-
Notifications
You must be signed in to change notification settings - Fork 17
/
Copy pathla_learner.py
172 lines (136 loc) · 7.42 KB
/
la_learner.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
'''Local Aggregation Learner for the fungi dataset, a child of `_Learner`
Written by: Anders Ohrn, October 2020
'''
import sys
import numpy as np
import torch
from _learner import _Learner, progress_bar
from cluster_utils import MemoryBank, LocalAggregationLoss
from ae_deep import EncoderVGGMerged, AutoEncoderVGG
class LALearner(_Learner):
'''Local Aggregation Learner class applied to the fungi image dataset for clustering of images
Args:
To be written
'''
def __init__(self, run_label='', random_seed=None, f_out=sys.stdout,
raw_csv_toc=None, raw_csv_root=None,
save_tmp_name='model_in_training',
selector=None, iselector=None,
dataset_type='full basic idx',
loader_batch_size=16, num_workers=0,
show_batch_progress=True, deterministic=True,
lr_init=0.01, momentum=0.9,
scheduler_step_size=15, scheduler_gamma=0.1,
k_nearest_neighbours=None, clustering_repeats=None, number_of_centroids=None,
temperature=None, memory_mixing=None, n_samples=None,
code_merger='mean',
img_input_dim=224, img_n_splits=6, crop_step_size=32, crop_dim=64):
dataset_kwargs = {'img_input_dim': img_input_dim, 'img_n_splits': img_n_splits,
'crop_step_size': crop_step_size, 'crop_dim': crop_dim}
super(LALearner, self).__init__(run_label=run_label, random_seed=random_seed, f_out=f_out,
raw_csv_toc=raw_csv_toc, raw_csv_root=raw_csv_root,
save_tmp_name=save_tmp_name,
selector=selector, iselector=iselector,
dataset_type=dataset_type, dataset_kwargs=dataset_kwargs,
loader_batch_size=loader_batch_size, num_workers=num_workers,
show_batch_progress=show_batch_progress,
deterministic=deterministic)
self.inp_k_nearest_neighbours = k_nearest_neighbours
self.inp_clustering_repeats = clustering_repeats
self.inp_number_of_centroids = number_of_centroids
self.inp_temperature = temperature
self.inp_memory_mixing = memory_mixing
self.inp_code_merger = code_merger
self.inp_lr_init = lr_init
self.inp_momentum = momentum
self.inp_scheduler_step_size = scheduler_step_size
self.inp_scheduler_gamma = scheduler_gamma
self.model = EncoderVGGMerged(merger_type=code_merger)
memory_bank = MemoryBank(n_vectors=n_samples, dim_vector=self.model.channels_code,
memory_mixing_rate=self.inp_memory_mixing)
self.criterion = LocalAggregationLoss(memory_bank=memory_bank,
temperature=self.inp_temperature,
k_nearest_neighbours=self.inp_k_nearest_neighbours,
clustering_repeats=self.inp_clustering_repeats,
number_of_centroids=self.inp_number_of_centroids)
self.set_sgd_optim(lr=self.inp_lr_init,
scheduler_step_size=self.inp_scheduler_step_size,
scheduler_gamma=self.inp_scheduler_gamma,
parameters=self.model.parameters())
self.print_inp()
def load_model(self, model_path):
'''Load encoder from saved state dictionary
The method dynamically determines if the state dictionary is from an encoder or an auto-encoder. In the latter
case the decoder part of the state dictionary is removed.
Args:
model_path (str): Path to the saved model to load
'''
saved_dict = torch.load('{}.tar'.format(model_path))[self.STATE_KEY_SAVE]
if any(['decoder' in key for key in saved_dict.keys()]):
encoder_state_dict = AutoEncoderVGG.state_dict_mutate('encoder', saved_dict)
else:
encoder_state_dict = saved_dict
self.model.load_state_dict(encoder_state_dict)
def save_model(self, model_path):
'''Save encoder state dictionary
Args:
model_path (str): Path and name to file to save state dictionary to. The filename on disk is this argument
appended with suffix `.tar`
'''
torch.save({self.STATE_KEY_SAVE: self.model.state_dict()},
'{}.tar'.format(model_path))
def train(self, n_epochs):
'''Train model for set number of epochs
Args:
n_epochs (int): Number of epochs to train the model for
'''
self.model.train()
for epoch in range(n_epochs):
print('Epoch {}/{}...'.format(epoch, n_epochs - 1), file=self.inp_f_out)
running_loss = 0.0
n_instances = 0
for inputs in self.dataloader:
size_batch = inputs[self.dataset.returnkey.image].size(0)
image = inputs[self.dataset.returnkey.image].to(self.device)
idx = inputs[self.dataset.returnkey.idx].detach().numpy()
# zero the parameter gradients
self.optimizer.zero_grad()
# Compute loss
output = self.model(image)
loss = self.criterion(output, idx)
# Back-propagate and optimize
loss.backward()
self.optimizer.step()
self.lr_scheduler.step()
# Update aggregates and reporting
running_loss += loss.item() * size_batch
if self.inp_show_batch_progress:
n_instances += size_batch
progress_bar(n_instances, self.dataset_size)
running_loss = running_loss / self.dataset_size
print('\nLoss: {:.4f}'.format(running_loss), file=self.inp_f_out)
self.save_model(self.inp_save_tmp_name)
def eval(self, clusterer, clusterer_kwargs={}, dloader=None):
'''Evaluate cluster properties for the data provided by data loader
Args:
clusterer (callable): Function that given a collection of feature vectors of shape (n_samples, n_features)
evaluates for each sample the cluster label. A function with these features are most `fit_predict`
methods of the clustering classes of `sklearn.cluster`.
clusterer_kwargs (dict, optional): Named argument dictionary for clusterer. Defaults to empty dictionary.
dloader (optional): Dataloader to collect data with. Defaults to `None`, in which case the Dataloader of
`self` is used.
Returns:
cluster_labels: The output of `clusterer` applied to the codes
'''
self.model.eval()
if dloader is None:
dloader = self.dataloader
all_output = None
for inputs in dloader:
image = inputs[self.dataset.returnkey.image].to(self.device)
output = self.model(image)
if all_output is None:
all_output = output.detach().numpy()
else:
all_output = np.append(all_output, output.detach().numpy(), axis=0)
return clusterer(all_output, **clusterer_kwargs)