torchviz生成一個pdf,pdf怎樣命名還不知道,或許只能默認命名。
注意,torchviz我試了下將模型和偽造輸入移到cuda上就會報錯,所以這種可視化直接在cpu運行上即可。
還有tensorboardX方式,兩種方式代碼:
1 from tensorboardX import SummaryWriter 2 from torchviz import make_dot 3 4 # 准備數據 5 inputs_fake = torch.randn(13, 1, 32, 32) 6 model = glimpse_cnn # 模型實例 7 8 # TensorboardX方式 9 with SummaryWriter(comment='vgg') as w: 10 w.add_graph(model, (inputs_fake,)) 11 12 # torchviz方式. inputs_fake后面跟.requires_grad_(True)顯示x輸入形狀框 13 vis_graph = make_dot(model(inputs_fake.requires_grad_(True)), 14 params=dict(list(model.named_parameters()) + [('x', inputs_fake)])) 15 vis_graph.view()
也可以直接保存為文本格式,很詳細,方便后面查看:
1 with open('model.txt', 'w') as w: 2 w.write(str(model))