\n
![\"\"](\"https:\/\/lingbo.online\/wp-content\/uploads\/2023\/07\/36-7976-1.jpg\")
nn.LSTM<\/h3>\n__init__<\/h4>\n\n- input_size\uff1a\u8f93\u5165\u7279\u5f81x\u7684\u7ef4\u6570\uff0c\u5373\u6bcf\u4e00\u884c\u8f93\u5165\u5143\u7d20\u7684\u4e2a\u6570\u3002<\/li>\n
- hidden_size\uff1a\u9690\u85cf\u5c42\u72b6\u6001h\u7684\u7ef4\u6570\uff0c\u5373\u9690\u85cf\u5c42\u8282\u70b9\u7684\u4e2a\u6570\u3002\u9690\u85cf\u5c42\u7684\u7ef4\u6570\u503c\u662f\u81ea\u5b9a\u4e49\u7684\uff0c\u548c\u8f93\u5165\u7684\u7ef4\u5ea6\u6ca1\u6709\u5173\u7cfb\u3002<\/li>\n
- num_layers\uff1aLSTM\u5806\u53e0\u7684\u5c42\u6570\uff0c\u9ed8\u8ba4\u503c\u662f1\u5c42\uff0c\u5982\u679c\u8bbe\u7f6e\u4e3a2\uff0c\u5c06\u88ab\u5904\u7406\u4e3a\u4e00\u4e2a
stacked LSTM<\/code>\uff0c\u5373\u7b2c\u4e8c\u4e2aLSTM\u63a5\u6536\u7b2c\u4e00\u4e2aLSTM\u7684\u8ba1\u7b97\u7ed3\u679c\uff0c\u5982\uff0c\u7b2c\u4e00\u5c42\u8f93\u5165[X_0,X_1,X_2,...,X_t]<\/span>\uff0c\u8ba1\u7b97\u51fa[h_0,h_1,h_2,...,h_t]<\/span>\uff0c\u7b2c\u4e8c\u5c42\u5c06 [h_0,h_1,h_2,...,h_t]<\/span>\uff0c\u4f5c\u4e3a[X_0,X_1,X_2,...,X_t]<\/span>\u8f93\u5165\u518d\u6b21\u8ba1\u7b97\uff0c\u8f93\u51fa\u6700\u540e\u7684[h_0,h_1,h_2,...,h_t]<\/span>\u3002\u9ed8\u8ba4\u503c\u4e3a1\u3002<\/li>\n<\/ul>\nLSTM.forward()<\/h4>\n
\u4e0d\u540c\u4e8eRNN\uff0cLSTM\u5728\u524d\u5411\u4f20\u64ad\u65f6\u8fd4\u56de\u4e24\u4e2a\u4e2d\u95f4\u53d8\u91cfh_t,c_t<\/span><\/p>\nout,(ht,ct)=lstm(x,[ht_1,ct_1])\n\n#\u6848\u4f8b\uff1a\u5efa\u7acb\u4e00\u4e2aLSTM\uff0c\u8f93\u5165\u7ef4\u6570\u4e3a100\uff0c\u9690\u85cf\u5c42\u4e3a20,LSTM\u5c42\u6570\u4e3a4\nlstm=nn.LSTM(input_size=100,hidden_size=20,num_layers=4)\nprint(lstm)\nx=torch.randn(10,3,100) # x:[seq,b,vec],\u53ef\u4ee5\u7406\u89e3\u4e3a3\u4e2a\u53e5\u5b50\uff0c\u6bcf\u4e2a\u53e5\u5b50\u6709\u5341\u4e2a\u5355\u8bcd\uff0c\u6bcf\u4e2a\u5355\u8bcdencoding\u6210\u4e3a\u957f\u5ea6\u4e3a100\u7684vector\nout,(h,c)=lstm(x) #[h_0,c_0]\nprint(out.shape,h.shape,c.shape)\n\n#torch.Size([10,3,20]) out:[seq,b,h]\n#torch.Size([4,3,20]) h\/c:[num_layer,b,h]\n#torch.Size([4,3,20]) h\/c:[num_layer,b,h]<\/code><\/pre>\nnn.LSTMCell<\/h3>\n__init__<\/h4>\n\n- input_size\uff1a\u8f93\u5165\u7ef4\u5ea6x<\/li>\n
- hidden_size\uff1a\u9690\u85cf\u5c42h<\/li>\n
- num_layers\uff1aLSTM\u5806\u53e0\u7684\u5c42\u6570<\/li>\n<\/ul>\n
LSTMCell.forward()<\/h4>\nht,ct=lstmcell(xt,[ht_1,ct_1])\n\n#Single layer\nprint('one layer lstm')\ncell=nn.LSTMCell(input_size=100,hidden_size=20)\nh=torch.zeros(3,20)\nc=torch.zeros(3,20) #1\u5c42\uff0c3\u53e5\u8bdd\uff0chidden=20\nfor xt in x:\n h,c=cell(xt,[h,c]) #xt:[b,vec]\nprint(h.shape,c.shape)\n\ntorch.Size([3,20])\ntorch.Size([3,20])#ht\/ct:[b,h]\n\n#Two Layers\nprint('two layer lstm')\ncell1=nn.LSTMCell(input_size=100,hidden_size=30)\ncell2=nn.LSTMCell(input_size=30,hidden_size=20)\nh1=torch.zeros(3,30)\nc1=torch.zeros(3,30)\nh2=torch.zeros(3,20)\nc2=torch.zeros(3,20)\nfor xt in x:\n h1,c1=cell1(xt,[ht,c1])\n h2,c2=cell2(h1,[h2,c2])\n\nprint(h2.shape,c2.shape)\ntorch.Size(3,20)\ntorch.Size(3,20)<\/code><\/pre>\n\u6848\u4f8b\u5206\u6790<\/h3>\n\u80cc\u666f\u4ecb\u7ecd<\/h4>\n
- \n
- input_size\uff1a\u8f93\u5165\u7279\u5f81x\u7684\u7ef4\u6570\uff0c\u5373\u6bcf\u4e00\u884c\u8f93\u5165\u5143\u7d20\u7684\u4e2a\u6570\u3002<\/li>\n
- hidden_size\uff1a\u9690\u85cf\u5c42\u72b6\u6001h\u7684\u7ef4\u6570\uff0c\u5373\u9690\u85cf\u5c42\u8282\u70b9\u7684\u4e2a\u6570\u3002\u9690\u85cf\u5c42\u7684\u7ef4\u6570\u503c\u662f\u81ea\u5b9a\u4e49\u7684\uff0c\u548c\u8f93\u5165\u7684\u7ef4\u5ea6\u6ca1\u6709\u5173\u7cfb\u3002<\/li>\n
- num_layers\uff1aLSTM\u5806\u53e0\u7684\u5c42\u6570\uff0c\u9ed8\u8ba4\u503c\u662f1\u5c42\uff0c\u5982\u679c\u8bbe\u7f6e\u4e3a2\uff0c\u5c06\u88ab\u5904\u7406\u4e3a\u4e00\u4e2a
stacked LSTM<\/code>\uff0c\u5373\u7b2c\u4e8c\u4e2aLSTM\u63a5\u6536\u7b2c\u4e00\u4e2aLSTM\u7684\u8ba1\u7b97\u7ed3\u679c\uff0c\u5982\uff0c\u7b2c\u4e00\u5c42\u8f93\u5165[X_0,X_1,X_2,...,X_t]<\/span>\uff0c\u8ba1\u7b97\u51fa[h_0,h_1,h_2,...,h_t]<\/span>\uff0c\u7b2c\u4e8c\u5c42\u5c06 [h_0,h_1,h_2,...,h_t]<\/span>\uff0c\u4f5c\u4e3a[X_0,X_1,X_2,...,X_t]<\/span>\u8f93\u5165\u518d\u6b21\u8ba1\u7b97\uff0c\u8f93\u51fa\u6700\u540e\u7684[h_0,h_1,h_2,...,h_t]<\/span>\u3002\u9ed8\u8ba4\u503c\u4e3a1\u3002<\/li>\n<\/ul>\nLSTM.forward()<\/h4>\n
\u4e0d\u540c\u4e8eRNN\uff0cLSTM\u5728\u524d\u5411\u4f20\u64ad\u65f6\u8fd4\u56de\u4e24\u4e2a\u4e2d\u95f4\u53d8\u91cfh_t,c_t<\/span><\/p>\n
out,(ht,ct)=lstm(x,[ht_1,ct_1])\n\n#\u6848\u4f8b\uff1a\u5efa\u7acb\u4e00\u4e2aLSTM\uff0c\u8f93\u5165\u7ef4\u6570\u4e3a100\uff0c\u9690\u85cf\u5c42\u4e3a20,LSTM\u5c42\u6570\u4e3a4\nlstm=nn.LSTM(input_size=100,hidden_size=20,num_layers=4)\nprint(lstm)\nx=torch.randn(10,3,100) # x:[seq,b,vec],\u53ef\u4ee5\u7406\u89e3\u4e3a3\u4e2a\u53e5\u5b50\uff0c\u6bcf\u4e2a\u53e5\u5b50\u6709\u5341\u4e2a\u5355\u8bcd\uff0c\u6bcf\u4e2a\u5355\u8bcdencoding\u6210\u4e3a\u957f\u5ea6\u4e3a100\u7684vector\nout,(h,c)=lstm(x) #[h_0,c_0]\nprint(out.shape,h.shape,c.shape)\n\n#torch.Size([10,3,20]) out:[seq,b,h]\n#torch.Size([4,3,20]) h\/c:[num_layer,b,h]\n#torch.Size([4,3,20]) h\/c:[num_layer,b,h]<\/code><\/pre>\nnn.LSTMCell<\/h3>\n
__init__<\/h4>\n
- \n
- input_size\uff1a\u8f93\u5165\u7ef4\u5ea6x<\/li>\n
- hidden_size\uff1a\u9690\u85cf\u5c42h<\/li>\n
- num_layers\uff1aLSTM\u5806\u53e0\u7684\u5c42\u6570<\/li>\n<\/ul>\n
LSTMCell.forward()<\/h4>\n
ht,ct=lstmcell(xt,[ht_1,ct_1])\n\n#Single layer\nprint('one layer lstm')\ncell=nn.LSTMCell(input_size=100,hidden_size=20)\nh=torch.zeros(3,20)\nc=torch.zeros(3,20) #1\u5c42\uff0c3\u53e5\u8bdd\uff0chidden=20\nfor xt in x:\n h,c=cell(xt,[h,c]) #xt:[b,vec]\nprint(h.shape,c.shape)\n\ntorch.Size([3,20])\ntorch.Size([3,20])#ht\/ct:[b,h]\n\n#Two Layers\nprint('two layer lstm')\ncell1=nn.LSTMCell(input_size=100,hidden_size=30)\ncell2=nn.LSTMCell(input_size=30,hidden_size=20)\nh1=torch.zeros(3,30)\nc1=torch.zeros(3,30)\nh2=torch.zeros(3,20)\nc2=torch.zeros(3,20)\nfor xt in x:\n h1,c1=cell1(xt,[ht,c1])\n h2,c2=cell2(h1,[h2,c2])\n\nprint(h2.shape,c2.shape)\ntorch.Size(3,20)\ntorch.Size(3,20)<\/code><\/pre>\n\u6848\u4f8b\u5206\u6790<\/h3>\n
\u80cc\u666f\u4ecb\u7ecd<\/h4>\n
![\"\"](\"https:\/\/lingbo.online\/wp-content\/uploads\/2023\/07\/\u6570\u636e\u683c\u5f0f.png\")