-
Notifications
You must be signed in to change notification settings - Fork 31
/
trainlstm.py
395 lines (325 loc) · 16.8 KB
/
trainlstm.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
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
import numpy as np
import csv, gzip, os, sys, gc
import math
import torch
from torch import nn
import torch.optim as optim
from torch.nn import functional as F
import logging
import datetime
import optparse
import pandas as pd
import os
from sklearn.metrics import log_loss
import ast
from torch.utils.data import Dataset
from sklearn.metrics import log_loss
from torch.utils.data import DataLoader
from scipy.ndimage import uniform_filter
from torch.optim.lr_scheduler import StepLR
from apex.parallel import DistributedDataParallel as DDP
from apex.fp16_utils import *
from apex import amp, optimizers
from apex.multi_tensor_apply import multi_tensor_applier
# Print info about environments
parser = optparse.OptionParser()
parser.add_option('-s', '--seed', action="store", dest="seed", help="model seed", default="1234")
parser.add_option('-o', '--fold', action="store", dest="fold", help="Fold for split", default="0")
parser.add_option('-p', '--nbags', action="store", dest="nbags", help="Number of bags for averaging", default="4")
parser.add_option('-e', '--epochs', action="store", dest="epochs", help="epochs", default="10")
parser.add_option('-b', '--batchsize', action="store", dest="batchsize", help="batch size", default="4")
parser.add_option('-r', '--rootpath', action="store", dest="rootpath", help="root directory", default="/share/dhanley2/rsna/")
parser.add_option('-i', '--imgpath', action="store", dest="imgpath", help="root directory", default="data/mount/512X512X6/")
parser.add_option('-w', '--workpath', action="store", dest="workpath", help="Working path", default="densenetv1/weights")
parser.add_option('-f', '--weightsname', action="store", dest="weightsname", help="Weights file name", default="pytorch_model.bin")
parser.add_option('-l', '--lr', action="store", dest="lr", help="learning rate", default="0.00005")
parser.add_option('-g', '--logmsg', action="store", dest="logmsg", help="root directory", default="Recursion-pytorch")
parser.add_option('-c', '--size', action="store", dest="size", help="model size", default="512")
parser.add_option('-a', '--globalepoch', action="store", dest="globalepoch", help="root directory", default="3")
parser.add_option('-n', '--loadcsv', action="store", dest="loadcsv", help="Convert csv embeddings to numpy", default="F")
parser.add_option('-j', '--lstm_units', action="store", dest="lstm_units", help="Lstm units", default="128")
parser.add_option('-d', '--dropout', action="store", dest="dropout", help="LSTM input spatial dropout", default="0.3")
parser.add_option('-z', '--decay', action="store", dest="decay", help="Weight Decay", default="0.0")
parser.add_option('-m', '--lrgamma', action="store", dest="lrgamma", help="Scheduler Learning Rate Gamma", default="1.0")
parser.add_option('-k', '--ttahflip', action="store", dest="ttahflip", help="Bag with horizontal flip on and off", default="F")
parser.add_option('-q', '--ttatranspose', action="store", dest="ttatranspose", help="Bag with horizontal flip on and off", default="F")
parser.add_option('-x', '--datapath', action="store", dest="datapath", help="Data path", default="data")
options, args = parser.parse_args()
package_dir = options.rootpath
sys.path.append(package_dir)
sys.path.insert(0, 'scripts')
from logs import get_logger
from utils import dumpobj, loadobj, GradualWarmupScheduler
# Print info about environments
logger = get_logger(options.logmsg, 'INFO') # noqa
logger.info('Cuda set up : time {}'.format(datetime.datetime.now().time()))
device=torch.device('cuda')
logger.info('Device : {}'.format(torch.cuda.get_device_name(0)))
logger.info('Cuda available : {}'.format(torch.cuda.is_available()))
n_gpu = torch.cuda.device_count()
logger.info('Cuda n_gpus : {}'.format(n_gpu ))
logger.info('Load params : time {}'.format(datetime.datetime.now().time()))
for (k,v) in options.__dict__.items():
logger.info('{}{}'.format(k.ljust(20), v))
SEED = int(options.seed)
SIZE = int(options.size)
EPOCHS = int(options.epochs)
GLOBALEPOCH=int(options.globalepoch)
n_epochs = EPOCHS
lr=float(options.lr)
lrgamma=float(options.lrgamma)
DECAY=float(options.decay)
batch_size = int(options.batchsize)
ROOT = options.rootpath
path_data = os.path.join(ROOT, options.datapath)
path_img = os.path.join(ROOT, options.imgpath)
WORK_DIR = os.path.join(ROOT, options.workpath)
path_emb = os.path.join(ROOT, options.imgpath)
WEIGHTS_NAME = options.weightsname
fold = int(options.fold)
LOADCSV= options.loadcsv=='T'
LSTM_UNITS=int(options.lstm_units)
nbags=int(options.nbags)
DROPOUT=float(options.dropout)
TTAHFLIP= 'T' if options.ttahflip=='T' else ''
TTATRANSPOSE= 'P' if options.ttatranspose=='T' else ''
n_classes = 6
label_cols = ['epidural', 'intraparenchymal', 'intraventricular', 'subarachnoid', 'subdural', 'any']
def makeSub(ypred, imgs):
imgls = np.array(imgs).repeat(len(label_cols))
icdls = pd.Series(label_cols*ypred.shape[0])
yidx = ['{}_{}'.format(i,j) for i,j in zip(imgls, icdls)]
subdf = pd.DataFrame({'ID' : yidx, 'Label': ypred.flatten()})
return subdf
class SpatialDropout(nn.Dropout2d):
def forward(self, x):
x = x.unsqueeze(2) # (N, T, 1, K)
x = x.permute(0, 3, 2, 1) # (N, K, 1, T)
x = super(SpatialDropout, self).forward(x) # (N, K, 1, T), some features are masked
x = x.permute(0, 3, 2, 1) # (N, T, 1, K)
x = x.squeeze(2) # (N, T, K)
return x
def criterion(data, targets, criterion = torch.nn.BCEWithLogitsLoss()):
''' Define custom loss function for weighted BCE on 'target' column '''
loss_all = criterion(data, targets)
loss_any = criterion(data[:,-1:], targets[:,-1:])
return (loss_all*6 + loss_any*1)/7
class IntracranialDataset(Dataset):
def __init__(self, df, mat, labels=label_cols):
self.data = df
self.mat = mat
self.labels = labels
self.patients = df.SliceID.unique()
self.data = self.data.set_index('SliceID')
def __len__(self):
return len(self.patients)
def __getitem__(self, idx):
patidx = self.patients[idx]
patdf = self.data.loc[patidx].sort_values('seq')
patemb = self.mat[patdf['embidx'].values]
patdeltalag = np.zeros(patemb.shape)
patdeltalead = np.zeros(patemb.shape)
patdeltalag [1:] = patemb[1:]-patemb[:-1]
patdeltalead[:-1] = patemb[:-1]-patemb[1:]
patemb = np.concatenate((patemb, patdeltalag, patdeltalead), -1)
ids = torch.tensor(patdf['embidx'].values)
if self.labels:
labels = torch.tensor(patdf[label_cols].values)
return {'emb': patemb, 'embidx' : ids, 'labels': labels}
else:
return {'emb': patemb, 'embidx' : ids}
def predict(loader):
valls = []
imgls = []
imgdf = loader.dataset.data.reset_index().set_index('embidx')[['Image']].copy()
for step, batch in enumerate(loader):
inputs = batch["emb"]
mask = batch['mask'].to(device, dtype=torch.int)
inputs = inputs.to(device, dtype=torch.float)
logits = model(inputs)
# get the mask for masked labels
maskidx = mask.view(-1)==1
# reshape for
logits = logits.view(-1, n_classes)[maskidx]
valls.append(torch.sigmoid(logits).detach().cpu().numpy())
# Get the list of images
embidx = batch["embidx"].detach().cpu().numpy().astype(np.int32)
embidx = embidx.flatten()[embidx.flatten()>-1]
images = imgdf.loc[embidx].Image.tolist()
imgls += images
return np.concatenate(valls, 0), imgls
# Print info about environments
logger.info('Cuda set up : time {}'.format(datetime.datetime.now().time()))
# Get image sequences
trnmdf = pd.read_csv(os.path.join(path_data, 'train_metadata.csv'))
tstmdf = pd.read_csv(os.path.join(path_data, 'test_metadata.csv'))
trnmdf['SliceID'] = trnmdf[['PatientID', 'SeriesInstanceUID', 'StudyInstanceUID']].apply(lambda x: '{}__{}__{}'.format(*x.tolist()), 1)
tstmdf['SliceID'] = tstmdf[['PatientID', 'SeriesInstanceUID', 'StudyInstanceUID']].apply(lambda x: '{}__{}__{}'.format(*x.tolist()), 1)
poscols = ['ImagePos{}'.format(i) for i in range(1, 4)]
trnmdf[poscols] = pd.DataFrame(trnmdf['ImagePositionPatient']\
.apply(lambda x: list(map(float, ast.literal_eval(x)))).tolist())
tstmdf[poscols] = pd.DataFrame(tstmdf['ImagePositionPatient']\
.apply(lambda x: list(map(float, ast.literal_eval(x)))).tolist())
trnmdf = trnmdf.sort_values(['SliceID']+poscols)\
[['PatientID', 'SliceID', 'SOPInstanceUID']+poscols].reset_index(drop=True)
tstmdf = tstmdf.sort_values(['SliceID']+poscols)\
[['PatientID', 'SliceID', 'SOPInstanceUID']+poscols].reset_index(drop=True)
trnmdf['seq'] = (trnmdf.groupby(['SliceID']).cumcount() + 1)
tstmdf['seq'] = (tstmdf.groupby(['SliceID']).cumcount() + 1)
keepcols = ['PatientID', 'SliceID', 'SOPInstanceUID', 'seq']
trnmdf = trnmdf[keepcols]
tstmdf = tstmdf[keepcols]
trnmdf.columns = tstmdf.columns = ['PatientID', 'SliceID', 'Image', 'seq']
# Load Data Frames
trndf = loadobj(os.path.join(path_emb, 'loader_trn_size{}_fold{}_ep{}'.format(SIZE, fold, GLOBALEPOCH))).dataset.data
valdf = loadobj(os.path.join(path_emb, 'loader_val_size{}_fold{}_ep{}'.format(SIZE, fold, GLOBALEPOCH))).dataset.data
tstdf = loadobj(os.path.join(path_emb, 'loader_tst_size{}_fold{}_ep{}'.format(SIZE, fold, GLOBALEPOCH))).dataset.data
trndf['embidx'] = range(trndf.shape[0])
valdf['embidx'] = range(valdf.shape[0])
tstdf['embidx'] = range(tstdf.shape[0])
trndf = trndf.merge(trnmdf.drop('PatientID', 1), on = 'Image')
valdf = valdf.merge(trnmdf.drop('PatientID', 1), on = 'Image')
tstdf = tstdf.merge(tstmdf, on = 'Image')
# Load embeddings
embnm='emb_sz256_wt256_fold{}_epoch{}'.format(fold, GLOBALEPOCH)
logger.info('Load npy..')
def loademb(TYPE, SIZE, fold, GLOBALEPOCH, TTA=''):
return np.load(os.path.join(path_emb, 'emb{}_{}_size{}_fold{}_ep{}.npz'.format(TTA, TYPE, SIZE, fold, GLOBALEPOCH)))['arr_0']
logger.info('Load embeddings...')
trnembls = [loademb('trn', SIZE, fold, GLOBALEPOCH)]
valembls = [loademb('val', SIZE, fold, GLOBALEPOCH)]
tstembls = [loademb('tst', SIZE, fold, GLOBALEPOCH)]
if TTAHFLIP=='T':
logger.info('Load hflip...')
trnembls.append(loademb('trn', SIZE, fold, GLOBALEPOCH, TTA='T'))
valembls.append(loademb('val', SIZE, fold, GLOBALEPOCH, TTA='T'))
tstembls.append(loademb('tst', SIZE, fold, GLOBALEPOCH, TTA='T'))
if TTATRANSPOSE=='P':
logger.info('Load transpose...')
trnembls.append(loademb('trn', SIZE, fold, GLOBALEPOCH, TTA='P'))
valembls.append(loademb('val', SIZE, fold, GLOBALEPOCH, TTA='P'))
tstembls.append(loademb('tst', SIZE, fold, GLOBALEPOCH, TTA='P'))
trnemb = sum(trnembls)/len(trnembls)
valemb = sum(valembls)/len(valembls)
tstemb = sum(tstembls)/len(tstembls)
logger.info('Trn shape {} {}'.format(*trnemb.shape))
logger.info('Val shape {} {}'.format(*valemb.shape))
logger.info('Tst shape {} {}'.format(*tstemb.shape))
logger.info('Add stg1 test labels to train')
del trnembls, valembls, tstembls
gc.collect()
# a simple custom collate function, just to show the idea
def collatefn(batch):
maxlen = max([l['emb'].shape[0] for l in batch])
embdim = batch[0]['emb'].shape[1]
withlabel = 'labels' in batch[0]
if withlabel:
labdim= batch[0]['labels'].shape[1]
for b in batch:
masklen = maxlen-len(b['emb'])
b['emb'] = np.vstack((np.zeros((masklen, embdim)), b['emb']))
b['embidx'] = torch.cat((torch.ones((masklen),dtype=torch.long)*-1, b['embidx']))
b['mask'] = np.ones((maxlen))
b['mask'][:masklen] = 0.
if withlabel:
b['labels'] = np.vstack((np.zeros((maxlen-len(b['labels']), labdim)), b['labels']))
outbatch = {'emb' : torch.tensor(np.vstack([np.expand_dims(b['emb'], 0) \
for b in batch])).float()}
outbatch['mask'] = torch.tensor(np.vstack([np.expand_dims(b['mask'], 0) \
for b in batch])).float()
outbatch['embidx'] = torch.tensor(np.vstack([np.expand_dims(b['embidx'], 0) \
for b in batch])).float()
if withlabel:
outbatch['labels'] = torch.tensor(np.vstack([np.expand_dims(b['labels'], 0) for b in batch])).float()
return outbatch
logger.info('Create loaders...')
trndataset = IntracranialDataset(trndf, trnemb, labels=True)
trnloader = DataLoader(trndataset, batch_size=batch_size, shuffle=True, num_workers=8, collate_fn=collatefn)
valdataset = IntracranialDataset(valdf, valemb, labels=False)
tstdataset = IntracranialDataset(tstdf, tstemb, labels=False)
tstloader = DataLoader(tstdataset, batch_size=batch_size*4, shuffle=False, num_workers=8, collate_fn=collatefn)
valloader = DataLoader(valdataset, batch_size=batch_size*4, shuffle=False, num_workers=8, collate_fn=collatefn)
# https://www.kaggle.com/bminixhofer/speed-up-your-rnn-with-sequence-bucketing
class NeuralNet(nn.Module):
def __init__(self, embed_size=trnemb.shape[-1]*3, LSTM_UNITS=64, DO = 0.3):
super(NeuralNet, self).__init__()
self.embedding_dropout = SpatialDropout(0.0) #DO)
self.lstm1 = nn.LSTM(embed_size, LSTM_UNITS, bidirectional=True, batch_first=True)
self.lstm2 = nn.LSTM(LSTM_UNITS * 2, LSTM_UNITS, bidirectional=True, batch_first=True)
self.linear1 = nn.Linear(LSTM_UNITS*2, LSTM_UNITS*2)
self.linear2 = nn.Linear(LSTM_UNITS*2, LSTM_UNITS*2)
self.linear = nn.Linear(LSTM_UNITS*2, n_classes)
def forward(self, x, lengths=None):
h_embedding = x
h_embadd = torch.cat((h_embedding[:,:,:2048], h_embedding[:,:,:2048]), -1)
h_lstm1, _ = self.lstm1(h_embedding)
h_lstm2, _ = self.lstm2(h_lstm1)
h_conc_linear1 = F.relu(self.linear1(h_lstm1))
h_conc_linear2 = F.relu(self.linear2(h_lstm2))
hidden = h_lstm1 + h_lstm2 + h_conc_linear1 + h_conc_linear2 + h_embadd
output = self.linear(hidden)
return output
logger.info('Create model')
model = NeuralNet(LSTM_UNITS=LSTM_UNITS, DO = DROPOUT)
model = model.to(device)
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
plist = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': DECAY},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
optimizer = optim.Adam(plist, lr=lr)
scheduler = StepLR(optimizer, 1, gamma=lrgamma, last_epoch=-1)
model, optimizer = amp.initialize(model, optimizer, opt_level="O1")
ypredls = []
ypredtstls = []
for epoch in range(EPOCHS):
tr_loss = 0.
for param in model.parameters():
param.requires_grad = True
model.train()
for step, batch in enumerate(trnloader):
y = batch['labels'].to(device, dtype=torch.float)
mask = batch['mask'].to(device, dtype=torch.int)
x = batch['emb'].to(device, dtype=torch.float)
x = torch.autograd.Variable(x, requires_grad=True)
y = torch.autograd.Variable(y)
logits = model(x).to(device, dtype=torch.float)
# get the mask for masked labels
maskidx = mask.view(-1)==1
y = y.view(-1, n_classes)[maskidx]
logits = logits.view(-1, n_classes)[maskidx]
# Get loss
loss = criterion(logits, y)
tr_loss += loss.item()
optimizer.zero_grad()
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
optimizer.step()
if step%1000==0:
logger.info('Trn step {} of {} trn lossavg {:.5f}'. \
format(step, len(trnloader), (tr_loss/(1+step))))
scheduler.step()
model.eval()
logger.info('Prep val score...')
ypred, imgval = predict(valloader)
ypredls.append(ypred)
yvalpred = sum(ypredls[-nbags:])/len(ypredls[-nbags:])
yvalout = makeSub(yvalpred, imgval)
yvalp = makeSub(ypred, imgval)
# get Val score
weights = ([1, 1, 1, 1, 1, 2] * ypred.shape[0])
yact = valloader.dataset.data[label_cols].values#.flatten()
yact = makeSub(yact, valloader.dataset.data['Image'].tolist())
yact = yact.set_index('ID').loc[yvalout.ID].reset_index()
valloss = log_loss(yact['Label'].values, yvalp['Label'].values.clip(.00001,.99999) , sample_weight = weights)
vallossavg = log_loss(yact['Label'].values, yvalout['Label'].values.clip(.00001,.99999) , sample_weight = weights)
logger.info('Epoch {} val logloss {:.5f} bagged {:.5f}'.format(epoch, valloss, vallossavg))
logger.info('Prep test sub...')
ypred, imgtst = predict(tstloader)
ypredtstls.append(ypred)
logger.info('Write out bagged prediction to preds folder')
ytstpred = sum(ypredtstls[-nbags:])/len(ypredtstls[-nbags:])
ytstout = makeSub(ytstpred, imgtst)
ytstout.to_csv('preds/lstm{}{}{}_{}_epoch{}_sub_{}.csv.gz'.format(TTAHFLIP, TTATRANSPOSE, LSTM_UNITS, WORK_DIR, epoch, embnm), \
index = False, compression = 'gzip')