simple_convnet4.py 5.96 KB
import sys, os
sys.path.append(os.pardir)
import pickle
import numpy as cp
import numpy as np
from collections import OrderedDict
from layers import *


class SimpleConvNet:
    def __init__(self, input_dim=(3, 32, 32),
                 conv_param={'filter_num':(32, 32, 64), 'filter_size':3, 'pad':1, 'stride':1},
                 hidden_size=512, output_size=10, weight_init_std=0.01, pretrained=False):
        filter_num = conv_param['filter_num']
        filter_size = conv_param['filter_size']
        filter_pad = conv_param['pad']
        filter_stride = conv_param['stride']
        input_size = input_dim[1]
        conv_output_size = (input_size - filter_size + 2*filter_pad) / filter_stride + 1
        conv_data_size    = int(filter_num[0] *  conv_output_size    *  conv_output_size   )
        pool1_output_size = int(filter_num[1] * (conv_output_size/2) * (conv_output_size/2))
        pool2_output_size = int(filter_num[2] * (conv_output_size/4) * (conv_output_size/4))
        pool3_output_size = int(filter_num[2] * (conv_output_size/8) * (conv_output_size/8))

        self.params = {}
        if pretrained:
            weights = self.load_weights()
            self.params['W1'] = cp.array(weights['W1'], dtype=np.float32)
            self.params['W2'] = cp.array(weights['W2'], dtype=np.float32)
            self.params['W3'] = cp.array(weights['W3'], dtype=np.float32)
            self.params['W4'] = cp.array(weights['W4'], dtype=np.float32)
            self.params['W5'] = cp.array(weights['W5'], dtype=np.float32)
            self.params['W6'] = cp.array(weights['W6'], dtype=np.float32)
            self.params['W7'] = cp.array(weights['W7'], dtype=np.float32)
        else:
            self.params['W1'] = cp.array( weight_init_std * \
                                cp.random.randn(filter_num[0], input_dim[0], filter_size, filter_size), dtype=np.float32)

            self.params['W2'] = cp.array( weight_init_std * \
                                cp.random.randn(filter_num[1], filter_num[0], 1, 1), dtype=np.float32)

            self.params['W3'] = cp.array( weight_init_std * \
                                cp.random.randn(filter_num[1], 1, filter_size, filter_size), dtype=np.float32)

            self.params['W4'] = cp.array( weight_init_std * \
                                cp.random.randn(filter_num[2], filter_num[1], 1, 1), dtype=np.float32)

            self.params['W5'] = cp.array( weight_init_std * \
                                cp.random.randn(filter_num[2], 1, filter_size, filter_size), dtype=np.float32)

            self.params['W6'] = cp.array( weight_init_std * \
                                cp.random.randn(pool3_output_size, hidden_size), dtype=np.float32)

            self.params['W7'] = cp.array( weight_init_std * \
                                cp.random.randn(hidden_size, output_size), dtype=np.float32)

        self.layers = OrderedDict()
        self.layers['Conv1'] = Convolution(self.params['W1'],
                                           conv_param['stride'], conv_param['pad'])
        self.layers['LightNorm1'] = LightNormalization()
        self.layers['Relu1'] = Relu()
        self.layers['Pool1'] = Pooling(pool_h=2, pool_w=2, stride=2)

        self.layers['Conv2'] = Convolution(self.params['W2'],
                                           1, 0)
        self.layers['LightNorm2'] = LightNormalization()
        self.layers['Relu2'] = Relu()
        self.layers['Conv3'] = DW_Convolution(self.params['W3'],
                                           conv_param['stride'], conv_param['pad'])
        self.layers['LightNorm3'] = LightNormalization()
        self.layers['Relu3'] = Relu()
        self.layers['Pool2'] = Pooling(pool_h=2, pool_w=2, stride=2)

        self.layers['Conv4'] = Convolution(self.params['W4'],
                                           1, 0)
        self.layers['LightNorm4'] = LightNormalization()
        self.layers['Relu4'] = Relu()
        self.layers['Conv5'] = DW_Convolution(self.params['W5'],
                                           conv_param['stride'], conv_param['pad'])
        self.layers['LightNorm5'] = LightNormalization()
        self.layers['Relu5'] = Relu()
        self.layers['Pool3'] = Pooling(pool_h=2, pool_w=2, stride=2)

        self.layers['Affine4'] = Affine(self.params['W6'])
        self.layers['LightNorm6'] = LightNormalization()
        self.layers['Relu6'] = Relu()

        self.layers['Affine5'] = Affine(self.params['W7'])

        self.last_layer = SoftmaxWithLoss()

    def predict(self, x):
        for layer in self.layers.values():
            x = layer.forward(x)
        return x

    def accuracy(self, x, t, batch_size=100):
        if t.ndim != 1 : t = np.argmax(t, axis=1)

        acc = 0.0

        for i in range(int(x.shape[0] / batch_size)):
            tx = x[i*batch_size:(i+1)*batch_size]
            tt = t[i*batch_size:(i+1)*batch_size]
            y = self.predict(tx)
            y = np.argmax(y, axis=1)
            print("answer : ", tt)
            print("predict : ", y)
            acc += np.sum(y == tt) #numpy

        return acc / x.shape[0]

    def gradient(self, x, t):
        self.loss(x, t)

        dout = 1
        dout = self.last_layer.backward(dout)

        layers = list(self.layers.values())
        layers.reverse()
        for layer in layers:
            dout = layer.backward(dout)

        grads = {}
        grads['W1'] = self.layers['Conv1'].dW
        grads['W2'] = self.layers['Conv2'].dW
        grads['W3'] = self.layers['Conv3'].dW
        grads['W4'] = self.layers['Conv4'].dW
        grads['W5'] = self.layers['Conv5'].dW
        grads['W6'] = self.layers['Affine4'].dW
        grads['W7'] = self.layers['Affine5'].dW
        return grads

    # ���� ����ġ �ҷ����� / SimpleconvNet�� pretrained ���� �߰��� : True�� ����ġ �о� ����
    def load_weights(self, file_name='params.pkl'):
        weights = []
        with open(file_name, 'rb') as f:
            weights = pickle.load(f)
            return weights