File size: 7,302 Bytes
9e14838 | 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 | # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import os
import math
from typing import Iterable, Optional
import torch
import torch.distributed as dist
from timm.data import Mixup
from timm.utils import accuracy, ModelEma
import utils
from utils import adjust_learning_rate
from scipy.special import softmax
from sklearn.metrics import (
average_precision_score,
accuracy_score
)
import numpy as np
def train_one_epoch(model: torch.nn.Module, criterion: torch.nn.Module,
data_loader: Iterable, optimizer: torch.optim.Optimizer,
device: torch.device, epoch: int, loss_scaler, max_norm: float = 0,
model_ema: Optional[ModelEma] = None, mixup_fn: Optional[Mixup] = None,
log_writer=None, args=None):
model.train(True)
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter('lr', utils.SmoothedValue(window_size=1, fmt='{value:.6f}'))
header = 'Epoch: [{}]'.format(epoch)
print_freq = 20
update_freq = args.update_freq
use_amp = args.use_amp
optimizer.zero_grad()
for data_iter_step, (samples, targets) in enumerate(metric_logger.log_every(data_loader, print_freq, header)):
# we use a per iteration (instead of per epoch) lr scheduler
if data_iter_step % update_freq == 0:
adjust_learning_rate(optimizer, data_iter_step / len(data_loader) + epoch, args)
samples = samples.to(device, non_blocking=True)
targets = targets.to(device, non_blocking=True)
if mixup_fn is not None:
samples, targets = mixup_fn(samples, targets)
if use_amp:
with torch.cuda.amp.autocast():
output = model(samples)
loss = criterion(output, targets)
else: # full precision
output = model(samples)
loss = criterion(output, targets)
loss_value = loss.item()
if not math.isfinite(loss_value):
print("Loss is {}, stopping training".format(loss_value))
assert math.isfinite(loss_value)
if use_amp:
# this attribute is added by timm on one optimizer (adahessian)
is_second_order = hasattr(optimizer, 'is_second_order') and optimizer.is_second_order
loss /= update_freq
grad_norm = loss_scaler(loss, optimizer, clip_grad=max_norm,
parameters=model.parameters(), create_graph=is_second_order,
update_grad=(data_iter_step + 1) % update_freq == 0)
if (data_iter_step + 1) % update_freq == 0:
optimizer.zero_grad()
if model_ema is not None:
model_ema.update(model)
else: # full precision
loss /= update_freq
loss.backward()
if (data_iter_step + 1) % update_freq == 0:
optimizer.step()
optimizer.zero_grad()
if model_ema is not None:
model_ema.update(model)
torch.cuda.synchronize()
if mixup_fn is None:
class_acc = (output.max(-1)[-1] == targets).float().mean()
else:
class_acc = None
metric_logger.update(loss=loss_value)
metric_logger.update(class_acc=class_acc)
min_lr = 10.
max_lr = 0.
for group in optimizer.param_groups:
min_lr = min(min_lr, group["lr"])
max_lr = max(max_lr, group["lr"])
metric_logger.update(lr=max_lr)
metric_logger.update(min_lr=min_lr)
weight_decay_value = None
for group in optimizer.param_groups:
if group["weight_decay"] > 0:
weight_decay_value = group["weight_decay"]
metric_logger.update(weight_decay=weight_decay_value)
if use_amp:
metric_logger.update(grad_norm=grad_norm)
if log_writer is not None:
log_writer.update(loss=loss_value, head="loss")
log_writer.update(class_acc=class_acc, head="loss")
log_writer.update(lr=max_lr, head="opt")
log_writer.update(min_lr=min_lr, head="opt")
log_writer.update(weight_decay=weight_decay_value, head="opt")
if use_amp:
log_writer.update(grad_norm=grad_norm, head="opt")
log_writer.set_step()
# gather the stats from all processes
metric_logger.synchronize_between_processes()
print("Averaged stats:", metric_logger)
return {k: meter.global_avg for k, meter in metric_logger.meters.items()}
@torch.no_grad()
def evaluate(data_loader, model, device, use_amp=False):
criterion = torch.nn.CrossEntropyLoss()
metric_logger = utils.MetricLogger(delimiter=" ")
header = 'Test:'
# switch to evaluation mode
model.eval()
for index, batch in enumerate(metric_logger.log_every(data_loader, 10, header)):
images = batch[0]
target = batch[-1]
images = images.to(device, non_blocking=True)
target = target.to(device, non_blocking=True)
# compute output
if use_amp:
with torch.cuda.amp.autocast(dytpe=torch.bfloat16):
output = model(images)
if isinstance(output, dict):
output = output['logits']
loss = criterion(output, target)
else:
output = model(images) #[bs, num_cls]
if isinstance(output, dict):
output = output['logits']
loss = criterion(output, target)
if index == 0:
predictions = output
labels = target
else:
predictions = torch.cat((predictions, output), 0)
labels = torch.cat((labels, target), 0)
torch.cuda.synchronize()
acc1, acc5 = accuracy(output, target, topk=(1, 2))
batch_size = images.shape[0]
metric_logger.update(loss=loss.item())
metric_logger.meters['acc1'].update(acc1.item(), n=batch_size)
metric_logger.meters['acc5'].update(acc5.item(), n=batch_size)
# gather the stats from all processes
metric_logger.synchronize_between_processes()
print('* Acc@1 {top1.global_avg:.3f} Acc@5 {top5.global_avg:.3f} loss {losses.global_avg:.3f}'
.format(top1=metric_logger.acc1, top5=metric_logger.acc5, losses=metric_logger.loss))
output_ddp = [torch.zeros_like(predictions) for _ in range(utils.get_world_size())]
dist.all_gather(output_ddp, predictions)
labels_ddp = [torch.zeros_like(labels) for _ in range(utils.get_world_size())]
dist.all_gather(labels_ddp, labels)
output_all = torch.concat(output_ddp, dim=0)
labels_all = torch.concat(labels_ddp, dim=0)
y_pred = softmax(output_all.detach().cpu().numpy(), axis=1)[:, 1]
y_true = labels_all.detach().cpu().numpy()
y_true = y_true.astype(int)
acc = accuracy_score(y_true, y_pred > 0.5)
ap = average_precision_score(y_true, y_pred)
return {k: meter.global_avg for k, meter in metric_logger.meters.items()}, acc, ap |