
|
215 |
+
data_img2 = loadmat(file)['data_img2']
|
216 |
+
|
217 |
+
L_post = np.double(L_post)
|
218 |
+
L_post[L_post==0]=-0.8
|
219 |
+
L_post[L_post==1]=0
|
220 |
+
L_post[L_post==0]=-0.2
|
221 |
+
|
222 |
+
L_pre = np.double(L_pre)
|
223 |
+
L_pre[L_pre==0]=-0.8
|
224 |
+
L_pre[L_pre==1]=0
|
225 |
+
L_pre[L_pre==0]=-0.2
|
226 |
+
|
227 |
+
|
228 |
+
data_t1 = data_t1[:L_post.shape[0],:L_post.shape[1],:]
|
229 |
+
data_t2 = data_t2[:L_post.shape[0],:L_post.shape[1],:]
|
230 |
+
data_cb1 = np.zeros(shape=(L_post.shape[0],L_post.shape[1],11),dtype=np.float32)
|
231 |
+
data_cb2 = np.zeros(shape=(L_post.shape[0],L_post.shape[1],11),dtype=np.float32)
|
232 |
+
data_cb1[:,:,:10]=data_t1
|
233 |
+
data_cb1[:,:,10]=L_pre
|
234 |
+
data_cb2[:,:,:10]=data_t2
|
235 |
+
data_cb2[:,:,10]=L_post
|
236 |
+
height, width, band = data_cb1.shape
|
237 |
+
height=height-4
|
238 |
+
width = width-4
|
239 |
+
x1 = patchify(data_cb1, (5, 5, 11), step=1).reshape(-1,5*5, 11)
|
240 |
+
x2 = patchify(data_cb2, (5, 5, 11), step=1).reshape(-1,5*5, 11)
|
241 |
+
|
242 |
+
# create model
|
243 |
+
model = SSTViT(
|
244 |
+
image_size = 5,
|
245 |
+
near_band = args.band_patches,
|
246 |
+
num_patches = 11,
|
247 |
+
num_classes = 3,
|
248 |
+
dim = 32,
|
249 |
+
depth = 2,
|
250 |
+
heads = 4,
|
251 |
+
dim_head=16,
|
252 |
+
mlp_dim = 8,
|
253 |
+
b_dim = 512,
|
254 |
+
b_depth = 3,
|
255 |
+
b_heads = 8,
|
256 |
+
b_dim_head= 32,
|
257 |
+
b_mlp_head = 8,
|
258 |
+
dropout = 0.2,
|
259 |
+
emb_dropout = 0.1,
|
260 |
+
)
|
261 |
+
model.load_state_dict(torch.load("model/lsstformer.pth",map_location=torch.device("cpu")))
|
262 |
+
|
263 |
+
x1_true_band=torch.from_numpy(x1.transpose(0,2,1)).type(torch.FloatTensor)
|
264 |
+
x2_true_band=torch.from_numpy(x1.transpose(0,2,1)).type(torch.FloatTensor)
|
265 |
+
Label_true=Data.TensorDataset(x1_true_band,x2_true_band)
|
266 |
+
label_true_loader=Data.DataLoader(Label_true,batch_size=100,shuffle=False)
|
267 |
+
model.eval()
|
268 |
+
# output classification maps
|
269 |
+
pre_u = test_epoch(model, label_true_loader)
|
270 |
+
prediction_matrix = pre_u.reshape(height,width)
|
271 |
+
|
272 |
+
x1_true_band=torch.from_numpy(x1.transpose(0,2,1)).type(torch.FloatTensor)
|
273 |
+
x2_true_band=torch.from_numpy(x2.transpose(0,2,1)).type(torch.FloatTensor)
|
274 |
+
Label_true=Data.TensorDataset(x1_true_band,x2_true_band)
|
275 |
+
label_true_loader=Data.DataLoader(Label_true,batch_size=100,shuffle=False)
|
276 |
+
model.eval()
|
277 |
+
# output classification maps
|
278 |
+
pre_u = test_epoch(model, label_true_loader)
|
279 |
+
prediction_matrix2 = pre_u.reshape(height,width)
|
280 |
+
A = prediction_matrix.reshape(-1)
|
281 |
+
B = prediction_matrix2.reshape(-1)
|
282 |
+
mg = np.array(np.where(A==2))
|
283 |
+
mg1 = np.array(np.where(B==2))
|
284 |
+
mgls = np.array(np.where(B==1))
|
285 |
+
class_counts = count_pixel(encode_masks_to_rgb(prediction_matrix))
|
286 |
+
class_counts1 = count_pixel1(encode_masks_to_rgb(prediction_matrix2))
|
287 |
+
|
288 |
+
with st.container():
|
289 |
+
st.subheader("Prediction Result")
|
290 |
+
col1, col2 = st.columns(2)
|
291 |
+
with col1:
|
292 |
+
with st.container():
|
293 |
+
fig = plt.figure(figsize=(10, 10))
|
294 |
+
plt.subplot(121)
|
295 |
+
plt.imshow(prediction_matrix, cmap=cmap)
|
296 |
+
plt.title('Before',fontsize=25, fontweight='bold')
|
297 |
+
plt.xticks([])
|
298 |
+
plt.yticks([])
|
299 |
+
plt.subplot(122)
|
300 |
+
plt.imshow(prediction_matrix2, cmap=cmap)
|
301 |
+
plt.title('After',fontsize=25, fontweight='bold')
|
302 |
+
plt.xticks([])
|
303 |
+
plt.yticks([])
|
304 |
+
plt.show()
|
305 |
+
st.pyplot(fig)
|
306 |
+
buf = io.BytesIO()
|
307 |
+
fig.savefig(buf, format="png")
|
308 |
+
with col2:
|
309 |
+
with st.container():
|
310 |
+
table_data = {
|
311 |
+
"Total mangrove before":f"{mg.shape[1]*100} m\u00B2",
|
312 |
+
"Total mangrove after":f"{mg1.shape[1]*100} m\u00B2",
|
313 |
+
"Total mangrove loss":f"{mgls.shape[1]*100} m\u00B2",
|
314 |
+
}
|
315 |
+
df = pd.DataFrame(list(table_data.items()), columns=['Key', 'Value'])
|
316 |
+
|
317 |
+
MIN_HEIGHT = 100
|
318 |
+
MAX_HEIGHT = 180
|
319 |
+
ROW_HEIGHT = 50
|
320 |
+
|
321 |
+
# st.dataframe(df, hide_index=True, use_container_width=True)
|
322 |
+
st_aggrid.AgGrid(df,fit_columns_on_grid_load=True, height=min(MIN_HEIGHT + len(df) * ROW_HEIGHT, MAX_HEIGHT))
|
323 |
+
with st.container():
|
324 |
+
st.subheader("Pixel Distribution")
|
325 |
+
|
326 |
+
|
327 |
+
df = class_counts
|
328 |
+
df = df.drop(0)
|
329 |
+
df1 = df.drop(1)
|
330 |
+
|
331 |
+
|
332 |
+
|
333 |
+
df2 = class_counts1
|
334 |
+
df3 = df2.drop(0)
|
335 |
+
vertical_concat = pd.concat([df1, df3], axis=0)
|
336 |
+
MIN_HEIGHT = 100
|
337 |
+
MAX_HEIGHT = 180
|
338 |
+
ROW_HEIGHT = 50
|
339 |
+
vertical_concat = vertical_concat.iloc[[0,2,1],:]
|
340 |
+
|
341 |
+
|
342 |
+
st_aggrid.AgGrid(vertical_concat,fit_columns_on_grid_load=True, height=min(MIN_HEIGHT + len(vertical_concat) * ROW_HEIGHT, MAX_HEIGHT))
|
343 |
+
fig = px.bar(vertical_concat, x='percentage', y='class', color='class', orientation='h',
|
344 |
+
color_discrete_sequence=["green","green", "red", "blue"],
|
345 |
+
category_orders={"class": ["Mangrove Before","Mangrove After", "Mangrove Loss", "Non Mangrove",]}
|
346 |
+
)
|
347 |
+
|
348 |
+
st.plotly_chart(fig,use_container_width=False)
|
349 |
+
end = time.time()
|
350 |
+
process = end-start
|
351 |
+
st.write('process',process)
|
352 |
+
|
353 |
+
|
354 |
+
show_file = st.empty()
|
355 |
+
|
356 |
+
if not file:
|
357 |
+
url = "https://drive.usercontent.google.com/download?id=1u48pMzRWQ2Etfjaq5A0CUjRtGKZaJoJy&export=download&authuser=2&confirm=t&uuid=52b0e01e-377f-42cb-8412-c84aa38a1740&at=APZUnTXslmuCCV1drJ2WWtkZr9BR%3A1710357675310"
|
358 |
+
show_file.info("""
|
359 |
+
The model was trained using Sentinel-2 imagery, users can upload MAT files to perform LSST-Former for mangrove loss detection models that have been trained in this research. Tool for generate from Sentinel-2 to MAT file i will create later, please download demo dataset bellow. for better in mobile phone, se desktop mode.
|
360 |
+
""")
|
361 |
+
st.write("download demo datasets this [link](%s)" % url)
|
module.py
ADDED
@@ -0,0 +1,275 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import numpy as np
|
4 |
+
from einops import rearrange, repeat
|
5 |
+
|
6 |
+
class Residual(nn.Module):
|
7 |
+
def __init__(self, fn):
|
8 |
+
super().__init__()
|
9 |
+
self.fn = fn
|
10 |
+
def forward(self, x, **kwargs):
|
11 |
+
return self.fn(x, **kwargs) + x
|
12 |
+
|
13 |
+
class PreNorm(nn.Module):
|
14 |
+
def __init__(self, dim, fn):
|
15 |
+
super().__init__()
|
16 |
+
self.norm = nn.LayerNorm(dim)
|
17 |
+
self.fn = fn
|
18 |
+
def forward(self, x, **kwargs):
|
19 |
+
return self.fn(self.norm(x), **kwargs)
|
20 |
+
|
21 |
+
class FeedForward(nn.Module):
|
22 |
+
def __init__(self, dim, hidden_dim, dropout = 0.):
|
23 |
+
super().__init__()
|
24 |
+
self.net = nn.Sequential(
|
25 |
+
nn.Linear(dim, hidden_dim),
|
26 |
+
nn.GELU(),
|
27 |
+
nn.Dropout(dropout),
|
28 |
+
nn.Linear(hidden_dim, dim),
|
29 |
+
nn.Dropout(dropout)
|
30 |
+
)
|
31 |
+
def forward(self, x):
|
32 |
+
return self.net(x)
|
33 |
+
|
34 |
+
class Attention(nn.Module):
|
35 |
+
def __init__(self, dim, heads, dim_head, dropout):
|
36 |
+
super().__init__()
|
37 |
+
inner_dim = dim_head * heads
|
38 |
+
self.heads = heads
|
39 |
+
self.scale = dim_head ** -0.5
|
40 |
+
|
41 |
+
self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False)
|
42 |
+
self.to_out = nn.Sequential(
|
43 |
+
nn.Linear(inner_dim, dim),
|
44 |
+
nn.Dropout(dropout)
|
45 |
+
)
|
46 |
+
def forward(self, x, mask = None):
|
47 |
+
# x:[b,n,dim]
|
48 |
+
b, n, _, h = *x.shape, self.heads
|
49 |
+
|
50 |
+
# get qkv tuple:([b,n,head_num*head_dim],[...],[...])
|
51 |
+
qkv = self.to_qkv(x).chunk(3, dim = -1)
|
52 |
+
# split q,k,v from [b,n,head_num*head_dim] -> [b,head_num,n,head_dim]
|
53 |
+
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = h), qkv)
|
54 |
+
# transpose(k) * q / sqrt(head_dim) -> [b,head_num,n,n]
|
55 |
+
dots = torch.einsum('bhid,bhjd->bhij', q, k) * self.scale
|
56 |
+
mask_value = -torch.finfo(dots.dtype).max
|
57 |
+
|
58 |
+
# mask value: -inf
|
59 |
+
if mask is not None:
|
60 |
+
mask = F.pad(mask.flatten(1), (1, 0), value = True)
|
61 |
+
assert mask.shape[-1] == dots.shape[-1], 'mask has incorrect dimensions'
|
62 |
+
mask = mask[:, None, :] * mask[:, :, None]
|
63 |
+
dots.masked_fill_(~mask, mask_value)
|
64 |
+
del mask
|
65 |
+
|
66 |
+
# softmax normalization -> attention matrix
|
67 |
+
attn = dots.softmax(dim=-1)
|
68 |
+
# value * attention matrix -> output
|
69 |
+
out = torch.einsum('bhij,bhjd->bhid', attn, v)
|
70 |
+
# cat all output -> [b, n, head_num*head_dim]
|
71 |
+
out = rearrange(out, 'b h n d -> b n (h d)')
|
72 |
+
out = self.to_out(out)
|
73 |
+
return out
|
74 |
+
|
75 |
+
class CrossAttention(nn.Module):
|
76 |
+
def __init__(self, dim, heads, dim_head, dropout):
|
77 |
+
super().__init__()
|
78 |
+
inner_dim = dim_head * heads
|
79 |
+
project_out = not (heads == 1 and dim_head == dim)
|
80 |
+
|
81 |
+
self.heads = heads
|
82 |
+
self.scale = dim_head ** -0.5
|
83 |
+
|
84 |
+
self.to_k = nn.Linear(dim, inner_dim , bias=False)
|
85 |
+
self.to_v = nn.Linear(dim, inner_dim , bias = False)
|
86 |
+
self.to_q = nn.Linear(dim, inner_dim, bias = False)
|
87 |
+
|
88 |
+
self.to_out = nn.Sequential(
|
89 |
+
nn.Linear(inner_dim, dim),
|
90 |
+
nn.Dropout(dropout)
|
91 |
+
) if project_out else nn.Identity()
|
92 |
+
|
93 |
+
def forward(self, x_qkv):
|
94 |
+
b, n, _, h = *x_qkv.shape, self.heads
|
95 |
+
|
96 |
+
k = self.to_k(x_qkv)
|
97 |
+
k = rearrange(k, 'b n (h d) -> b h n d', h = h)
|
98 |
+
|
99 |
+
v = self.to_v(x_qkv)
|
100 |
+
v = rearrange(v, 'b n (h d) -> b h n d', h = h)
|
101 |
+
|
102 |
+
q = self.to_q(x_qkv[:, 0].unsqueeze(1))
|
103 |
+
q = rearrange(q, 'b n (h d) -> b h n d', h = h)
|
104 |
+
|
105 |
+
dots = torch.einsum('b h i d, b h j d -> b h i j', q, k) * self.scale
|
106 |
+
|
107 |
+
attn = dots.softmax(dim=-1)
|
108 |
+
|
109 |
+
out = torch.einsum('b h i j, b h j d -> b h i d', attn, v)
|
110 |
+
out = rearrange(out, 'b h n d -> b n (h d)')
|
111 |
+
out = self.to_out(out)
|
112 |
+
return out
|
113 |
+
|
114 |
+
class Transformer(nn.Module):
|
115 |
+
def __init__(self, dim, depth, heads, dim_head, mlp_head, dropout, num_channel):
|
116 |
+
super().__init__()
|
117 |
+
|
118 |
+
self.layers = nn.ModuleList([])
|
119 |
+
for _ in range(depth):
|
120 |
+
self.layers.append(nn.ModuleList([
|
121 |
+
Residual(PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout))),
|
122 |
+
Residual(PreNorm(dim, FeedForward(dim, mlp_head, dropout = dropout)))
|
123 |
+
]))
|
124 |
+
|
125 |
+
self.skipcat = nn.ModuleList([])
|
126 |
+
for _ in range(depth-2):
|
127 |
+
self.skipcat.append(nn.Conv2d(num_channel+1, num_channel+1, [1, 2], 1, 0))
|
128 |
+
|
129 |
+
def forward(self, x, mask = None):
|
130 |
+
for attn, ff in self.layers:
|
131 |
+
x = attn(x, mask = mask)
|
132 |
+
x = ff(x)
|
133 |
+
return x
|
134 |
+
|
135 |
+
class SSTransformer(nn.Module):
|
136 |
+
def __init__(self, dim, depth, heads, dim_head, mlp_head, b_dim, b_depth, b_heads, b_dim_head, b_mlp_head, num_patches, dropout):
|
137 |
+
super().__init__()
|
138 |
+
|
139 |
+
self.layers = nn.ModuleList([])
|
140 |
+
self.k_layers = nn.ModuleList([])
|
141 |
+
self.channels_to_embedding = nn.Linear(num_patches, b_dim)
|
142 |
+
self.cls_token = nn.Parameter(torch.randn(1, 1, b_dim))
|
143 |
+
for _ in range(depth):
|
144 |
+
self.layers.append(nn.ModuleList([
|
145 |
+
Residual(PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout))),
|
146 |
+
Residual(PreNorm(dim, FeedForward(dim, mlp_head, dropout = dropout)))
|
147 |
+
]))
|
148 |
+
for _ in range(b_depth):
|
149 |
+
self.k_layers.append(nn.ModuleList([
|
150 |
+
Residual(PreNorm(b_dim, Attention(dim=b_dim, heads=b_heads, dim_head=b_dim_head, dropout = dropout))),
|
151 |
+
Residual(PreNorm(b_dim, FeedForward(b_dim, b_mlp_head, dropout = dropout)))
|
152 |
+
]))
|
153 |
+
|
154 |
+
def forward(self, x, mask = None):
|
155 |
+
for attn, ff in self.layers:
|
156 |
+
x = attn(x, mask = mask)
|
157 |
+
x = ff(x)
|
158 |
+
x = rearrange(x, 'b n d -> b d n')
|
159 |
+
x = self.channels_to_embedding(x)
|
160 |
+
b, d, n = x.shape
|
161 |
+
cls_tokens = repeat(self.cls_token, '() n d -> b n d', b = b)
|
162 |
+
x = torch.cat((cls_tokens, x), dim = 1)
|
163 |
+
for attn, ff in self.k_layers:
|
164 |
+
x = attn(x, mask = mask)
|
165 |
+
x = ff(x)
|
166 |
+
return x
|
167 |
+
|
168 |
+
class SSTransformer_pyramid(nn.Module):
|
169 |
+
def __init__(self, dim, depth, heads, dim_head, mlp_head, b_dim, b_depth, b_heads, b_dim_head, b_mlp_head, num_patches, dropout):
|
170 |
+
super().__init__()
|
171 |
+
|
172 |
+
self.layers = nn.ModuleList([])
|
173 |
+
self.k_layers = nn.ModuleList([])
|
174 |
+
self.channels_to_embedding = nn.Linear(num_patches, b_dim)
|
175 |
+
self.cls_token = nn.Parameter(torch.randn(1, 1, b_dim))
|
176 |
+
for _ in range(depth):
|
177 |
+
self.layers.append(nn.ModuleList([
|
178 |
+
Residual(PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout))),
|
179 |
+
Residual(PreNorm(dim, FeedForward(dim, mlp_head, dropout = dropout)))
|
180 |
+
]))
|
181 |
+
for _ in range(b_depth):
|
182 |
+
self.k_layers.append(nn.ModuleList([
|
183 |
+
Residual(PreNorm(b_dim, Attention(dim=b_dim, heads=b_heads, dim_head=b_dim_head, dropout = dropout))),
|
184 |
+
Residual(PreNorm(b_dim, FeedForward(b_dim, b_mlp_head, dropout = dropout)))
|
185 |
+
]))
|
186 |
+
|
187 |
+
def forward(self, x, mask = None):
|
188 |
+
for attn, ff in self.layers:
|
189 |
+
x = attn(x, mask = mask)
|
190 |
+
x = ff(x)
|
191 |
+
out_feature = x
|
192 |
+
x = rearrange(x, 'b n d -> b d n')
|
193 |
+
x = self.channels_to_embedding(x)
|
194 |
+
b, d, n = x.shape
|
195 |
+
cls_tokens = repeat(self.cls_token, '() n d -> b n d', b = b)
|
196 |
+
x = torch.cat((cls_tokens, x), dim = 1)
|
197 |
+
for attn, ff in self.k_layers:
|
198 |
+
x = attn(x, mask = mask)
|
199 |
+
x = ff(x)
|
200 |
+
return x, out_feature
|
201 |
+
|
202 |
+
class ViT(nn.Module):
|
203 |
+
def __init__(self, image_size, near_band, num_patches, num_classes, dim, depth, heads, mlp_dim, pool='cls', channel_dim=1, dim_head = 16, dropout=0., emb_dropout=0., mode='ViT'):
|
204 |
+
super().__init__()
|
205 |
+
|
206 |
+
patch_dim = image_size ** 2 * near_band
|
207 |
+
|
208 |
+
self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, dim))
|
209 |
+
self.patch_to_embedding = nn.Linear(channel_dim, dim)
|
210 |
+
self.cls_token = nn.Parameter(torch.randn(1, 1, dim))
|
211 |
+
|
212 |
+
self.dropout = nn.Dropout(emb_dropout)
|
213 |
+
self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim, dropout, num_patches, mode)
|
214 |
+
|
215 |
+
self.pool = pool
|
216 |
+
self.to_latent = nn.Identity()
|
217 |
+
|
218 |
+
self.mlp_head = nn.Sequential(
|
219 |
+
nn.LayerNorm(dim),
|
220 |
+
nn.Linear(dim, num_classes)
|
221 |
+
)
|
222 |
+
def forward(self, x, mask = None):
|
223 |
+
# patchs[batch, patch_num, patch_size*patch_size*c] [batch,200,145*145]
|
224 |
+
# x = rearrange(x, 'b c h w -> b c (h w)')
|
225 |
+
## embedding every patch vector to embedding size: [batch, patch_num, embedding_size]
|
226 |
+
|
227 |
+
x = self.patch_to_embedding(x) #[b,n,dim]
|
228 |
+
b, n, _ = x.shape
|
229 |
+
|
230 |
+
# add position embedding
|
231 |
+
cls_tokens = repeat(self.cls_token, '() n d -> b n d', b = b) #[b,1,dim]
|
232 |
+
x = torch.cat((cls_tokens, x), dim = 1) #[b,n+1,dim]
|
233 |
+
x += self.pos_embedding[:, :(n + 1)]
|
234 |
+
x = self.dropout(x)
|
235 |
+
# transformer: x[b,n + 1,dim] -> x[b,n + 1,dim]
|
236 |
+
x = self.transformer(x, mask)
|
237 |
+
# classification: using cls_token output
|
238 |
+
x = self.to_latent(x[:,0])
|
239 |
+
|
240 |
+
# MLP classification layer
|
241 |
+
return self.mlp_head(x)
|
242 |
+
|
243 |
+
class SSFormer_v4(nn.Module):
|
244 |
+
def __init__(self, dim, depth, heads, dim_head, mlp_head, b_dim, b_depth, b_heads, b_dim_head, b_mlp_head, num_patches, dropout, mode):
|
245 |
+
super().__init__()
|
246 |
+
|
247 |
+
self.layers = nn.ModuleList([])
|
248 |
+
self.k_layers = nn.ModuleList([])
|
249 |
+
self.channels_to_embedding = nn.Linear(num_patches, b_dim)
|
250 |
+
self.cls_token = nn.Parameter(torch.randn(1, 1, b_dim))
|
251 |
+
for _ in range(depth):
|
252 |
+
self.layers.append(nn.ModuleList([
|
253 |
+
Residual(PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout))),
|
254 |
+
Residual(PreNorm(dim, FeedForward(dim, mlp_head, dropout = dropout)))
|
255 |
+
]))
|
256 |
+
for _ in range(b_depth):
|
257 |
+
self.k_layers.append(nn.ModuleList([
|
258 |
+
Residual(PreNorm(b_dim, Attention(dim=b_dim, heads=b_heads, dim_head=b_dim_head, dropout = dropout))),
|
259 |
+
Residual(PreNorm(b_dim, FeedForward(b_dim, b_mlp_head, dropout = dropout)))
|
260 |
+
]))
|
261 |
+
self.mode = mode
|
262 |
+
|
263 |
+
def forward(self, x, c, mask = None):
|
264 |
+
for attn, ff in self.layers:
|
265 |
+
x = attn(x, mask = mask)
|
266 |
+
x = ff(x)
|
267 |
+
x = rearrange(x, 'b n d -> b d n')
|
268 |
+
x = self.channels_to_embedding(x)
|
269 |
+
b, d, n = x.shape
|
270 |
+
cls_tokens = repeat(c, '() n d -> b n d', b = b)
|
271 |
+
x = torch.cat((cls_tokens, x), dim = 1)
|
272 |
+
for attn, ff in self.k_layers:
|
273 |
+
x = attn(x, mask = mask)
|
274 |
+
x = ff(x)
|
275 |
+
return x
|
requirements.txt
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
einops
|
2 |
+
patchify
|
3 |
+
argparse
|
4 |
+
scipy
|
5 |
+
scikit-learn
|
6 |
+
torch
|
7 |
+
streamlit-aggrid
|
8 |
+
plotly
|
9 |
+
collection
|
sstvit.py
ADDED
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from torch import nn, einsum
|
3 |
+
import torch.nn.functional as F
|
4 |
+
from einops import rearrange, repeat
|
5 |
+
from einops.layers.torch import Rearrange
|
6 |
+
from module import Attention, PreNorm, FeedForward, CrossAttention, SSTransformer
|
7 |
+
import numpy as np
|
8 |
+
|
9 |
+
class SSTTransformerEncoder(nn.Module):
|
10 |
+
|
11 |
+
def __init__(self, dim, depth, heads, dim_head, mlp_dim, b_dim, b_depth, b_heads, b_dim_head, b_mlp_head, num_patches, cross_attn_depth=3, cross_attn_heads=8, dropout = 0):
|
12 |
+
super().__init__()
|
13 |
+
|
14 |
+
self.transformer = SSTransformer(dim, depth, heads, dim_head, mlp_dim, b_dim, b_depth, b_heads, b_dim_head, b_mlp_head, num_patches, dropout)
|
15 |
+
|
16 |
+
self.cross_attn_layers = nn.ModuleList([])
|
17 |
+
for _ in range(cross_attn_depth):
|
18 |
+
self.cross_attn_layers.append(PreNorm(b_dim, CrossAttention(b_dim, heads = cross_attn_heads, dim_head=dim_head, dropout=0)))
|
19 |
+
|
20 |
+
def forward(self, x1, x2):
|
21 |
+
x1 = self.transformer(x1)
|
22 |
+
x2 = self.transformer(x2)
|
23 |
+
|
24 |
+
for cross_attn in self.cross_attn_layers:
|
25 |
+
x1_class = x1[:, 0]
|
26 |
+
x1 = x1[:, 1:]
|
27 |
+
x2_class = x2[:, 0]
|
28 |
+
x2 = x2[:, 1:]
|
29 |
+
|
30 |
+
# Cross Attn
|
31 |
+
cat1_q = x1_class.unsqueeze(1)
|
32 |
+
cat1_qkv = torch.cat((cat1_q, x2), dim=1)
|
33 |
+
cat1_out = cat1_q+cross_attn(cat1_qkv)
|
34 |
+
x1 = torch.cat((cat1_out, x1), dim=1)
|
35 |
+
cat2_q = x2_class.unsqueeze(1)
|
36 |
+
cat2_qkv = torch.cat((cat2_q, x1), dim=1)
|
37 |
+
cat2_out = cat2_q+cross_attn(cat2_qkv)
|
38 |
+
x2 = torch.cat((cat2_out, x2), dim=1)
|
39 |
+
|
40 |
+
return cat1_out, cat2_out
|
41 |
+
|
42 |
+
class SSTViT(nn.Module):
|
43 |
+
def __init__(self, image_size, near_band, num_patches, num_classes, dim, depth, heads, mlp_dim, b_dim, b_depth, b_heads, b_dim_head, b_mlp_head, pool='cls', channels=1, dim_head = 16, dropout=0., emb_dropout=0., multi_scale_enc_depth=1):
|
44 |
+
super().__init__()
|
45 |
+
|
46 |
+
patch_dim = image_size ** 2 * near_band
|
47 |
+
self.num_patches = num_patches+1
|
48 |
+
self.pos_embedding = nn.Parameter(torch.randn(1, self.num_patches, dim))
|
49 |
+
self.patch_to_embedding = nn.Linear(patch_dim, dim)
|
50 |
+
self.cls_token_t1 = nn.Parameter(torch.randn(1, 1, dim))
|
51 |
+
self.cls_token_t2 = nn.Parameter(torch.randn(1, 1, dim))
|
52 |
+
|
53 |
+
self.dropout = nn.Dropout(emb_dropout)
|
54 |
+
|
55 |
+
self.multi_scale_transformers = nn.ModuleList([])
|
56 |
+
for _ in range(multi_scale_enc_depth):
|
57 |
+
self.multi_scale_transformers.append(SSTTransformerEncoder(dim, depth, heads, dim_head, mlp_dim,b_dim, b_depth, b_heads, b_dim_head, b_mlp_head, self.num_patches,
|
58 |
+
dropout = 0.))
|
59 |
+
|
60 |
+
self.pool = pool
|
61 |
+
self.to_latent = nn.Identity()
|
62 |
+
|
63 |
+
self.mlp_head = nn.Sequential(
|
64 |
+
nn.LayerNorm(b_dim),
|
65 |
+
nn.Linear(b_dim, num_classes)
|
66 |
+
)
|
67 |
+
def forward(self, x1, x2):
|
68 |
+
# patchs[batch, patch_num, patch_size*patch_size*c] [batch,200,145*145]
|
69 |
+
# x = rearrange(x, 'b c h w -> b c (h w)')
|
70 |
+
## embedding every patch vector to embedding size: [batch, patch_num, embedding_size]
|
71 |
+
x1 = self.patch_to_embedding(x1) #[b,n,dim]
|
72 |
+
x2 = self.patch_to_embedding(x2)
|
73 |
+
b, n, _ = x1.shape
|
74 |
+
# add position embedding
|
75 |
+
cls_tokens_t1 = repeat(self.cls_token_t1, '() n d -> b n d', b = b) #[b,1,dim]
|
76 |
+
cls_tokens_t2 = repeat(self.cls_token_t2, '() n d -> b n d', b = b)
|
77 |
+
|
78 |
+
x1 = torch.cat((cls_tokens_t1, x1), dim = 1) #[b,n+1,dim]
|
79 |
+
x1 += self.pos_embedding[:, :(n + 1)]
|
80 |
+
x1 = self.dropout(x1)
|
81 |
+
x2 = torch.cat((cls_tokens_t2, x2), dim = 1) #[b,n+1,dim]
|
82 |
+
x2 += self.pos_embedding[:, :(n + 1)]
|
83 |
+
x2 = self.dropout(x2)
|
84 |
+
# transformer: x[b,n + 1,dim] -> x[b,n + 1,dim]
|
85 |
+
for multi_scale_transformer in self.multi_scale_transformers:
|
86 |
+
out1, out2 = multi_scale_transformer(x1, x2)
|
87 |
+
# classification: using cls_token output
|
88 |
+
out1 = self.to_latent(out1[:,0])
|
89 |
+
out2 = self.to_latent(out2[:,0])
|
90 |
+
out = out1+out2
|
91 |
+
# MLP classification layer
|
92 |
+
return self.mlp_head(out)
|
93 |
+
|
94 |
+
|