def _test_save_load(self, attention):
inputs = keras.layers.Input(shape=(None,), name='Input')
embd = keras.layers.Embedding(input_dim=3,
output_dim=5,
mask_zero=True,
name='Embedding')(inputs)
lstm = keras.layers.Bidirectional(keras.layers.LSTM(units=7,
return_sequences=True),
name='Bi-LSTM')(embd)
if attention.return_attention:
layer, weights = attention(lstm)
else:
layer = attention(lstm)
dense = keras.layers.Dense(units=2, activation='softmax', name='Softmax')(layer)
loss = {'Softmax': 'sparse_categorical_crossentropy'}
if attention.return_attention:
outputs = [dense, weights]
loss[attention.name] = 'mse'
else:
outputs = dense
model = keras.models.Model(inputs=inputs, outputs=outputs)
model.compile(optimizer='adam', loss=loss)
model_path = os.path.join(tempfile.gettempdir(), 'keras_weighted_att_test_sl_%f.h5' % np.random.random())
model.save(model_path)
model = keras.models.load_model(model_path, custom_objects=Attention.get_custom_objects())
model.summary(line_length=100)
if attention.return_attention:
self.assertEqual(2, len(model.outputs))
else:
self.assertEqual(1, len(model.outputs))
def get_model(args):
model = None
modelname = args.model_id
# workaround for getting vis for attention
if modelname == "model_folk100k_melody_2lstm32_attention":
# (100000, 64, 58)
model = att_model(32, False, 2, True)
elif modelname == "model_folk100k_melody_bi3lstm64_attention":
model = att_model(64, True, 3, True)
else:
json_model = open(os.path.join(modelname, "model.json"), "r").read()
model = keras.models.model_from_json(
json_model,
custom_objects=SeqWeightedAttention.get_custom_objects())
model.load_weights(os.path.join(modelname, "model.h5"))
print(model.summary(line_length=100))
return model
def get_model(args, dshape):
model_dir = get_model_dir(args)
model = None
loss = 'categorical_crossentropy'
optimizer = keras.optimizers.Adam(lr=0.005)
if args.new:
print('generating NEW model...')
model = new_architecture(dshape[1] - 1, dshape[2], args.layers,
args.bi, args.att, args.cells)
# copy arch to folder
shutil.copy('architecture.py', model_dir)
model_json = model.to_json()
model_json_path = os.path.join(model_dir, "model.json")
print('storing model json in %s' % model_json_path)
with open(model_json_path, "w") as json_file:
json_file.write(model_json)
# delete epoch counters
delete_epoch_counters(model_dir)
model.compile(loss=loss, optimizer=optimizer)
else:
print('using existing model...')
model_json_path = os.path.join(model_dir, "model.json")
model = keras.models.model_from_json(
open(model_json_path, "r").read(),
custom_objects=SeqWeightedAttention.get_custom_objects())
model_weights_path = os.path.join(model_dir, "model.h5")
print('loading existing weights from %s...' % model_weights_path)
model.load_weights(model_weights_path)
model.compile(loss=loss, optimizer=optimizer)
print(model.summary())
return model, model_dir
:param text_list:
:param token_dict:
:return:
"""
X1 = []
X2 = []
tokenizer = Tokenizer(token_dict)
for line in text_list:
x1, x2 = tokenizer.encode(first=line)
X1.append(x1)
X2.append(x2)
X1 = sequence.pad_sequences(X1, maxlen=maxlen, padding='post', truncating='post')
X2 = sequence.pad_sequences(X2, maxlen=maxlen, padding="post", truncating='post')
return [X1, X2]
if __name__=="__main__":
maxlen=100
text_list=["TW 0:02 / 41:54 Mind Your Language Season 3 Episode 2 Who Loves Ya Baby? | Funny TV Show (GM)","I have a dream"]
token_dict=get_token_dict(dict_path)
[X1,X2]=get_encode(text_list,token_dict)
print(X1)
wordvec=build_bert_model(X1,X2)
print(wordvec)
#with open("test_keras_bert2.yml","r") as f:
# yaml_string=yaml.load(f)
#model=keras.models.load_model(yaml_string,custom_objects=SeqSelfAttention.get_custom_objects())
print("loading weights")
model=keras.models.load_model("test_keras_bert4.h5",custom_objects=SeqWeightedAttention.get_custom_objects())
result=model.predict(wordvec)
print(result)
del model