使用pyecharts做自己的visualDL

# 安装pyecharts! pip install -q pyecharts

# 持久化安装地图包# ! mkdir /home/aistudio/maps# ! pip install echarts-countries-pypkg -t /home/aistudio/maps# ! pip install echarts-china-provinces-pypkg -t /home/aistudio/maps# ! pip install echarts-china-cities-pypkg -t /home/aistudio/maps# ! pip install echarts-china-counties-pypkg -t /home/aistudio/maps# ! pip install echarts-china-misc-pypkg -t /home/aistudio/maps# ! pip install echarts-cities-pypkg -t /home/aistudio/maps# import sys # sys.path.append('/home/aistudio/maps')

import numpy as npimport paddlefrom paddle.vision.datasets import Cifar10import paddle.vision.transforms as Tfrom paddle.io import DataLoaderfrom paddle.nn.functional import cross_entropyfrom paddle.metric import accuracyfrom paddle.vision.models import mobilenet_v2from paddle.optimizer import Adam

# 配置batch_size = 2048learning_rate = 2e-6epoch_num = 10eval_epoch = 1  # 多少轮评估一次

# 数据normalize = T.ToTensor()  # 据大佬发现，totensor带有简单的normalize的功能train_data = Cifar10(mode='train', transform=normalize)
eval_data = Cifar10(mode='test', transform=normalize)
train_reader = DataLoader(train_data, batch_size=batch_size, shuffle=True, drop_last=True)
eval_reader = DataLoader(eval_data, batch_size=batch_size, shuffle=True, drop_last=True)# 模型model = mobilenet_v2(pretrained=True, num_classes=10)# 优化器opt = Adam(learning_rate=learning_rate, parameters=model.parameters())

from pyecharts import options as optsfrom pyecharts.charts import Line, Gauge  # 折线图和仪表图

# 训练Loss和Accdef show_train_laa(iters, loss, acc):
    c = (
        Line(init_opts=opts.InitOpts(width="512px", height="512px"))
        .add_xaxis(xaxis_data=list(range(iters)))
        .add_yaxis(
            series_name="TRAIN_LOSS",
            y_axis=list(loss),
            label_opts=opts.LabelOpts(is_show=False),
            yaxis_index=0,
            color="#d14a61"
        )
        .add_yaxis(
            series_name="TRAIN_ACC",
            y_axis=list(acc),
            label_opts=opts.LabelOpts(is_show=False),
            yaxis_index=1,
            color="#000000"
        )
        .extend_axis(
            yaxis=opts.AxisOpts(
                name="ACC",
                type_="value",
                position="right",
                axisline_opts=opts.AxisLineOpts(
                    linestyle_opts=opts.LineStyleOpts(color="#d14a61")
                ),
            )
        )
        .extend_axis(
            yaxis=opts.AxisOpts(
                name="LOSS",
                type_="value",
                position="left",
                axisline_opts=opts.AxisLineOpts(
                    linestyle_opts=opts.LineStyleOpts(color="#000000")
                ),
            )
        )
        .set_global_opts(
            title_opts=opts.TitleOpts(title="TRAIN_STATE"),
            tooltip_opts=opts.TooltipOpts(trigger="axis"),
            xaxis_opts=opts.AxisOpts(type_="category", boundary_gap=False),
        )
    )    return c# 评估Accdef show_eval_acc(eval_acc):
    c = (
        Gauge(init_opts=opts.InitOpts(width="512px", height="512px"))
        .add("", [("", (100 * eval_acc))])
        .set_global_opts(title_opts=opts.TitleOpts(title="EVAL_ACC"))
    )    return c

iters = 9loss = [45, 14, 7.2, 5.4, 2.2, 1, 0.99, 0.95, 0.75]
acc = [0.2, 0.41, 0.61, 0.66, 0.57, 0.68, 0.69, 0.71, 0.55]
show_train_laa(iters, loss, acc).render_notebook()  # 需要自己运行才能看到

eval_acc = 0.782show_eval_acc(eval_acc).render_notebook()  # render_notebook可以在notebook环境显示

# 刷新htmldef f5_html(html_path):
    with open(html_path, 'a+') as f:
        refesh = '<meta http-equiv="Refresh" content="3";/>'  # 每3秒钟刷新一次
        f.write(refesh)

losses = []
acces = []for e in range(epoch_num):    for i, (image, label) in enumerate(train_reader()):
        out = model(image)
        loss = cross_entropy(out, label)
        avg_loss = paddle.mean(loss)
        lab = paddle.reshape(label, (-1, 1))  # accuracy需要label是[*, 1]大小的
        acc = accuracy(out, lab)
        avg_loss.backward()
        opt.minimize(avg_loss)
        model.clear_gradients()
        losses.append(float(loss.numpy()[0]))
        acces.append(float(acc.numpy()[0]))
        show_train_laa(len(losses), losses, acces).render('train.html')  # 存入html
        f5_html('train.html')  # 刷新html
        # print("[train] Epoch {} batch {}: loss = {} acc = {}".format(e, i, np.mean(loss.numpy()), acc.numpy()[0]))
        if e % eval_epoch == 0 and i == ((50000 // batch_size) - 1):
            eval_accs = []            for ee, (eimage, elabel) in enumerate(eval_reader()):
                eout = model(eimage)
                elab = paddle.reshape(elabel, (-1, 1))
                eacc = accuracy(eout, elab).numpy()[0]
                eval_accs.append(float(eacc))                # print("[Eval] batch {}: acc = {}".format(ee, eacc))
            eval_acc = np.mean(eval_accs)
            show_eval_acc(eval_acc).render('eval.html')
            f5_html('eval.html')