def sign_of_life(cls,
job,
num_already_present_imgs,
outdir,
start_time,
force_rewrite=False):
# Time for sign of life?
now_time = datetime.datetime.now()
time_duration = now_time - start_time
# Every 3 seconds, but at least 3:
if force_rewrite \
or (time_duration.seconds > 0 and time_duration.seconds % 3 == 0):
# A human readable duration st down to minutes:
duration_str = FileUtils.time_delta_str(time_duration,
granularity=4)
# Get current and new spectro imgs in outdir:
num_now_present_imgs = len(
Utils.find_in_dir_tree(outdir, pattern="*.png"))
num_newly_present_imgs = num_now_present_imgs - num_already_present_imgs
# Keep printing number of done snippets in the same
# terminal line:
print((f"{job.name}---Number of spectros: {num_now_present_imgs} "
f"({num_newly_present_imgs} new) after {duration_str}"),
end='\r')
return num_newly_present_imgs
else:
return num_already_present_imgs
def validate_split(self, step):
'''
Validate one split, using that split's
validation fold. Return time taken. Record
results for tensorboard and other record keeping.
:param step: current combination of epoch and
split
:type step: int
:return: number of epoch seconds needed for the validation
:rtype: int
'''
# Validation
self.log.debug(
f"Start of validation: \n{'none--on CPU' if self.fastest_device.type == 'cpu' else torch.cuda.memory_summary()}"
)
start_time = datetime.datetime.now()
self.log.info(f"Starting validation for step {step}")
self.model.eval()
with torch.no_grad():
for img_tensor, target in self.train_loader.validation_samples():
expanded_img_tensor = unsqueeze(img_tensor, dim=0)
expanded_target = unsqueeze(target, dim=0)
# Update sanity record:
self.class_coverage[int(target)]['val'] += 1
images = FileUtils.to_device(expanded_img_tensor, 'gpu')
label = FileUtils.to_device(expanded_target, 'gpu')
outputs = self.model(images)
loss = self.loss_fn(outputs, label)
images = FileUtils.to_device(images, 'cpu')
outputs = FileUtils.to_device(outputs, 'cpu')
label = FileUtils.to_device(label, 'cpu')
loss = FileUtils.to_device(loss, 'cpu')
self.remember_results(LearningPhase.VALIDATING, step, outputs,
label, loss)
del images
del outputs
del label
del loss
torch.cuda.empty_cache()
end_time = datetime.datetime.now()
val_time_duration = end_time - start_time
# A human readable duration st down to minues:
duration_str = FileUtils.time_delta_str(val_time_duration,
granularity=4)
self.log.info(f"Done validation (duration: {duration_str})")
return val_time_duration
def train(self):
overall_start_time = datetime.datetime.now()
# Just for sanity: keep track
# of number of batches...
total_batch_num = 0
# Note: since we are cross validating, the
# data loader's set_epoch() method is only
# called once (automatically) during instantiation
# of the associated sampler. Moving from split
# to split includes shuffling if the caller
# specified that.
# Training
for split_num in range(self.train_loader.num_folds):
split_start_time = datetime.datetime.now()
self.initialize_model()
for epoch in range(self.max_epochs):
# Set model to train mode:
self.model.train()
epoch_start_time = datetime.datetime.now()
self.log.info(f"Starting epoch {epoch} training")
# Sanity check record: will record
# how many samples from each class were
# used:
self.class_coverage = {}
# Sanity records: will record number
# of samples of each class that are used
# during training and validation:
label_distrib = {}
batch_num = 0
self.log.info(
f"Train epoch {epoch}/{self.max_epochs} split {split_num}/{self.train_loader.num_folds}"
)
try:
for batch, targets in self.train_loader:
# Update the sanity check
# num of batches seen, and distribution
# of samples across classes:
batch_num += 1
total_batch_num += 1
# Update sanity check records:
for lbl in targets:
lbl = int(lbl)
try:
label_distrib[lbl] += 1
except KeyError:
label_distrib[lbl] = 1
try:
self.class_coverage[lbl]['train'] += 1
except KeyError:
self.class_coverage[lbl] = {
'train': 1,
'val': 0
}
self.log.debug(
f"Top of training loop: \n{'none--on CPU' if self.fastest_device.type == 'cpu' else torch.cuda.memory_summary()}"
)
images = FileUtils.to_device(batch, 'gpu')
labels = FileUtils.to_device(targets, 'gpu')
outputs = self.model(images)
loss = self.loss_fn(outputs, labels)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# Remember the last batch's train result of this
# split (results for earlier batches of
# the same split will be overwritten). This statement
# must sit before deleting output and labels:
step_num = self.step_number(epoch, split_num,
self.num_folds)
self.remember_results(LearningPhase.TRAINING, step_num,
outputs, labels, loss)
self.log.debug(
f"Just before clearing gpu: \n{'none--on CPU' if self.fastest_device.type == 'cpu' else torch.cuda.memory_summary()}"
)
images = FileUtils.to_device(images, 'cpu')
outputs = FileUtils.to_device(outputs, 'cpu')
labels = FileUtils.to_device(labels, 'cpu')
loss = FileUtils.to_device(loss, 'cpu')
del images
del outputs
del labels
del loss
torch.cuda.empty_cache()
self.log.debug(
f"Just after clearing gpu: \n{'none--on CPU' if self.fastest_device.type == 'cpu' else torch.cuda.memory_summary()}"
)
except EndOfSplit:
end_time = datetime.datetime.now()
train_time_duration = end_time - epoch_start_time
# A human readable duration st down to minutes:
duration_str = FileUtils.time_delta_str(
train_time_duration, granularity=4)
self.log.info(
f"Done training epoch {epoch} of split {split_num} (duration: {duration_str})"
)
#***********
#print(f"****** num_batches in split: {batch_num}" )
#print(f"****** LblDist: {label_distrib}")
#***********
self.validate_split(step_num)
self.visualize_step(step_num)
# Save model, keeping self.model_archive_size models:
self.model_archive.save_model(self.model, epoch)
self.log.debug(
f"After eval: \n{'none--on CPU' if self.fastest_device.type == 'cpu' else torch.cuda.memory_summary()}"
)
# Next Epoch
continue
end_time = datetime.datetime.now()
train_time_duration = end_time - split_start_time
# A human readable duration st down to minutes:
duration_str = FileUtils.time_delta_str(train_time_duration,
granularity=4)
self.log.info(
f"Done training split {split_num} (duration: {duration_str})")
# Next split
continue
end_time = datetime.datetime.now()
epoch_duration = end_time - epoch_start_time
epoch_dur_str = FileUtils.time_delta_str(epoch_duration, granularity=4)
cumulative_dur = end_time - overall_start_time
cum_dur_str = FileUtils.time_delta_str(cumulative_dur, granularity=4)
msg = f"Done epoch {epoch} (epoch duration: {epoch_dur_str}; cumulative: {cum_dur_str})"
self.log.info(msg)
#******self.scheduler.step()
# Fresh results tallying
#self.results.clear()
self.log.info(
f"Training complete after {self.train_loader.num_folds} splits")
# Report the sanity checks:
self.log.info(f"Total batches processed: {total_batch_num}")
for cid in self.class_coverage.keys():
train_use, val_use = self.class_coverage[cid].items()
self.log.info(
f"{self.class_names[cid]} Training: {train_use}, Validation: {val_use}"
)
# All seems to have gone well. Report the
# overall result of the final epoch for the
# hparms config used in this process:
self.report_hparams_summary(self.latest_result)
# The final epoch number:
return epoch
def run_inference(self, gpu_to_use=0):
'''
Runs model over dataloader. Along
the way: creates ResultTally for each
batch, and maintains dict instance variable
self.raw_results for later conversion of
logits to class IDs under different threshold
assumptions.
self.raw_results:
{'all_outputs' : <arr>,
'all_labels' : <arr>
}
Returns a ResultCollection with the
ResultTally instances of each batch.
:param gpu_to_use: which GPU to deploy to (if it is available)
:type gpu_to_use: int
:return: collection of tallies, one for each batch,
or None if something went wrong.
:rtype: {None | ResultCollection}
'''
# Just in case the loop never runs:
batch_num = -1
overall_start_time = datetime.datetime.now()
try:
try:
if torch.cuda.is_available():
self.model.load_state_dict(torch.load(self.model_path))
FileUtils.to_device(self.model, 'gpu', gpu_to_use)
else:
self.model.load_state_dict(
torch.load(self.model_path,
map_location=torch.device('cpu')))
except RuntimeError as e:
emsg = repr(e)
if emsg.find("size mismatch for conv1") > -1:
emsg += " Maybe model was trained with to_grayscale=False, but local net created for grayscale?"
raise RuntimeError(emsg) from e
loss_fn = nn.CrossEntropyLoss()
result_coll = ResultCollection()
# Save all per-class logits for ability
# later to use different thresholds for
# conversion to class IDs:
all_outputs = []
all_labels = []
self.model.eval()
num_test_samples = len(self.loader.dataset)
self.log.info(
f"Begin inference ({num_test_samples} test samples)...")
samples_processed = 0
loop_start_time = overall_start_time
with torch.no_grad():
for batch_num, (batch, targets) in enumerate(self.loader):
if torch.cuda.is_available():
images = FileUtils.to_device(batch, 'gpu')
labels = FileUtils.to_device(targets, 'gpu')
else:
images = batch
labels = targets
outputs = self.model(images)
loss = loss_fn(outputs, labels)
images = FileUtils.to_device(images, 'cpu')
outputs = FileUtils.to_device(outputs, 'cpu')
labels = FileUtils.to_device(labels, 'cpu')
loss = FileUtils.to_device(loss, 'cpu')
#**********
max_logit = outputs[0].max().item()
max_idx = (outputs.squeeze() == max_logit).nonzero(
as_tuple=False).item()
smpl_id = torch.utils.data.dataloader.sample_id_seq[-1]
lbl = labels[0].item()
pred_cl = max_idx
self.curr_dict[smpl_id] = (smpl_id, lbl, pred_cl)
#**********
# Specify the batch_num in place
# of an epoch, which is not applicatble
# during testing:
tally = ResultTally(batch_num, LearningPhase.TESTING,
outputs, labels, loss,
self.num_classes, self.batch_size)
result_coll.add(tally, step=None)
all_outputs.append(outputs)
all_labels.append(labels)
samples_processed += len(labels)
del images
del outputs
del labels
del loss
torch.cuda.empty_cache()
time_now = datetime.datetime.now()
# Sign of life every 6 seconds:
if (time_now - loop_start_time).seconds >= 5:
self.log.info(
f"GPU{gpu_to_use} processed {samples_processed}/{num_test_samples} samples"
)
loop_start_time = time_now
finally:
#*********
print(f"Sample seq: {torch.utils.data.dataloader.sample_id_seq}")
torch.utils.data.dataloader.sample_id_seq = []
#*********
time_now = datetime.datetime.now()
test_time_duration = time_now - overall_start_time
# A human readable duration st down to minutes:
duration_str = FileUtils.time_delta_str(test_time_duration,
granularity=4)
self.log.info(
f"Done with inference: {samples_processed} test samples; {duration_str}"
)
# Total number of batches we ran:
num_batches = 1 + batch_num # b/c of zero-base
# If loader delivered nothing, the loop
# never ran; warn, and get out:
if num_batches == 0:
self.log.warn(
f"Dataloader delivered no data from {self.samples_path}")
self.close()
return None
# Var all_outputs is now:
# [tensor([pred_cl0, pred_cl1, pred_cl<num_classes - 1>], # For sample0
# tensor([pred_cl0, pred_cl1, pred_cl<num_classes - 1>], # For sample1
# ...
# ]
# Make into one tensor: (num_batches, batch_size, num_classes),
# unless an exception was raised at some point,
# throwing us into this finally clause:
if len(all_outputs) == 0:
self.log.info(
f"No outputs were produced; thus no results to report")
return None
self.all_outputs_tn = torch.stack(all_outputs)
# Be afraid...be very afraid:
assert(self.all_outputs_tn.shape == \
torch.Size([num_batches,
self.batch_size,
self.num_classes])
)
# Var all_labels is now num-batches tensors,
# each containing batch_size labels:
assert (len(all_labels) == num_batches)
# list of single-number tensors. Make
# into one tensor:
self.all_labels_tn = torch.stack(all_labels)
assert(self.all_labels_tn.shape == \
torch.Size([num_batches, self.batch_size])
)
# And equivalently:
assert(self.all_labels_tn.shape == \
(self.all_outputs_tn.shape[0],
self.all_outputs_tn.shape[1]
)
)
self.report_results(result_coll)
self.close()
return result_coll