def parse_log(path_to_log):
"""Parse log file
Returns (train_dict_list, test_dict_list)
train_dict_list and test_dict_list are lists of dicts that define the table
rows
"""
regex_iteration = re.compile('Iteration (\d+)')
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]+)?)')
# Pick out lines of interest
iteration = -1
learning_rate = float('NaN')
train_dict_list = []
test_dict_list = []
train_row = None
test_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)
if iteration_match:
iteration = float(iteration_match.group(1))
if iteration == -1:
# Only start parsing for other stuff if we've found the first
# iteration
continue
try:
time = extract_seconds.extract_datetime_from_line(line,
logfile_year)
except ValueError:
# Skip lines with bad formatting, for example when resuming solver
continue
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))
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
def 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_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]+)?)')
# Pick out lines of interest
iteration = -1
learning_rate = float('NaN')
train_dict_list = []
test_dict_list = []
train_row = None
test_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)
if iteration_match:
iteration = float(iteration_match.group(1))
# if iteration == -1:
# # Only start parsing for other stuff if we've found the first
# # iteration
# continue
# print('Find in line: ', line.find('RuntimeWarning'))
# print('Find in line: ', line.find('>>>'))
if line.find('RuntimeWarning:') != -1 or line.find('>>>') != -1:
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))
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
def 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+) (-3.53679e+10 iter/s, 0.216956s/500 iters), loss = 0.0236502')
regex_iteration = re.compile('Iteration (\d+)')
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]+)?)')
# Pick out lines of interest
iteration = -1
learning_rate = float('NaN')
train_dict_list = []
test_dict_list = []
train_row = None
test_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)
if iteration_match:
iteration = float(iteration_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))
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
def parse_log(path_to_log):
"""Parse log file
Returns (train_dict_list, test_dict_list)
train_dict_list and test_dict_list are lists of dicts that define the table
rows
"""
regex_iteration = re.compile('Iteration (\d+)')
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_time_stamp = re.compile('I(\d){4}')
# Pick out lines of interest
iteration = -1
learning_rate = float('NaN')
train_dict_list = []
test_dict_list = []
train_row = None
test_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)
if iteration_match:
iteration = float(iteration_match.group(1))
if iteration == -1:
# Only start parsing for other stuff if we've found the first
# iteration
continue
# skip line not starting with time stamps
if not regex_time_stamp.match(line):
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))
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
def 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_batch = re.compile('Batch (\d+)')
regex_test_output = re.compile('\, (\S+) = ([\.\deE+-]+)')
# Pick out lines of interest
batch = -1
learning_rate = float('NaN')
test_dict_list = []
test_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_batch.search(line)
if iteration_match:
batch = float(iteration_match.group(1))
if batch == -1:
# Only start parsing for other stuff if we've found the first
# batch
continue
time = extract_seconds.extract_datetime_from_line(
line, logfile_year)
seconds = (time - start_time).total_seconds()
test_dict_list, test_row = parse_line_for_net_output(
regex_test_output, test_row, test_dict_list, line, batch,
seconds)
return test_dict_list
def parse_log(path_to_log):
"""Parse log file
Returns (train_dict_list, test_dict_list)
train_dict_list and test_dict_list are lists of dicts that define the table
rows
"""
# 设置正则化表达式
regex_iteration = re.compile('Iteration (\d+)')
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]+)?)')
# Pick out lines of interest
# 选出我们感兴趣的行
iteration = -1
learning_rate = float('NaN')
train_dict_list = []
test_dict_list = []
train_row = None
test_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)
last_time = start_time
for line in f:
iteration_match = regex_iteration.search(line)
if iteration_match:
# group(0) 是获取整个匹配字符串；group(1) 是获取匹配字符串中（）中的内容
# 比如，m = re.match(r'www\.(.+)\.com', 'www.google.com')
# m.group(0) 输出 www.google.com
# m.group(1) 输出 google
iteration = float(iteration_match.group(1))
if iteration == -1:
# Only start parsing for other stuff if we've found the first
# iteration
# 只有我们找到第一个 iteration, 我们才开始解析其他 stuff
continue
try:
time = extract_seconds.extract_datetime_from_line(
line, logfile_year)
except ValueError:
# Skip lines with bad formatting, for example when resuming solver
# 跳过匹配错误的行，比如重新启动的 solver
continue
# if it's another year
# 如果出现跨年的情况
if time.month < last_time.month:
logfile_year += 1
time = extract_seconds.extract_datetime_from_line(
line, logfile_year)
last_time = time
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))
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
def 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
"""
re_iteration = re.compile('Iteration (\d+)')
#re_accuracy = re.compile('output #\d+: accuracy = ([\.\d]+)')
re_accuracy = re.compile('output #\d+: loss3/top-1 = ([\.\d]+)')
#re_train_loss = re.compile('Iteration \d+, loss = ([\.\d]+)')
re_train_loss = re.compile('Iteration \d+, loss3/loss3 = ([\.\d]+)')
#re_output_loss = re.compile('output #\d+: loss = ([\.\d]+)')
re_output_loss = re.compile('output #\d+: loss3/loss3 = ([\.\d]+)')
re_lr = re.compile('lr = ([\.\d]+)')
# Pick out lines of interest
iteration = -1
test_accuracy = -1
learning_rate = float('NaN')
train_dict_list = []
test_dict_list = []
train_dict_names = ('NumIters', 'Seconds', 'TrainingLoss', 'LearningRate')
test_dict_names = ('NumIters', 'Seconds', 'TestAccuracy', 'TestLoss')
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 = re_iteration.search(line)
if iteration_match:
iteration = float(iteration_match.group(1))
if iteration == -1:
# Only look 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()
lr_match = re_lr.search(line)
if lr_match:
learning_rate = float(lr_match.group(1))
accuracy_match = re_accuracy.search(line)
if accuracy_match and get_line_type(line) == 'test':
test_accuracy = float(accuracy_match.group(1))
train_loss_match = re_train_loss.search(line)
if train_loss_match:
train_loss = float(train_loss_match.group(1))
train_dict_list.append({
'NumIters': iteration,
'Seconds': seconds,
'TrainingLoss': train_loss,
'LearningRate': learning_rate
})
output_loss_match = re_output_loss.search(line)
if output_loss_match and get_line_type(line) == 'test':
test_loss = float(output_loss_match.group(1))
# NOTE: we assume that (1) accuracy always comes right before
# loss for test data so the test_accuracy variable is already
# correctly populated and (2) there's one and only one output
# named "accuracy" for the test net
test_dict_list.append({
'NumIters': iteration,
'Seconds': seconds,
'TestAccuracy': test_accuracy,
'TestLoss': test_loss
})
return train_dict_list, train_dict_names, test_dict_list, test_dict_names
def parse_log(path_to_log):
"""Parse log file
Returns (df_train, df_test)
df_train and df_test are pandas DataFrame with data from log
"""
re_correct_line = re.compile('^\w+\d+')
re_iteration = re.compile('Iteration (\d+)')
# alexnet
#re_output_loss = re.compile('output #\d+: loss = ([\.\d]+)')
#re_output_acc = re.compile('output #\d+: accuracy = ([\.\d]+)')
#googlenet
re_output_loss = re.compile('output #\d+: loss3\/loss3 = ([\.\d]+)')
re_output_acc = re.compile('output #\d+: loss3\/top-1 = ([\.\d]+)')
re_lr = re.compile('lr = ([\.\d]+)')
# Pick out lines of interest
iteration = -1
test_accuracy = -1
learning_rate = float('NaN')
acc = float('NaN')
train_dict_list = []
test_dict_list = []
train_dict_names = ('NumIters', 'Loss', 'Accuracy', 'LearningRate', 'Seconds')
test_dict_names = ('NumIters', 'Loss', 'Accuracy')
is_test_group = False
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:
if not re_correct_line.match(line):
continue
iteration_match = re_iteration.search(line)
if iteration_match:
iteration = int(iteration_match.group(1))
if iteration == -1:
# Only look 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()
lr_match = re_lr.search(line)
if lr_match:
learning_rate = float(lr_match.group(1))
output_acc_match = re_output_acc.search(line)
if output_acc_match:
acc = float(output_acc_match.group(1))
output_loss_match = re_output_loss.search(line)
if output_loss_match:
if get_line_type(line) == 'test':
test_loss = float(output_loss_match.group(1))
test_dict_list.append({'NumIters': iteration,
'Loss': test_loss,
'Accuracy': acc})
else:
train_loss = float(output_loss_match.group(1))
train_dict_list.append({'NumIters': iteration,
'Loss': train_loss,
'Accuracy': acc,
'LearningRate': learning_rate,
'Seconds': seconds})
df_train = pd.DataFrame(columns=train_dict_names)
df_test = pd.DataFrame(columns=test_dict_names)
for col in train_dict_names:
df_train[col] = [d[col] for d in train_dict_list]
for col in test_dict_names:
df_test[col] = [d[col] for d in test_dict_list]
return df_train, df_test
def parse_log(path_to_log):
"""Parse log file
Returns (train_dict_list, test_dict_list)
train_dict_list and test_dict_list are lists of dicts that define the table
rows
"""
regex_iteration = re.compile('Iteration (\d+)')
regex_train_output = re.compile('Train net output #(\d+): (\S+) = ([\.\deE+-]+)')
regex_test_output = re.compile('loss = ([-+]?[0-9]*\.?[0-9]+([eE]?[-+]?[0-9]+)?)')
regex_learning_rate = re.compile('lr = ([-+]?[0-9]*\.?[0-9]+([eE]?[-+]?[0-9]+)?)')
# Pick out lines of interest
iteration = -1
learning_rate = float('NaN')
train_dict_list = []
test_dict_list = []
train_row = None
test_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)
if iteration_match:
iteration = float(iteration_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))
train_dict_list, train_row = parse_line_for_net_output(
regex_train_output, train_row, train_dict_list,
line, iteration, seconds, learning_rate
)
test_rate_match = regex_test_output.search(line)
if test_rate_match:
if not test_row or test_row['NumIters'] != iteration:
# Push the last row and start a new one
if test_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
test_dict_list.append(test_row)
test_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_val = test_rate_match.group(1)
test_row["loss"] = float(output_val)
if test_row and len(test_dict_list) >= 1 and len(test_row) == len(test_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
test_dict_list.append(test_row)
test_row = None
fix_initial_nan_learning_rate(train_dict_list)
fix_initial_nan_learning_rate(test_dict_list)
return train_dict_list, test_dict_list
def parse_log(path_to_log):
"""Parse log file
Returns (train_dict_list, test_dict_list)
train_dict_list and test_dict_list are lists of dicts that define the table
rows
"""
regex_iteration = re.compile('batch_mini_iter: (\d+)')
regex_train_output = re.compile('loss: ([\.\deE+-]+)')
regex_obj_score = re.compile('avg_obj: ([\.\deE+-]+)')
regex_noobj = re.compile('avg_noobj: ([\.\deE+-]+)')
regex_iou = re.compile('avg_iou: ([\.\deE+-]+)')
regex_cat = re.compile('avg_cat: ([\.\deE+-]+)')
regex_recall = re.compile('recall: ([\.\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]+)?)')
# Pick out lines of interest
iteration = -1
learning_rate = float('NaN')
train_dict_list = []
test_dict_list = []
train_row = None
test_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)
if iteration_match:
iteration = float(iteration_match.group(1))
if iteration == -1:
# Only start parsing for other stuff if we've found the first
# iteration
continue
try:
time = extract_seconds.extract_datetime_from_line(
line, logfile_year)
except ValueError:
# Skip lines with bad formatting, for example when resuming solver
continue
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))
train_dict_list, train_row = parse_line_for_net_output(
regex_train_output, regex_obj_score, regex_noobj, regex_iou,
regex_cat, regex_recall, train_row, train_dict_list, line,
iteration, seconds)
#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
def 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_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]+)?)')
# Pick out lines of interest
iteration = -1
learning_rate = float('NaN')
train_dict_list = []
test_dict_list = []
train_row = None
test_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:
if line[:5] == "speed" or line[:5] == "Wrote":
continue
if line[:4] == "done":
break
iteration_match = regex_iteration.search(line)
if iteration_match:
iteration = float(iteration_match.group(1))
if iteration == -1:
# Only start parsing for other stuff if we've found the first
# iteration
continue
try:
time = extract_seconds.extract_datetime_from_line(
line, logfile_year)
except ValueError:
break
except:
print "Unexpected error reading line: "
print line
print "I will go on though"
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))
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
def 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
"""
re_iteration = re.compile('Iteration (\d+)')
re_top_output = re.compile('Iteration \d+, (\w+) = ([\.\d]+)')
re_output = re.compile('(Test|Train) net output #\d+: '
'(\w+) = ([+-]*[\.\d]+(e[+-][\d]+)*)')
# Pick out lines of interest
iteration = -1
train_dict_list = []
test_dict_list = []
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 = re_iteration.search(line)
if iteration_match:
parsed_iteration = float(iteration_match.group(1))
if parsed_iteration != iteration:
train_dict_list += [{'NumIters': parsed_iteration}]
if (not test_dict_list or
len(test_dict_list[-1].keys()) != 1):
test_dict_list += [{'NumIters': parsed_iteration}]
else:
test_dict_list[-1]['NumIters'] = parsed_iteration
iteration = parsed_iteration
if iteration == -1:
# Only look for other stuff if we've found the first iteration
continue
try:
time = extract_seconds.extract_datetime_from_line(line,
logfile_year)
except:
continue
seconds = (time - start_time).total_seconds()
top_output_match = re_top_output.search(line)
if top_output_match:
top_output_name = top_output_match.group(1)
top_output_value = float(top_output_match.group(2))
train_dict_list[-1][top_output_name] = top_output_value
output_match = re_output.search(line)
if output_match:
is_test_output = output_match.group(1).lower() == 'test'
if is_test_output:
dict_list = test_dict_list
else:
dict_list = train_dict_list
output_name = output_match.group(2)
output_value = float(output_match.group(3))
dict_list[-1][output_name] = output_value
dict_list[-1]['Seconds'] = seconds
train_dict_names = train_dict_list[0].keys()
test_dict_names = test_dict_list[0].keys()
# print train_dict_list[: 3]
return train_dict_list, train_dict_names, test_dict_list, test_dict_names
def 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
"""
re_iteration = re.compile('Iteration (\d+)')
#re_accuracy = re.compile('output #\d+: accuracy = ([\.\d]+)')
re_accuracy = re.compile('output #\d+: loss3/top-1 = ([\.\d]+)')
#re_train_loss = re.compile('Iteration \d+, loss = ([\.\d]+)')
re_train_loss = re.compile('Iteration \d+, loss3/loss3 = ([\.\d]+)')
#re_output_loss = re.compile('output #\d+: loss = ([\.\d]+)')
re_output_loss = re.compile('output #\d+: loss3/loss3 = ([\.\d]+)')
re_lr = re.compile('lr = ([\.\d]+)')
# Pick out lines of interest
iteration = -1
test_accuracy = -1
learning_rate = float('NaN')
train_dict_list = []
test_dict_list = []
train_dict_names = ('NumIters', 'Seconds', 'TrainingLoss', 'LearningRate')
test_dict_names = ('NumIters', 'Seconds', 'TestAccuracy', 'TestLoss')
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 = re_iteration.search(line)
if iteration_match:
iteration = float(iteration_match.group(1))
if iteration == -1:
# Only look 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()
lr_match = re_lr.search(line)
if lr_match:
learning_rate = float(lr_match.group(1))
accuracy_match = re_accuracy.search(line)
if accuracy_match and get_line_type(line) == 'test':
test_accuracy = float(accuracy_match.group(1))
train_loss_match = re_train_loss.search(line)
if train_loss_match:
train_loss = float(train_loss_match.group(1))
train_dict_list.append({'NumIters': iteration,
'Seconds': seconds,
'TrainingLoss': train_loss,
'LearningRate': learning_rate})
output_loss_match = re_output_loss.search(line)
if output_loss_match and get_line_type(line) == 'test':
test_loss = float(output_loss_match.group(1))
# NOTE: we assume that (1) accuracy always comes right before
# loss for test data so the test_accuracy variable is already
# correctly populated and (2) there's one and only one output
# named "accuracy" for the test net
test_dict_list.append({'NumIters': iteration,
'Seconds': seconds,
'TestAccuracy': test_accuracy,
'TestLoss': test_loss})
return train_dict_list, train_dict_names, test_dict_list, test_dict_names
def parse_log(path_to_log):
"""Parse log file
Returns (train_dict_list, test_dict_list)
train_dict_list and test_dict_list are lists of dicts that define the table
rows
"""
regex_float = '([-+]?[0-9]*\.?[0-9]+([eE]?[-+]?[0-9]+)?)'
regex_iteration = re.compile('Iteration (\d+)')
regex_train_output = re.compile('Train net output #(\d+): (\S+) = ([\.\deE+-]+)')
regex_learning_rate = re.compile('lr = %s'%regex_float)
regex_end_of_phase = re.compile('Wrote snapshot to')
regex_ignore_rows = re.compile('speed: [0-9\.]*s / iter')
regex_mean_ap = re.compile('Mean AP = %s'%regex_float)
# Pick out lines of interest
iteration = -1
learning_rate = float('NaN')
phases=[]
train_dict_list=[]
train_row = None
test_row = None
logfile_year = extract_seconds.get_log_created_year(path_to_log)
mean_ap = None
with open(path_to_log) as f:
start_time = extract_seconds.get_start_time(f, logfile_year)
last_time = start_time
for line in f:
if regex_ignore_rows.search(line):
continue
if regex_end_of_phase.search(line):
# start a new learning phase
fix_initial_nan_learning_rate(train_dict_list)
phases.append(train_dict_list)
train_dict_list=[]
iteration=-1
continue
mean_ap_match=regex_mean_ap.search(line)
if mean_ap_match:
mean_ap=float(mean_ap_match.group(1))
iteration_match = regex_iteration.search(line)
if iteration_match:
iteration = float(iteration_match.group(1))
continue
if iteration == -1:
# Only start parsing for other stuff if we've found the first
# iteration
continue
try:
time = extract_seconds.extract_datetime_from_line(line,
logfile_year)
except ValueError:
# Skip lines with bad formatting, for example when resuming solver
continue
# if it's another year
if time.month < last_time.month:
logfile_year += 1
time = extract_seconds.extract_datetime_from_line(line, logfile_year)
last_time = time
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))
continue
train_dict_list, train_row, train_match = parse_line_for_net_output(
regex_train_output, train_row, train_dict_list,
line, iteration, seconds, learning_rate
)
return phases, mean_ap
def 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
"""
# I0526 13:50:19.771880 28824 solver.cpp:319] Test loss: 2.65821
# I0526 13:50:19.772948 28824 solver.cpp:332] Test net output #0: cross_entropy_loss = 65.4734
# I0526 13:50:19.772971 28824 solver.cpp:332] Test net output #1: l2_error = 2.65821 (* 1 = 2.65821 loss)
# I0526 13:50:19.772979 28824 solver.cpp:283] Iteration 24000, Testing net (#1)
# I0526 13:50:20.337695 28824 solver.cpp:319] Test loss: 2.78423
# I0526 13:50:20.337750 28824 solver.cpp:332] Test net output #0: cross_entropy_loss = 65.8907
# I0526 13:50:20.337774 28824 solver.cpp:332] Test net output #1: l2_error = 2.78423 (* 1 = 2.78423 loss)
re_batch_size = re.compile('batch_size: (\d+)')
re_iteration = re.compile('Iteration (\d+)')
re_train_loss = re.compile('Iteration \d+, loss = ([\.\d\-+ena]+)')
re_accuracy = re.compile('output #\d+: (accuracy|l2_error) = ([\.\d\-+ena]+)')
re_output_loss = re.compile('output #\d+: (loss|cross_entropy_loss) = ([\.\d\-+ena]+)')
re_lr = re.compile('lr = ([\.\d\-+ena]+)')
re_grad_norm = re.compile('avg_grad_norm = ([\.\d\-+enan]+)')
re_step_norm = re.compile('avg_step_norm = ([\.\d\-+enan]+)')
re_eff_lr = re.compile('avg_effective_learning_rate = ([\.\d\-+enan]+)')
re_test_start_seconds = re.compile('Testing net')
# Pick out lines of interest
iteration = -1
test_accuracy = -1
test_start_seconds = float('NaN')
learning_rate = float('NaN')
avg_grad_norm = float('NaN')
avg_step_norm = float('NaN')
eff_lr = float('NaN')
batch_size = None
train_dict_list = []
test_dict_list = []
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)
with open(path_to_log) as f:
for line in f:
# print 'parse_log.py line: ', line
if batch_size is None:
batch_size_match = re_batch_size.search(line)
if batch_size_match:
batch_size = float(batch_size_match.group(1))
iteration_match = re_iteration.search(line)
if iteration_match:
iteration = float(iteration_match.group(1))
if iteration == -1:
# Only look 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()
lr_match = re_lr.search(line)
if lr_match:
learning_rate = float(lr_match.group(1))
grad_norm_match = re_grad_norm.search(line)
if grad_norm_match:
avg_grad_norm = float(grad_norm_match.group(1))
step_norm_match = re_step_norm.search(line)
if step_norm_match:
avg_step_norm = float(step_norm_match.group(1))
eff_lr_match = re_eff_lr.search(line)
if eff_lr_match:
eff_lr = float(eff_lr_match.group(1))
test_start_match = re_test_start_seconds.search(line)
if test_start_match:
test_start_seconds = seconds
accuracy_match = re_accuracy.search(line)
if accuracy_match and get_line_type(line) == 'test':
test_accuracy = float(accuracy_match.group(2))
# ce_match = re_ce.search(line)
# if ce_match and get_line_type(line) == 'test':
# test_accuracy = float(ce_match.group(1))
train_loss_match = re_train_loss.search(line)
if train_loss_match:
train_loss = float(train_loss_match.group(1))
train_dict_list.append({'NumIters': iteration,
'Seconds': seconds,
'TrainingLoss': train_loss,
'LearningRate': learning_rate,
'AvgGradientNorm': avg_grad_norm,
# 'AvgStepNorm': avg_step_norm})
'AvgStepNorm': avg_step_norm,
'EffectiveLearningRate': eff_lr})
output_loss_match = re_output_loss.search(line)
if output_loss_match and get_line_type(line) == 'test':
test_loss = float(output_loss_match.group(2))
# NOTE: we assume that (1) accuracy always comes right before
# loss for test data so the test_accuracy variable is already
# correctly populated and (2) there's one and only one output
# named "accuracy" for the test net
test_dict_list.append({'NumIters': iteration,
'SecondsAtStart': test_start_seconds,
'SecondsAtEnd': seconds,
'TestAccuracy': test_accuracy,
'TestLoss': test_loss})
return train_dict_list, test_dict_list, batch_size
def 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+), loss = ([-+]?[0-9]*\.?[0-9]+([eE]?[-+]?[0-9]+)?)')
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]+)?)')
# Pick out lines of interest
iteration = -1
learning_rate = float('NaN')
loss = 0
train_dict_list = []
test_dict_list = []
train_row = None
test_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)
with open(path_to_log) as f:
for line in f:
iteration_match = regex_iteration.search(line)
if iteration_match:
iteration = float(iteration_match.group(1))
loss = float(iteration_match.group(2))
#print ('iteration = ', iteration, 'loss = ', loss)
if (iteration == -1 or line.find('speed') != -1 or line.find('Wrote') != -1 or line.find('None') != -1):
# Only start parsing for other stuff if we've found the first
# iteration
continue
#FRDC lxq
if line.find('done solving') != -1:
# Only start parsing for other stuff if we've found the first
# iteration
break
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))
train_dict_list, train_row = parse_line_for_net_output(
regex_train_output, train_row, train_dict_list,
line, iteration, seconds, learning_rate, loss
)
test_dict_list, test_row = parse_line_for_net_output(
regex_test_output, test_row, test_dict_list,
line, iteration, seconds, learning_rate, loss
)
fix_initial_nan_learning_rate(train_dict_list)
fix_initial_nan_learning_rate(test_dict_list)
return train_dict_list, test_dict_list
def 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_valid_line = re.compile('I(\d\d\d\d)')
regex_iteration = re.compile('Iteration (\d+)')
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]+)?)')
# Pick out lines of interest
iteration = -1
learning_rate = float('NaN')
train_dict_list = []
test_dict_list = []
train_row = None
test_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:
#line = line.strip()
valid_line_match = regex_valid_line.match(line)
if not valid_line_match:
#if len(line)<1 or not line.startswith('I'):
#line.startswith('speed') or line.startswith('Wrote') or line.startswith('done'):#added by zhaoyafei 20161124
#if :
continue
iteration_match = regex_iteration.search(line)
if iteration_match:
iteration = float(iteration_match.group(1))
#else:#added by zhaoyafei 20161124
#print 'No iteration_match: continue'
#print line
#sys.exit(0)
#continue
if iteration == -1:
# Only start parsing for other stuff if we've found the first
# iteration
#print 'iteration == -1:continue'
#sys.exit(0)
continue
print line
#sys.exit(0)
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))
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
def parse_log(path_to_log):
"""Parse log file
Returns (train_dict_list, test_dict_list)
train_dict_list and test_dict_list are lists of dicts that define the table
rows
"""
regex_iteration = re.compile('Iteration (\d+)')
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]+)?)')
# Pick out lines of interest
iteration = -1
learning_rate = float('NaN')
train_dict_list = []
test_dict_list = []
train_row = None
test_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)
last_time = start_time
for line in f:
iteration_match = regex_iteration.search(line)
if iteration_match:
iteration = float(iteration_match.group(1))
if iteration == -1:
# Only start parsing for other stuff if we've found the first
# iteration
continue
try:
time = extract_seconds.extract_datetime_from_line(
line, logfile_year)
except ValueError:
# Skip lines with bad formatting, for example when resuming solver
continue
# if it's another year
if time.month < last_time.month:
logfile_year += 1
time = extract_seconds.extract_datetime_from_line(
line, logfile_year)
last_time = time
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))
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
def parse_log(path_to_log):
"""Parse log file
Returns (train_dict_list, test_dict_list)
train_dict_list and test_dict_list are lists of dicts that define the table
rows
"""
regex_iteration = re.compile('Iteration (\d+)')
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]+)?)')
# Pick out lines of interest
iteration = -1
learning_rate = float('NaN')
train_dict_list = [[] for i in range(4)]
test_dict_list = [[] for i in range(4)]
train_row = None
test_row = None
stage_num = 0
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)
line_s = line.strip()
if len(line_s) == 0:
continue
if line.find(
'Stage 1 Fast R-CNN using RPN proposals, init from ImageNet model'
) != -1:
stage_num = 1
if line.find(
'Stage 2 RPN, init from stage 1 Fast R-CNN model') != -1:
stage_num = 2
if line.find(
'Stage 2 Fast R-CNN, init from stage 2 RPN R-CNN model'
) != -1:
stage_num = 3
if iteration_match:
iteration = float(iteration_match.group(1))
if iteration == -1:
# Only start parsing for other stuff if we've found the first
# iteration
continue
try:
time = extract_seconds.extract_datetime_from_line(
line, logfile_year)
except ValueError:
# Skip lines with bad formatting, for example when resuming solver
continue
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))
train_dict_list[stage_num], train_row = parse_line_for_net_output(
regex_train_output, train_row, train_dict_list[stage_num],
line, iteration, seconds, learning_rate)
test_dict_list[stage_num], test_row = parse_line_for_net_output(
regex_test_output, test_row, test_dict_list[stage_num], line,
iteration, seconds, learning_rate)
for i in range(4):
fix_initial_nan_learning_rate(train_dict_list[i])
fix_initial_nan_learning_rate(test_dict_list[i])
return train_dict_list, test_dict_list
def 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
"""
re_iteration = re.compile('Iteration (\d+)')
re_test_accuracy = re.compile('output #\d+: accuracy = ([\.\d]+)')
re_test_loss = re.compile('output #\d+: loss = ([\.\d]+)')
re_loss = re.compile('Iteration \d+, loss = ([\.\d]+)')
re_lr = re.compile('lr = ([\d]+e-[\d]+|[\.\d]+)')
# Pick out lines of interest
iteration = -1
current_line_iteration = -1
iteration_type = None
iteration_dict = None
accuracy = -1
learning_rate = float('NaN')
train_dict_list = []
test_dict_list = []
train_dict_names = ('Iterations', 'Seconds', 'Loss', 'LearningRate')
test_dict_names = ('Iterations', 'Seconds', 'Loss', 'Accuracy')
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 = re_iteration.search(line)
if iteration_match:
current_line_iteration = int(iteration_match.group(1))
if current_line_iteration == -1:
# Only look for other stuff if we've found the first iteration
continue
#print("{0} {1} {2}".format(iteration_type, iteration, iteration_dict['Loss'] if iteration_dict and 'Loss' in iteration_dict else ""))
# new iteration or switching from test to train
if(iteration < current_line_iteration or
(get_line_type(line) and (get_line_type(line) != iteration_type))):
iteration = current_line_iteration
# new iteration
if(iteration > 0 or (iteration_type and get_line_type(line) != iteration_type)):
# log previous iteration
if(iteration_type == 'train'):
train_dict_list.append(iteration_dict)
else:
test_dict_list.append(iteration_dict)
time = extract_seconds.extract_datetime_from_line(line,
logfile_year)
seconds = (time - start_time).total_seconds()
iteration_dict = {'Iterations': '{:d}'.format(iteration),
'Seconds': '{:f}'.format(seconds)}
iteration_type = get_line_type(line) or 'train'
lr_match = re_lr.search(line)
if lr_match:
iteration_dict['LearningRate'] = float(lr_match.group(1))
accuracy_match = re_test_accuracy.search(line)
if accuracy_match:
iteration_dict['Accuracy'] = float(accuracy_match.group(1))
loss_test_match = re_test_loss.search(line)
if loss_test_match:
iteration_dict['Loss'] = float(loss_test_match.group(1))
loss_match = re_loss.search(line)
if loss_match:
iteration_dict['Loss'] = float(loss_match.group(1))
# log last iteration
if(iteration_dict and iteration_type == 'train'):
train_dict_list.append(iteration_dict)
elif(iteration_dict):
test_dict_list.append(iteration_dict)
return train_dict_list, train_dict_names, test_dict_list, test_dict_names
def parse_log(path_to_log):
"""Parse log file
Returns (train_dict_list, test_dict_list)
train_dict_list and test_dict_list are lists of dicts that define the table
rows
"""
regex_iteration = re.compile('Iteration (\d+), loss = ([\.\deE+-]+)')
regex_train_output = re.compile(
'Train net output #(\d+): (\S+) = ([\.\deE+-]+)')
regex_learning_rate = re.compile('lr = ([\.\deE+-]+)')
regex_test_output = re.compile(
'Test net output #(\d+): detection_eval = ([\.\deE+-]+)')
# Pick out lines of interest
iteration = 0
loss = -1
learning_rate = 0.001
train_dict_list = []
train_row = None
test_score = 0.0
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)
last_time = start_time
for line in f:
iteration_match = regex_iteration.search(line)
if iteration_match:
iteration = float(iteration_match.group(1))
loss = float(iteration_match.group(2))
try:
time = extract_seconds.extract_datetime_from_line(
line, logfile_year)
except:
# Skip lines with bad formatting, for example when resuming solver
continue
# if it's another year
if time.month < last_time.month:
logfile_year += 1
time = extract_seconds.extract_datetime_from_line(
line, logfile_year)
last_time = time
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))
test_score_match = regex_test_output.search(line)
if test_score_match:
test_score = float(test_score_match.group(2))
train_dict_list, train_row = parse_line_for_net_output(
regex_train_output, train_row, train_dict_list, line,
iteration, seconds, learning_rate, loss, test_score)
return train_dict_list
def parse_log(path_to_log):
"""Parse log file
Returns (train_dict_list, test_dict_list)
train_dict_list and test_dict_list are lists of dicts that define the table
rows
"""
regex_phase = re.compile('\+ ./tools/(\S+)')
regex_iteration = re.compile('Iteration (\d+)')
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]+)?)')
# Pick out lines of interest
iteration = -1
learning_rate = float('NaN')
train_dict_list = []
test_dict_list = []
train_row = None
test_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)
time = 0.
seconds = 0.
phase = 0
for line in f:
try:
phase_match = regex_phase.search(line)
if phase_match:
phase = phase + 1
start_time = time
continue
iteration_match = regex_iteration.search(line)
if iteration_match:
iteration = float(iteration_match.group(1))
time = extract_seconds.extract_datetime_from_line(
line, logfile_year)
seconds = (time - start_time).total_seconds()
continue
lr_match = regex_learning_rate.search(line)
if lr_match:
learning_rate = float(learning_rate_match.group(1))
train_match = regex_train_output.search(line)
if train_match:
train_dict_list, train_row = parse_line_for_net_output(
regex_train_output, train_row, train_dict_list, line,
iteration, seconds, learning_rate, phase)
test_match = regex_test_output.search(line)
if test_match:
test_dict_list, test_row = parse_line_for_net_output(
regex_test_output, test_row, test_dict_list, line,
iteration, seconds, learning_rate, phase)
except ValueError:
print("Oops!")
print line
continue
fix_initial_nan_learning_rate(train_dict_list)
fix_initial_nan_learning_rate(test_dict_list)
#print train_dict_list
return train_dict_list, test_dict_list
def parse_log(path_to_log):
"""Parse log file
Returns (train_dict_list, train_dict_names, test_dict_list,
test_dict_names, debug_info_names, debug_info_dict_list)
If the debug info wasn't enabled for the run, debug_info_dict_list is
empty
train_dict_list, test_dict_list and debug_info_dict_list are lists of
dicts that define the table rows
train_dict_names, test_dict_names and debug_info_names are ordered
tuples of the column names for the two dict_lists
"""
re_iteration = re.compile('Iteration (\d+)')
re_train_loss = re.compile('Iteration (\d+), loss = (' + FLOAT_RE + ')')
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]+)?)')
# For parsing debug info
re_forward_data = re.compile(
'\[Forward\] Layer (\S+), top blob (\S+) data: (' + FLOAT_RE + ')')
re_backward_diff = re.compile(
'\[Backward\] Layer (\S+), bottom blob (\S+) diff: (' + FLOAT_RE + ')')
re_backward_param_diff = re.compile(
'\[Backward\] Layer (\S+), param blob (\d+) diff: (' + FLOAT_RE + ')')
re_forward_param_data = re.compile(
'\[Forward\] Layer (\S+), param blob (\d+) data: (' + FLOAT_RE + ')')
was_in_backward = False
# Pick out lines of interest
iteration = -1
fb_iteration = -1 # iter # used for timing forward\backward
debug_flag = False
max_param_count = -1
learning_rate = float('NaN')
train_dict_list = []
test_dict_list = []
debug_info_dict_list = []
debug_layer_dict = {}
train_row = None
test_row = None
train_dict_names = ('NumIters', 'Seconds', 'TrainingLoss', 'LearningRate')
test_dict_names = ('NumIters', 'Seconds', 'TestAccuracy', 'TestLoss')
debug_info_dict_list = []
debug_info_names_list = [
'NumIters', 'LayerName', 'Activation', 'BackPropBottomDiff'
]
debug_info_names = tuple(debug_info_names_list)
logfile_year = extract_seconds.get_log_created_year(path_to_log)
with open(path_to_log) as f:
display_interval = get_display_interval(f)
debug_flag = is_debug_mode(f)
if debug_flag:
max_param_count = getMaxParamCount(f)
additional_header_list = []
backward_param_headers = [
'BackPropDiff' + '_param' + str(i)
for i in range(max_param_count + 1)
]
additional_header_list += backward_param_headers
for i in range(max_param_count + 1):
additional_header_list.append('param' + str(i) + '_Data')
additional_header_list.append('param' + str(i) + '_Change')
# adding new headers for each of the parameters
debug_info_names_list += additional_header_list
debug_info_names = tuple(debug_info_names_list)
f.seek(0) # return to head of file
start_time = extract_seconds.get_start_time(f, logfile_year)
for line in f:
iteration_match = re_iteration.search(line)
if iteration_match:
iteration = float(iteration_match.group(1))
if iteration == -1:
# Only look for other stuff if we've found the first iteration
continue
# Try to extract date and time from line, assuming there exists one in
# the expected format
try:
time = extract_seconds.extract_datetime_from_line(
line, logfile_year)
except:
continue
seconds = (time - start_time).total_seconds()
train_loss_match = re_train_loss.search(line)
if train_loss_match:
fb_iteration = float(train_loss_match.group(1))
learning_rate_match = regex_learning_rate.search(line)
if learning_rate_match:
learning_rate = float(learning_rate_match.group(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)
# Only extract debug information if debug_info is true
if not debug_flag:
continue
forward_match = re_forward_data.search(line)
if forward_match:
# If was_in_update flag was on, we are starting a new forward
# pass so we will save last iteration info and
# initialize the iteration specific variables
if was_in_backward:
debug_info_dict_list += debug_layer_dict.values()
debug_layer_dict = {}
was_in_backward = False
layer_name = forward_match.group(1)
activation_val = extended_float(forward_match.group(3))
if not debug_layer_dict.has_key(layer_name):
debug_layer_dict[layer_name] = dict.fromkeys(
debug_info_names)
debug_layer_dict[layer_name]['LayerName'] = layer_name
debug_layer_dict[layer_name]['NumIters'] = \
(fb_iteration != -1) * (fb_iteration + display_interval)
debug_layer_dict[layer_name]['Activation'] = activation_val
forward_param_data_match = re_forward_param_data.search(line)
if forward_param_data_match:
layer_name = forward_param_data_match.group(1)
param_num = forward_param_data_match.group(2)
param_header = 'param' + param_num
param_data = extended_float(forward_param_data_match.group(3))
debug_layer_dict[layer_name][param_header +
'_Data'] = param_data
backward_match = re_backward_diff.search(line)
if backward_match:
layer_name = backward_match.group(1)
back_prop_val = extended_float(backward_match.group(3))
if not debug_layer_dict.has_key(layer_name):
debug_layer_dict[layer_name] = dict.fromkeys(
debug_info_names)
debug_layer_dict[layer_name][
'BackPropBottomDiff'] = back_prop_val
backward_param_match = re_backward_param_diff.search(line)
if backward_param_match:
was_in_backward = True
layer_name = backward_param_match.group(1)
param_num = backward_param_match.group(2)
param_header = '_param' + param_num
back_prop_param_val = extended_float(
backward_param_match.group(3))
if not debug_layer_dict.has_key(layer_name):
debug_layer_dict[layer_name] = dict.fromkeys(
debug_info_names)
debug_layer_dict[layer_name][
'BackPropDiff' + param_header] = back_prop_param_val
# add last iteration information if it exists
if debug_flag and debug_layer_dict:
debug_info_dict_list += debug_layer_dict.values()
return train_dict_list, train_dict_names, test_dict_list, test_dict_names, \
debug_info_dict_list, debug_info_names
