-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathneuston_models.py
190 lines (159 loc) · 8.32 KB
/
neuston_models.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
"""a module for defining model architecture"""
# built in imports
import argparse
# 3rd party imports
import torch
import torch.nn as nn
from torch.optim import Adam
from torch.nn.functional import softmax
import torchvision.models as MODEL_MODULE
from torchvision.models.inception import InceptionOutputs
import pytorch_lightning as ptl
from sklearn import metrics
import numpy as np
import ifcb
# project imports #
from neuston_data import IfcbBinDataset
def get_namebrand_model(model_name, num_o_classes, pretrained=False):
if model_name == 'inception_v3':
model = MODEL_MODULE.inception_v3(pretrained) #, num_classes=num_o_classes, aux_logits=False)
model.AuxLogits.fc = nn.Linear(model.AuxLogits.fc.in_features, num_o_classes)
model.fc = nn.Linear(model.fc.in_features, num_o_classes)
elif model_name == 'alexnet':
model = getattr(MODEL_MODULE, model_name)(pretrained)
model.classifier[6] = nn.Linear(model.classifier[6].in_features, num_o_classes)
elif model_name == 'squeezenet':
model = getattr(MODEL_MODULE, model_name+'1_1')(pretrained)
model.classifier[1] = nn.Conv2d(512, num_o_classes, kernel_size=(1, 1), stride=(1, 1))
model.num_classes = num_o_classes
elif model_name.startswith('vgg'):
model = getattr(MODEL_MODULE, model_name)(pretrained)
model.classifier[6] = nn.Linear(model.classifier[6].in_features, num_o_classes)
elif model_name.startswith('resnet'):
model = getattr(MODEL_MODULE, model_name)(pretrained)
model.fc = nn.Linear(model.fc.in_features, num_o_classes)
elif model_name.startswith('densenet'):
model = getattr(MODEL_MODULE, model_name)(pretrained)
model.classifier = nn.Linear(model.classifier.in_features, num_o_classes)
else:
raise KeyError("model unknown!")
return model
class NeustonModel(ptl.LightningModule):
def __init__(self, hparams):
super().__init__()
if isinstance(hparams,dict):
hparams = argparse.Namespace(**hparams)
self.save_hyperparameters(hparams)
self.criterion = nn.CrossEntropyLoss()
self.model = get_namebrand_model(hparams.MODEL, len(hparams.classes), hparams.pretrained)
# Instance Variables
self.best_val_loss = np.inf
self.best_epoch = 0
self.agg_train_loss = 0.0
def configure_optimizers(self):
return Adam(self.parameters(), lr=0.001)
def forward(self, inputs):
outputs = self.model(inputs)
return outputs
def loss(self, inputs, outputs):
if isinstance(outputs,tuple) and len(outputs)==2: # inception_v3
outputs, aux_outputs = outputs
loss1 = self.criterion(outputs, inputs)
loss2 = self.criterion(aux_outputs, inputs)
batch_loss = loss1+0.4*loss2
else:
batch_loss = self.criterion(outputs, inputs)
return batch_loss
# TRAINING #
def training_step(self, batch, batch_nb):
input_data, input_classes, input_src = batch
outputs = self.forward(input_data)
batch_loss = self.loss(input_classes, outputs)
self.agg_train_loss += batch_loss.item()
return dict(loss=batch_loss)
def training_epoch_end(self, steps):
train_loss = torch.stack([batch['loss'] for batch in steps]).sum().item()
#print('training_epoch_end: self.agg_train_loss={:.5f}, train_loss={:.5f}, DIFF={:.9f}'.format(self.agg_train_loss, train_loss, self.agg_train_loss-train_loss), end='\n\n')
#return dict(train_loss=train_loss)
# Validation #
def validation_step(self, batch, batch_idx):
input_data, input_classes, input_src = batch
outputs = self.forward(input_data)
val_batch_loss = self.loss(input_classes, outputs)
outputs = outputs.logits if isinstance(outputs,InceptionOutputs) else outputs
outputs = softmax(outputs,dim=1)
return dict(val_batch_loss=val_batch_loss,
val_outputs=outputs,
val_input_classes=input_classes,
val_input_srcs=input_src)
def validation_epoch_end(self, steps):
print(end='\n\n') # give space for progress bar
if self.current_epoch==0: self.best_val_loss = np.inf # takes care of any lingering val_loss from sanity checks
validation_loss = torch.stack([batch['val_batch_loss'] for batch in steps]).sum()
#eoe0 = 'validation_epoch_end: best_val_loss={}, curr_val_loss={}, curr<best={}, curr-best (neg is good)={}'
#eoe0 = eoe0.format(self.best_val_loss, validation_loss.item(), validation_loss.item()<self.best_val_loss, validation_loss.item()-self.best_val_loss)
#print(eoe0)
if validation_loss.item()<self.best_val_loss:
self.best_val_loss = validation_loss.item()
self.best_epoch = self.current_epoch
outputs = torch.cat([batch['val_outputs'] for batch in steps],dim=0).detach().cpu().numpy()
output_classes = np.argmax(outputs, axis=1)
input_classes = torch.cat([batch['val_input_classes'] for batch in steps],dim=0).detach().cpu().numpy()
input_srcs = [item for sublist in [batch['val_input_srcs'] for batch in steps] for item in sublist]
f1_weighted = metrics.f1_score(input_classes, output_classes, average='weighted')
f1_macro = metrics.f1_score(input_classes, output_classes, average='macro')
eoe = 'Best Epoch: {}, train_loss: {:.3f}, val_loss: {:.3f}, val_f1_w={:02.1f}%, val_f1_m={:02.1f}%'
eoe = eoe.format(True if self.current_epoch==self.best_epoch else self.best_epoch+1, self.agg_train_loss, validation_loss, 100*f1_weighted, 100*f1_macro)
print(eoe, flush=True, end='\n\n') # so slurm output can be followed along
# used by callbacks and logger
self.log('epoch', self.current_epoch, on_epoch=True)
self.log('best', self.best_epoch==self.current_epoch, on_epoch=True)
self.log('train_loss', self.agg_train_loss, on_epoch=True)
self.log('val_loss', validation_loss, on_epoch=True)
# csv_logger logger hacked to not include these in epochs.csv output
self.log('input_classes', input_classes, on_epoch=True)
self.log('output_classes', output_classes, on_epoch=True)
self.log('input_srcs', input_srcs, on_epoch=True)
self.log('outputs', outputs, on_epoch=True)
# these will apppear in epochs.csv, but are not used by callbacks
self.log('f1_macro',f1_macro, on_epoch=True)
self.log('f1_weighted',f1_weighted, on_epoch=True)
# Cleanup
self.agg_train_loss = 0.0
return dict(hiddens=dict(outputs=outputs))
# RUNNING the model #
def test_step(self, batch, batch_idx, dataloader_idx=None):
input_data, input_srcs = batch
outputs = self.forward(input_data)
outputs = outputs.logits if isinstance(outputs,InceptionOutputs) else outputs
outputs = softmax(outputs, dim=1)
return dict(test_outputs=outputs, test_srcs=input_srcs)
def test_epoch_end(self, steps):
# handle single and multiple test dataloaders
datasets = self.test_dataloader()
if isinstance(datasets, list): datasets = [ds.dataset for ds in datasets]
else: datasets = [datasets.dataset]
if isinstance(steps[0],dict):
steps = [steps]
RRs = []
for steps,dataset in zip(steps,datasets):
outputs = torch.cat([batch['test_outputs'] for batch in steps],dim=0).detach().cpu().numpy()
images = [batch['test_srcs'] for batch in steps]
images = [item for sublist in images for item in sublist] # flatten list
if isinstance(dataset, IfcbBinDataset):
input_obj = dataset.bin.pid
else:
input_obj = dataset.input_src # a path string
rr = self.RunResults(inputs=images, outputs=outputs, input_obj=input_obj)
RRs.append(rr)
self.log('RunResults',RRs)
#return dict(RunResults=RRs)
class RunResults:
def __init__(self, inputs, outputs, input_obj):
self.inputs = inputs
self.outputs = outputs
self.input_obj = input_obj
self.type = 'Bin' if isinstance(input_obj,ifcb.Pid) else 'ImgDir'
def __repr__(self):
rep = '{}: {} ({} imgs)'.format(self.type, self.input_obj, len(self.inputs))
return repr(rep)