fast-reid/projects/DistillReID/kdreid/kd_trainer.py

136 lines
4.4 KiB
Python

# encoding: utf-8
"""
@author: l1aoxingyu
@contact: sherlockliao01@gmail.com
"""
import logging
import time
import torch
import torch.nn.functional as F
from torch import nn
from torch.nn.parallel import DistributedDataParallel
from fastreid.engine import DefaultTrainer
from fastreid.modeling.meta_arch import build_model
from fastreid.utils.checkpoint import Checkpointer
from .config import update_model_teacher_config
class KDTrainer(DefaultTrainer):
"""
A knowledge distillation trainer for person reid of task.
"""
def __init__(self, cfg):
"""
Args:
cfg (CfgNode):
"""
super().__init__(cfg)
model_t = self.build_model_teacher(self.cfg)
for param in model_t.parameters():
param.requires_grad = False
logger = logging.getLogger('fastreid.' + __name__)
# Load pre-trained teacher model
logger.info("Loading teacher model ...")
Checkpointer(model_t).load(cfg.MODEL.TEACHER_WEIGHTS)
model_t.eval()
self.model_t = model_t
def run_step(self):
"""
Implement the moco training logic described above.
"""
assert self.model.training, "[KDTrainer] base model was changed to eval mode!"
start = time.perf_counter()
"""
If your want to do something with the data, you can wrap the dataloader.
"""
data = next(self._data_loader_iter)
data_time = time.perf_counter() - start
outs = self.model(data)
# Compute reid loss
if isinstance(self.model, DistributedDataParallel):
loss_dict = self.model.module.losses(outs)
else:
loss_dict = self.model.losses(outs)
with torch.no_grad():
outs_t = self.model_t(data)
q_logits = outs["outputs"]["pred_class_logits"]
t_logits = outs_t["outputs"]["pred_class_logits"].detach()
# k_logits = outs_k["outputs"]["pred_class_logits"].detach()
loss_dict['loss_kl'] = self.distill_loss(q_logits, t_logits, t=16)
losses = sum(loss_dict.values())
with torch.cuda.stream(torch.cuda.Stream()):
metrics_dict = loss_dict
metrics_dict["data_time"] = data_time
self._write_metrics(metrics_dict)
self._detect_anomaly(losses, loss_dict)
"""
If you need accumulate gradients or something similar, you can
wrap the optimizer with your custom `zero_grad()` method.
"""
self.optimizer.zero_grad()
losses.backward()
"""
If you need gradient clipping/scaling or other processing, you can
wrap the optimizer with your custom `step()` method.
"""
self.optimizer.step()
@classmethod
def build_model_teacher(cls, cfg) -> nn.Module:
cfg_t = update_model_teacher_config(cfg)
model_t = build_model(cfg_t)
return model_t
@staticmethod
def pkt_loss(output_net, target_net, eps=0.0000001):
# Normalize each vector by its norm
output_net_norm = torch.sqrt(torch.sum(output_net ** 2, dim=1, keepdim=True))
output_net = output_net / (output_net_norm + eps)
output_net[output_net != output_net] = 0
target_net_norm = torch.sqrt(torch.sum(target_net ** 2, dim=1, keepdim=True))
target_net = target_net / (target_net_norm + eps)
target_net[target_net != target_net] = 0
# Calculate the cosine similarity
model_similarity = torch.mm(output_net, output_net.transpose(0, 1))
target_similarity = torch.mm(target_net, target_net.transpose(0, 1))
# Scale cosine similarity to 0..1
model_similarity = (model_similarity + 1.0) / 2.0
target_similarity = (target_similarity + 1.0) / 2.0
# Transform them into probabilities
model_similarity = model_similarity / torch.sum(model_similarity, dim=1, keepdim=True)
target_similarity = target_similarity / torch.sum(target_similarity, dim=1, keepdim=True)
# Calculate the KL-divergence
loss = torch.mean(target_similarity * torch.log((target_similarity + eps) / (model_similarity + eps)))
return loss
@staticmethod
def distill_loss(y_s, y_t, t=4):
p_s = F.log_softmax(y_s / t, dim=1)
p_t = F.softmax(y_t / t, dim=1)
loss = F.kl_div(p_s, p_t, reduction='sum') * (t ** 2) / y_s.shape[0]
return loss