In [3]:
def models_plot(histories, a=None, b=None,
metric='loss',
plot_validation=True):
"""
Prints performance plot from a, to b on a list with object names in strings:
i.e., ["model00", "model01", "modelltsm00", ..]
Inputs:
histories: an array with dicts containing the metrics key
a: epoch start
b. last epoch
metric: plot this metric.
plot_validation: boolean indicating if validation data should be plotted.
a: from this epoch
b: to this epoch
"""
"""
# Plot loss
plt.subplot(imgrows, 2, 1)
plt.title('Loss')
plt.plot(history['loss'][a:b], label='Training', linewidth=2)
if plot_validation:
plt.plot(history['val_loss'][a:b], label='Validation', linewidth=2)
plt.legend()
plt.xlabel('Epoch')
plt.ylabel(f'Loss')
### quantiles
plt.xticks(ticks=quantiles-a,
labels=quantiles)
plt.grid(True)
# Plot accuracy
"""
models = [globals()[h] for h in histories]
max_epochs = np.max([len(o[metric]) for o in models])
if a is None:
a = 0
if b is None:
b = max_epochs
a = np.min((a,b))
b = np.max((a,b))
#print(f"a={a}, b={b}")
imgrows = (len(histories)) / 2.
imgrows = np.round(imgrows + .1, 0)
imgrows = int(imgrows) #+ 1
#print(f"length={len(histories)}")
#print(f"imgrows={imgrows}")
# x ticks
quantiles = np.quantile(range(a, b),
[.2, .4, .6, .8]).round(0).astype(int)
quantiles = np.insert(quantiles, 0, [a])
quantiles += 1
quantiles = np.append(quantiles, [b-1])
if plot_validation:
model_order = {h: np.mean(v["val_" + metric]) for h, v in zip(histories, models)}
else:
model_order = {h: np.mean(v[metric]) for h, v in zip(histories, models)}
model_order = {m: model_order[m] for m in sorted(model_order,
key=model_order.get)}
# init the plot
plt.figure(figsize=(14., 4.*imgrows), frameon=False, tight_layout=True)
plt.suptitle(metric, va='top', weight='bold', size='x-large')
# create plot for every model
#for i, (model_name, model) in enumerate(zip(histories, models)):
for i, (model_name) in enumerate(model_order.keys()):
model = globals()[model_name]
plt.subplot(imgrows, 2, i+1)
plt.title(model_name)
if plot_validation:
mean_label="Mean Validation"
else:
mean_label="Mean Training"
plt.hlines(y=model_order[model_name], xmin=0, linestyles="dashed",
xmax=b-a, label=mean_label, color="green")
plt.plot(model[metric][a:b], label='Training',
linewidth=2)
if plot_validation:
plt.plot(model["val_" + metric][a:b],
label='Validation', linewidth=2)
plt.legend()
plt.xlabel('Epoch')
plt.ylabel(metric)
plt.xticks(ticks=quantiles-a,
labels=quantiles)
plt.grid(True)
plt.show()
model_histories = [o
for o in object_names
if o.endswith("_hist")
]
model_metrics = [k
for k in globals()[model_histories[0]].keys()
if re.search("^(val_|loss)", k) is None
]
for model_metric in model_metrics:
models_plot(model_histories, metric=model_metric)
