Spaces:
Runtime error
Runtime error
File size: 6,923 Bytes
7900c16 |
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 |
"""
This script provides an exmaple to wrap TencentPretrain for image classification.
"""
import sys
import os
import random
import argparse
import torch
import torch.nn as nn
import torchvision.datasets as dest
from torchvision import transforms
from torchvision.io import read_image
from torchvision.io.image import ImageReadMode
tencentpretrain_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
sys.path.append(tencentpretrain_dir)
from tencentpretrain.layers import *
from tencentpretrain.encoders import *
from tencentpretrain.utils.vocab import Vocab
from tencentpretrain.utils.constants import *
from tencentpretrain.utils import *
from tencentpretrain.utils.optimizers import *
from tencentpretrain.utils.config import load_hyperparam
from tencentpretrain.utils.misc import ZeroOneNormalize, count_lines
from tencentpretrain.utils.seed import set_seed
from tencentpretrain.model_saver import save_model
from tencentpretrain.opts import finetune_opts
from finetune.run_classifier import *
def data_loader(args, path):
transform = transforms.Compose([
transforms.Resize((args.image_height, args.image_width)),
ZeroOneNormalize()
])
dataset, columns = [], {}
with open(path, mode="r", encoding="utf-8") as f:
src_batch, tgt_batch, seg_batch = [], [], []
for line_id, line in enumerate(f):
if line_id == 0:
for i, column_name in enumerate(line.rstrip("\r\n").split("\t")):
columns[column_name] = i
continue
line = line.rstrip("\r\n").split("\t")
tgt = int(line[columns["label"]])
path = line[columns["path"]]
image = read_image(path, ImageReadMode.RGB)
image = image.to(args.device)
src = transform(image)
seg = [1] * ((src.size()[1] // args.patch_size) * (src.size()[2] // args.patch_size) + 1)
src_batch.append(src)
tgt_batch.append(tgt)
seg_batch.append(seg)
if len(src_batch) == args.batch_size:
yield torch.stack(src_batch, 0), \
torch.LongTensor(tgt_batch), \
torch.LongTensor(seg_batch)
src_batch, tgt_batch, seg_batch = [], [], []
if len(src_batch) > 0:
yield torch.stack(src_batch, 0), \
torch.LongTensor(tgt_batch), \
torch.LongTensor(seg_batch)
def evaluate(args, dataset_path):
correct, instances_num = 0, 0
# Confusion matrix.
confusion = torch.zeros(args.labels_num, args.labels_num, dtype=torch.long)
args.model.eval()
for i, (src_batch, tgt_batch, seg_batch) in enumerate(data_loader(args, dataset_path)):
src_batch = src_batch.to(args.device)
tgt_batch = tgt_batch.to(args.device)
seg_batch = seg_batch.to(args.device)
with torch.no_grad():
_, logits = args.model(src_batch, tgt_batch, seg_batch)
pred = torch.argmax(nn.Softmax(dim=1)(logits), dim=1)
gold = tgt_batch
for j in range(pred.size()[0]):
confusion[pred[j], gold[j]] += 1
correct += torch.sum(pred == gold).item()
instances_num += len(pred)
args.logger.info("Confusion matrix:")
args.logger.info(confusion)
args.logger.info("Report precision, recall, and f1:")
eps = 1e-9
for i in range(confusion.size()[0]):
p = confusion[i, i].item() / (confusion[i, :].sum().item() + eps)
r = confusion[i, i].item() / (confusion[:, i].sum().item() + eps)
f1 = 2 * p * r / (p + r + eps)
args.logger.info("Label {}: {:.3f}, {:.3f}, {:.3f}".format(i, p, r, f1))
args.logger.info("Acc. (Correct/Total): {:.4f} ({}/{}) ".format(correct / instances_num, correct, instances_num))
return correct / instances_num, confusion
def main():
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
finetune_opts(parser)
tokenizer_opts(parser)
adv_opts(parser)
args = parser.parse_args()
# Load the hyperparameters from the config file.
args = load_hyperparam(args)
args.soft_targets, args.soft_alpha = False, 0
# Count the number of labels.
args.labels_num = count_labels_num(args.train_path)
instances_num = count_lines(args.train_path) - 1
# Build tokenizer.
args.tokenizer = str2tokenizer["virtual"](args)
set_seed(args.seed)
# Build classification model.
model = Classifier(args)
# Load or initialize parameters.
load_or_initialize_parameters(args, model)
# Get logger.
args.logger = init_logger(args)
args.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(args.device)
# Training phase.
batch_size = args.batch_size
args.train_steps = int(instances_num * args.epochs_num / batch_size) + 1
args.logger.info("Batch size: {}".format(batch_size))
args.logger.info("The number of training instances: {}".format(instances_num))
optimizer, scheduler = build_optimizer(args, model)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
args.amp = amp
if torch.cuda.device_count() > 1:
args.logger.info("{} GPUs are available. Let's use them.".format(torch.cuda.device_count()))
model = torch.nn.DataParallel(model)
args.model = model
if args.use_adv:
args.adv_method = str2adv[args.adv_type](model)
total_loss, result, best_result = 0.0, 0.0, 0.0
args.logger.info("Start training.")
for epoch in range(1, args.epochs_num + 1):
model.train()
for i, (src_batch, tgt_batch, seg_batch) in enumerate(data_loader(args, args.train_path)):
loss = train_model(args, model, optimizer, scheduler, src_batch, tgt_batch, seg_batch)
total_loss += loss.item()
if (i + 1) % args.report_steps == 0:
args.logger.info("Epoch id: {}, Training steps: {}, Avg loss: {:.3f}".format(epoch, i + 1, total_loss / args.report_steps))
total_loss = 0.0
result = evaluate(args, args.dev_path)
if result[0] > best_result:
best_result = result[0]
save_model(model, args.output_model_path)
# Evaluation phase.
if args.test_path is not None:
args.logger.info("Test set evaluation.")
if torch.cuda.device_count() > 1:
args.model.module.load_state_dict(torch.load(args.output_model_path))
else:
args.model.load_state_dict(torch.load(args.output_model_path))
evaluate(args, args.test_path)
if __name__ == "__main__":
main()
|