def after_epoch(self):
self.learner.final_record = self.log[1:].copy()
self.values.append(self.learner.final_record)
self.log.append(format_time(time.time() - self.start_epoch))
self.logger(self.log)
self.iters.append(self.smooth_loss.count)
def after_epoch(self):
stats = [str(self.epoch)]
for o in [self.train_metrics, self.valid_metrics]:
for m in o:
stats += [f"{m.value:.6f}"]
stats += [format_time(time.time() - self.start_time)]
self.logger(stats)
def after_epoch(self):
stats = [str(self.epoch)]
for o in [self.train_stats, self.valid_stats]:
#now storing stats in array, and pass to logger
stats += [f'{v:.6f}' for v in o.avg_stats]
stats += [format_time(time.time() - self.start_time)]
self.logger(stats)
def format_stats(self, stats:TensorOrNumList)->None:
"Format stats before printing."
str_stats = []
for name,stat in zip(self.names,stats):
str_stats.append('#na#' if stat is None else str(stat) if isinstance(stat, int) else f'{stat:.6f}')
if self.add_time: str_stats.append(format_time(time() - self.start_epoch))
if not self.silent: self.pbar.write(str_stats, table=True)
def after_epoch(self):
"Store and log the loss/metric values"
self.learn.final_record = self.log[1:].copy()
self.values.append(self.learn.final_record)
if self.add_time: self.log.append(format_time(time.time() - self.start_epoch))
self.logger(self.log)
self.iters.append(self.smooth_loss.count)
def after_epoch(self):
"Store and log the loss/metric values"
self.learn.final_record = self.log[1:].copy() # first element is epoch
self.values.append(self.learn.final_record)
if self.show_time:
self.log.append(format_time(time.time() - self.epoch_start_time))
self.logger(self.log)
self.iters.append(self.train_loss.count)
def on_epoch_end(self, epoch: int, smooth_loss: Tensor, last_metrics: MetricsList, **kwargs: Any) -> bool:
"Add a line with `epoch` number, `smooth_loss` and `last_metrics`."
last_metrics = ifnone(last_metrics, [])
stats = [str(stat) if isinstance(stat, int) else '#na#' if stat is None else f'{stat:.6f}'
for name, stat in zip(self.learn.recorder.names, [epoch, smooth_loss] + last_metrics)]
if self.add_time: stats.append(format_time(time() - self.start_epoch))
str_stats = ','.join(stats)
self.file.write(str_stats + '\n')
def _format_stats(self, stats: TensorOrNumList) -> None:
"""Format stats before printing. Note, this does the same thing as Recorder's"""
str_stats = []
for name, stat in zip(self.names, stats):
str_stats.append("#na#" if stat is None else str(stat)
if isinstance(stat, int) else f"{stat:.6f}")
str_stats.append(format_time(time() - self.start_epoch))
self.pbar.write(str_stats, table=True)
def after_epoch(self, e: Event):
#print(self.train_stats)
#print(self.valid_stats)
stats = [str(e.learn.epoch)]
for o in [self.train_stats, self.valid_stats]:
stats += [f'{v:.6f}' for v in o.avg_stats]
stats += [format_time(time.time() - self.start_time)]
e.learn.logger(stats)
def fit(self, num_epochs, args, device='cuda:0'):
"""
Fit the PyTorch model
:param num_epochs: number of epochs to train (int)
:param args:
:param device: str (defaults to 'cuda:0')
"""
optimizer, scheduler, step_scheduler_on_batch = self.optimizer(args)
self.model = self.model.to(device)
pbar = master_bar(range(num_epochs))
headers = [
'Train_Loss', 'Val_Loss', 'F1-Macro', 'F1-Micro', 'JS', 'Time'
]
pbar.write(headers, table=True)
for epoch in pbar:
epoch += 1
start_time = time.time()
self.model.train()
overall_training_loss = 0.0
for step, batch in enumerate(
progress_bar(self.train_data_loader, parent=pbar)):
loss, num_rows, _, _ = self.model(batch, device)
overall_training_loss += loss.item() * num_rows
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
max_norm=1.0)
optimizer.step()
if step_scheduler_on_batch:
scheduler.step()
optimizer.zero_grad()
if not step_scheduler_on_batch:
scheduler.step()
overall_training_loss = overall_training_loss / len(
self.train_data_loader.dataset)
overall_val_loss, pred_dict = self.predict(device, pbar)
y_true, y_pred = pred_dict['y_true'], pred_dict['y_pred']
str_stats = []
stats = [
overall_training_loss, overall_val_loss,
f1_score(y_true, y_pred, average="macro"),
f1_score(y_true, y_pred, average="micro"),
jaccard_score(y_true, y_pred, average="samples")
]
for stat in stats:
str_stats.append('NA' if stat is None else str(stat)
if isinstance(stat, int) else f'{stat:.4f}')
str_stats.append(format_time(time.time() - start_time))
print('epoch#: ', epoch)
pbar.write(str_stats, table=True)
self.early_stop(overall_val_loss, self.model)
if self.early_stop.early_stop:
print("Early stopping")
break
def on_epoch_end(self, epoch: int, smooth_loss: Tensor, last_metrics: MetricsList, **kwargs: Any) -> bool:
"Add a line with `epoch` number, `smooth_loss` and `last_metrics`."
last_metrics = ifnone(last_metrics, [])
stats = [str(stat) if isinstance(stat, int) else '#na#' if stat is None else f'{stat:.6f}'
for name, stat in zip(self.learn.recorder.names, [epoch, smooth_loss] + last_metrics)]
if self.add_time: stats.append(format_time(time() - self.start_epoch))
str_stats = ','.join(stats)
if not self.gpu: self.file.write(str_stats + '\n')
self.file.flush()
os.fsync(self.file.fileno())
def after_epoch(self):
stats = [str(self.epoch)] # Begin with the epoch's number.
for o in [self.train_stats, self.valid_stats]:
# And nicely format all other metrics to be displayed.
stats += [f'{v:.6f}' for v in o.avg_stats]
# Also ensure total duration of is displayed.
stats += [format_time(time.time() - self.start_time)]
# Use Learner's self.logger function to display the metrics.
# print is the default but we will pass fastprogress'
# write() function to self.logger.
self.logger(stats)
def after_epoch(self):
stats = [str(self.epoch)]
for o in [self.train_stats, self.valid_stats]:
stats += [f"{v:.6f}" for v in o.avg_stats]
stats += [format_time(time.time() - self.start_time)]
writer.add_scalar("Loss/train", float(stats[1]), int(self.epoch))
writer.add_scalar("Accuracy/train", float(stats[2]), int(self.epoch))
writer.add_scalar("Loss/valid", float(stats[3]), int(self.epoch))
writer.add_scalar("Accuracy/train", float(stats[4]), int(self.epoch))
writer.add_scalar(
"Time/epoch", float(self.time_seconds(str(stats[5]))), int(self.epoch)
)
def on_epoch_end(self, epoch, smooth_loss, last_metrics, **kwargs):
"Add a line with `epoch` number, `smooth_loss` and `last_metrics`."
last_metrics = last_metrics if last_metrics is not None else []
metrics = zip(self.learn.recorder.names,
[epoch, smooth_loss] + last_metrics)
stats = [
str(stat) if isinstance(stat, int) else
'#na#' if stat is None else f'{stat:.6f}' for name, stat in metrics
]
if self.add_time:
stats.append(format_time(time.time() - self.start_epoch))
str_stats = ','.join(stats)
with self.path.open('a') as f:
f.write(str_stats + '\n')
def after_epoch(self):
try:
rec = self.learn.recorder
names = [str(m) for m in rec.metric_names]
values = [rec.epoch] + [f"{r:.3f}" for r in rec.final_record] + [
format_time(time() - self.start_epoch_time)
]
if len(names) == len(values) + 1:
names = names[:1] + names[
2:] # remove train_loss when validating only
msg = '\n'.join(
[f"{name}: {value}" for name, value in zip(names, values)])
msg = self.pre_msg_txt + '\n\n' + msg + '\n-------------\n'
self.async_send_message(msg)
except Exception as e:
warn("Could not deliver message. Error: " + str(e), RuntimeWarning)
def predict(self, device='cuda:0', pbar=None):
"""
Evaluate the model on a validation set
:param device: str (defaults to 'cuda:0')
:param pbar: fast_progress progress bar (defaults to None)
:returns: None
"""
self.model.to(device).load_state_dict(torch.load(self.model_path))
self.model.eval()
current_size = len(self.test_data_loader.dataset)
preds_dict = {
'y_true': np.zeros([current_size, 11]),
'y_pred': np.zeros([current_size, 11])
}
start_time = time.time()
with torch.no_grad():
index_dict = 0
for step, batch in enumerate(
progress_bar(self.test_data_loader,
parent=pbar,
leave=(pbar is not None))):
_, num_rows, y_pred, targets = self.model(batch, device)
current_index = index_dict
preds_dict['y_true'][current_index:current_index +
num_rows, :] = targets
preds_dict['y_pred'][current_index:current_index +
num_rows, :] = y_pred
index_dict += num_rows
y_true, y_pred = preds_dict['y_true'], preds_dict['y_pred']
str_stats = []
stats = [
f1_score(y_true, y_pred, average="macro"),
f1_score(y_true, y_pred, average="micro"),
jaccard_score(y_true, y_pred, average="samples")
]
for stat in stats:
str_stats.append('NA' if stat is None else str(stat) if isinstance(
stat, int) else f'{stat:.4f}')
str_stats.append(format_time(time.time() - start_time))
headers = ['F1-Macro', 'F1-Micro', 'JS', 'Time']
print(' '.join('{}: {}'.format(*k) for k in zip(headers, str_stats)))
def on_epoch_end(self, epoch: int, smooth_loss: Tensor,
last_metrics: MetricsList, **kwargs) -> bool:
"Add a line with `epoch` number, `smooth_loss` and `last_metrics`."
msg = ','.join(
self.learn.recorder.names[:(None if self.add_time else -1)]) + '\n'
last_metrics = ifnone(last_metrics, [])
stats = [
str(stat) if isinstance(stat, int) else
'#na#' if stat is None else f'{stat:.6f}' for name, stat in zip(
self.learn.recorder.names, [epoch, smooth_loss] + last_metrics)
]
if self.add_time: stats.append(format_time(time() - self.start_epoch))
str_stats = ','.join(stats)
msg = msg + str_stats + '\n'
try:
self.bot.send_message(chat_id=self.chat_id, text=msg)
except Exception as e:
warn("Could not deliver message. Error: " + str(e), RuntimeWarning)
def after_epoch(self):
if not getattr(self.learn,'inner_xla',False):
return # skip if not spawned
if 'recorder' not in self.learn.cbs.attrgot('name'):
all_metrics = {
'train_mets': L([]),
'valid_mets': L([]),
}
else:
all_metrics = {
'train_mets': self.recorder._train_mets,
'valid_mets': self.recorder._valid_mets,
}
# send metrics data to sync ranks across spawned processes
device = self.learn.xla_training.pdevice
packed_metrics = pack_metrics(all_metrics, device) # convert metrics to tensor list on TPU
reduced_metrics = xm.all_reduce(xm.REDUCE_SUM, packed_metrics)
xm.mark_step()
if xm.is_master_ordinal():
all_metrics = restore_metrics(reduced_metrics, all_metrics) # convert list to metric objects
for m in self.recorder._train_mets:
self.sync_log += _maybe_item(m)
for m in self.recorder._valid_mets:
self.sync_log += _maybe_item(m)
self.learn.final_record = self.sync_log[1:].copy()
del self.recorder.values[-1] # remove last entry added by recorder
self.recorder.values.append(self.learn.final_record) # add updated metrics
if self.recorder.add_time:
updated_time = (time.time() - self.recorder.start_epoch)
self.sync_log.append(format_time(updated_time))
self.recorder.log = self.sync_log
self._sync_stats_log(self.sync_log) # write_stats to output
self.learn.logger = self.orig_logger # restore orig logger after skipping recorder.logger(log)
def after_epoch(self):
"Store and log the loss/metric values"
self.values.append(self.log[1:].copy())
if self.add_time: self.log.append(format_time(time.time() - self.start_epoch))
self.logger(self.log)
def after_epoch(self):
stats = [str(self.epoch)]
for o in [self.train_stats, self.valid_stats]:
stats += [f'{v:.6f}' for v in o.avg_stats]
stats += [format_time(time.time() - self.start_time)]
self.logger(stats)
def fit(self,
num_epochs,
scheduler,
step_scheduler_on_batch,
loss_function,
optimizer,
device='cuda:0',
best_model=False):
"""
Fit the PyTorch model
Parameters
----------
num_epochs: int
number of epochs to train
scheduler: torch.optim.lr_scheduler
PyTorch learning rate scheduler
step_scheduler_on_batch: bool
flag of whether to step scheduler on batch (if True) or on epoch (if False)
loss_function: function
function to calculate loss with in model
optimizer: torch.optim
PyTorch optimzer
device: str (defaults to 'cuda:0')
device to run calculations on
best_model: bool (defaults to `False`)
flag to save best model from a single `fit` training run based on validation loss
The default is `False`, which will keep the final model from the training run.
`True` will keep the best model from the training run instead of the model
from the final epoch of the training cycle.
"""
self.model = self.model.to(device)
current_best_loss = np.iinfo(np.intp).max
pbar = master_bar(range(num_epochs))
headers = ['train_loss', 'val_loss']
for task in self.tasks:
headers.append(f'{task}_train_loss')
headers.append(f'{task}_val_loss')
headers.append(f'{task}_acc')
headers.append('time')
pbar.write(headers, table=True)
for epoch in pbar:
start_time = time.time()
self.model.train()
training_loss_dict = {task: 0.0 for task in self.tasks}
overall_training_loss = 0.0
for step, batch in enumerate(
progress_bar(self.train_dataloader, parent=pbar)):
task_type, (x, y) = batch
x = self._return_input_on_device(x, device)
y = y.type(torch.LongTensor)
y = y.to(device)
num_rows = self._get_num_rows(x)
output = self.model(x)
current_loss = loss_function(output[task_type], y)
scaled_loss = current_loss.item() * num_rows
training_loss_dict[task_type] += scaled_loss
overall_training_loss += scaled_loss
optimizer.zero_grad()
current_loss.backward()
optimizer.step()
if step_scheduler_on_batch:
scheduler.step()
if not step_scheduler_on_batch:
scheduler.step()
overall_training_loss = overall_training_loss / self.train_dataloader.total_samples
for task in self.tasks:
training_loss_dict[task] = (
training_loss_dict[task] /
len(self.train_dataloader.loader_dict[task].dataset))
overall_val_loss, val_loss_dict, accuracies = self.validate(
loss_function, device, pbar)
str_stats = []
stats = [overall_training_loss, overall_val_loss]
for stat in stats:
str_stats.append('NA' if stat is None else str(stat)
if isinstance(stat, int) else f'{stat:.6f}')
for task in self.tasks:
str_stats.append(f'{training_loss_dict[task]:.6f}')
str_stats.append(f'{val_loss_dict[task]:.6f}')
str_stats.append(f"{accuracies[task]['accuracy']:.6f}")
str_stats.append(format_time(time.time() - start_time))
pbar.write(str_stats, table=True)
if best_model and overall_val_loss < current_best_loss:
current_best_loss = overall_val_loss
best_model_wts = copy.deepcopy(self.model.state_dict())
best_model_epoch = epoch
if best_model:
self.model.load_state_dict(best_model_wts)
print(
f'Epoch {best_model_epoch} best model saved with loss of {current_best_loss}'
)
def fit(
self,
num_epochs,
scheduler,
step_scheduler_on_batch,
optimizer,
device='cuda:0',
best_model=False,
smooth_loss_alpha=0.2
):
"""
Fit the PyTorch model
Parameters
----------
num_epochs: int
number of epochs to train
scheduler: torch.optim.lr_scheduler
PyTorch learning rate scheduler
step_scheduler_on_batch: bool
flag of whether to step scheduler on batch (if True) or on epoch (if False)
optimizer: torch.optim
PyTorch optimzer
device: str (defaults to 'cuda:0')
device to run calculations on
best_model: bool (defaults to `False`)
flag to save best model from a single `fit` training run based on validation loss
The default is `False`, which will keep the final model from the training run.
`True` will keep the best model from the training run instead of the model
from the final epoch of the training cycle.
smooth_loss_alpha: float
Training loss values displayed during fitting and at the end of each epoch are
exponentially weighted moving averages over batches. Using an exponentially weighted
average over batches is a compromise between reporting the value from the most recent
batch, which is highly relevant but noisy, and reporting a simple average over batches,
which is more stable but reflects the value of the loss at the beginning of the epoch
as much as at the end. `smooth_loss_alpha` controls how much weight is given to the
current batch. It must be in the `(0, 1]` interval. Higher values are more like
reporting only the most recent batch, while lower values are more like giving all
batches equal weight, so this value controls the tradeoff between stability and
relevance.
"""
if not 0 < smooth_loss_alpha <= 1:
raise ValueError("`smooth_loss_alpha` must be in (0, 1]")
self.model = self.model.to(device)
current_best_loss = np.iinfo(np.intp).max
pbar = master_bar(range(num_epochs))
headers = ['train_loss', 'val_loss']
for task in self.tasks:
headers.append(f'{task}_train_loss')
headers.append(f'{task}_val_loss')
headers.append(f'{task}'+'_'+self.metric_function_dict[task].__name__)
headers.append('time')
pbar.write(headers, table=True)
self.smooth_training_loss_dict = {}
for epoch in pbar:
start_time = time.time()
self.model.train()
overall_training_loss = 0.0
subpbar = progress_bar(self.train_dataloader, parent=pbar)
for _, batch in enumerate(subpbar):
task_type, (x, y) = batch
x = self._return_input_on_device(x, device)
y = y.to(device)
num_rows = self._get_num_rows(x)
if num_rows == 1:
# skip batches of size 1
continue
output = self.model(x)
current_loss = self.loss_function_dict[task_type](output[task_type], y)
optimizer.zero_grad()
current_loss.backward()
optimizer.step()
if step_scheduler_on_batch:
scheduler.step()
self._update_smooth_training_loss_dict(
task_type, current_loss.item(), smooth_loss_alpha
)
subpbar.comment = self._report_smooth_training_loss()
overall_val_loss, val_loss_dict, metrics_scores = self.validate(
device,
pbar
)
if not step_scheduler_on_batch:
if isinstance(scheduler, torch.optim.lr_scheduler.ReduceLROnPlateau):
scheduler.step(overall_val_loss)
else:
scheduler.step()
overall_training_loss = self._calculate_overall_loss()
stats = [overall_training_loss, overall_val_loss]
str_stats = []
for stat in stats:
str_stats.append(
'NA' if stat is None else str(stat) if isinstance(stat, int) else f'{stat:.6f}'
)
for task in self.tasks:
str_stats.append(f'{self.smooth_training_loss_dict[task]:.6f}')
str_stats.append(f'{val_loss_dict[task]:.6f}')
str_stats.append(
f"{metrics_scores[task][self.metric_function_dict[task].__name__]:.6f}"
)
str_stats.append(format_time(time.time() - start_time))
pbar.write(str_stats, table=True)
if best_model and overall_val_loss < current_best_loss:
current_best_loss = overall_val_loss
best_model_wts = copy.deepcopy(self.model.state_dict())
best_model_epoch = epoch
if best_model:
self.model.load_state_dict(best_model_wts)
print(f'Epoch {best_model_epoch} best model saved with loss of {current_best_loss}')
def after_epoch(self):
stats = [str(self.epoch)]
for o in [self.train_stats, self.valid_stats]: stats += [f'{v:.6f}' for v in o.avg_stats]
stats += [format_time(time.time() - self.start_time)]
self.mbar.write(stats, table=True)