File size: 4,208 Bytes
0240c6e |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 |
import tensorflow as tf
import tf_utils as tfu
from graph.base import GraphPartBase
class MainCls(GraphPartBase):
def __init__(self, for_usage, global_settings, current_settings, optimiser, reset_optimiser):
super().__init__(for_usage, global_settings, current_settings, optimiser, reset_optimiser, 'main', ['Loss', 'Accuracy'])
self.checks = []
self.weights = []
self.keep_drops = []
self.losses = []
self.probs = []
self.results = []
self.ys = []
self.drops = []
self.top_k = global_settings['main_class_k']
def __build_graph_for_device__(self, x, seq_len, gram_probs, gram_drop):
self.xs.append(x)
self.x_seq_lens.append(seq_len)
self.drops.append(gram_drop)
y = tf.placeholder(dtype=tf.int32, shape=(None, self.main_classes_count), name='Y')
weights = tf.placeholder(dtype=tf.float32, shape=(None,), name='Weight')
x_emd_init = tf.random_normal((self.chars_count, self.settings['char_vector_size']))
x_emb = tf.get_variable("Embeddings", initializer=x_emd_init)
rnn_input = tf.nn.embedding_lookup(x_emb, x)
if self.for_usage:
cls_keep_drop = tf.constant(1, dtype=tf.float32, name='KeepDrop')
else:
cls_keep_drop = tf.placeholder(dtype=tf.float32, name='KeepDrop')
self.keep_drops.append(cls_keep_drop)
init_state = tf.concat(gram_probs, 1)
with tf.variable_scope("InitRnnState", reuse=tf.AUTO_REUSE) as scope:
rez_size = self.settings['rnn_state_size']
w_softmax = tf.get_variable("W", (init_state.shape[1], rez_size))
b_softmax = tf.get_variable("b", [rez_size])
init_state = tf.matmul(init_state, w_softmax) + b_softmax
rnn_logits = tfu.build_rnn(rnn_input,
cls_keep_drop,
seq_len,
self.settings,
init_state,
init_state,
self.for_usage)
logits = tfu.rnn_top('RnnTop',
rnn_logits,
self.settings,
self.main_classes_count)
float_y = tf.cast(y, tf.float32)
errors = tf.nn.sigmoid_cross_entropy_with_logits(logits=logits,
labels=float_y)
errors = tf.reduce_sum(errors, axis=1)
errors = errors * weights
if not self.for_usage:
self.checks.append(tf.check_numerics(errors, "ErrorNullCheck"))
probs = tf.nn.softmax(logits)
result = tf.math.top_k(probs, self.top_k, name="Results")
loss = tf.reduce_sum(errors)
if not self.for_usage:
self.checks.append(tf.check_numerics(errors, "LossNullCheck"))
vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=self.main_scope_name)
grads = self.optimiser.compute_gradients(loss, var_list=vars)
self.losses.append(loss)
self.probs.append(probs)
self.results.append(result)
self.ys.append(y)
self.dev_grads.append(grads)
self.weights.append(weights)
# metrics
self.create_mean_metric(0, loss)
labels = tf.math.argmax(y, axis=1)
predictions = tf.math.argmax(probs, axis=1)
self.create_accuracy_metric(1, labels, predictions)
def __update_feed_dict__(self, op_name, feed_dict, batch, dev_num):
for gram_drop in self.drops[dev_num]:
feed_dict[gram_drop] = 1
feed_dict[self.keep_drops[dev_num]] = 1 if op_name == 'test' else self.settings['keep_drop']
feed_dict[self.ys[dev_num]] = batch['y']
feed_dict[self.weights[dev_num]] = batch['weight']
def __load_dataset__(self, operation_name):
return list(
tfu.load_cls_dataset(
self.dataset_path,
self.devices_count,
operation_name,
self.settings['train_batch_size'],
self.settings['use_weights']
)
)
|