def visualize_context_vector(model: Model,
dataset_id,
dataset_prefix,
cutoff=None,
limit=None,
normalize_timeseries=False,
visualize_sequence=True,
visualize_classwise=False):
X_train, y_train, X_test, y_test, is_timeseries = load_dataset_at(
dataset_id, normalize_timeseries=normalize_timeseries)
_, sequence_length = calculate_dataset_metrics(X_train)
if sequence_length != MAX_SEQUENCE_LENGTH_LIST[dataset_id]:
if cutoff is None:
choice = cutoff_choice(dataset_id, sequence_length)
else:
assert cutoff in [
'pre', 'post'
], 'Cutoff parameter value must be either "pre" or "post"'
choice = cutoff
if choice not in ['pre', 'post']:
return
else:
X_train, X_test = cutoff_sequence(X_train, X_test, choice,
dataset_id, sequence_length)
attn_lstm_layer = [(i, layer) for (i, layer) in enumerate(model.layers)
if layer.__class__.__name__ == 'AttentionLSTM']
if len(attn_lstm_layer) == 0:
raise AttributeError('Provided model does not have an Attention layer')
else:
i, attn_lstm_layer = attn_lstm_layer[
0] # use first attention lstm layer only
attn_lstm_layer.return_attention = True
model.layers[i] = attn_lstm_layer
model.load_weights("./weights/%s_weights.h5" % dataset_prefix)
attention_output = model.layers[i].call(model.input)
eval_functions = build_function(model,
attn_lstm_layer.name,
outputs=[attention_output])
attention_vectors = []
for i in range(X_train.shape[0]):
activations = get_outputs(model,
X_train[i, :, :][np.newaxis, ...],
eval_functions,
verbose=False)[0]
attention_vector = np.sum(activations, axis=1).squeeze()
attention_vectors.append(attention_vector)
attention_vectors = np.array(attention_vectors)
attention_vector_final = np.mean(attention_vectors, axis=0)
if visualize_sequence:
# plot input sequence part that is paid attention too in detail
attention_vector_final = attention_vector_final.reshape(
(1, attention_vector_final.shape[0]))
X_train_attention = np.zeros_like(X_train)
X_test_attention = np.zeros_like(X_test)
for i in range(X_train.shape[0]):
X_train_attention[
i, :, :] = attention_vector_final * X_train[i, :, :]
for i in range(X_test.shape[0]):
X_test_attention[
i, :, :] = attention_vector_final * X_test[i, :, :]
plot_dataset(dataset_id,
seed=1,
limit=limit,
cutoff=cutoff,
normalize_timeseries=normalize_timeseries,
plot_data=(X_train, y_train, X_test, y_test,
X_train_attention, X_test_attention),
type='Context',
plot_classwise=visualize_classwise)
else:
# plot only attention chart
train_df = pd.DataFrame({'attention (%)': attention_vector_final},
index=range(attention_vector_final.shape[0]))
train_df.plot(kind='bar',
title='Attention Mechanism (Train) as '
'a function of input'
' dimensions.')
plt.show()
def visualize_context_vector(model: Model,
series_values,
labels,
run_prefix,
cutoff=None,
limit=None,
val_split=1 / 3,
random_state=0,
visualize_sequence=True,
visualize_classwise=False):
"""
Visualize the Context Vector of the Attention LSTM.
Args:
model: an Attention LSTM-FCN Model.
dataset_id: Integer id representing the dataset index containd in
`utils/constants.py`.
dataset_prefix: Name of the dataset. Used for weight saving.
batch_size: Size of each batch for evaluation.
test_data_subset: Optional integer id to subset the test set. To be used if
the test set evaluation time is significantly.
cutoff: Optional integer which slices of the first `cutoff` timesteps
from the input signal.
limit: Number of samples to be visualized in one plot.
normalize_timeseries: Bool / Integer. Determines whether to normalize
the timeseries.
If False, does not normalize the time series.
If True / int not equal to 2, performs standard sample-wise
z-normalization.
If 2: Performs full dataset z-normalization.
visualize_sequence: Bool flag, whetehr to visualize the sequence attended to
by the Context Vector or just the Context Vector itself.
visualize_classwise: Bool flag. Wheter to visualize the samples
seperated by class. When doing so, `limit` is multiplied by
the number of classes so it is better to set `limit` to 1 in
such cases.
"""
inds = np.arange(series_values.shape[0])
np.random.seed(random_state)
np.random.shuffle(inds)
series_values = series_values[inds]
labels = labels[inds]
val_split = int(val_split * series_values.shape[0])
X_train, y_train = series_values[:-val_split], labels[:-val_split]
X_test, y_test = series_values[-val_split:], labels[-val_split:]
sequence_length = series_values.shape[1]
attn_lstm_layer = [(i, layer) for (i, layer) in enumerate(model.layers)
if layer.__class__.__name__ == 'AttentionLSTM']
if len(attn_lstm_layer) == 0:
raise AttributeError('Provided model does not have an Attention layer')
else:
i, attn_lstm_layer = attn_lstm_layer[
0] # use first attention lstm layer only
attn_lstm_layer.return_attention = True
model.layers[i] = attn_lstm_layer
model.load_weights("./weights/%s_weights.h5" % run_prefix)
attention_output = model.layers[i].call(model.input)
eval_functions = build_function(model,
attn_lstm_layer.name,
outputs=[attention_output])
train_attention_vectors = []
test_attention_vectors = []
output_shape = [X_train.shape[-1], 1, 1]
for i in range(X_train.shape[0]):
activations = get_outputs(model,
X_train[i, :, :][np.newaxis, ...],
eval_functions,
verbose=False)[0]
# print("activations", activations.shape)
attention_vector = activations.reshape((-1, 1, 1))
attention_vector = (attention_vector - attention_vector.min()) / (
attention_vector.max() - attention_vector.min())
attention_vector = (attention_vector * 2.) - 1.
attention_vector = resize(attention_vector,
output_shape,
mode='reflect',
anti_aliasing=True)
attention_vector = attention_vector.reshape([1, -1])
train_attention_vectors.append(attention_vector)
for i in range(X_test.shape[0]):
activations = get_outputs(model,
X_test[i, :, :][np.newaxis, ...],
eval_functions,
verbose=False)[0]
# print("activations", activations.shape)
attention_vector = activations.reshape((-1, 1, 1))
attention_vector = (attention_vector - attention_vector.min()) / (
attention_vector.max() - attention_vector.min())
attention_vector = (attention_vector * 2.) - 1.
attention_vector = resize(attention_vector,
output_shape,
mode='reflect',
anti_aliasing=True)
attention_vector = attention_vector.reshape([1, -1])
test_attention_vectors.append(attention_vector)
train_attention_vectors = np.array(train_attention_vectors)
test_attention_vectors = np.array(test_attention_vectors)
print("Train Attention Vectors Shape :", train_attention_vectors.shape)
print("Test Attentin Vectors Shape :", test_attention_vectors.shape)
if visualize_sequence:
# plot input sequence part that is paid attention too in detail
X_train_attention = train_attention_vectors * X_train
X_test_attention = test_attention_vectors * X_test
plot_dataset(series_values,
labels,
run_prefix,
val_split=val_split,
seed=1,
limit=limit,
cutoff=cutoff,
plot_data=(X_train, y_train, X_test, y_test,
X_train_attention, X_test_attention),
type='Context',
plot_classwise=visualize_classwise)
else:
# plot only attention chart
choice = np.random.randint(0, train_attention_vectors.shape[0])
train_df = pd.DataFrame(
{'attention (%)': train_attention_vectors[choice, 0]},
index=range(train_attention_vectors.shape[-1]))
train_df.plot(kind='bar',
title='Attention Mechanism (Train) as '
'a function of input'
' dimensions. Class = %d' % (y_train[choice]))
plt.show()