缺少yaml、tensorboardX、pandas、seaborn
debug的时候会出现错误,显示没有“yaml”库,在pycharm中安装yaml也不行,经过查阅资料发现需要安装一个pyyaml库。
来到settings中点击右边的加号搜索并安装库。
第一步:
打印信息,打印出加载的参数都是什么。
其实就是做一些配置。
第二步:
指定路径
在debug环境下,把鼠标移到参数上可以查看具体信息。
只有当网络模型和预训练模型并且名字一致时,那些层才会加载,所以有些层是不会加载的。
要冻住哪些层可以自己指定,冻住的层不参与更新。
# Freeze
freeze = [] # parameter names to freeze (full or partial)
for k, v in model.named_parameters():
v.requires_grad = True # train all layers
if any(x in k for x in freeze):
print('freezing %s' % k)
v.requires_grad = False
累加梯度,间接地做梯度下降。
# Optimizer
nbs = 4 # nominal batch size
accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing
hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay
logger.info(f"Scaled weight_decay = {hyp['weight_decay']}")
设置三种参数,用相同的方法更新,但是不同的层用不同的衰减策略。
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_modules():
if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter):
pg2.append(v.bias) # biases
if isinstance(v, nn.BatchNorm2d):
pg0.append(v.weight) # no decay
elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter):
pg1.append(v.weight) # apply decay
if hasattr(v, 'im'):
if hasattr(v.im, 'implicit'):
pg0.append(v.im.implicit)
else:
for iv in v.im:
pg0.append(iv.implicit)
if hasattr(v, 'imc'):
if hasattr(v.imc, 'implicit'):
pg0.append(v.imc.implicit)
else:
for iv in v.imc:
pg0.append(iv.implicit)
if hasattr(v, 'imb'):
if hasattr(v.imb, 'implicit'):
pg0.append(v.imb.implicit)
else:
for iv in v.imb:
pg0.append(iv.implicit)
if hasattr(v, 'imo'):
if hasattr(v.imo, 'implicit'):
pg0.append(v.imo.implicit)
else:
for iv in v.imo:
pg0.append(iv.implicit)
if hasattr(v, 'ia'):
if hasattr(v.ia, 'implicit'):
pg0.append(v.ia.implicit)
else:
for iv in v.ia:
pg0.append(iv.implicit)
if hasattr(v, 'attn'):
if hasattr(v.attn, 'logit_scale'):
pg0.append(v.attn.logit_scale)
if hasattr(v.attn, 'q_bias'):
pg0.append(v.attn.q_bias)
if hasattr(v.attn, 'v_bias'):
pg0.append(v.attn.v_bias)
if hasattr(v.attn, 'relative_position_bias_table'):
pg0.append(v.attn.relative_position_bias_table)
if hasattr(v, 'rbr_dense'):
if hasattr(v.rbr_dense, 'weight_rbr_origin'):
pg0.append(v.rbr_dense.weight_rbr_origin)
if hasattr(v.rbr_dense, 'weight_rbr_avg_conv'):
pg0.append(v.rbr_dense.weight_rbr_avg_conv)
if hasattr(v.rbr_dense, 'weight_rbr_pfir_conv'):
pg0.append(v.rbr_dense.weight_rbr_pfir_conv)
if hasattr(v.rbr_dense, 'weight_rbr_1x1_kxk_idconv1'):
pg0.append(v.rbr_dense.weight_rbr_1x1_kxk_idconv1)
if hasattr(v.rbr_dense, 'weight_rbr_1x1_kxk_conv2'):
pg0.append(v.rbr_dense.weight_rbr_1x1_kxk_conv2)
if hasattr(v.rbr_dense, 'weight_rbr_gconv_dw'):
pg0.append(v.rbr_dense.weight_rbr_gconv_dw)
if hasattr(v.rbr_dense, 'weight_rbr_gconv_pw'):
pg0.append(v.rbr_dense.weight_rbr_gconv_pw)
if hasattr(v.rbr_dense, 'vector'):
pg0.append(v.rbr_dense.vector)
if opt.adam:
optimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum
else:
optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)
optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
学习率衰减策略:
if opt.linear_lr:
lf = lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp['lrf']) + hyp['lrf'] # linear,线性衰减
else:
lf = one_cycle(1, hyp['lrf'], epochs) # cosine 1->hyp['lrf'], one cycle的学习率变化过程是从lr=0.01呈余弦变化衰减到lr0*lrf = 0.01 * 0.1 = 0.001
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)