-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
337 lines (301 loc) · 16.7 KB
/
train.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
from time import time
from tqdm import tqdm
import torch
import torch.nn
import torch.optim
import numpy as np
import data
import argparse
import math
from torch.utils.data import DataLoader, Subset
import os
import yaml
from architecture import get_model_by_params
from datetime import date
import pprint
from torchvision import transforms
from PIL import ImageFile
import matplotlib.pyplot as plt
import socket
from autoencoder import AutoEncoder
def parse_yaml(file_path: str, create_folder: bool = True) -> dict:
"""
Create dictionary from yaml file
:param file_path: path of params yaml file
:param create_folder: create folder for saved_models, in case has not been created yet
:return: dictionary of model parameters and training progress, which were loaded from saved_models folder
"""
print("Reading paramfile {}".format(file_path))
try:
with open(file_path) as f:
param = yaml.load(f, Loader=yaml.FullLoader)
except:
raise (RuntimeError("Could not load model parameters from " + file_path + "."))
if create_folder:
# Find save directory
if not os.path.exists("saved_models"):
os.mkdir("saved_models")
save_dir = os.path.join("saved_models", str(param["model_name"]) + "_" + "".join(str(date.today()).split("-")[1:]))
save_dir_mod = save_dir + "_0"
i = 1
while os.path.exists(save_dir_mod):
save_dir_mod = save_dir + "_" + str(i)
i += 1
os.mkdir(save_dir_mod)
param["save_dir"] = save_dir_mod
# Copy yaml to training dir
with open(os.path.join (save_dir_mod, 'modelcopy.yaml'), 'w') as outfile:
yaml.dump(param, outfile, default_flow_style=True)
else:
raise RuntimeError("No folders can be created in read-only mode.")
return param
def get_transform():
return transforms.Resize((64, 64))
def create_dataloaders(data_path, batch_size, test_ratio, split=None, only_classes=None, only_one_sample=False, load_on_request=False, add_data=None, p=1, bw=False, color=False):
"""
Create data loaders from ImageDataSet according parameters. If split is provided, it is used by the data loader.
:param data_path: path of root directory of data set, while directories 'photo' and 'sketch' are sub directories
:param batch_size:
:param test_ratio: 0.1 means 10% test data
:param split: optional train/test split
:param only_classes: optional list of folder names to retrieve training data from
:param only_one_sample: Load only one sketch and one image
:param num_workers: number of workers threads for loading sketches and images from drive
:return: train and test dataloaders and train and test split
"""
if bw and color:
raise(RuntimeError("Can't do both black and white and coloring at the same time."))
if only_one_sample:
test_ratio = 0.5
batch_size = 1
data_set = data.ImageDataSet(root_dir=data_path, transform=get_transform(), only_classes=only_classes, only_one_sample=only_one_sample, load_on_request=load_on_request, bw=bw, color=color)
if split is None:
perm = torch.randperm(len(data_set))
train_split, test_split = perm[:math.ceil(len(data_set) * (1-test_ratio))], perm[math.ceil(len(data_set) * (1-test_ratio)):]
else:
train_split, test_split = split[0], split[1]
if add_data:
data_set_add = data.ImageDataSet(root_dir=add_data, transform=get_transform(), only_classes=only_classes, only_one_sample=only_one_sample, load_on_request=load_on_request, bw=bw, color=color)
dataloader_train = data.CompositeDataloader(DataLoader(Subset(data_set, train_split), batch_size=batch_size, shuffle=True, num_workers=0, drop_last=True),
DataLoader(data_set_add, batch_size=batch_size, shuffle=True, num_workers=0, drop_last=True),
p=p,
anneal_rate=0.99)
else:
dataloader_train = DataLoader(Subset(data_set, train_split), batch_size=batch_size, shuffle=True, num_workers=0, drop_last=True)
dataloader_train.p = p
dataloader_train.anneal_p = lambda *args : None
dataloader_test = DataLoader(Subset(data_set, test_split), batch_size=batch_size, shuffle=False, num_workers=0)
return dataloader_train, dataloader_test, train_split, test_split
def get_optimizer(param, trainables):
optimizer = getattr(torch.optim, param.get("optimizer"))(trainables, **param.get("optimizer_params"))
scheduler = torch.optim.lr_scheduler.ExponentialLR(
optimizer,
gamma=0.98
)
return scheduler, optimizer
def get_model(param):
return get_model_by_params(param).to(device)#cINN(**param.get("model_params")).to(device)
def load_state(param):
"""
Provide model, optimizer, epoch and split obejcts from given parameter dictionary
:param param: parameter dictionary of params yaml file
:return: model, optimizer, epoch, data split
"""
model = get_model(param)
scheduler, optimizer = get_optimizer(param, model.parameters())
try:
print("saved_models/" + param.get("model_name")+f"/{param.get('model_name')}.tar")
state_dicts = torch.load("saved_models/" + param.get("model_name")+f"/{param.get('model_name')}.tar", map_location=device)
except:
raise (RuntimeError("Could not load training state parameters for " + param.get('model_name') + "."))
model.model.load_state_dict(state_dicts["model_state_dict"])
optimizer.load_state_dict(state_dicts["optimizer_state_dict"])
epoch = state_dicts["epoch"] + 1
opt_param = param.get("optimizer_params")
if not opt_param.get("weight_decay") is None and not opt_param.get("weight_decay") == optimizer.param_groups[0]["weight_decay"]:
optimizer.param_groups[0]["weight_decay"] = opt_param["weight_decay"]
if not opt_param.get("lr") is None and not opt_param.get("lr") == optimizer.param_groups[0]["lr"]:
optimizer.param_groups[0]["lr"] = opt_param["lr"]
split = (state_dicts["train_split"], state_dicts["test_split"])
try:
scheduler.load_state_dict(state_dicts["scheduler_state_dict"])
except:
print("Warning, could not load scheduler state dict, continuing with default values")
try:
p = state_dicts["prob_data1"]
except:
p=param.get("prob_data1")
return model, optimizer, epoch, split, scheduler, p
def save_state(param, model_state, optim_state, scheduler_state, epoch, running_loss, split, p, overwrite_chkpt=True):
"""
Save state of training into yaml file in folder saved_models
:param param: dictionary of used parameters
:param overwrite_chkpt: If true, save to file without _epoch extension. Otherwise save to file with _epoch extension.
:return:
"""
ImageFile.LOAD_TRUNCATED_IMAGES = True
if not overwrite_chkpt:
path = os.path.join(param["save_dir"], param["model_name"] + "_" + str(epoch))
else:
path = os.path.join(param["save_dir"], param["model_name"])
torch.save({
'epoch': epoch,
'model_state_dict': model_state,
'optimizer_state_dict': optim_state,
'loss': running_loss,
'train_split': split[0],
'test_split': split[1],
'model_params': param['model_params'],
'batch_size': param['batch_size'],
'test_ratio': param['test_ratio'],
'only_classes': param.get('only_classes', None),
'only_one_sample': param.get('only_one_sample', False),
'scheduler_state_dict': scheduler_state,
'architecture': param.get("architecture"),
'data_path': param.get("data_path"),
'prob_data1': p
}, f"{path}.tar")
def validate(model, dataloader_test):
model.eval()
with torch.no_grad():
val_loss = 0
for batch, (sketch, real, label) in enumerate(tqdm(dataloader_test)):
sketch, real, label = sketch.to(device), real.to(device), label.to(device)
if params.get("half_precision", False):
sketch, real = sketch.half(), real.half()
gauss_output = model(real, sketch)
loss = torch.mean(gauss_output**2/2) - torch.mean(model.log_jacobian()) / gauss_output.shape[1]
val_loss += loss/len(dataloader_test)
model.train()
return val_loss
def train_ae(encoder_sizes, decoder_sizes, train_loader, num_epochs=1, bw=False):
"""
Pre-train Autoencoder on condition sketches
:return trained encoder nn.Module
"""
AE = AutoEncoder(encoder_sizes, decoder_sizes, bw=bw).to(device)
optimizer = torch.optim.Adam(AE.parameters(), lr=0.0005, weight_decay=3e-6)
for epoch in tqdm(range(num_epochs), "Encoder pretraining"):
epoch_loss = 0
for batch, (sketch, real, label) in enumerate(tqdm(dataloader_train, "Batch")):
optimizer.zero_grad()
sketch, real, label = sketch.to(device), real.to(device), label.to(device)
recon = AE(sketch)
loss = torch.mean((sketch - recon) ** 2)
loss.backward()
epoch_loss += loss.item()/len(dataloader_train)
optimizer.step()
print("AutoEncoder pretraining: Epoch {} Loss: {}".format(epoch, epoch_loss))
del AE.decoder
optimizer.zero_grad()
del optimizer
return AE.encoder
if __name__ == "__main__":
# Determine, whether cuda will be enabled
# use x.to_device(args.device)
parser = argparse.ArgumentParser(description='PyTorch')
parser.add_argument('modelnames', nargs='+', help='model names to train')
parser.add_argument('--nocuda', help='Disable CUDA', action='store_true')
parser.add_argument('--nopreload', help='Disable pre-loading of image data', action='store_true')
parser.add_argument('--nocheckpoints', action='store_true',
help='Disable storing training state checkpoints. Model will be saved at the end of training.')
args = parser.parse_args()
device = None
if not args.nocuda and torch.cuda.is_available():
device = torch.device('cuda')
print("CUDA enabled.")
else:
device = torch.device('cpu')
print("CUDA disabled.")
# Define dictionary of hyper parameters
if not args.modelnames is None:
list_hyper_params = args.modelnames
else:
print("No model name specified in command line arguments. Will use hard-coded model list...")
list_hyper_params = ["clamp_glow"]
# Loop over hyper parameter configurations
pp = pprint.PrettyPrinter(indent=4)
for param_name in list_hyper_params:
if not param_name.endswith(".yaml"):
param_name = "{}.yaml".format(param_name)
params = parse_yaml(os.path.join("params", param_name))
if not 'only_one_sample' in params:
params['only_one_sample'] = False
if args.nocheckpoints:
print("No checkpoint mode active. No checkpoint is created after every batch. Model will be saved to {} at the end of trainig.".format(params["save_dir"]))
if params['only_one_sample']:
print("!!!!!!!!!!!!!!!! ONLY ONE SAMPLE MODE IS ACITVE !!!!!!!!!!!!!!!!")
if not params.get("data_path", False):
params["data_path"] = "dataset/SketchyDatabase/256x256"
if params.get("load_model", False):
# Load training progress from existing split
model, optimizer, epoch, split, scheduler, p = load_state(params)
dataloader_train, dataloader_test, train_split, test_split = create_dataloaders(params["data_path"],
params["batch_size"],
params["test_ratio"],
split=split,
only_classes=params.get('only_classes', None),
only_one_sample=params.get('only_one_sample', False),
load_on_request=args.nopreload,
add_data=params.get("add_data"),
p=p,
bw=params["model_params"].get("bw"),
color=params["model_params"].get("color"))
else:
# Init new training with new split
model = get_model(params)
scheduler, optimizer = get_optimizer(params, model.parameters())
epoch = 0
dataloader_train, dataloader_test, train_split, test_split = create_dataloaders(params["data_path"],
params["batch_size"],
params["test_ratio"],
only_classes=params.get('only_classes', None),
only_one_sample=params.get('only_one_sample', False),
load_on_request=args.nopreload,
add_data=params.get("add_data"),
p=params.get("prob_data1"),
bw=params["model_params"].get("bw"),
color=params["model_params"].get("color"))
split = (train_split, test_split)
if params.get("half_precision", False):
model.half()
t_start = time()
loss_summary = np.zeros(0)
print(socket.gethostname())
print("Starting training for params:")
pp.pprint(params)
if params["model_params"].get("pretrain_cond", False) and not params.get("load_model", False):
del model.cond_net
model.cond_net = train_ae(params["model_params"]["encoder_sizes"], params["model_params"]["decoder_sizes"], dataloader_train, bw = params['model_params'].get("bw"))
for e in range(params["n_epochs"]):
epoch += 1
epoch_loss = 0
for batch, (sketch, real, label) in enumerate(tqdm(dataloader_train)):
optimizer.zero_grad()
sketch, real, label = sketch.to(device), real.to(device), label.to(device)
if params.get("half_precision", False):
sketch, real = sketch.half(), real.half()
gauss_output = model(real, sketch)
loss = torch.mean(gauss_output**2/2) - torch.mean(model.log_jacobian()) / gauss_output.shape[1]
loss.backward()
epoch_loss += loss.item()/len(dataloader_train)
loss_summary = np.append(loss_summary, loss.item())
optimizer.step()
dataloader_train.anneal_p()
scheduler.step()
np.savetxt(os.path.join(params["save_dir"], 'summary_{}_epoch{}'.format(params["model_name"], str(epoch))), loss_summary, fmt='%1.3f')
print("Epoch {} / {} Training Loss: {}, Validation Loss: {}".format(e + 1, params["n_epochs"], epoch_loss, validate(model, dataloader_test)))
if not args.nocheckpoints:
save_state(params, model.model.state_dict(), optimizer.state_dict(), scheduler.state_dict(), epoch, epoch_loss, split, dataloader_train.p)
np.savetxt(os.path.join(params["save_dir"], 'summary_{}_epoch{}_FINAL'.format(params["model_name"], str(epoch))),
loss_summary, fmt='%1.3f')
f = plt.figure()
plt.plot(loss_summary)
plt.xlabel('Batch')
plt.ylabel('Batch Loss')
plt.savefig(os.path.join(params["save_dir"], 'batchloss_{}.pdf'.format(params["model_name"]) ))
plt.close()
if args.nocheckpoints:
save_state(params, model.model.state_dict(), optimizer.state_dict(), scheduler.state_dict(), epoch, epoch_loss, split, dataloader_train.p)
print("Model is saved to {}".format(params["save_dir"]))
print('%.3i \t%.6f min' % (epoch, (time() - t_start) / 60.))