def __init__(self, config, models, tensor_dict=None):
self.config = config
self.model = models[0]
self.models = models
self.global_step = models[0].global_step
self.yp = models[0].yp
self.yp2 = models[0].yp2
self.yp_list = models[0].decoder_inference
self.yp_mat = models[0].decoder_train_softmax
self.loss = models[0].loss
self.tensor_dict = {} if tensor_dict is None else tensor_dict
self.y = models[0].y
#word2index = config.word2idx.copy()
#word2index.update(config.new_word2idx)
#self.index2word = {v: k for k, v in word2index.items()}
with tf.name_scope("eval_concat"):
N, M, JX = config.batch_size, config.max_num_sents, config.max_sent_size
self.yp = tf.concat(
[padded_reshape(model.yp, [N, M, JX]) for model in models], 0)
self.yp2 = tf.concat(
[padded_reshape(model.yp2, [N, M, JX]) for model in models], 0)
self.loss = tf.add_n([model.loss
for model in models]) / len(models)
def __init__(self, config, models, tensor_dict=None):
super(MultiGPUF1Evaluator, self).__init__(config,
models[0],
tensor_dict=tensor_dict)
self.models = models
with tf.name_scope("eval_concat"):
if config.split_supports == True:
N, M, JX = config.batch_size, 1, config.max_para_size
else:
N, M, JX = config.batch_size, config.max_num_sents, config.max_sent_size
self.yp = tf.concat(axis=0,
values=[
padded_reshape(model.yp, [N, M, JX])
for model in models
])
self.yp2 = tf.concat(axis=0,
values=[
padded_reshape(model.yp2, [N, M, JX])
for model in models
])
self.wy = tf.concat(axis=0,
values=[
padded_reshape(model.wy, [N, M, JX])
for model in models
])
self.loss = tf.add_n([model.get_loss()
for model in models]) / len(models)
def __init__(self, config, models, tensor_dict=None):
super(MultiGPUF1Evaluator, self).__init__(config, models[0], tensor_dict=tensor_dict)
self.models = models
with tf.name_scope("eval_concat"):
N, M, JX = config.batch_size, config.max_num_sents, config.max_sent_size
self.yp = tf.concat(0, [padded_reshape(model.yp, [N, M, JX]) for model in models])
self.yp2 = tf.concat(0, [padded_reshape(model.yp2, [N, M, JX]) for model in models])
self.loss = tf.add_n([model.loss for model in models])/len(models)
def __init__(self, config, models, tensor_dict=None):
super(MultiGPUF1Evaluator, self).__init__(config, models[0], tensor_dict=tensor_dict)
self.models = models
with tf.name_scope("eval_concat"):
N, M, JX = config.batch_size, config.max_num_sents, config.max_sent_size
self.yp = tf.concat(0, [padded_reshape(model.yp, [N, M, JX]) for model in models])
self.yp2 = tf.concat(0, [padded_reshape(model.yp2, [N, M, JX]) for model in models])
self.loss = tf.add_n([model.loss for model in models])/len(models)
def __init__(self, config, models, tensor_dict=None):
super(MultiGPUClassificationAccuracyEvaluator, self).__init__(config, models[0], tensor_dict=tensor_dict)
self.models = models
with tf.name_scope("eval_concat"):
N, M, JX = config.batch_size, config.max_num_sents, config.max_sent_size
C = 3 if config.data_dir.startswith('data/snli') else 2
each_yp = lambda model: padded_reshape(model.yp0, [N, M, C])
self.yp = tf.concat(axis=0, values=[each_yp(model) for model in models])
self.loss = tf.add_n([model.loss for model in models])/len(models)
self.p = tf.concat(axis=0, values=[model.tensor_dict['a_u'] for model in models])
def __init__(self, config, models, tensor_dict=None):
super(MultiGPUF1Evaluator, self).__init__(config,
models[0],
tensor_dict=tensor_dict)
self.models = models
with tf.name_scope("eval_concat"):
N, M, JX = config.batch_size, config.max_num_sents, config.max_sent_size
self.yp = tf.concat(axis=0,
values=[
padded_reshape(model.yp, [N, M, JX])
for model in models
])
self.yp2 = tf.concat(axis=0,
values=[
padded_reshape(model.yp2, [N, M, JX])
for model in models
])
self.wy = tf.concat(axis=0,
values=[
padded_reshape(model.wy, [N, M, JX])
for model in models
])
self.loss = tf.add_n([model.get_loss()
for model in models]) / len(models)
if config.dataset == 'hotpotqa':
self.yp3 = tf.concat(axis=0,
values=[
padded_reshape(model.yp3, [N, 1, 2])
for model in models
])
self.yp3_yesno = tf.concat(axis=0,
values=[
padded_reshape(
model.yp3_yesno, [N, 1, 2])
for model in models
])
def __init__(self, config, models, tensor_dict=None):
super(MultiGPUF1CandidateDocSelEvaluator,
self).__init__(config, models[0], tensor_dict=tensor_dict)
self.models = models
with tf.name_scope("eval_concat"):
if config.split_supports == True:
N, M, JX = config.batch_size, 1, tf.reduce_max(
[tf.shape(model.yp)[2] for model in models])
else:
N, M, JX = config.batch_size, config.max_num_sents, tf.reduce_max(
[tf.shape(model.yp)[2] for model in models])
self.yp = tf.concat(axis=0,
values=[
padded_reshape(model.yp, [N, M, JX])
for model in models
])
self.loss = tf.add_n([model.get_loss()
for model in models]) / len(models)
