File size: 13,837 Bytes
0f8411f a070253 0f8411f | 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 | import os
import argparse
import time
import torch
import torch.nn as nn
import numpy as np
import pandas as pd
import utils
from .models.bert_labeler import bert_labeler
from .datasets.impressions_dataset import ImpressionsDataset
from .constants import *
def collate_fn_labels(sample_list):
"""Custom collate function to pad reports in each batch to the max len
@param sample_list (List): A list of samples. Each sample is a dictionary with
keys 'imp', 'label', 'len' as returned by the __getitem__
function of ImpressionsDataset
@returns batch (dictionary): A dictionary with keys 'imp', 'label', 'len' but now
'imp' is a tensor with padding and batch size as the
first dimension. 'label' is a stacked tensor of labels
for the whole batch with batch size as first dim. And
'len' is a list of the length of each sequence in batch
"""
tensor_list = [s['imp'] for s in sample_list]
batched_imp = torch.nn.utils.rnn.pad_sequence(tensor_list,
batch_first=True,
padding_value=PAD_IDX)
label_list = [s['label'] for s in sample_list]
batched_label = torch.stack(label_list, dim=0)
len_list = [s['len'] for s in sample_list]
batch = {'imp': batched_imp, 'label': batched_label, 'len': len_list}
return batch
def load_data(train_csv_path, train_list_path, dev_csv_path,
dev_list_path, train_weights=None, batch_size=BATCH_SIZE,
shuffle=True, num_workers=NUM_WORKERS):
""" Create ImpressionsDataset objects for train and test data
@param train_csv_path (string): path to training csv file containing labels
@param train_list_path (string): path to list of encoded impressions for train set
@param dev_csv_path (string): same as train_csv_path but for dev set
@param dev_list_path (string): same as train_list_path but for dev set
@param train_weights (torch.Tensor): Tensor of shape (train_set_size) containing weights
for each training example, for the purposes of batch
sampling with replacement
@param batch_size (int): the batch size. As per the BERT repository, the max batch size
that can fit on a TITAN XP is 6 if the max sequence length
is 512, which is our case. We have 3 TITAN XP's
@param shuffle (bool): Whether to shuffle data before each epoch, ignored if train_weights
is not None
@param num_workers (int): How many worker processes to use to load data
@returns dataloaders (tuple): tuple of two ImpressionsDataset objects, for train and dev sets
"""
collate_fn = collate_fn_labels
train_dset = ImpressionsDataset(train_csv_path, train_list_path)
dev_dset = ImpressionsDataset(dev_csv_path, dev_list_path)
if train_weights is None:
train_loader = torch.utils.data.DataLoader(train_dset, batch_size=batch_size, shuffle=shuffle,
num_workers=num_workers, collate_fn=collate_fn)
else:
sampler = torch.utils.data.WeightedRandomSampler(weights=train_weights,
num_samples=len(train_weights),
replacement=True)
train_loader = torch.utils.data.DataLoader(train_dset,
batch_size=batch_size,
num_workers=num_workers,
collate_fn=collate_fn,
sampler=sampler)
dev_loader = torch.utils.data.DataLoader(dev_dset, batch_size=batch_size, shuffle=shuffle,
num_workers=num_workers, collate_fn=collate_fn)
dataloaders = (train_loader, dev_loader)
return dataloaders
def load_test_data(test_csv_path, test_list_path, batch_size=BATCH_SIZE,
num_workers=NUM_WORKERS, shuffle=False):
""" Create ImpressionsDataset object for the test set
@param test_csv_path (string): path to test csv file containing labels
@param test_list_path (string): path to list of encoded impressions
@param batch_size (int): the batch size. As per the BERT repository, the max batch size
that can fit on a TITAN XP is 6 if the max sequence length
is 512, which is our case. We have 3 TITAN XP's
@param num_workers (int): how many worker processes to use to load data
@param shuffle (bool): whether to shuffle the data or not
@returns test_loader (dataloader): dataloader object for test set
"""
collate_fn = collate_fn_labels
test_dset = ImpressionsDataset(test_csv_path, test_list_path)
test_loader = torch.utils.data.DataLoader(test_dset, batch_size=batch_size, shuffle=shuffle,
num_workers=num_workers, collate_fn=collate_fn)
return test_loader
def train(save_path, dataloaders, f1_weights, model=None, device=None,
optimizer=None, lr=LEARNING_RATE, log_every=LOG_EVERY,
valid_niter=VALID_NITER, best_metric=0.0):
""" Main training loop for the labeler
@param save_path (string): Directory in which model weights are stored
@param model (nn.Module): the labeler model to train, if applicable
@param device (torch.device): device for the model. If model is not None, this
parameter is required
@param dataloaders (tuple): tuple of dataloader objects as returned by load_data
@param f1_weights (dictionary): maps conditions to weights for blank, negation,
uncertain and positxive f1 task averaging
@param optimizer (torch.optim.Optimizer): the optimizer to use, if applicable
@param lr (float): learning rate to use in the optimizer, ignored if optimizer
is not None
@param log_every (int): number of iterations to log after
@param valid_niter (int): number of iterations after which to evaluate the model and
save it if it is better than old best model
@param best_metric (float): save checkpoints only if dev set performance is higher
than best_metric
"""
if model and not device:
print("train function error: Model specified but not device")
return
if model is None:
model = bert_labeler(pretrain_path=PRETRAIN_PATH)
model.train() #put the model into train mode
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
if torch.cuda.device_count() > 1:
print("Using", torch.cuda.device_count(), "GPUs!")
model = nn.DataParallel(model) #to utilize multiple GPU's
model = model.to(device)
else:
model.train()
if optimizer is None:
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
begin_time = time.time()
report_examples = 0
report_loss = 0.0
train_ld = dataloaders[0]
dev_ld = dataloaders[1]
loss_func = nn.CrossEntropyLoss(reduction='sum')
print('begin labeler training')
for epoch in range(NUM_EPOCHS):
for i, data in enumerate(train_ld, 0):
batch = data['imp'] #(batch_size, max_len)
batch = batch.to(device)
label = data['label'] #(batch_size, 14)
label = label.permute(1, 0).to(device)
src_len = data['len']
batch_size = batch.shape[0]
attn_mask = utils.generate_attention_masks(batch, src_len, device)
optimizer.zero_grad()
out = model(batch, attn_mask) #list of 14 tensors
batch_loss = 0.0
for j in range(len(out)):
batch_loss += loss_func(out[j], label[j])
report_loss += batch_loss
report_examples += batch_size
loss = batch_loss / batch_size
loss.backward()
optimizer.step()
if (i+1) % log_every == 0:
print('epoch %d, iter %d, avg_loss %.3f, time_elapsed %.3f sec' % (epoch+1, i+1, report_loss/report_examples,
time.time() - begin_time))
report_loss = 0.0
report_examples = 0
if (i+1) % valid_niter == 0:
print('\n begin validation')
metrics = utils.evaluate(model, dev_ld, device, f1_weights)
weighted = metrics['weighted']
kappas = metrics['kappa']
for j in range(len(CONDITIONS)):
print('%s kappa: %.3f' % (CONDITIONS[j], kappas[j]))
print('average: %.3f' % (np.mean(kappas)))
#for j in range(len(CONDITIONS)):
# print('%s weighted_f1: %.3f' % (CONDITIONS[j], weighted[j]))
#print('average of weighted_f1: %.3f' % (np.mean(weighted)))
for j in range(len(CONDITIONS)):
print('%s blank_f1: %.3f, negation_f1: %.3f, uncertain_f1: %.3f, positive: %.3f' % (CONDITIONS[j],
metrics['blank'][j],
metrics['negation'][j],
metrics['uncertain'][j],
metrics['positive'][j]))
metric_avg = np.mean(kappas)
if metric_avg > best_metric: #new best network
print("saving new best network!\n")
best_metric = metric_avg
path = os.path.join(save_path, "model_epoch%d_iter%d" % (epoch+1, i+1))
torch.save({'epoch': epoch+1,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict()},
path)
def model_from_ckpt(model, ckpt_path):
"""Load up model checkpoint
@param model (nn.Module): the module to be loaded
@param ckpt_path (string): path to a checkpoint. If this is None, then
model is trained from scratch
@return (tuple): tuple containing the model, optimizer and device
"""
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
if torch.cuda.device_count() > 1:
print("Using", torch.cuda.device_count(), "GPUs!")
model = nn.DataParallel(model) #to utilize multiple GPU's
model = model.to(device)
optimizer = torch.optim.Adam(model.parameters())
checkpoint = torch.load(ckpt_path)
model.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
return (model, optimizer, device)
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Train BERT-base model on task of labeling 14 medical conditions.')
parser.add_argument('--train_csv', type=str, nargs='?', required=True,
help='path to csv containing train reports.')
parser.add_argument('--dev_csv', type=str, nargs='?', required=True,
help='path to csv containing dev reports.')
parser.add_argument('--train_imp_list', type=str, nargs='?', required=True,
help='path to list of tokenized train set report impressions')
parser.add_argument('--dev_imp_list', type=str, nargs='?', required=True,
help='path to list of tokenized dev set report impressions')
parser.add_argument('--output_dir', type=str, nargs='?', required=True,
help='path to output directory where checkpoints will be saved')
parser.add_argument('--checkpoint', type=str, nargs='?', required=False,
help='path to existing checkpoint to initialize weights from')
args = parser.parse_args()
train_csv_path = args.train_csv
dev_csv_path = args.dev_csv
train_imp_path = args.train_imp_list
dev_imp_path = args.dev_imp_list
out_path = args.output_dir
checkpoint_path = args.checkpoint
if checkpoint_path:
model, optimizer, device = model_from_ckpt(bert_labeler(), checkpoint_path)
else:
model, optimizer, device = None, None, None
f1_weights = utils.get_weighted_f1_weights(dev_csv_path)
dataloaders = load_data(train_csv_path, train_imp_path, dev_csv_path, dev_imp_path)
train(save_path=out_path,
dataloaders=dataloaders,
model=model,
optimizer=optimizer,
device=device,
f1_weights=f1_weights)
|