-
Notifications
You must be signed in to change notification settings - Fork 17
/
Copy pathDDFFTrainer.py
214 lines (194 loc) · 9.99 KB
/
DDFFTrainer.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
#! /usr/bin/python3
import os
import numpy as np
import torch
import torch.nn as nn
import torchvision
from torch import optim
from torch.utils.data import DataLoader
import ddff.models.DDFFNet as DDFFNet
import ddff.dataproviders.datareaders.FocalStackDDFFH5Reader as FocalStackDDFFH5Reader
from ddff.trainers.BaseTrainer import BaseTrainer
class DDFFTrainer(BaseTrainer):
def __init__(self, stack_size, learning_rate=0.001, cliprange=[0.0202, 0.2825],
cc1_enabled=False,
cc2_enabled=False,
cc3_enabled=True,
cc4_enabled=False,
cc5_enabled=False,
pretrained='no_bn',
scheduler_step_size=4,
scheduler_gama=0.9,
deterministic=False,
optimizer='sgd',
normalize_loss=False):
#Define model
net = DDFFNet.DDFFNet(stack_size, cc1_enabled=cc1_enabled, cc2_enabled=cc2_enabled, cc3_enabled=cc3_enabled, cc4_enabled=cc4_enabled, cc5_enabled=cc5_enabled, pretrained=pretrained)
#Define optimizer
if optimizer == 'sgd':
opt = self.create_optimizer(net, {"algorithm":'sgd', "learning_rate":learning_rate, "weight_decay": 0.0005, "momentum":0.9})
else:
opt = self.create_optimizer(net, {"algorithm":'adam', "learning_rate":learning_rate, "weight_decay": 0.0005})
#Define scheduler
scheduler = optim.lr_scheduler.StepLR(opt, step_size=scheduler_step_size, gamma=scheduler_gama)
#Define training loss
training_loss = self.MaskedLoss(nn.MSELoss(reduction="elementwise_mean" if normalize_loss else "sum"), valid_cond=lambda x : x >= cliprange[0])
#Call parent constructor
super(DDFFTrainer, self).__init__(net, opt, training_loss, deterministic, scheduler=scheduler)
@classmethod
def from_h5_data(cls,root_dir,
learning_rate=0.001,
cc1_enabled=False,
cc2_enabled=False,
cc3_enabled=True,
cc4_enabled=False,
cc5_enabled=False,
training_crop_size=None,
validation_crop_size=None,
pretrained='no_bn',
normalize_mean=[0.485, 0.456, 0.406],
normalize_std=[0.229, 0.224, 0.225],
scheduler_step_size=4,
scheduler_gama=0.9,
max_gradient=5.0,
deterministic=False,
optimizer='sgd',
normalize_loss=False,
epochs=20,
batch_size=2,
num_workers=4,
checkpoint_file=None,
checkpoint_frequency=50):
#Create data loaders
transform_train = cls.__create_preprocessing(cls, crop_size=training_crop_size, mean=normalize_mean, std=normalize_std)
transform_validation = cls.__create_preprocessing(cls, crop_size=validation_crop_size, mean=normalize_mean, std=normalize_std)
#Create h5 reader
dataset_train = FocalStackDDFFH5Reader.FocalStackDDFFH5Reader(root_dir, transform=transform_train, stack_key="stack_train", disp_key="disp_train")
dataset_validation = FocalStackDDFFH5Reader.FocalStackDDFFH5Reader(root_dir, transform=transform_validation, stack_key="stack_val", disp_key="disp_val")
#Create data loader
dataloader_train = DataLoader(dataset_train, batch_size=batch_size, shuffle=True, num_workers=num_workers)
dataloader_validation = DataLoader(dataset_validation, batch_size=1, shuffle=True, num_workers=0)
#Call constructor
instance = cls(dataset_train.get_stack_size(), learning_rate=learning_rate,
cc1_enabled=cc1_enabled,
cc2_enabled=cc2_enabled,
cc3_enabled=cc3_enabled,
cc4_enabled=cc4_enabled,
cc5_enabled=cc5_enabled,
pretrained=pretrained,
scheduler_step_size=scheduler_step_size,
scheduler_gama=scheduler_gama,
deterministic=deterministic,
optimizer=optimizer,
normalize_loss=normalize_loss)
#Save instances
instance.dataloader_validation = dataloader_validation
#Load checkpoint if ther already exists a file
if os.path.isfile(checkpoint_file):
start_epoch = instance.load_checkpoint(checkpoint_file)
if start_epoch is None:
start_epoch = 0
else:
start_epoch = 0
#Fit instance
epoch_losses = instance.train(dataloader_train, epochs, checkpoint_file=checkpoint_file, checkpoint_frequency=checkpoint_frequency, max_gradient=max_gradient)
print("Losses per epoch: " + str(epoch_losses))
return instance
@classmethod
def from_checkpoint(cls, checkpoint_file, stack_size,
cc1_enabled=False,
cc2_enabled=False,
cc3_enabled=True,
cc4_enabled=False,
cc5_enabled=False,
deterministic=False,
optimizer='sgd',
normalize_loss=False):
#Call constructor
instance = cls(stack_size,
cc1_enabled=cc1_enabled,
cc2_enabled=cc2_enabled,
cc3_enabled=cc3_enabled,
cc4_enabled=cc4_enabled,
cc5_enabled=cc5_enabled,
deterministic=deterministic,
optimizer=optimizer,
normalize_loss=normalize_loss)
#Load checkpoint
instance.load_checkpoint(checkpoint_file)
return instance
@classmethod
def from_tflearn(cls, checkpoint_file, stack_size,
cc1_enabled=False,
cc2_enabled=False,
cc3_enabled=True,
cc4_enabled=False,
cc5_enabled=False,
deterministic=False,
optimizer='sgd'):
#Call constructor
instance = cls(stack_size,
cc1_enabled=cc1_enabled,
cc2_enabled=cc2_enabled,
cc3_enabled=cc3_enabled,
cc4_enabled=cc4_enabled,
cc5_enabled=cc5_enabled,
deterministic=deterministic,
optimizer=optimizer,
pretrained=None)
#Load checkpoint
instance.load_tflearn(checkpoint_file)
return instance
def load_tflearn(self, checkpoint_file):
#Load dict
pretrained_dict = np.load(checkpoint_file)
#Update according to generated mapping
pretrained_dict = {self.__translate_tflearn_key(k): v for k, v in pretrained_dict.items()}
#Transpose all weight tensors since tflearn stores them transposed
# Tensorflow 2D Conv layer: h * w * in_channels * out_channels
# PyTorch 2D Conv layer: out_channels * in_channels * h * w
#Same logic was also implemented in https://github.com/ruotianluo/pytorch-mobilenet-from-tf/blob/master/convert.py
pretrained_dict = {k:(v.transpose((3, 2, 0, 1)) if (k.startswith("conv") or k.startswith("upconv")) and v.ndim == 4 else v) for k, v in pretrained_dict.items()}
pretrained_dict = {("scoring" + k[len("conv_disp"):] if k.startswith("conv_disp") else "autoencoder." + k):v for k, v in pretrained_dict.items()}
#Convert weight arrays to torch tensors
pretrained_dict = {k:torch.from_numpy(v).float() for k, v in pretrained_dict.items()}
#Load weights
model_state_dict = self.model.state_dict()
model_state_dict.update(pretrained_dict)
self.model.load_state_dict(model_state_dict)
def __translate_tflearn_key(self, key):
if key.endswith("/W:0"):
return key[:-len("/W:0")] + ".weight"
if key.endswith("/up_filter:0"):
return key[:-len("/up_filter:0")] + ".weight"
if key.endswith("/gamma:0"):
return key[:-len("/gamma:0")] + ".weight"
if key.endswith("/beta:0"):
return key[:-len("/beta:0")] + ".bias"
if key.endswith("/moving_mean:0"):
return key[:-len("/moving_mean:0")] + ".running_mean"
if key.endswith("/moving_variance:0"):
return key[:-len("/moving_variance:0")] + ".running_var"
def __create_preprocessing(self, crop_size=None, cliprange=[0.0202, 0.2825], mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]):
transform = [FocalStackDDFFH5Reader.FocalStackDDFFH5Reader.ToTensor()]
if cliprange is not None:
transform += [FocalStackDDFFH5Reader.FocalStackDDFFH5Reader.ClipGroundTruth(cliprange[0], cliprange[1])]
if crop_size is not None:
transform += [FocalStackDDFFH5Reader.FocalStackDDFFH5Reader.RandomCrop(crop_size)]
if mean is not None and std is not None:
transform += [FocalStackDDFFH5Reader.FocalStackDDFFH5Reader.Normalize(mean_input=mean, std_input=std)]
transform = torchvision.transforms.Compose(transform)
return transform
def create_validation_loader(self):
try:
return self.dataloader_validation
except AttributeError:
return None
class MaskedLoss(nn.Module):
def __init__(self, loss, valid_cond=lambda x : x > 0.0):
super(DDFFTrainer.MaskedLoss, self).__init__()
self.loss = loss
self.valid_cond = valid_cond
def forward(self, inputs, outputs):
mask = self.valid_cond(outputs)
return self.loss(inputs[mask], outputs[mask])