-
Notifications
You must be signed in to change notification settings - Fork 36
/
train_post_clip.py
374 lines (290 loc) · 17.5 KB
/
train_post_clip.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
import os
import os.path as osp
import logging
import argparse
from tqdm import tqdm
from sklearn.metrics import accuracy_score, confusion_matrix
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
from mpl_toolkits.mplot3d import Axes3D
import torch
from torchvision import transforms
from torch.optim import lr_scheduler
import torch.nn.functional as F
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
from utils import helper
from utils import visualization
from dataset import shapenet_dataset
from train_autoencoder import experiment_name, parsing
from networks import autoencoder, latent_flows
import clip
###################################### Experiment Utils########################################################
def experiment_name2(args):
tokens = ["Clip_Conditioned", args.flow_type, args.num_blocks, args.checkpoint, args.num_views, args.clip_model_type, args.num_hidden, args.seed_nf]
if args.noise != "add":
tokens.append("no_noise")
return "_".join(map(str, tokens))
def get_clip_model(args):
if args.clip_model_type == "B-16":
print("Bigger model is being used B-16")
clip_model, clip_preprocess = clip.load("ViT-B/16", device=args.device)
cond_emb_dim = 512
elif args.clip_model_type == "RN50x16":
print("Using the RN50x16 model")
clip_model, clip_preprocess = clip.load("RN50x16", device=args.device)
cond_emb_dim = 768
else:
clip_model, clip_preprocess = clip.load("ViT-B/32", device=args.device)
cond_emb_dim = 512
input_resolution = clip_model.visual.input_resolution
#train_cond_embs_length = clip_model.train_cond_embs_length
vocab_size = clip_model.vocab_size
#cond_emb_dim = clip_model.embed_dim
#print("Model parameters:", f"{np.sum([int(np.prod(p.shape)) for p in clip_model.parameters()]):,}")
print("cond_emb_dim:", cond_emb_dim)
print("Input resolution:", input_resolution)
#print("train_cond_embs length:", train_cond_embs_length)
print("Vocab size:", vocab_size)
args.n_px = input_resolution
args.cond_emb_dim = cond_emb_dim
return args, clip_model
###################################### Experiment Utils########################################################
############################################# data loader #################################################
def get_dataloader(args, split="train", dataset_flag=False):
dataset_name = args.dataset_name
if dataset_name == "Shapenet":
pointcloud_field = shapenet_dataset.PointCloudField("pointcloud.npz")
points_field = shapenet_dataset.PointsField("points.npz", unpackbits=True)
voxel_fields = shapenet_dataset.VoxelsField("model.binvox")
if split == "train":
image_field = shapenet_dataset.ImagesField("img_choy2016", random_view=True, n_px=args.n_px)
else:
image_field = shapenet_dataset.ImagesField("img_choy2016", random_view=False, n_px=args.n_px)
fields = {}
fields['pointcloud'] = pointcloud_field
fields['points'] = points_field
fields['voxels'] = voxel_fields
fields['images'] = image_field
def my_collate(batch):
batch = list(filter(lambda x : x is not None, batch))
return torch.utils.data.dataloader.default_collate(batch)
if split == "train":
dataset = shapenet_dataset.Shapes3dDataset(args.dataset_path, fields, split=split,
categories=args.categories, no_except=True, transform=None, num_points=args.num_points)
dataloader = DataLoader(dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers, drop_last=True, collate_fn=my_collate)
total_shapes = len(dataset)
else:
dataset = shapenet_dataset.Shapes3dDataset(args.dataset_path, fields, split=split,
categories=args.categories, no_except=True, transform=None, num_points=args.num_points)
dataloader = DataLoader(dataset, batch_size=args.test_batch_size, shuffle=True, num_workers=args.num_workers, drop_last=False, collate_fn=my_collate)
total_shapes = len(dataset)
if dataset_flag == True:
return dataloader, total_shapes, dataset
return dataloader, total_shapes
else:
raise ValueError("Dataset name is not defined {}".format(dataset_name))
######################################## data loader ########################################
######################################## Pre-compute stuff ########################################
#### Get the clip embedding and shape embedding. Done to be more efficent
def get_condition_embeddings(args, model, clip_model, dataloader, times=5):
model.eval()
clip_model.eval()
shape_embeddings = []
cond_embeddings = []
with torch.no_grad():
for i in range(0, times):
for data in tqdm(dataloader):
pc = data['pc_org'].type(torch.FloatTensor).to(args.device)
query_points, occ = data['points'], data['points.occ']
data_index = data['idx'].to(args.device)
image = data['images'].type(torch.FloatTensor).to(args.device)
query_points = query_points.type(torch.FloatTensor).to(args.device)
occ = occ.type(torch.FloatTensor).to(args.device)
if args.input_type == "Voxel":
data_input = data['voxels'].type(torch.FloatTensor).to(args.device)
elif args.input_type == "Pointcloud":
data_input = data['pc_org'].type(torch.FloatTensor).to(args.device).transpose(-1, 1)
shape_emb = model.encoder(data_input)
image_features = clip_model.encode_image(image)
image_features = image_features / image_features.norm(dim=-1, keepdim=True)
shape_embeddings.append(shape_emb.detach().cpu().numpy())
cond_embeddings.append(image_features.detach().cpu().numpy())
#break
logging.info("Number of views done: {}/{}".format(i, times))
shape_embeddings = np.concatenate(shape_embeddings)
cond_embeddings = np.concatenate(cond_embeddings)
return shape_embeddings, cond_embeddings
######################################## Pre-compute stuff ########################################
###################################### Generating stuff ###############################################
def generate_on_query_text(args, clip_model, autoencoder, latent_flow_model):
autoencoder.eval()
latent_flow_model.eval()
clip_model.eval()
save_loc = args.generate_dir + "/"
count = 1
num_figs = 3
with torch.no_grad():
voxel_size = 32
shape = (voxel_size, voxel_size, voxel_size)
p = visualization.make_3d_grid([-0.5] * 3, [+0.5] * 3, shape).type(torch.FloatTensor).to(args.device)
query_points = p.expand(num_figs, *p.size())
for text_in in args.text_query:
text = clip.tokenize([text_in]).to(args.device)
text_features = clip_model.encode_text(text)
text_features = text_features / text_features.norm(dim=-1, keepdim=True)
noise = torch.Tensor(num_figs, args.emb_dims).normal_().to(args.device)
decoder_embs = latent_flow_model.sample(num_figs, noise=noise, cond_inputs=text_features.repeat(num_figs,1))
out = autoencoder.decoding(decoder_embs, query_points)
if args.output_type == "Implicit":
voxels_out = (out.view(num_figs, voxel_size, voxel_size, voxel_size) > args.threshold).detach().cpu().numpy()
visualization.multiple_plot_voxel(voxels_out, save_loc=save_loc +"{}_text_query.png".format(text_in))
elif args.output_type == "Pointcloud":
pred = out.detach().cpu().numpy()
visualization.multiple_plot(pred, save_loc=save_loc +"{}_text_query.png".format(text_in))
latent_flow_model.train()
###################################### Generating stuff ###############################################
###################################### train and validation ###########################################
def train_one_epoch(args, latent_flow_model, train_dataloader, optimizer, epoch):
loss_prob_array = []
loss_array = []
latent_flow_model.train()
for data in train_dataloader:
optimizer.zero_grad()
train_embs, train_cond_embs = data
train_embs = train_embs.type(torch.FloatTensor).to(args.device)
train_cond_embs = train_cond_embs.type(torch.FloatTensor).to(args.device)
if args.noise == "add":
train_embs = train_embs + 0.1 * torch.randn(train_embs.size(0), args.emb_dims).to(args.device)
loss_log_prob = - latent_flow_model.log_prob(train_embs, train_cond_embs).mean()
loss = loss_log_prob
loss.backward()
optimizer.step()
loss_array.append(loss.item())
loss_prob_array.append(loss_log_prob.item())
loss_array = np.asarray(loss_array)
loss_prob_array = np.asarray(loss_prob_array)
logging.info("[Train] Epoch {} Train loss {} Prob loss {} ".format(epoch, np.mean(loss_array), np.mean(loss_prob_array)))
def val_one_epoch(args, latent_flow_model, val_dataloader, epoch):
loss_prob_array = []
loss_array = []
latent_flow_model.eval()
with torch.no_grad():
for data in val_dataloader:
train_embs, train_cond_embs = data
train_embs = train_embs.type(torch.FloatTensor).to(args.device)
train_cond_embs = train_cond_embs.type(torch.FloatTensor).to(args.device)
loss_log_prob = - latent_flow_model.log_prob(train_embs, train_cond_embs).mean()
loss = loss_log_prob
loss_array.append(loss.item())
loss_prob_array.append(loss_log_prob.item())
loss_array = np.asarray(loss_array)
loss_prob_array = np.asarray(loss_prob_array)
logging.info("[VAL] Epoch {} Train loss {} Prob loss {} ".format(epoch, np.mean(loss_array), np.mean(loss_prob_array)))
return np.mean(loss_array)
###################################### train and validation ###########################################
######################################## main and parser stuff ##########################################
def get_local_parser(mode="args"):
parser = parsing(mode="parser")
parser.add_argument("--num_blocks", type=int, default=5, help='Num of blocks for prior')
parser.add_argument("--flow_type", type=str, default='realnvp_half', help='flow type: mf, glow, realnvp ')
parser.add_argument("--num_hidden", type=int, default=1024, help='Number of parameter for flow model')
parser.add_argument("--latent_load_checkpoint", type=str, default=None, help='Checkpoint to load latent flow model')
parser.add_argument("--text_query", nargs='+', default=None, metavar='N', help='text query array')
parser.add_argument("--num_views", type=int, default=5, metavar='N', help='Number of views')
parser.add_argument("--clip_model_type", type=str, default='B-32', metavar='N', help='what model to use')
parser.add_argument("--noise", type=str, default='add', metavar='N', help='add or remove')
parser.add_argument("--seed_nf", type=int, default=1, metavar='N', help='add or remove')
parser.add_argument("--images_type", type=str, default=None, help='img_choy13 or img_custom')
parser.add_argument("--n_px", type=int, default=224, help='Resolution of the image')
if mode == "args":
args = parser.parse_args()
return args
else:
return parser
def main():
args = get_local_parser()
exp_name = experiment_name(args)
exp_name_2 = experiment_name2(args)
manualSeed = args.seed_nf
helper.set_seed(manualSeed)
# Create directories for checkpoints and logging
log_filename = osp.join('exps', exp_name, exp_name_2, 'log.txt')
args.experiment_dir = osp.join('exps', exp_name, exp_name_2)
args.experiment_dir_base = osp.join('exps', exp_name)
args.checkpoint_dir = osp.join('exps', exp_name, exp_name_2, 'checkpoints')
args.checkpoint_dir_base = osp.join('exps', exp_name, 'checkpoints')
args.vis_dir = osp.join('exps', exp_name, exp_name_2, 'vis_dir') + "/"
args.generate_dir = osp.join('exps', exp_name, exp_name_2, 'generate_dir') + "/"
helper.create_dir(args.checkpoint_dir)
helper.create_dir(args.vis_dir)
helper.create_dir(args.generate_dir)
if args.train_mode != "test":
helper.setup_logging(log_filename, args.log_level, 'w')
else:
test_log_filename = osp.join('exps', exp_name, exp_name_2, 'test_log.txt')
helper.setup_logging(test_log_filename, args.log_level, 'w')
args.query_generate_dir = osp.join('exps', exp_name, exp_name_2, 'query_generate_dir') + "/"
helper.create_dir(args.query_generate_dir)
args.vis_gen_dir = osp.join('exps', exp_name, exp_name_2, 'vis_gen_dir') + "/"
helper.create_dir(args.vis_gen_dir)
logging.info("Experiment name: {} and Experiment name 2 {}".format(exp_name, exp_name_2))
logging.info("{}".format(args))
device, gpu_array = helper.get_device(args)
args.device = device
args, clip_model = get_clip_model(args)
logging.info("#############################")
train_dataloader, total_shapes = get_dataloader(args, split="train")
args.total_shapes = total_shapes
logging.info("Train Dataset size: {}".format(total_shapes))
val_dataloader, total_shapes_val = get_dataloader(args, split="val")
logging.info("Test Dataset size: {}".format(total_shapes_val))
logging.info("#############################")
net = autoencoder.get_model(args).to(args.device)
checkpoint = torch.load(args.checkpoint_dir_base +"/"+ args.checkpoint +".pt", map_location=args.device)
net.load_state_dict(checkpoint['model'])
net.eval()
logging.info("#############################")
logging.info("Getting train shape embeddings and condition embedding")
train_shape_embeddings, train_cond_embeddings = get_condition_embeddings(args, net, clip_model, train_dataloader, times=args.num_views)
logging.info("Train Embedding Shape {}, Train Condition Embedding {}".format(train_shape_embeddings.shape, train_cond_embeddings.shape))
train_dataset_new = torch.utils.data.TensorDataset(torch.from_numpy(train_shape_embeddings), torch.from_numpy(train_cond_embeddings))
train_dataloader_new = DataLoader(train_dataset_new, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers, drop_last=True)
logging.info("Getting val shape embeddings and condition embedding")
val_shape_embeddings, val_cond_embeddings = get_condition_embeddings(args, net, clip_model, val_dataloader, times=1)
logging.info("Val Embedding Shape {}, Val Condition Embedding {}".format(val_shape_embeddings.shape, val_cond_embeddings.shape))
val_dataset_new = torch.utils.data.TensorDataset(torch.from_numpy(val_shape_embeddings), torch.from_numpy(val_cond_embeddings))
val_dataloader_new = DataLoader(val_dataset_new, batch_size=args.test_batch_size, shuffle=True, num_workers=args.num_workers, drop_last=False)
logging.info("#############################")
latent_flow_network = latent_flows.get_generator(args.emb_dims, args.cond_emb_dim, device, flow_type=args.flow_type, num_blocks=args.num_blocks, num_hidden=args.num_hidden)
if args.train_mode == "test":
pass
else:
optimizer = torch.optim.Adam(latent_flow_network.parameters(), lr=0.00003)
scheduler = lr_scheduler.CosineAnnealingLR(optimizer, args.num_iterations, 0.000001)
start_epoch = 0
if args.latent_load_checkpoint is not None:
checkpoint_dir = args.new_checkpoint_dir + "/{}.pt".format(args.latent_load_checkpoint)
checkpoint = torch.load(checkpoint_dir, map_location=args.device)
latent_flow_network.load_state_dict(checkpoint['model'])
start_epoch = checkpoint['current_epoch']
best_loss = 100000
for epoch in range(start_epoch, args.epochs):
logging.info("#############################")
if (epoch + 1) % 5 == True:
if args.text_query is not None:
generate_on_query_text(args, clip_model, net, latent_flow_network)
train_one_epoch(args, latent_flow_network, train_dataloader_new, optimizer, epoch)
val_loss = val_one_epoch(args, latent_flow_network, val_dataloader_new, epoch)
filename = '{}.pt'.format(args.checkpoint_dir + "/last")
logging.info("Saving Model........{}".format(filename))
torch.save({'model': latent_flow_network.state_dict(), 'args': args, "current_epoch": epoch}, '{}'.format(filename))
if best_loss > val_loss:
best_loss = val_loss
filename = '{}.pt'.format(args.checkpoint_dir + "/best")
logging.info("Saving Model........{}".format(filename))
torch.save({'model': latent_flow_network.state_dict(), 'args': args, "current_epoch": epoch}, '{}'.format(filename))
if __name__ == "__main__":
main()