def test(net,
val_data,
use_cuda,
calc_weight_count=False,
calc_flops=False,
extended_log=False):
acc_top1 = AverageMeter()
acc_top5 = AverageMeter()
tic = time.time()
err_top1_val, err_top5_val = validate(
acc_top1=acc_top1,
acc_top5=acc_top5,
net=net,
val_data=val_data,
use_cuda=use_cuda)
if calc_weight_count:
weight_count = calc_net_weight_count(net)
logging.info('Model: {} trainable parameters'.format(weight_count))
if calc_flops:
n_flops, n_params = measure_model(net, 224, 224)
logging.info('Params: {} ({:.2f}M), FLOPs: {} ({:.2f}M)'.format(
n_params, n_params / 1e6, n_flops, n_flops / 1e6))
if extended_log:
logging.info('Test: err-top1={top1:.4f} ({top1})\terr-top5={top5:.4f} ({top5})'.format(
top1=err_top1_val, top5=err_top5_val))
else:
logging.info('Test: err-top1={top1:.4f}\terr-top5={top5:.4f}'.format(
top1=err_top1_val, top5=err_top5_val))
logging.info('Time cost: {:.4f} sec'.format(
time.time() - tic))
def test(net,
val_data,
use_cuda,
input_image_size,
in_channels,
calc_weight_count=False,
calc_flops=False,
calc_flops_only=True,
extended_log=False):
if not calc_flops_only:
acc_top1 = AverageMeter()
acc_top5 = AverageMeter()
tic = time.time()
err_top1_val, err_top5_val = validate(acc_top1=acc_top1,
acc_top5=acc_top5,
net=net,
val_data=val_data,
use_cuda=use_cuda)
if extended_log:
logging.info(
'Test: err-top1={top1:.4f} ({top1})\terr-top5={top5:.4f} ({top5})'
.format(top1=err_top1_val, top5=err_top5_val))
else:
logging.info(
'Test: err-top1={top1:.4f}\terr-top5={top5:.4f}'.format(
top1=err_top1_val, top5=err_top5_val))
logging.info('Time cost: {:.4f} sec'.format(time.time() - tic))
if calc_weight_count:
weight_count = calc_net_weight_count(net)
if not calc_flops:
logging.info('Model: {} trainable parameters'.format(weight_count))
if calc_flops:
num_flops, num_macs, num_params = measure_model(
net, in_channels, input_image_size)
assert (not calc_weight_count) or (weight_count == num_params)
stat_msg = "Params: {params} ({params_m:.2f}M), FLOPs: {flops} ({flops_m:.2f}M)," \
" FLOPs/2: {flops2} ({flops2_m:.2f}M), MACs: {macs} ({macs_m:.2f}M)"
logging.info(
stat_msg.format(params=num_params,
params_m=num_params / 1e6,
flops=num_flops,
flops_m=num_flops / 1e6,
flops2=num_flops / 2,
flops2_m=num_flops / 2 / 1e6,
macs=num_macs,
macs_m=num_macs / 1e6))
def test(net,
test_data,
metric,
use_cuda,
input_image_size,
in_channels,
calc_weight_count=False,
calc_flops=False,
calc_flops_only=True,
extended_log=False):
if not calc_flops_only:
tic = time.time()
validate(
metric=metric,
net=net,
val_data=test_data,
use_cuda=use_cuda)
accuracy_msg = report_accuracy(
metric=metric,
extended_log=extended_log)
logging.info("Test: {}".format(accuracy_msg))
logging.info("Time cost: {:.4f} sec".format(
time.time() - tic))
if calc_weight_count:
weight_count = calc_net_weight_count(net)
if not calc_flops:
logging.info("Model: {} trainable parameters".format(weight_count))
if calc_flops:
num_flops, num_macs, num_params = measure_model(net, in_channels, input_image_size)
assert (not calc_weight_count) or (weight_count == num_params)
stat_msg = "Params: {params} ({params_m:.2f}M), FLOPs: {flops} ({flops_m:.2f}M)," \
" FLOPs/2: {flops2} ({flops2_m:.2f}M), MACs: {macs} ({macs_m:.2f}M)"
logging.info(stat_msg.format(
params=num_params, params_m=num_params / 1e6,
flops=num_flops, flops_m=num_flops / 1e6,
flops2=num_flops / 2, flops2_m=num_flops / 2 / 1e6,
macs=num_macs, macs_m=num_macs / 1e6))
def test(net,
test_data,
use_cuda,
input_image_size,
in_channels,
num_classes,
calc_weight_count=False,
calc_flops=False,
calc_flops_only=True,
extended_log=False,
dataset_metainfo=None):
assert (dataset_metainfo is not None)
if not calc_flops_only:
metric = []
pix_acc_macro_average = False
metric.append(
PixelAccuracyMetric(vague_idx=dataset_metainfo["vague_idx"],
use_vague=dataset_metainfo["use_vague"],
macro_average=pix_acc_macro_average))
mean_iou_macro_average = False
metric.append(
MeanIoUMetric(num_classes=num_classes,
vague_idx=dataset_metainfo["vague_idx"],
use_vague=dataset_metainfo["use_vague"],
bg_idx=dataset_metainfo["background_idx"],
ignore_bg=dataset_metainfo["ignore_bg"],
macro_average=mean_iou_macro_average))
tic = time.time()
accuracy_info = validate1(accuracy_metrics=metric,
net=net,
val_data=test_data,
use_cuda=use_cuda)
pix_acc = accuracy_info[0][1]
mean_iou = accuracy_info[1][1]
pix_macro = "macro" if pix_acc_macro_average else "micro"
iou_macro = "macro" if mean_iou_macro_average else "micro"
if extended_log:
logging.info(
"Test: {pix_macro}-pix_acc={pix_acc:.4f} ({pix_acc}), "
"{iou_macro}-mean_iou={mean_iou:.4f} ({mean_iou})".format(
pix_macro=pix_macro,
pix_acc=pix_acc,
iou_macro=iou_macro,
mean_iou=mean_iou))
else:
logging.info(
"Test: {pix_macro}-pix_acc={pix_acc:.4f}, {iou_macro}-mean_iou={mean_iou:.4f}"
.format(pix_macro=pix_macro,
pix_acc=pix_acc,
iou_macro=iou_macro,
mean_iou=mean_iou))
logging.info('Time cost: {:.4f} sec'.format(time.time() - tic))
if calc_weight_count:
weight_count = calc_net_weight_count(net)
if not calc_flops:
logging.info('Model: {} trainable parameters'.format(weight_count))
if calc_flops:
num_flops, num_macs, num_params = measure_model(
net, in_channels, input_image_size)
assert (not calc_weight_count) or (weight_count == num_params)
stat_msg = "Params: {params} ({params_m:.2f}M), FLOPs: {flops} ({flops_m:.2f}M)," \
" FLOPs/2: {flops2} ({flops2_m:.2f}M), MACs: {macs} ({macs_m:.2f}M)"
logging.info(
stat_msg.format(params=num_params,
params_m=num_params / 1e6,
flops=num_flops,
flops_m=num_flops / 1e6,
flops2=num_flops / 2,
flops2_m=num_flops / 2 / 1e6,
macs=num_macs,
macs_m=num_macs / 1e6))
def test(net,
test_data,
metric,
use_cuda,
input_image_size,
in_channels,
calc_weight_count=False,
calc_flops=False,
calc_flops_only=True,
extended_log=False,
show_bad_samples=False):
"""
Main test routine.
Parameters:
----------
net : Module
Model.
test_data : DataLoader
Data loader.
metric : EvalMetric
Metric object instance.
use_cuda : bool
Whether to use CUDA.
input_image_size : tuple of 2 ints
Spatial size of the expected input image.
in_channels : int
Number of input channels.
calc_weight_count : bool, default False
Whether to calculate count of weights.
calc_flops : bool, default False
Whether to calculate FLOPs.
calc_flops_only : bool, default True
Whether to only calculate FLOPs without testing.
extended_log : bool, default False
Whether to log more precise accuracy values.
show_bad_samples : bool, default False
Whether to log file names for bad samples.
"""
if not calc_flops_only:
tic = time.time()
validate(
metric=metric,
net=net,
val_data=test_data,
use_cuda=use_cuda)
accuracy_msg = report_accuracy(
metric=metric,
extended_log=extended_log)
logging.info("Test: {}".format(accuracy_msg))
logging.info("Time cost: {:.4f} sec".format(
time.time() - tic))
if calc_weight_count:
weight_count = calc_net_weight_count(net)
if not calc_flops:
logging.info("Model: {} trainable parameters".format(weight_count))
if calc_flops:
num_flops, num_macs, num_params = measure_model(net, in_channels, input_image_size)
assert (not calc_weight_count) or (weight_count == num_params)
stat_msg = "Params: {params} ({params_m:.2f}M), FLOPs: {flops} ({flops_m:.2f}M)," \
" FLOPs/2: {flops2} ({flops2_m:.2f}M), MACs: {macs} ({macs_m:.2f}M)"
logging.info(stat_msg.format(
params=num_params, params_m=num_params / 1e6,
flops=num_flops, flops_m=num_flops / 1e6,
flops2=num_flops / 2, flops2_m=num_flops / 2 / 1e6,
macs=num_macs, macs_m=num_macs / 1e6))
if show_bad_samples:
store_misses = StoreMisses()
validate(
metric=store_misses,
net=net,
val_data=test_data,
use_cuda=use_cuda)
_, misses_list = store_misses.get()
if len(misses_list) > 0:
dataset = test_data.iterable.dataset if isinstance(test_data, tqdm) else test_data.dataset
for i, miss_ind in enumerate(misses_list):
logging.info("Miss [{}]: {}".format(i, dataset.get_file_name(miss_ind)))
def calc_model_accuracy(net,
test_data,
metric,
use_cuda,
input_image_size,
in_channels,
calc_weight_count=False,
calc_flops=False,
calc_flops_only=True,
extended_log=False,
ml_type="cls"):
"""
Estimating particular model accuracy.
Parameters:
----------
net : Module
Model.
test_data : DataLoader
Data loader.
metric : EvalMetric
Metric object instance.
use_cuda : bool
Whether to use CUDA.
input_image_size : tuple of 2 ints
Spatial size of the expected input image.
in_channels : int
Number of input channels.
calc_weight_count : bool, default False
Whether to calculate count of weights.
calc_flops : bool, default False
Whether to calculate FLOPs.
calc_flops_only : bool, default True
Whether to only calculate FLOPs without testing.
extended_log : bool, default False
Whether to log more precise accuracy values.
ml_type : str, default 'cls'
Machine learning type.
Returns:
-------
list of floats
Accuracy values.
"""
if not calc_flops_only:
tic = time.time()
validate(
metric=metric,
net=net,
val_data=test_data,
use_cuda=use_cuda)
accuracy_msg = report_accuracy(
metric=metric,
extended_log=extended_log)
logging.info("Test: {}".format(accuracy_msg))
logging.info("Time cost: {:.4f} sec".format(
time.time() - tic))
acc_values = metric.get()[1]
acc_values = acc_values if type(acc_values) == list else [acc_values]
else:
acc_values = []
if calc_weight_count:
weight_count = calc_net_weight_count(net)
if not calc_flops:
logging.info("Model: {} trainable parameters".format(weight_count))
if calc_flops:
in_shapes = [(1, 640 * 25 * 5), (1,)] if ml_type == "asr" else\
[(1, in_channels, input_image_size[0], input_image_size[1])]
num_flops, num_macs, num_params = measure_model(
model=net,
in_shapes=in_shapes)
assert (not calc_weight_count) or (weight_count == num_params)
stat_msg = "Params: {params} ({params_m:.2f}M), FLOPs: {flops} ({flops_m:.2f}M)," \
" FLOPs/2: {flops2} ({flops2_m:.2f}M), MACs: {macs} ({macs_m:.2f}M)"
logging.info(stat_msg.format(
params=num_params, params_m=num_params / 1e6,
flops=num_flops, flops_m=num_flops / 1e6,
flops2=num_flops / 2, flops2_m=num_flops / 2 / 1e6,
macs=num_macs, macs_m=num_macs / 1e6))
return acc_values
