feat(model): 实现自定义层数和大小的神经网络FCNet

Author: zjZSTU

pynet/models/FCNet.py

@@ -0,0 +1,134 @@
+# -*- coding: utf-8 -*-
+
+# @Time    : 19-7-8 下午1:36
+# @Author  : zj
+
+"""
+参考 cs231n assignment2 FullyConnectedNet，实现自定义层数和大小的神经网络
+
+网络结构为
+
+{FC - [batch/layer norm] - RELU - [dropout]} * (L - 1) - FC
+"""
+
+__all__ = ['FCNet']
+
+import numpy as np
+from pynet import nn
+from .Net import Net
+
+
+class FCNet(Net):
+    """
+    实现2层神经网络
+    """
+
+    def __init__(self, hidden_dims, input_dim=3 * 32 * 32, num_classes=10,
+                 dropout=1.0, weight_scale=1e-2, dtype=np.double):
+        super(FCNet, self).__init__()
+        self.hidden_dims = hidden_dims
+        self.input_dim = input_dim
+        self.num_classes = num_classes
+        self.weight_scale = weight_scale
+        self.dtype = dtype
+
+        self.use_dropout = dropout != 1
+        self.num_layers = 1 + len(hidden_dims)
+        self.relu = nn.ReLU()
+
+        self.fcs = self._get_fcs()
+        self.params = self._get_params()
+        self.caches = self._get_caches()
+
+        # 转换参数为指定数据类型
+        for k, v in self.params.items():
+            self.params[k] = v.astype(dtype)
+
+        self.use_dropout = dropout != 1.0
+        self.dropout_param = {}
+        if self.use_dropout:
+            self.dropout_param['mode'] = 'train'
+            self.dropout_param['p'] = dropout
+            self.dropout = nn.Dropout()
+            self.U1 = None
+            self.U2 = None
+
+    def __call__(self, inputs):
+        return self.forward(inputs)
+
+    def forward(self, inputs):
+        inputs = inputs.reshape(inputs.shape[0], -1)
+        inputs = inputs.astype(self.dtype)
+
+        x = None
+        for i in range(self.num_layers):
+            w = self.params['W%d' % (i + 1)]
+            b = self.params['b%d' % (i + 1)]
+
+            if i == 0:
+                x = inputs
+            self.caches['z%d' % (i + 1)], self.caches['z%d_cache' % (i + 1)] = self.fcs[i].forward(x, w, b)
+
+            if i != (self.num_layers - 1):
+                x = self.relu(self.caches['z%d' % (i + 1)])
+
+        return self.caches['z%d' % self.num_layers]
+
+    def backward(self, grad_out):
+        grad = dict()
+
+        da = None
+        for i in reversed(range(self.num_layers)):
+            z = self.caches['z%d' % (i + 1)]
+            cache = self.caches['z%d_cache' % (i + 1)]
+
+            if i == (self.num_layers - 1):
+                dz = grad_out
+            else:
+                dz = self.relu.backward(da, z)
+
+            grad['W%d' % (i + 1)], grad['b%d' % (i + 1)], da = self.fcs[i].backward(dz, cache)
+
+        self.caches = self._get_caches()
+        return grad
+
+    def _get_fcs(self):
+        fcs = list()
+        if self.hidden_dims is None:
+            fcs.append(nn.FC(self.input_dim, self.num_classes, weight_scale=self.weight_scale))
+        else:
+            for i in range(self.num_layers):
+                if i == 0:
+                    num_in = self.input_dim
+                else:
+                    num_in = self.hidden_dims[i - 1]
+
+                if i == (self.num_layers - 1):
+                    num_out = self.num_classes
+                else:
+                    num_out = self.hidden_dims[i]
+
+                fcs.append(nn.FC(num_in, num_out))
+
+        return fcs
+
+    def _get_params(self):
+        params = dict()
+        for i, fc in enumerate(self.fcs):
+            params['W%d' % (i + 1)], params['b%d' % (i + 1)] = fc.get_params()
+        return params
+
+    def _get_caches(self):
+        caches = dict()
+        for i in range(1, self.num_layers):
+            caches['z%d' % i] = None
+            caches['z%d_cache' % i] = None
+        return caches
+
+    def train(self):
+        if self.use_dropout:
+            self.dropout_param['mode'] = 'train'
+
+    def eval(self):
+        if self.use_dropout:
+            self.dropout_param['mode'] = 'test'

pynet/models/__init__.py

@@ -1,5 +1,6 @@
 from .TwoLayerNet import *
 from .ThreeLayerNet import *
+from .FCNet import *
 from .LeNet5 import *
 from .AlexNet import *
 from .NIN import *