读Linkage Based Face Clustering via Graph Convolution代码复现

+-------------------+---------------------------------------+
|     Parameter     |                 Value                 |
+===================+=======================================+
| Active connection | 10                                    |
+-------------------+---------------------------------------+
| Batch size        | 16                                    |
+-------------------+---------------------------------------+
| Epochs            | 4                                     |
+-------------------+---------------------------------------+
| Feat path         | facedata/CASIA.feas.npy               |
+-------------------+---------------------------------------+
| K at hop          | [200, 10]                             |
+-------------------+---------------------------------------+
| Knn graph path    | facedata/knn.graph.CASIA.kdtree.npy   |
+-------------------+---------------------------------------+
| Label path        | facedata/CASIA.labels.npy             |
+-------------------+---------------------------------------+
| Logs dir          | D:\jupyter\gcn_clustering-master\logs |
+-------------------+---------------------------------------+
| Lr                | 0.010                                 |
+-------------------+---------------------------------------+
| Momentum          | 0.900                                 |
+-------------------+---------------------------------------+
| Print freq        | 200                                   |
+-------------------+---------------------------------------+
| Seed              | 1                                     |
+-------------------+---------------------------------------+
| Weight decay      | 0.000                                 |
+-------------------+---------------------------------------+
| Workers           | 16                                    |对于我这普通电脑还是得调成0啊
+-------------------+---------------------------------------+

依然是用dataset和dataloader这俩包直接划分好数据
读取CASIA.feas.npy得到(454590, 512)的features，knn.graph.CASIA.kdtree.npy得到(454590, 201)的knn_graph，CASIA.labels.npy得到454590的labels
每轮一开始先adjust_lr，自适应调整学习率，从1轮还是每轮乘个0.1（不断减小学习率啊）
训练刚开始，通过AverageMeter类初始化一些性能指标
遍历dataloader划分出的batch，有[16, 2201, 512]的feat，[16, 2201, 2201]的adj，[16, 1]的cid，[16, 200]的h1id，[16, 200]的gtmat（前4个好像原本是包在一个tuple里的）

输入：
	feat->x				[16, 2201, 512]
	adj	->A				[16, 2201, 2201]
	h1id->one_hop_idcs	[16, 200]
gcn(
  #把x view成[35216, 512]
  (bn0): BatchNorm1d(512, eps=1e-05, momentum=0.1, affine=False, track_running_stats=True)
  #再view回[16, 2201, 512]
  (conv1): GraphConv(
    (agg): MeanAggregator()
    	#[16,2201,2201]x[16,2201,512]->[16,2201,512]agg_feats
    #拼接前后两个[16,2201,512]得到cat_feats[16,2201,1024]
    #通过einsum，[16,2201,1024]x[1024,512]->[16,2201,512]
    #加偏重后relu
  )
  (conv2): GraphConv(
    (agg): MeanAggregator()
  )
  (conv3): GraphConv(
    (agg): MeanAggregator()
    #[16,2201,1024]x[1024,256]->[16,2201,256]
  )
  (conv4): GraphConv(
    (agg): MeanAggregator()
    #[16,2201,512]x[512,256]->[16,2201,256]
  )
  #初始化一个全0的[16,200,256]的edge_feat
  for b in range(B):	#16
    edge_feat[b,:,:] = x[b, one_hop_idcs[b]]
  #再view成[3200,256]
  (classifier): Sequential(
    (0): Linear(in_features=256, out_features=256, bias=True)
    (1): PReLU(num_parameters=256)
    (2): Linear(in_features=256, out_features=2, bias=True)
  )
  #得到[3200,2]
)
SGD (
Parameter Group 0
    dampening: 0
    lr: 0.01
    momentum: 0.9
    nesterov: False
    weight_decay: 0.0001
)

把[16,200]的gtmat view成3200的labels和pred交叉熵
再计算准确率（这会才发现gtmat里只有0和1，这怎么就成了二分类？）从而有平均准确率acc0.5884，precision0.07675753228120516，召回率recall0.781021897810219

然后再opt.zero_grad()，loss.backward()，opt.step()（也就是说这之前计算的变量都没有梯度？这样一说，好像这样反传更加合理？）
————————————————————
（又看了下关于二分类的事，是在一开始dataset类里实现的，总之就是和我想象的节点聚类不太一样，应该是对我没啥帮助了）