def _get_input():
"""
Get tf.data.Dataset object according to command-line flags for testing
using tf.estimator.Estimator
Returns:
dataset : elements structured as [features, labels]
feature structure can be seen in postbatch_fn
in mjsynth.py
"""
# WARNING: More than two filters causes SEVERE throughput slowdown
filter_fn = filters.input_filter_fn \
( min_image_width=FLAGS.min_image_width,
max_image_width=FLAGS.max_image_width,
min_string_length=FLAGS.min_string_length,
max_string_length=FLAGS.max_string_length )
# Get data according to flags
dataset = pipeline.get_data(use_static_data=True,
base_dir=FLAGS.test_path,
file_patterns=str.split(
FLAGS.filename_pattern, ','),
num_threads=FLAGS.num_input_threads,
batch_size=FLAGS.batch_size,
filter_fn=filter_fn,
num_epochs=1)
return dataset
def _get_input():
"""
Get tf.data.Dataset according to command-line flags for training
using tf.estimator.Estimator
Note: Default behavior is bucketing according to default bucket boundaries
listed in pipeline.get_data
Returns:
dataset : elements structured as [features, labels]
feature structure can be seen in postbatch_fn
in mjsynth.py or maptextsynth.py for static or dynamic
data pipelines respectively
"""
# WARNING: More than two filters causes SEVERE throughput slowdown
filter_fn = filters.input_filter_fn \
( min_image_width=FLAGS.min_image_width,
max_image_width=FLAGS.max_image_width,
min_string_length=FLAGS.min_string_length,
max_string_length=FLAGS.max_string_length,
check_input=(not FLAGS.static_data) )
gpu_batch_size = FLAGS.batch_size // FLAGS.num_gpus
# Pack keyword arguments into dictionary
data_args = {
'num_threads': FLAGS.num_input_threads,
'batch_size': gpu_batch_size,
'filter_fn': filter_fn
}
if FLAGS.static_data: # Pack data stream-specific parameters
data_args['base_dir'] = FLAGS.train_path
data_args['file_patterns'] = str.split(FLAGS.filename_pattern, ',')
else:
data_args['synth_config_file'] = FLAGS.synth_config_file
data_args['use_ipc_synth'] = FLAGS.ipc_synth
if not FLAGS.bucket_data:
data_args['boundaries'] = None # Turn off bucketing (on by default)
elif not FLAGS.static_data: # Extra buckets for the wider synthetic data
data_args['boundaries'] = [
32, 64, 96, 128, 160, 192, 224, 256, 288, 320, 352, 384, 416, 448,
480, 512
]
# Get data according to flags
dataset = pipeline.get_data(FLAGS.static_data, **data_args)
return dataset
def run_results_analysis(run_log_df):
model_id_list = list(run_log_df.model_id)
model_id_list_str = "'" + "','".join(map(str, model_id_list)) + "'"
results_stats_tablename = "model.results_stats"
query_results_stats = ('select *'
' from ' + results_stats_tablename +
' where model_id in (' + model_id_list_str + ')')
results_stats_df = get_data(query_results_stats)
valid_tablename = "model.valid"
query_valid = ('select *'
' from ' + valid_tablename + ' where model_id in (' +
model_id_list_str + ')')
#valid_df= get_data(query_valid)
features_tablename = "model.feature_importances"
query_features = ('select *'
' from ' + features_tablename + ' where model_id in (' +
model_id_list_str + ')')
#features_df= get_data(query_features)
model_top_precision = list(
results_stats_df.sort_values('precision_at_1',
ascending=False)['model_id'])[0]
valid_top_precision_tablename = "model.valid"
query_valid_top_precision = ('select *'
' from ' + valid_top_precision_tablename +
" where model_id in ('" +
model_top_precision + "')")
print query_valid_top_precision
valid_top_precision_df = get_data(query_valid_top_precision)
gen_decile_table(model_top_precision, valid_top_precision_df)
def _get_input():
"""
Get tf.data.Dataset object according to command-line flags for evaluation
using tf.estimator.Estimator
Note: Default behavior is bucketing according to default bucket boundaries
listed in pipeline.get_data
Returns:
features, labels
feature structure can be seen in postbatch_fn
in mjsynth.py or maptextsynth.py for static or dynamic
data pipelines respectively
"""
# WARNING: More than two filters causes SEVERE throughput slowdown
filter_fn = filters.input_filter_fn \
( min_image_width=FLAGS.min_image_width,
max_image_width=FLAGS.max_image_width,
min_string_length=FLAGS.min_string_length,
max_string_length=FLAGS.max_string_length )
# Pack keyword arguments into dictionary
data_args = {
'base_dir': FLAGS.test_path,
'file_patterns': str.split(FLAGS.filename_pattern, ','),
'num_threads': FLAGS.num_input_threads,
'batch_size': FLAGS.batch_size,
'filter_fn': filter_fn
}
if not FLAGS.bucket_data: # Turn off bucketing (on by default in pipeline)
data_args['boundaries'] = None
# Get data according to flags
dataset = pipeline.get_data(use_static_data=True, **data_args)
return dataset
def run_feature_heatmap():
tablename = "model.results_stats"
break_window = "'3Year'"
past_year = "6"
static_features = "'diameters,pipe_age,install_year_imputed,musym,rocktype1,rocktype2,zone_name,material_imputed'"
query_models = ('select model_id,model_name'
' from ' + tablename + ' where break_window = ' +
break_window + ' and past_year = ' + past_year +
' and static_features = ' + static_features)
print query_models
model_df = get_data(query_models)
model_name_list = list(model_df.model_name)
model_id_list = list(model_df.model_id.values.ravel())
model_id_list_str = "'" + "','".join(map(str, model_id_list)) + "'"
tablename = "model.feature_importances"
query_features = ('select *'
' from ' + tablename + ' where model_id in (' +
model_list_str + ')')
print query_features
feature_imp = get_data(query_features)
#Check if all models have the same features
for model_id in model_id_list:
for i in (
feature_imp.feature[feature_imp.model_id == model_id].
sort_values(inplace=False) == feature_imp.feature[
feature_imp.model_id == '1471225699.39507699'].sort_values(
inplace=False)):
if i == False:
print model_id
feature_names_list = list(
feature_imp.feature[feature_imp.model_id ==
'1471225699.39507699'].sort_values(inplace=False))
feature_imp_matrix = pd.DataFrame
for model_id in model_list[:1]:
feature_imp_matrix = feature_imp[
feature_imp.model_id == model_id].sort_values(
"feature", inplace=False).importance.values
for model_id in model_list[1:]:
b = feature_imp[feature_imp.model_id == model_id].sort_values(
"feature", inplace=False).importance.values
feature_imp_matrix = np.vstack((feature_imp_matrix, b))
feature_imp_matrix_normd = scale(np.transpose(feature_imp_matrix),
axis=0,
with_mean=True,
with_std=True,
copy=True)
feature_imp_matrix_normd.shape
plt.figure(figsize=(50, 50))
sns.heatmap(
feature_imp_matrix_normd
#, annot=True
,
square=True,
linewidths=0.5,
yticklabels=feature_names_list,
xticklabels=model_names_list)
plt.savefig('feature_model_heatmap.png', bbox_inches='tight')
plt.close()