用python去读取模型保存为tensorboard格式（.v2）的文件源代码

from tensorboard.backend.event_processing import event_accumulator import matplotlib.pyplot as plt import numpy as np import tensorflow as tf ## 首先直接使用此方法 def read_tensorboard_data(tensorboard_path, name): """读取tensorboard数据， tensorboard_path是tensorboard数据地址val_name是需要读取的变量名称""" ea = event_accumulator.EventAccumulator(tensorboard_path) ea.Reload() tag = ea.Tags() print(tag) if tag['scalars'] != []: key = ea.scalars.Keys() print(key) val = ea.scalars.Items(name) val = [i.value for i in val] #遍历val_lo的值，得到step、value的值 else: # tag['scalars'] == [] but tag['tensors'] !=[] 用上面同样的方法行不通，可以试一下看看 val = [] for event in tf.compat.v1.train.summary_iterator(tensorboard_val_path): for value in event.summary.value: if value.tag == name: val.append(tf.make_ndarray(value.tensor)) #结果只读取到loss的值 return val def draw_plt(val, val_name): """将数据绘制成曲线图，val是数据，val_name是变量名称""" plt.figure() plt.plot([i.step for i in val], [j.value for j in val], label=val_name) """横坐标是step，迭代次数 纵坐标是变量值""" plt.xlabel('step') plt.ylabel(val_name) plt.show() def getTwoDimensionListIndex(L,value): """获得二维列表某个值的一维索引值 思想：先选出包含value值的一维列表，然后判断此一维列表在二维列表中的索引 """ data = [data for data in L if data[1]==value] #data=[(53, 1016.1)] index = int(np.argwhere((L==data[0]).all(axis=1))) return index # 另一种，只适用这一种的方法 def getindex(L, value): """ 一种只适用于读取tensorboard保存的v2版本的文件 因为step想当于index """ data = [data for data in L if data[1]==value] index = data[0][0] return index """ 本代码两种都可以使用 """ if __name__ == "__main__": tensorboard_train_path = './loop_noise5_cbamconcat_rms10_batchsize512/train/events.out.tfevents.1666350020.67d731bb51ab.314.0.v2' # tensorboard_train_path = './loop_noise5_cbamconcat_rms1_batchsize512/train/events.out.tfevents.1666014996.MS-IJHPFNLAGMLT.11164.17195.v2' tensorboard_val_path = './loop_noise5_cbamconcat_rms10_batchsize512/validation/events.out.tfevents.1666350039.67d731bb51ab.314.1.v2' # tensorboard_val_path = './loop_noise5_cbamconcat_rms1_batchsize512/validation/events.out.tfevents.1666015011.MS-IJHPFNLAGMLT.11164.37325.v2' # 读取epcoh_loss, 因为模型是以val中的epoch_loss为指标保存的 val_epoch_loss_name = 'epoch_loss' val_lo = read_tensorboard_data(tensorboard_val_path, val_epoch_loss_name) #读取tendorboard版本的文件的内容 val_loss = np.array(val_lo) val_loss_min = np.min(val_loss) #得到val_loss中的最小值 print("val_epoch_loss:{}".format(val_loss_min)) idx_min = int(np.argwhere(val_loss==val_loss_min)) #得到val_loss最小值对应的索引，即step的值，方便val_epoch_euclidean、train_loss、train_epoch_euclidean得到对应step的值 print(idx_min) # draw_plt(val_lo, val_epoch_loss_name) # # 读取val中的epoch_Euclidean值，并根据val_epoch_loss中最小值的索引找到对应的epoch_Euclidean的值 val_epoch_Euclidean_name = 'epoch_Euclidean' val_eucl = read_tensorboard_data(tensorboard_val_path, val_epoch_Euclidean_name) val_e_value = [i for i in val_eucl][idx_min] print("val_epoch_euclidean".format(val_e_value)) # draw_plt(val_eucl, val_epoch_Euclidean_name) # 读取train中的epoch_loss值，并根据val_epoch_loss中最小值的索引找到对应的train_epoch_loss的值 train_epoch_loss_name = 'epoch_loss' train_lo = read_tensorboard_data(tensorboard_train_path, train_epoch_loss_name) train_loss = [i for i in train_lo][idx_min] print("train_epoch_loss:{}".format(train_loss)) # draw_plt(train_lo, train_epoch_loss_name) # 读取train中的epoch_Euclidean值，并根据val_epoch_loss中最小值的索引找到对应的train_epoch_Euclidean的值 train_epoch_Euclidean_name = 'epoch_Euclidean' train_eucl = read_tensorboard_data(tensorboard_train_path, train_epoch_Euclidean_name) train_e_value = [i for i in train_eucl][idx_min] print("train_epoch_euclidean:{}".format(train_e_value)) # draw_plt(train_eucl, train_epoch_Euclidean_name)