def before_fit(self, **kwargs):
self.train_pca = ifnone(
self.train_pca,
self.do_pca(get_xy(self.learn.dls)[0], do_train=True))
self.train_trace = scatter(self.train_pca,
name='Training data',
mode='markers',
marker_color='#539dcc',
marker_size=1.5 if self.is_3d else 4)
self.weight_pca = self.do_pca(self.learn.model.weights)
self.weight_trace = scatter(self.weight_pca,
name='SOM weights',
mode='markers',
marker_color='#e58368',
marker_size=3 if self.is_3d else 6)
expl_var = str(
tuple(
map(lambda pct: f'{pct:.0f}%',
self.pca.explained_variance_ratio_ * 100)))[1:-1]
layout = go.Layout(
title=f"SOM Visualization ({expl_var} explained variance)")
self.fig = go.FigureWidget([self.train_trace, self.weight_trace],
layout=layout)
self.fig.update_layout(margin=dict(l=20, r=20, t=20, b=20),
paper_bgcolor="LightSteelBlue")
def batch_stats(self, funcs=None):
"Grab a batch of data and call reduction function `func` per channel"
funcs = ifnone(funcs, [torch.mean, torch.std])
ds_type = DatasetType.Valid if self.valid_dl else DatasetType.Train
x = self.one_batch(ds_type=ds_type, denorm=False)[0].cpu()
def multi_channel_view(x):
return x.transpose(3, 0).contiguous().view(x.shape[3], -1)
return [func(multi_channel_view(x), 1) for func in funcs]
def on_epoch_end(self, epoch: int, smooth_loss: Tensor,
last_metrics: MetricsList, **kwargs: Any) -> bool:
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)
]
SendData.sendMessage(key=self.key,
auth_token=self.auth_token,
params=stats,
ModelName=self.ModelName)
def before_fit(self):
self.train_pca = ifnone(
self.train_pca,
self.do_pca(get_xy(self.learn.dls)[0], do_train=True))
self.weight_pca = self.do_pca(self.learn.model.weights)
plt.ion()
self.fig = plt.figure()
self.train_axis = self.fig.add_subplot(
111, projection='3d' if self.is_3d else None)
# Draw weights
self.weight_scatter = self.train_axis.scatter(*tuple(
[self.weight_pca[:, i] for i in range(self.weight_pca.shape[-1])]),
c="#e58368",
zorder=100)
# Draw training data (once)
self.train_axis.scatter(*tuple(
[self.train_pca[:, i] for i in range(self.train_pca.shape[-1])]),
c="#539dcc")
def on_epoch_end(self, epoch: int, smooth_loss: Tensor,
last_metrics: MetricsList, **kwargs: Any) -> bool:
if (time.time() - self.start_time > 3000):
self.start_time = time.time()
self.key, self.auth_token = SendData.signin(email=self.email,
password=self.password)
last_metrics = ifnone(last_metrics, [])
self.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)
]
SendData.sendMessage(key=self.key,
auth_token=self.auth_token,
params=self.stats,
ModelName=self.ModelName)