def save_model_tfjs(model: ks.models.Model, save_loc: str) -> None:
word_pred_model = ks.models.Model(
inputs=model.inputs,
outputs=[model.get_layer('word_preds').output],
)
word_pred_model.compile(optimizer='sgd', loss='mse')
tfjs.converters.save_keras_model(
word_pred_model,
save_loc,
)
def get_excitations(model: ks.models.Model, num_sentences: int,
dataset: Dataset, save_loc: str) -> None:
conv_len = model.get_layer('convs').get_weights()[0].shape[0]
conv_output_model = ks.models.Model(
inputs=model.inputs,
outputs=[model.get_layer('convs').output],
)
preds = conv_output_model.predict(dataset.x_train[:num_sentences])
def get_preds(idxs: Matrix, preds: Matrix):
num_padding = np.sum(idxs == 0)
idxs, preds = idxs[num_padding:], preds[num_padding + 1:]
words = dataset.decode(idxs)
return {
'words': words,
'preds': [[float(j) for j in i] for i in preds.T],
}
excitations = {
'conv_len':
conv_len,
'activations': [
get_preds(i, p)
for i, p in zip(dataset.x_train[:num_sentences], preds)
],
}
utils.save_json(excitations, save_loc, 'excitations.json')
with open(os.path.join(save_loc, 'excitations_show.txt'), 'w') as f:
for example in excitations['activations']:
for i in range(3):
ordered_sentences = get_ordered_sentences(
example['words'],
example['preds'][i],
conv_len,
)
f.write('Layer {}\n'.format(i))
f.write('\n'.join('{}: {}'.format(*g)
for g in ordered_sentences))
f.write('\n\n')