代码来源：https://github.com/eriklindernoren/ML-From-Scratch
卷积神经网络中卷积层Conv2D（带stride、padding）的具体实现：https://www.cnblogs.com/xiximayou/p/12706576.html
激活函数的实现（sigmoid、softmax、tanh、relu、leakyrelu、elu、selu、softplus）：https://www.cnblogs.com/xiximayou/p/12713081.html
损失函数定义（均方误差、交叉熵损失）：https://www.cnblogs.com/xiximayou/p/12713198.html
优化器的实现（SGD、Nesterov、Adagrad、Adadelta、RMSprop、Adam）：https://www.cnblogs.com/xiximayou/p/12713594.html
卷积层反向传播过程：https://www.cnblogs.com/xiximayou/p/12713930.html
全连接层实现：https://www.cnblogs.com/xiximayou/p/12720017.html
批量归一化层实现：https://www.cnblogs.com/xiximayou/p/12720211.html
池化层实现：https://www.cnblogs.com/xiximayou/p/12720324.html
padding2D实现：https://www.cnblogs.com/xiximayou/p/12720454.html
Flatten层实现：https://www.cnblogs.com/xiximayou/p/12720518.html
上采样层UpSampling2D实现：https://www.cnblogs.com/xiximayou/p/12720558.html
Dropout层实现：https://www.cnblogs.com/xiximayou/p/12720589.html
激活层实现：https://www.cnblogs.com/xiximayou/p/12720622.html
首先是所有的代码：
from __future__ import print_function, division from terminaltables import AsciiTable import numpy as np import progressbar from mlfromscratch.utils import batch_iterator from mlfromscratch.utils.misc import bar_widgets class NeuralNetwork(): """Neural Network. Deep Learning base model. Parameters: ----------- optimizer: class The weight optimizer that will be used to tune the weights in order of minimizing the loss. loss: class Loss function used to measure the model's performance. SquareLoss or CrossEntropy. validation: tuple A tuple containing validation data and labels (X, y) """ def __init__(self, optimizer, loss, validation_data=None): self.optimizer = optimizer self.layers = [] self.errors = {"training": [], "validation": []} self.loss_function = loss() self.progressbar = progressbar.ProgressBar(widgets=bar_widgets) self.val_set = None if validation_data: X, y = validation_data self.val_set = {"X": X, "y": y} def set_trainable(self, trainable): """ Method which enables freezing of the weights of the network's layers. """ for layer in self.layers: layer.trainable = trainable def add(self, layer): """ Method which adds a layer to the neural network """ # If this is not the first layer added then set the input shape # to the output shape of the last added layer if self.layers: layer.set_input_shape(shape=self.layers[-1].output_shape()) # If the layer has weights that needs to be initialized if hasattr(layer, 'initialize'): layer.initialize(optimizer=self.optimizer) # Add layer to the network self.layers.append(layer) def test_on_batch(self, X, y): """ Evaluates the model over a single batch of samples """ y_pred = self._forward_pass(X, training=False) loss = np.mean(self.loss_function.loss(y, y_pred)) acc = self.loss_function.acc(y, y_pred) return loss, acc def train_on_batch(self, X, y): """ Single gradient update over one batch of samples """ y_pred = self._forward_pass(X) loss = np.mean(self.loss_function.loss(y, y_pred)) acc = self.loss_function.acc(y, y_pred) # Calculate the gradient of the loss function wrt y_pred loss_grad = self.loss_function.gradient(y, y_pred) # Backpropagate. Update weights self._backward_pass(loss_grad=loss_grad) return loss, acc def fit(self, X, y, n_epochs, batch_size): """ Trains the model for a fixed number of epochs """ for _ in self.progressbar(range(n_epochs)): batch_error = [] for X_batch, y_batch in batch_iterator(X, y, batch_size=batch_size): loss, _ = self.train_on_batch(X_batch, y_batch) batch_error.append(loss) self.errors["training"].append(np.mean(batch_error)) if self.val_set is not None: val_loss, _ = self.test_on_batch(self.val_set["X"], self.val_set["y"]) self.errors["validation"].append(val_loss) return self.errors["training"], self.errors["validation"] def _forward_pass(self, X, training=True): """ Calculate the output of the NN """ layer_output = X for layer in self.layers: layer_output = layer.forward_pass(layer_output, training) return layer_output def _backward_pass(self, loss_grad): """ Propagate the gradient 'backwards' and update the weights in each layer """ for layer in reversed(self.layers): loss_grad = layer.backward_pass(loss_grad) def summary(self, name="Model Summary"): # Print model name print (AsciiTable([[name]]).table) # Network input shape (first layer's input shape) print ("Input Shape: %s" % str(self.layers[0].input_shape)) # Iterate through network and get each layer's configuration table_data = [["Layer Type", "Parameters", "Output Shape"]] tot_params = 0 for layer in self.layers: layer_name = layer.layer_name() params = layer.parameters() out_shape = layer.output_shape() table_data.append([layer_name, str(params), str(out_shape)]) tot_params += params # Print network configuration table print (AsciiTable(table_data).table) print ("Total Parameters: %d\n" % tot_params) def predict(self, X): """ Use the trained model to predict labels of X """ return self._forward_pass(X, training=False)
接着我们来一个一个函数进行分析：
1、初始化__init__：这里面定义好优化器optimizer、模型层layers、错误errors、损失函数loss_function、用于显示进度条progressbar，这里从mlfromscratch.utils.misc中导入了bar_widgets，我们看看这是什么：
bar_widgets = [ 'Training: ', progressbar.Percentage(), ' ', progressbar.Bar(marker="-", left="[", right="]"), ' ', progressbar.ETA() ]
2、set_trainable()：用于设置哪些模型层需要进行参数的更新
3、add()：将一个模块放入到卷积神经网络中，例如卷积层、池化层、激活层等等。
4、test_on_batch()：使用batch进行测试，这里不需要进行反向传播。
5、train_on_batch()：使用batch进行训练，包括前向传播计算损失以及反向传播更新参数。
6、fit()：喂入数据进行训练或验证，这里需要定义好epochs和batch_size的大小，同时有一个读取数据的函数batch_iterator()，位于mlfromscratch.utils下的data_manipulation.py中：
def batch_iterator(X, y=None, batch_size=64): """ Simple batch generator """ n_samples = X.shape[0] for i in np.arange(0, n_samples, batch_size): begin, end = i, min(i+batch_size, n_samples) if y is not None: yield X[begin:end], y[begin:end] else: yield X[begin:end]
7、_forward_pass()：模型层的前向传播。
8、_backward_pass()：模型层的反向传播。
9、summary()：用于输出模型的每层的类型、参数数量以及输出大小。
10、predict()：用于输出预测值。
不难发现，该代码是借鉴了tensorflow中的一些模块的设计思想。