def _evaluate(data_loader,
model,
criterion,
writer,
epoch,
logging_label,
no_cuda=False,
log_interval=10,
**kwargs):
"""
The evaluation routine
Parameters
----------
data_loader : torch.utils.data.DataLoader
The dataloader of the evaluation set
model : torch.nn.module
The network model being used
criterion: torch.nn.loss
The loss function used to compute the loss of the model
writer : tensorboardX.writer.SummaryWriter
The tensorboard writer object. Used to log values on file for the tensorboard visualization.
epoch : int
Number of the epoch (for logging purposes)
logging_label : string
Label for logging purposes. Typically 'test' or 'valid'. Its prepended to the logging output path and messages.
no_cuda : boolean
Specifies whether the GPU should be used or not. A value of 'True' means the CPU will be used.
log_interval : int
Interval limiting the logging of mini-batches. Default value of 10.
Returns
-------
top1.avg : float
Accuracy of the model of the evaluated split
"""
multi_run = kwargs['run'] if 'run' in kwargs else None
# Instantiate the counters
batch_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
data_time = AverageMeter()
# Switch to evaluate mode (turn off dropout & such )
model.eval()
# Iterate over whole evaluation set
end = time.time()
# Empty lists to store the predictions and target values
preds = []
targets = []
multi_run = False
pbar = tqdm(enumerate(data_loader),
total=len(data_loader),
unit='batch',
ncols=150,
leave=False)
for batch_idx, (input, target) in pbar:
# todo: how to you implement sliding window accross batches
if len(input.size()) == 5:
multi_run = True
# input [64, 5, 3, 299, 299]
bs, ncrops, c, h, w = input.size()
# input.view leaves the 3rd 4th and 5th dimension as is, but multiplies the 1st and 2nd together
# result [320, 3, 299, 299]
# result = input.view(-1, c, h, w) # fuse batch size and ncrops
# result_avg = input.view(bs, -1, c, h, w).mean(1)
input = input.view(-1, c, h, w)
# If you are using tensor.max(1) then you get a tupel with two tensors, choose the first one
# which is a floattensor and what you need.
# input = result_avg[0]
# Measure data loading time
data_time.update(time.time() - end)
# Moving data to GPU
if not no_cuda:
input = input.cuda(async=True)
target = target.cuda(async=True)
# Convert the input and its labels to Torch Variables
# todo: check them out in debugger
input_var = torch.autograd.Variable(input, volatile=True)
target_var = torch.autograd.Variable(target, volatile=True)
# Compute output
output = model(input_var)
if multi_run:
output = output.view(bs, ncrops, -1).mean(1)
# Compute and record the loss
loss = criterion(output, target_var)
losses.update(loss.data[0], input.size(0))
# Compute and record the accuracy
acc1 = accuracy(output.data, target, topk=(1, ))[0]
top1.update(acc1[0], input.size(0))
# Get the predictions
_ = [
preds.append(item) for item in
[np.argmax(item) for item in output.data.cpu().numpy()]
]
_ = [targets.append(item) for item in target.cpu().numpy()]
# Add loss and accuracy to Tensorboard
if multi_run is None:
writer.add_scalar(logging_label + '/mb_loss', loss.data[0],
epoch * len(data_loader) + batch_idx)
writer.add_scalar(logging_label + '/mb_accuracy',
acc1.cpu().numpy(),
epoch * len(data_loader) + batch_idx)
else:
writer.add_scalar(logging_label + '/mb_loss_{}'.format(multi_run),
loss.data[0],
epoch * len(data_loader) + batch_idx)
writer.add_scalar(
logging_label + '/mb_accuracy_{}'.format(multi_run),
acc1.cpu().numpy(),
epoch * len(data_loader) + batch_idx)
# Measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if batch_idx % log_interval == 0:
pbar.set_description(logging_label +
' epoch [{0}][{1}/{2}]\t'.format(
epoch, batch_idx, len(data_loader)))
pbar.set_postfix(
Time='{batch_time.avg:.3f}\t'.format(batch_time=batch_time),
Loss='{loss.avg:.4f}\t'.format(loss=losses),
Acc1='{top1.avg:.3f}\t'.format(top1=top1),
Data='{data_time.avg:.3f}\t'.format(data_time=data_time))
# Make a confusion matrix
try:
cm = confusion_matrix(y_true=targets, y_pred=preds)
confusion_matrix_heatmap = make_heatmap(cm,
data_loader.dataset.classes)
except ValueError:
logging.warning('Confusion Matrix did not work as expected')
confusion_matrix_heatmap = np.zeros((10, 10, 3))
# Logging the epoch-wise accuracy and confusion matrix
if multi_run is None:
writer.add_scalar(logging_label + '/accuracy', top1.avg, epoch)
# ERROR: save_image_and_log_tensorboard() got an unexpected keyword argument 'image_tensore'
# changed 'image_tensor=confusion_matrix_heattmap' to 'image=confusion_mastrix_heatmap'
save_image_and_log_to_tensorboard(writer,
tag=logging_label +
'/confusion_matrix',
image=confusion_matrix_heatmap,
global_step=epoch)
else:
writer.add_scalar(logging_label + '/accuracy_{}'.format(multi_run),
top1.avg, epoch)
save_image_and_log_to_tensorboard(
writer,
tag=logging_label + '/confusion_matrix_{}'.format(multi_run),
image=confusion_matrix_heatmap,
global_step=epoch)
logging.info(
_prettyprint_logging_label(logging_label) + ' epoch[{}]: '
'Acc@1={top1.avg:.3f}\t'
'Loss={loss.avg:.4f}\t'
'Batch time={batch_time.avg:.3f} ({data_time.avg:.3f} to load data)'.
format(epoch,
batch_time=batch_time,
data_time=data_time,
loss=losses,
top1=top1))
# Generate a classification report for each epoch
_log_classification_report(data_loader, epoch, preds, targets, writer)
return top1.avg
def _evaluate(data_loader,
model,
criterion,
observer,
observer_criterion,
writer,
epoch,
logging_label,
no_cuda=False,
log_interval=10,
**kwargs):
"""
The evaluation routine
Parameters
----------
:param data_loader : torch.utils.data.DataLoader
The dataloader of the evaluation set
:param model : torch.nn.module
The network model being used
:param criterion: torch.nn.loss
The loss function used to compute the loss of the model
:param writer : tensorboardX.writer.SummaryWriter
The tensorboard writer object. Used to log values on file for the tensorboard visualization.
:param epoch : int
Number of the epoch (for logging purposes)
:param logging_label : string
Label for logging purposes. Typically 'test' or 'valid'. Its prepended to the logging output path and messages.
:param no_cuda : boolean
Specifies whether the GPU should be used or not. A value of 'True' means the CPU will be used.
:param log_interval : int
Interval limiting the logging of mini-batches. Default value of 10.
:return:
None
"""
multi_run = kwargs['run'] if 'run' in kwargs else None
# Instantiate the counters
batch_time = AverageMeter()
losses = AverageMeter()
observer_loss_meter = AverageMeter()
top1 = AverageMeter()
observer_acc_meter = AverageMeter()
data_time = AverageMeter()
# Switch to evaluate mode (turn off dropout & such )
model.eval()
# Iterate over whole evaluation set
end = time.time()
# Empty lists to store the predictions and target values
preds = []
targets = []
pbar = tqdm(enumerate(data_loader),
total=len(data_loader),
unit='batch',
ncols=150,
leave=False)
for batch_idx, (input, target) in pbar:
# Measure data loading time
data_time.update(time.time() - end)
# Moving data to GPU
if not no_cuda:
input = input.cuda(async=True)
target = target.cuda(async=True)
# Convert the input and its labels to Torch Variables
input_var = torch.autograd.Variable(input, volatile=True)
target_var = torch.autograd.Variable(target, volatile=True)
# Compute output
output = model(input_var)
# Get the features from second last layer
input_features_var = torch.autograd.Variable(
model.module.features.data)
# Use observer on the features
observer_acc, observer_loss = evaluate_one_mini_batch(
observer, observer_criterion, input_features_var, target_var,
observer_loss_meter, observer_acc_meter)
# Compute and record the loss
loss = criterion(output, target_var)
losses.update(loss.data[0], input.size(0))
# Compute and record the accuracy
acc1 = accuracy(output.data, target, topk=(1, ))[0]
top1.update(acc1[0], input.size(0))
# Get the predictions
_ = [
preds.append(item) for item in
[np.argmax(item) for item in output.data.cpu().numpy()]
]
_ = [targets.append(item) for item in target.cpu().numpy()]
# Add loss and accuracy to Tensorboard
if multi_run is None:
writer.add_scalar(logging_label + '/mb_loss', loss.data[0],
epoch * len(data_loader) + batch_idx)
writer.add_scalar(logging_label + '/mb_accuracy',
acc1.cpu().numpy(),
epoch * len(data_loader) + batch_idx)
writer.add_scalar(logging_label + '/obs_mb_loss',
observer_loss.data[0],
epoch * len(data_loader) + batch_idx)
writer.add_scalar(logging_label + '/obs_mb_accuracy',
observer_acc.cpu().numpy(),
epoch * len(data_loader) + batch_idx)
else:
writer.add_scalar(logging_label + '/mb_loss_{}'.format(multi_run),
loss.data[0],
epoch * len(data_loader) + batch_idx)
writer.add_scalar(
logging_label + '/mb_accuracy_{}'.format(multi_run),
acc1.cpu().numpy(),
epoch * len(data_loader) + batch_idx)
writer.add_scalar(
logging_label + '/obs_mb_loss_{}'.format(multi_run),
observer_loss.data[0],
epoch * len(data_loader) + batch_idx)
writer.add_scalar(
logging_label + '/obs_mb_accuracy_{}'.format(multi_run),
observer_acc.cpu().numpy(),
epoch * len(data_loader) + batch_idx)
# Measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if batch_idx % log_interval == 0:
pbar.set_description(logging_label +
' epoch [{0}][{1}/{2}]\t'.format(
epoch, batch_idx, len(data_loader)))
pbar.set_postfix(
Time='{batch_time.avg:.3f}\t'.format(batch_time=batch_time),
Loss='{loss.avg:.4f}\t'.format(loss=losses),
Acc1='{top1.avg:.3f}\t'.format(top1=top1),
Data='{data_time.avg:.3f}\t'.format(data_time=data_time))
# Make a confusion matrix
try:
cm = confusion_matrix(y_true=targets, y_pred=preds)
confusion_matrix_heatmap = make_heatmap(cm,
data_loader.dataset.classes)
except ValueError:
logging.warning('Confusion Matrix did not work as expected')
confusion_matrix_heatmap = np.zeros((10, 10, 3))
# Logging the epoch-wise accuracy and confusion matrix
if multi_run is None:
writer.add_scalar(logging_label + '/accuracy', top1.avg, epoch)
writer.add_scalar(logging_label + '/obs_accuracy',
observer_acc_meter.avg, epoch)
save_image_and_log_to_tensorboard(
writer,
tag=logging_label + '/confusion_matrix',
image_tensor=confusion_matrix_heatmap,
global_step=epoch)
else:
writer.add_scalar(logging_label + '/accuracy_{}'.format(multi_run),
top1.avg, epoch)
writer.add_scalar(logging_label + '/obs_accuracy_{}'.format(multi_run),
observer_acc_meter.avg, epoch)
save_image_and_log_to_tensorboard(
writer,
tag=logging_label + '/confusion_matrix_{}'.format(multi_run),
image_tensor=confusion_matrix_heatmap,
global_step=epoch)
logging.info(
_prettyprint_logging_label(logging_label) + ' epoch[{}]: '
'Acc@1={top1.avg:.3f}\t'
'Loss={loss.avg:.4f}\t'
'Batch time={batch_time.avg:.3f} ({data_time.avg:.3f} to load data)'.
format(epoch,
batch_time=batch_time,
data_time=data_time,
loss=losses,
top1=top1))
# Generate a classification report for each epoch
_log_classification_report(data_loader, epoch, preds, targets, writer)
return top1.avg
def feature_extract(data_loader, model, writer, epoch, no_cuda, log_interval, classify, **kwargs):
"""
The evaluation routine
Parameters
----------
data_loader : torch.utils.data.DataLoader
The dataloader of the evaluation set
model : torch.nn.module
The network model being used
writer : tensorboardX.writer.SummaryWriter
The tensorboard writer object. Used to log values on file for the tensorboard visualization.
epoch : int
Number of the epoch (for logging purposes)
no_cuda : boolean
Specifies whether the GPU should be used or not. A value of 'True' means the CPU will be used.
log_interval : int
Interval limiting the logging of mini-batches. Default value of 10.
classify : boolean
Specifies whether to generate a classification report for the data or not.
Returns
-------
None
"""
logging_label = 'apply'
# Switch to evaluate mode (turn off dropout & such )
model.eval()
labels, features, preds, filenames = [], [], [], []
multi_crop = False
# Iterate over whole evaluation set
pbar = tqdm(enumerate(data_loader), total=len(data_loader), unit='batch', ncols=200)
with torch.no_grad():
for batch_idx, (data, label, filename) in pbar:
if len(data.size()) == 5:
multi_crop = True
bs, ncrops, c, h, w = data.size()
data = data.view(-1, c, h, w)
if not no_cuda:
data = data.cuda()
# Compute output
out = model(data)
if multi_crop:
out = out.view(bs, ncrops, -1).mean(1)
preds.append([np.argmax(item.data.cpu().numpy()) for item in out])
features.append(out.data.cpu().numpy())
labels.append(label)
filenames.append(filename)
# Log progress to console
if batch_idx % log_interval == 0:
pbar.set_description(logging_label + ' Epoch: {} [{}/{} ({:.0f}%)]'.format(
epoch, batch_idx * len(data), len(data_loader.dataset),
100. * batch_idx / len(data_loader)))
# Measure accuracy (FPR95)
num_tests = len(data_loader.dataset)
labels = np.concatenate(labels, 0).reshape(num_tests)
features = np.concatenate(features, 0)
preds = np.concatenate(preds, 0)
filenames = np.concatenate(filenames, 0)
if classify:
# Make a confusion matrix
try:
cm = confusion_matrix(y_true=labels, y_pred=preds)
confusion_matrix_heatmap = make_heatmap(cm, data_loader.dataset.classes)
save_image_and_log_to_tensorboard(writer, tag=logging_label + '/confusion_matrix',
image=confusion_matrix_heatmap, global_step=epoch)
except ValueError:
logging.warning('Confusion matrix received weird values')
# Generate a classification report for each epoch
logging.info('Classification Report for epoch {}\n'.format(epoch))
logging.info('\n' + classification_report(y_true=labels,
y_pred=preds,
target_names=[str(item) for item in data_loader.dataset.classes]))
else:
preds = None
return features, preds, labels, filenames
if batch_idx % log_interval == 0:
pbar.set_description(logging_label +
' epoch [{0}][{1}/{2}]\t'.format(
epoch, batch_idx, len(data_loader)))
pbar.set_postfix(
Time='{batch_time.avg:.3f}\t'.format(batch_time=batch_time),
Loss='{loss.avg:.4f}\t'.format(loss=losses),
Acc1='{top1.avg:.3f}\t'.format(top1=top1),
Data='{data_time.avg:.3f}\t'.format(data_time=data_time))
# Make a confusion matrix
try:
cm = confusion_matrix(y_true=targets, y_pred=preds)
confusion_matrix_heatmap = make_heatmap(cm,
data_loader.dataset.classes)
except ValueError:
logging.warning('Confusion Matrix did not work as expected')
confusion_matrix_heatmap = np.zeros((10, 10, 3))
# Logging the epoch-wise accuracy and confusion matrix
if multi_run is None:
writer.add_scalar(logging_label + '/accuracy', top1.avg, epoch)
save_image_and_log_to_tensorboard(
writer,
tag=logging_label + '/confusion_matrix',
image_tensor=confusion_matrix_heatmap,
global_step=epoch)
else:
writer.add_scalar(logging_label + '/accuracy_{}'.format(multi_run),
