-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_novel_shapenet.py
executable file
·426 lines (362 loc) · 20.8 KB
/
train_novel_shapenet.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
import argparse
import random
import torch
import torch.optim as optim
import torch.nn.functional as F
from tqdm import tqdm
import open_clip
from utils.mv_utils_zs_ver_2 import Realistic_Projection_Learnable_new as Realistic_Projection
from model.PointNet import PointNetfeat, feature_transform_regularizer, STN3d
from model.curvenet import *
from model.Transformation import Transformation
from utils.datautil_3D_memory_incremental_shapenet import *
from model.Relation import RelationNetwork
import os
import numpy as np
from matplotlib import pyplot as plt
from torch import nn
from utils.Loss import CombinedConstraintLoss
from model.Unet_dropout import UNetPlusPlus
from torchmetrics.functional.image import image_gradients
from configs.shapenet_info import task_ids_total as tid
import json
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
def set_random_seed(seed):
random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed)
# read a txt file line by line and save it in a list, and remove the empty lines
def read_txt_file(file):
with open(file, 'r') as f:
array = f.readlines()
array = ["A depth map of " + x.strip() for x in array]
array = list(filter(None, array))
return array
def read_txt_file_class_name(file):
with open(file, 'r') as f:
array = f.readlines()
array = [x.strip() for x in array]
array = list(filter(None, array))
return array
# read json file
def read_json_file(file):
with open(file, 'r') as f:
array = json.load(f)
return array
def accuracy(output, target, topk=(1,)):
pred = output.topk(max(topk), 1, True, True)[1].t()
correct = pred.eq(target.view(1, -1).expand_as(pred))
return [float(correct[:k].reshape(-1).float().sum(0, keepdim=True).cpu().numpy()) for k in topk]
# define the main function
def main(opt):
num_rotations = 1
fea_weight = 0.8
set_random_seed(opt.manualSeed)
# import pointnet model
#pointnet = PointNetfeat(global_feat=True, feature_transform=opt.feature_transform)
#pointnet = pointnet.to(device)
curvenet = CurveNet()
curvenet = curvenet.to(device)
curvenet.load_state_dict(torch.load('cls/shapenet/curvenet_40.pth', map_location=device))
# Step 1: Load CLIP model
clip_model, _, preprocess = open_clip.create_model_and_transforms('ViT-B-16', pretrained='laion2b_s34b_b88k')
clip_model.to(device)
for param in clip_model.parameters():
param.requires_grad = False
# Step 2: Load Realistic Projection object
proj = Realistic_Projection().to(device)
# Step 3: Load the Transformation model
transform = {str(i): STN3d() for i in range(num_rotations)}
for i in range(num_rotations):
transform[str(i)].to(device)
transform[str(i)].load_state_dict(torch.load('cls/shapenet/transform_40_%d.pth' % i, map_location=device))
# load the Unet model
unet = UNetPlusPlus().to(device)
unet.load_state_dict(torch.load('cls/shapenet/unet_40.pth', map_location=device))
# Step 4: Load the Relation Network
relation = RelationNetwork(1536, 2048, 1024)
relation = relation.to(device)
relation.load_state_dict(torch.load('cls/shapenet/relation_40.pth', map_location=device))
#load the text features
class_name = read_txt_file_class_name("class_name_shapenet.txt")
prompts = read_json_file("shapenet.json")
# define the optimizer
Parameters = [p for model in transform.values() for p in model.parameters()]
#optimizer = optim.Adam(Parameters + list(relation.parameters()) + list(unet.parameters()) + list(pointnet.parameters()), lr=0.001, betas=(0.9, 0.999))
optimizer = optim.Adam(relation.parameters(), lr=0.001, betas=(0.9, 0.999))
# load loss function
cross_entrpy = nn.BCELoss()
constraint_loss = CombinedConstraintLoss(num_rotations=num_rotations)
loss_orthogonal_weight = 0.01
mse_loss = nn.MSELoss()
# load memory bank as a numpy array
memory_bank = np.load('memory/memory_bank_shapenet.npy')
memory_bank_label = np.load('memory/memory_bank_label_shapenet.npy')
memory_bank = torch.from_numpy(memory_bank).to(device)
memory_bank_label = torch.from_numpy(memory_bank_label).to(device)
for t in range(0,7):
path=Path(opt.dataset_path)
print(path)
dataloader = DatasetGen(opt, root=path, fewshot=5)
dataset = dataloader.get(t,'training')
trainDataLoader = dataset[t]['train']
testDataLoader = dataset[t]['test']
num_category = 25 + t * 5
print('task:', t)
# train the model
clip_model.train()
for i in range(num_rotations):
transform[format(i)].train()
unet.train()
relation.train()
curvenet.train()
print("=> Start training the model")
# construct the memory bank
memory_bank_task = memory_bank[0:(num_category-5),:,:]
memory_bank_label_task = memory_bank_label[0:(num_category-5),:]
mm = 0
if t == 0:
nepoch = 0
else:
nepoch = opt.nepoch
for epoch in range(nepoch):
# define the loss
train_loss = 0
train_correct = 0
train_total = 0
for i, data in tqdm(enumerate(trainDataLoader, 0)):
points, target = data['pointclouds'].to(device).float(), data['labels'].to(device)
points, target = points.to(device), target.to(device)
if points.shape[0] < opt.batch_size:
continue
# Select 16 samples from memory_bank_task and memory_bank_label_task
indices = torch.randperm(memory_bank_task.shape[0])[:opt.batch_size]
memory_bank_task_samples = memory_bank_task[indices, :, :]
memory_bank_label_task_samples = memory_bank_label_task[indices, :]
points = torch.cat((points, memory_bank_task_samples), 0)
target = torch.cat((target, memory_bank_label_task_samples.squeeze(1)), 0)
optimizer.zero_grad()
points = points.transpose(2, 1)
# Forward samples to the PointNet model
points_embedding = curvenet(points)
# transformation module
trans = torch.zeros((points.shape[0], num_rotations, 3, 3), device=device)
for jj in range(num_rotations):
trans[:, jj, :, :] = transform[format(jj)](points)
loss_orthogonal = constraint_loss(trans).mean()
# depth map generation
points = points.transpose(2, 1)
depth_map = torch.zeros((points.shape[0] * num_rotations, 3, 224, 224)).to(device)
for jj in range(num_rotations):
depth_map_tmp = proj.get_img(points, trans[:,jj,:,:].view(-1, 9))
depth_map_tmp = torch.nn.functional.interpolate(depth_map_tmp, size=(224, 224), mode='bilinear', align_corners=True)
depth_map[jj * points.shape[0]:(jj + 1) * points.shape[0], :, :, :] = depth_map_tmp
loss_gradient = 0
RGB_map = torch.zeros((points.shape[0] * num_rotations, 3, 224, 224)).to(device)
for jj in range(num_rotations):
# unet model
depth_map_reverse = 1 - depth_map[jj * points.shape[0]:(jj + 1) * points.shape[0]]
mask = (depth_map_reverse != 0).float()
texture_map = unet(mask)
# loss for gradient
dy_init, dx_init = image_gradients(mask)
dy, dx = image_gradients(texture_map)
loss_gradient += mse_loss(dy, dy_init) + mse_loss(dx, dx_init)
RGB_map[jj * points.shape[0]:(jj + 1) * points.shape[0], :, :, :] = depth_map[jj * points.shape[0]:(jj + 1) * points.shape[0]] * texture_map
# Forward samples to the vision CLIP model
img_embedding_tmp = clip_model.encode_image(RGB_map).to(device)
img_embedding = 0
for jj in range(num_rotations):
img_embedding += img_embedding_tmp[jj * points.shape[0]:(jj + 1) * points.shape[0], :]/ num_rotations
# merge img_embedding and points_embedding
img_embedding = img_embedding / img_embedding.norm(dim=-1, keepdim=True)
points_embedding = points_embedding / points_embedding.norm(dim=-1, keepdim=True)
fea_embedding = torch.cat((img_embedding, points_embedding), 1)
# Sample prompts from prompts dictionary
tid_all = []
for h in range(t+1):
tid_all += tid[h]
prompts_batch = []
for j in range(num_category):
tmp_1 = (class_name[tid_all[j]])
tmp_1 = tmp_1.split(' ')
tmp_2 = prompts[tmp_1[1]]
random_idx = random.randint(0, len(tmp_2)-1)
prompts_batch.append(tmp_2[random_idx])
# Forward samples to the text CLIP model
text = open_clip.tokenize(prompts_batch)
text_embedding = clip_model.encode_text(text.to(device))
# normalize the text embedding
text_embedding = text_embedding / text_embedding.norm(dim=-1, keepdim=True)
# forwarding samples to the Relation module
text_embedding = text_embedding.unsqueeze(0).repeat((opt.batch_size)*2,1,1).to(device)
fea_embedding = fea_embedding.unsqueeze(0).repeat(num_category,1,1)
fea_embedding = torch.transpose(fea_embedding,0,1).to(device)
relation_pairs = torch.cat((text_embedding.float(),fea_embedding.float()),2).view(-1,1536)
relations = relation(relation_pairs.float()).view(-1, num_category).to(device)
# cllculate the loss
one_hot_labels = (torch.zeros((opt.batch_size)*2, num_category).to(device).scatter_(1, target.long().view(-1,1), 1))
loss_t = cross_entrpy(relations, one_hot_labels)
loss = loss_t + loss_orthogonal * loss_orthogonal_weight + loss_gradient
loss.backward(retain_graph=True)
optimizer.step()
# Calculating the accuracy
train_loss += loss.clone().detach().item()
prediction = relations.cpu().detach().numpy()
prediction = np.argmax(prediction, axis=1)
target = target.cpu().detach().numpy()
train_total += target.shape[0]
train_correct += np.sum(prediction == target)
# delete the variables to free the memory
del points, target, depth_map, img_embedding, text_embedding, loss
torch.cuda.empty_cache()
print('Relation Module','Point embedding + img _embedding:',loss_orthogonal_weight, 'number of view', num_rotations)
print(f"=> Epoch {epoch} loss: {train_loss:.2f} accuracy: {100 * train_correct / train_total:.2f}")
# evaluate the model
base_class_correct = 0
base_class_total = 0
for i in range(num_rotations):
transform[format(i)].eval()
relation.eval()
unet.eval()
clip_model.eval()
curvenet.eval()
#load the text features
prompts_test = read_txt_file("class_name_shapenet.txt")
text = open_clip.tokenize(prompts_test)
text_embedding_all_classes = clip_model.encode_text(text.to(device))
task1, task2, task3, task4, task5, task6, task7, task1_total, task2_total, task3_total, task4_total, task5_total, task6_total, task7_total = [0] * 12
tid_all = []
for h in range(t+1):
tid_all += tid[h]
for j, data in tqdm(enumerate(testDataLoader, 0)):
points, target = data['pointclouds'].to(device).float(), data['labels'].to(device)
points, target = points.to(device), target.to(device)
features_2D = torch.zeros((1, 512), device=device)
with torch.no_grad():
depth_map = torch.zeros((points.shape[0] * num_rotations, 3, 110, 110)).to(device)
# Forward samples to the PointNet model
points = points.transpose(2, 1)
points = points.repeat(2, 1, 1)
points_embedding = curvenet(points)
# transformation module
trans = torch.zeros((points.shape[0], num_rotations, 3, 3), device=device)
for jj in range(num_rotations):
trans[:, jj, :, :] = transform[format(jj)](points)
# depth map generation
points = points.transpose(2, 1)
depth_map = torch.zeros((points.shape[0] * num_rotations, 3, 224, 224)).to(device)
for jj in range(num_rotations):
depth_map_tmp = proj.get_img(points, trans[:,jj,:,:].view(-1, 9))
depth_map_tmp = torch.nn.functional.interpolate(depth_map_tmp, size=(224, 224), mode='bilinear', align_corners=True)
depth_map[jj * points.shape[0]:(jj + 1) * points.shape[0], :, :, :] = depth_map_tmp
RGB_map = torch.zeros((points.shape[0] * num_rotations, 3, 224, 224)).to(device)
for jj in range(num_rotations):
# unet model
depth_map_reverse = 1 - depth_map[jj * points.shape[0]:(jj + 1) * points.shape[0]]
mask = (depth_map_reverse != 0).float()
texture_map = unet(mask)
RGB_map[jj * points.shape[0]:(jj + 1) * points.shape[0], :, :, :] = depth_map[jj * points.shape[0]:(jj + 1) * points.shape[0]] * texture_map
# Forward samples to the CLIP model
img_embedding_tmp = clip_model.encode_image(RGB_map).to(device)
img_embedding = 0
for jj in range(num_rotations):
img_embedding += img_embedding_tmp[jj * points.shape[0]:(jj + 1) * points.shape[0], :]/ num_rotations
# merge img_embedding and points_embedding
img_embedding = img_embedding / img_embedding.norm(dim=-1, keepdim=True)
points_embedding = points_embedding / points_embedding.norm(dim=-1, keepdim=True)
fea_embedding = torch.cat((img_embedding, points_embedding), 1)
fea_embedding = fea_embedding[0,:].unsqueeze(0)
# Forward samples to the text CLIP model
text_embedding = text_embedding_all_classes[tid_all].to(device)
text_embedding = text_embedding / text_embedding.norm(dim=-1, keepdim=True)
# forwarding samples to the Relation module
text_embedding = text_embedding.unsqueeze(0).repeat(1,1,1).to(device)
fea_embedding = fea_embedding.unsqueeze(0).repeat(num_category,1,1).to(device)
fea_embedding = torch.transpose(fea_embedding,0,1).to(device)
relation_pairs = torch.cat((text_embedding.float(),fea_embedding.float()),2).view(-1,1536)
relations = relation(relation_pairs.float()).view(-1, num_category).to(device)
prediction = relations.cpu().detach().numpy()
prediction = np.argmax(prediction, axis=1)
if prediction == target.cpu().detach().numpy():
base_class_correct += 1
if prediction == target.cpu().detach().numpy() and target.cpu().detach().numpy() < 25:
task1 += 1
if prediction == target.cpu().detach().numpy() and target.cpu().detach().numpy() >= 25 and target.cpu().detach().numpy() < 30:
task2 += 1
if prediction == target.cpu().detach().numpy() and target.cpu().detach().numpy() >= 30 and target.cpu().detach().numpy() < 35:
task3 += 1
if prediction == target.cpu().detach().numpy() and target.cpu().detach().numpy() >= 35 and target.cpu().detach().numpy() < 40:
task4 += 1
if prediction == target.cpu().detach().numpy() and target.cpu().detach().numpy() >= 40 and target.cpu().detach().numpy() < 45:
task5 += 1
if prediction == target.cpu().detach().numpy() and target.cpu().detach().numpy() >= 45 and target.cpu().detach().numpy() < 50:
task6 += 1
if prediction == target.cpu().detach().numpy() and target.cpu().detach().numpy() >= 50 and target.cpu().detach().numpy() < 55:
task7 += 1
# tasks total number samples
if target.cpu().detach().numpy() < 25:
task1_total += 1
if target.cpu().detach().numpy() >= 25 and target.cpu().detach().numpy() < 30:
task2_total += 1
if target.cpu().detach().numpy() >= 30 and target.cpu().detach().numpy() < 35:
task3_total += 1
if target.cpu().detach().numpy() >= 35 and target.cpu().detach().numpy() < 40:
task4_total += 1
if target.cpu().detach().numpy() >= 40 and target.cpu().detach().numpy() < 45:
task5_total += 1
if target.cpu().detach().numpy() >= 45 and target.cpu().detach().numpy() < 50:
task6_total += 1
if target.cpu().detach().numpy() >= 50 and target.cpu().detach().numpy() < 55:
task7_total += 1
acc = (base_class_correct / testDataLoader.__len__()) * 100
if task1_total > 0:
print('task1:', task1/task1_total)
if task2_total > 0:
print('task2:', task2/task2_total)
if task3_total > 0:
print('task3:', task3/task3_total)
if task4_total > 0:
print('task4:', task4/task4_total)
if task5_total > 0:
print('task5:', task5/task5_total)
if task6_total > 0:
print('task6:', task6/task6_total)
if task7_total > 0:
print('task7:', task7/task7_total)
print(f"=> total accuracy: {acc:.2f}")
print('-------------------------------------------------------------------------')
# put the models in the training mode
for i in range(num_rotations):
transform[format(i)].train()
relation.train()
unet.train()
clip_model.train()
curvenet.train()
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--batch_size', type=int, default= 16, help='input batch size')
parser.add_argument('--num_points', type=int, default=2048, help='number of points in each input point cloud')
parser.add_argument('--workers', type=int, help='number of data loading workers', default=4)
parser.add_argument('--nepoch', type=int, default=14, help='number of epochs to train for')
parser.add_argument('--outf', type=str, default='cls', help='output folder to save results')
parser.add_argument('--model', type=str, default='cls/3D_model_249.pth', help='path to load a pre-trained model')
parser.add_argument('--feature_transform', action='store_true', help='use feature transform')
parser.add_argument('--manualSeed', type=int, default = 42, help='random seed')
parser.add_argument('--dataset_path', type=str, default= 'dataset/FSCIL/shapenet/', help="dataset path")
parser.add_argument('--ntasks', type=str, default= '5', help="number of tasks")
parser.add_argument('--nclasses', type=str, default= '25', help="number of classes")
parser.add_argument('--task', type=str, default= '0', help="task number")
parser.add_argument('--num_samples', type=str, default= '0', help="number of samples per class")
parser.add_argument('--process_data', action='store_true', default=False, help='save data offline')
parser.add_argument('--num_point', type=int, default=2048, help='Point Number')
parser.add_argument('--use_uniform_sample', action='store_true', default=False, help='use uniform sampiling')
parser.add_argument('--use_normals', action='store_true', default=False, help='use normals')
parser.add_argument('--num_category', default=25, type=int, choices=[20, 40], help='training on ModelNet10/40')
parser.add_argument('--sem_file', default=None, help='training on ModelNet10/40')
parser.add_argument('--use_memory', default=False, help='use_memory')
parser.add_argument('--herding', default=True, help='herding')
opt = parser.parse_args()
main(opt)
print("Done!")