caffe 输出信息分析+debug_info

1 caffe输出基本信息

caffe在训练的时候屏幕会输出程序运行的状态信息，通过查看状态信息方便查看程序运行是否正常，且方便查找bug.
caffe debug信息默认是不开启的，此时的输出信息的总体结构如下所示：

1.1 solver信息加载并显示

1.2 train网络结构输出

1.3 Train 各层创建状态信息

I0821 09:53:35.572999 10308 layer_factory.hpp:77] Creating layer mnist    ####创建第一层I0821 09:53:35.572999 10308 common.cpp:36] System entropy source not available, using fallback algorithm to generate seed instead.I0821 09:53:35.572999 10308 net.cpp:100] Creating Layer mnistI0821 09:53:35.572999 10308 net.cpp:418] mnist -> dataI0821 09:53:35.572999 10308 net.cpp:418] mnist -> labelI0821 09:53:35.572999 10308 data_transformer.cpp:25] Loading mean file from: ....../image_mean.binaryprotoI0821 09:53:35.579999 11064 common.cpp:36] System entropy source not available, using fallback algorithm to generate seed instead.I0821 09:53:35.580999 11064 db_lmdb.cpp:40] Opened lmdb ......./train/trainlmdbI0821 09:53:35.623999 10308 data_layer.cpp:41] output data size: 100,3,32,32  ###输出blob尺寸I0821 09:53:35.628999 10308 net.cpp:150] Setting up mnistI0821 09:53:35.628999 10308 net.cpp:157] Top shape: 100 3 32 32 (307200)I0821 09:53:35.628999 10308 net.cpp:157] Top shape: 100 (100)I0821 09:53:35.628999 10308 net.cpp:165] Memory required for data: 1229200I0821 09:53:35.628999 10308 layer_factory.hpp:77] Creating layer conv1   ##### 创建第二层I0821 09:53:35.628999 10308 net.cpp:100] Creating Layer conv1I0821 09:53:35.628999 10308 net.cpp:444] conv1 <- dataI0821 09:53:35.628999 10308 net.cpp:418] conv1 -> conv1I0821 09:53:35.629999  7532 common.cpp:36] System entropy source not available, using fallback algorithm to generate seed instead.I0821 09:53:35.909999 10308 net.cpp:150] Setting up conv1I0821 09:53:35.909999 10308 net.cpp:157] Top shape: 100 64 28 28 (5017600)  #### 输出blob尺寸I0821 09:53:35.909999 10308 net.cpp:165] Memory required for data: 21299600...I0821 09:53:35.914000 10308 layer_factory.hpp:77] Creating layer lossI0821 09:53:35.914000 10308 net.cpp:150] Setting up lossI0821 09:53:35.914000 10308 net.cpp:157] Top shape: (1)I0821 09:53:35.914000 10308 net.cpp:160]     with loss weight 1I0821 09:53:35.914000 10308 net.cpp:165] Memory required for data: 49322804I0821 09:53:35.914000 10308 net.cpp:226] loss needs backward computation.    ######## 各层反向传播信息I0821 09:53:35.914000 10308 net.cpp:226] ip2 needs backward computation.I0821 09:53:35.914000 10308 net.cpp:226] relu3 needs backward computation.I0821 09:53:35.914000 10308 net.cpp:226] ip1 needs backward computation.I0821 09:53:35.914000 10308 net.cpp:226] pool2 needs backward computation.I0821 09:53:35.914000 10308 net.cpp:226] relu2 needs backward computation.I0821 09:53:35.914000 10308 net.cpp:226] conv2 needs backward computation.I0821 09:53:35.914000 10308 net.cpp:226] pool1 needs backward computation.I0821 09:53:35.914000 10308 net.cpp:226] relu1 needs backward computation.I0821 09:53:35.914000 10308 net.cpp:226] conv1 needs backward computation.I0821 09:53:35.914000 10308 net.cpp:228] mnist does not need backward computation.I0821 09:53:35.914000 10308 net.cpp:270] This network produces output loss   ######## 网络输出节点个数及名称（重要），后续参数输出均是此节点的信息 ###############I0821 09:53:35.914000 10308 net.cpp:283] Network initialization done.  ###网络创建完成

1、通过查看网络创建信息科了解网络节点blob大小2、可知道网络后续最终输出信息3、test的创建过程与train类似，此处不再重复说明

1.4 Train 和Test网络的迭代信息输出

I0821 09:53:35.929999 10308 solver.cpp:60] Solver scaffolding done.I0821 09:53:35.929999 10308 caffe.cpp:252] Starting Optimization     ####### 开始网络训练I0821 09:53:35.929999 10308 solver.cpp:279] Solving LeNetI0821 09:53:35.929999 10308 solver.cpp:280] Learning Rate Policy: multistepI0821 09:53:35.930999 10308 solver.cpp:337] Iteration 0, Testing net (#0)                                                           #### Test（Iteration 0）I0821 09:53:35.993999 10308 blocking_queue.cpp:50] Data layer prefetch queue emptyI0821 09:53:36.180999 10308 solver.cpp:404]     Test net output #0: accuracy = 0.1121                                           #### Test（Iteration 0）网络输出节点0，accuracy信息。（由网络定义决定）I0821 09:53:36.180999 10308 solver.cpp:404]     Test net output #1: loss = 2.30972 (* 1 = 2.30972 loss)     #### Test（Iteration 0）网络输出节点1，loss信息。       （由网络定义决定）I0821 09:53:36.190999 10308 solver.cpp:228] Iteration 0, loss = 2.2891                                                                      #### Tain（Iteration 0） 网络loss值I0821 09:53:36.190999 10308 solver.cpp:244]     Train net output #0: loss = 2.2891 (* 1 = 2.2891 loss)    #### Tain（Iteration 0） 只有一个输出值I0821 09:53:36.190999 10308 sgd_solver.cpp:106] Iteration 0, lr = 0.001                                                                     #### Tain（Iteration 0）I0821 09:53:36.700999 10308 solver.cpp:228] Iteration 100, loss = 2.24716                                                                   #### Tain（Iteration 100）I0821 09:53:36.700999 10308 solver.cpp:244]     Train net output #0: loss = 2.24716 (* 1 = 2.24716 loss)    #### Tain（Iteration 100）I0821 09:53:36.700999 10308 sgd_solver.cpp:106] Iteration 100, lr = 0.001                                                                   #### Tain（Iteration 100）I0821 09:53:37.225999 10308 solver.cpp:228] Iteration 200, loss = 2.08563I0821 09:53:37.225999 10308 solver.cpp:244]     Train net output #0: loss = 2.08563 (* 1 = 2.08563 loss)I0821 09:53:37.225999 10308 sgd_solver.cpp:106] Iteration 200, lr = 0.001I0821 09:53:37.756000 10308 solver.cpp:228] Iteration 300, loss = 2.11631I0821 09:53:37.756000 10308 solver.cpp:244]     Train net output #0: loss = 2.11631 (* 1 = 2.11631 loss)I0821 09:53:37.756000 10308 sgd_solver.cpp:106] Iteration 300, lr = 0.001I0821 09:53:38.286999 10308 solver.cpp:228] Iteration 400, loss = 1.89424I0821 09:53:38.286999 10308 solver.cpp:244]     Train net output #0: loss = 1.89424 (* 1 = 1.89424 loss)I0821 09:53:38.286999 10308 sgd_solver.cpp:106] Iteration 400, lr = 0.001I0821 09:53:38.819999 10308 solver.cpp:337] Iteration 500, Testing net (#0)                                                             #### Test（Iteration 500）I0821 09:53:39.069999 10308 solver.cpp:404]     Test net output #0: accuracy = 0.3232                                           #### Test（Iteration 500）I0821 09:53:39.069999 10308 solver.cpp:404]     Test net output #1: loss = 1.87822 (* 1 = 1.87822 loss)     #### Test（Iteration 500）I0821 09:53:39.072999 10308 solver.cpp:228] Iteration 500, loss = 1.94478I0821 09:53:39.072999 10308 solver.cpp:244]     Train net output #0: loss = 1.94478 (* 1 = 1.94478 loss)I0821 09:53:39.072999 10308 sgd_solver.cpp:106] Iteration 500, lr = 0.001

从输出可以看出，Train和Test一次输出周期如下：

• Test 一次训练周期
  –Iteration 0, Testing net (#0)
  –Test net output #0: accuracy = 0.1121
  –Test net output #1: loss = 2.30972 (* 1 = 2.30972 loss)
• Train 一次训练周期
  –Iteration 0, loss = 2.2891
  –Train net output #0: loss = 2.2891 (* 1 = 2.2891 loss)
  –Iteration 0, lr = 0.001

2 log信息解析

2 debug info信息

• 在solver 中添加 debug_info:true
• 开启caffe的debug信息输出

import osimport reimport extract_secondsimport argparseimport csvfrom collections import OrderedDictdef get_datadiff_paradiff(line,data_row,para_row,data_list,para_list,L_list,L_row,top_list,top_row,iteration):    regex_data=re.compile('\[Backward\] Layer (\S+), bottom blob (\S+) diff: ([\.\deE+-]+)')    regex_para=re.compile('\[Backward\] Layer (\S+), param blob (\d+) diff: ([\.\deE+-]+)')    regex_L1L2=re.compile('All net params \(data, diff\): L1 norm = \(([\.\deE+-]+), ([\.\deE+-]+)\); L2 norm = \(([\.\deE+-]+), ([\.\deE+-]+)\)')    regex_topdata=re.compile('\[Forward\] Layer (\S+), (\S+) blob (\S+) data: ([\.\deE+-]+)')    #regex_toppara=re.compile('')    out_match_data=regex_data.search(line)    if out_match_data or iteration>-1:        if not data_row or iteration>-1 :            if data_row:                data_row['NumIters']=iteration                data_list.append(data_row)            data_row = OrderedDict()        if out_match_data :             layer_name=out_match_data.group(1)            blob_name=out_match_data.group(2)            data_diff_value=out_match_data.group(3)            key=layer_name+'-'+blob_name            data_row[key]=float(data_diff_value)    out_match_para=regex_para.search(line)    if out_match_para or iteration>-1:        if not para_row or iteration>-1:            if para_row:                para_row['NumIters']=iteration                para_list.append(para_row)            para_row=OrderedDict()        if out_match_para:            layer_name=out_match_para.group(1)            param_d=out_match_para.group(2)            para_diff_value=out_match_para.group(3)            layer_name=layer_name+'-blob'+'-'+param_d            para_row[layer_name]=para_diff_value    out_match_norm=regex_L1L2.search(line)    if out_match_norm or iteration>-1:        if not L_row or iteration>-1:            if L_row:                L_row['NumIters']=iteration                L_list.append(L_row)            L_row=OrderedDict()        if out_match_norm:            L_row['data-L1']=out_match_norm.group(1)            L_row['diff-L1']=out_match_norm.group(2)            L_row['data-L2']=out_match_norm.group(3)            L_row['diff-L2']=out_match_norm.group(4)    out_match_top=regex_topdata.search(line)    if out_match_top or iteration>-1:        if not top_row or iteration>-1:            if top_row:                top_row['NumIters']=iteration                top_list.append(top_row)            top_row=OrderedDict()        if out_match_top:            layer_name=out_match_top.group(1)            top_para=out_match_top.group(2)            blob_or_num=out_match_top.group(3)            key=layer_name+'-'+top_para+'-'+blob_or_num            data_value=out_match_top.group(4)            top_row[key]=float(data_value)    return data_list,data_row,para_list,para_row,L_list,L_row,top_list,top_rowdef parse_log(path_to_log):    """Parse log file    Returns (train_dict_list, train_dict_names, test_dict_list, test_dict_names)    train_dict_list and test_dict_list are lists of dicts that define the table    rows    train_dict_names and test_dict_names are ordered tuples of the column names    for the two dict_lists    """    regex_iteration = re.compile('Iteration (\d+)')    regex_train_iteration=re.compile('Iteration (\d+), loss')    regex_train_output = re.compile('Train net output #(\d+): (\S+) = ([\.\deE+-]+)')    regex_test_output = re.compile('Test net output #(\d+): (\S+) = ([\.\deE+-]+)')    regex_learning_rate = re.compile('lr = ([-+]?[0-9]*\.?[0-9]+([eE]?[-+]?[0-9]+)?)')    regex_backward = re.compile('\[Backward\] Layer ')    # Pick out lines of interest    iteration = -1    train_iter=-1    learning_rate = float('NaN')    train_dict_list = []    test_dict_list = []    train_row = None    test_row = None    data_diff_list=[]    para_diff_list=[]    L1L2_list=[]    top_list=[]    data_diff_row=None    para_diff_row=None    L1L2_row = None    top_row=None    logfile_year = extract_seconds.get_log_created_year(path_to_log)    with open(path_to_log) as f:        start_time = extract_seconds.get_start_time(f, logfile_year)        for line in f:            iteration_match = regex_iteration.search(line)            train_iter_match=regex_train_iteration.search(line)            if iteration_match:                iteration = float(iteration_match.group(1))            if train_iter_match:                train_iter=float(train_iter_match.group(1))            if iteration == -1:                # Only start parsing for other stuff if we've found the first                # iteration                continue            time = extract_seconds.extract_datetime_from_line(line,                                                              logfile_year)            seconds = (time - start_time).total_seconds()            learning_rate_match = regex_learning_rate.search(line)            if learning_rate_match:                learning_rate = float(learning_rate_match.group(1))            back_match=regex_backward.search(line)           # if back_match:            data_diff_list,data_diff_row,para_diff_list,para_diff_row,L1L2_list,L1L2_row,top_list,top_row=get_datadiff_paradiff(                    line,data_diff_row,para_diff_row,                    data_diff_list,para_diff_list,                    L1L2_list,L1L2_row,                    top_list,top_row,                    train_iter                    )            train_iter=-1            train_dict_list, train_row = parse_line_for_net_output(                regex_train_output, train_row, train_dict_list,                line, iteration, seconds, learning_rate            )            test_dict_list, test_row = parse_line_for_net_output(                regex_test_output, test_row, test_dict_list,                line, iteration, seconds, learning_rate            )    fix_initial_nan_learning_rate(train_dict_list)    fix_initial_nan_learning_rate(test_dict_list)    return train_dict_list, test_dict_list,data_diff_list,para_diff_list,L1L2_list,top_listdef parse_line_for_net_output(regex_obj, row, row_dict_list,                              line, iteration, seconds, learning_rate):    """Parse a single line for training or test output    Returns a a tuple with (row_dict_list, row)    row: may be either a new row or an augmented version of the current row    row_dict_list: may be either the current row_dict_list or an augmented    version of the current row_dict_list    """    output_match = regex_obj.search(line)    if output_match:        if not row or row['NumIters'] != iteration:            # Push the last row and start a new one            if row:                # If we're on a new iteration, push the last row                # This will probably only happen for the first row; otherwise                # the full row checking logic below will push and clear full                # rows                row_dict_list.append(row)            row = OrderedDict([                ('NumIters', iteration),                ('Seconds', seconds),                ('LearningRate', learning_rate)            ])        # output_num is not used; may be used in the future        # output_num = output_match.group(1)        output_name = output_match.group(2)        output_val = output_match.group(3)        row[output_name] = float(output_val)    if row and len(row_dict_list) >= 1 and len(row) == len(row_dict_list[0]):        # The row is full, based on the fact that it has the same number of        # columns as the first row; append it to the list        row_dict_list.append(row)        row = None    return row_dict_list, rowdef fix_initial_nan_learning_rate(dict_list):    """Correct initial value of learning rate    Learning rate is normally not printed until after the initial test and    training step, which means the initial testing and training rows have    LearningRate = NaN. Fix this by copying over the LearningRate from the    second row, if it exists.    """    if len(dict_list) > 1:        dict_list[0]['LearningRate'] = dict_list[1]['LearningRate']def save_csv_files(logfile_path, output_dir, train_dict_list, test_dict_list,data_diff_list, para_diff_list,L1L2_list,top_list,                   delimiter=',', verbose=False):    """Save CSV files to output_dir    If the input log file is, e.g., caffe.INFO, the names will be    caffe.INFO.train and caffe.INFO.test    """    log_basename = os.path.basename(logfile_path)    train_filename = os.path.join(output_dir, log_basename + '.train')    write_csv(train_filename, train_dict_list, delimiter, verbose)    test_filename = os.path.join(output_dir, log_basename + '.test')    write_csv(test_filename, test_dict_list, delimiter, verbose)    data_diff_filename=os.path.join(output_dir, log_basename + '.datadiff')    write_csv(data_diff_filename, data_diff_list, delimiter, verbose)    para_diff_filename=os.path.join(output_dir, log_basename + '.paradiff')    write_csv(para_diff_filename, para_diff_list, delimiter, verbose)    L1L2_filename=os.path.join(output_dir, log_basename + '.L1L2')    write_csv(L1L2_filename, L1L2_list, delimiter, verbose)    topdata_filename=os.path.join(output_dir, log_basename + '.topdata')    write_csv(topdata_filename, top_list, delimiter, verbose)def write_csv(output_filename, dict_list, delimiter, verbose=False):    """Write a CSV file    """    if not dict_list:        if verbose:            print('Not writing %s; no lines to write' % output_filename)        return    dialect = csv.excel    dialect.delimiter = delimiter    with open(output_filename, 'w') as f:        dict_writer = csv.DictWriter(f, fieldnames=dict_list[0].keys(),                                     dialect=dialect)        dict_writer.writeheader()        dict_writer.writerows(dict_list)    if verbose:        print 'Wrote %s' % output_filenamedef parse_args():    description = ('Parse a Caffe training log into two CSV files '                   'containing training and testing information')    parser = argparse.ArgumentParser(description=description)    parser.add_argument('logfile_path',                        help='Path to log file')    parser.add_argument('output_dir',                        help='Directory in which to place output CSV files')    parser.add_argument('--verbose',                        action='store_true',                        help='Print some extra info (e.g., output filenames)')    parser.add_argument('--delimiter',                        default=',',                        help=('Column delimiter in output files '                              '(default: \'%(default)s\')'))    args = parser.parse_args()    return argsdef main():    args = parse_args()    train_dict_list, test_dict_list,data_diff_list,para_diff_list,L1L2_list,top_list = parse_log(args.logfile_path)    save_csv_files(args.logfile_path, args.output_dir, train_dict_list,                   test_dict_list, data_diff_list, para_diff_list, L1L2_list,top_list,delimiter=args.delimiter)if __name__ == '__main__':    main()

参数变化趋势图。