File size: 8,008 Bytes
d6c2e08 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 |
import tensorflow as tf
from scipy.stats import ttest_rel
from tqdm import tqdm
import numpy as np
from attention_dynamic_model import AttentionDynamicModel
from attention_dynamic_model import set_decode_type
from utils import generate_data_onfly
def copy_of_tf_model(model, embedding_dim=128, graph_size=20):
"""Copy model weights to new model
"""
# https://stackoverflow.com/questions/56841736/how-to-copy-a-network-in-tensorflow-2-0
CAPACITIES = {10: 20.,
20: 30.,
50: 40.,
100: 50.
}
data_random = [tf.random.uniform((2, 2,), minval=0, maxval=1, dtype=tf.dtypes.float32),
tf.random.uniform((2, graph_size, 2), minval=0, maxval=1, dtype=tf.dtypes.float32),
tf.cast(tf.random.uniform(minval=1, maxval=10, shape=(2, graph_size),
dtype=tf.int32), tf.float32) / tf.cast(CAPACITIES[graph_size], tf.float32)]
new_model = AttentionDynamicModel(embedding_dim)
set_decode_type(new_model, "sampling")
_, _ = new_model(data_random)
for a, b in zip(new_model.variables, model.variables):
a.assign(b)
return new_model
def rollout(model, dataset, batch_size = 1000, disable_tqdm = False):
# Evaluate model in greedy mode
set_decode_type(model, "greedy")
costs_list = []
for batch in tqdm(dataset.batch(batch_size), disable=disable_tqdm, desc="Rollout greedy execution"):
cost, _ = model(batch)
costs_list.append(cost)
return tf.concat(costs_list, axis=0)
def validate(dataset, model, batch_size=1000):
"""Validates model on given dataset in greedy mode
"""
val_costs = rollout(model, dataset, batch_size=batch_size)
set_decode_type(model, "sampling")
mean_cost = tf.reduce_mean(val_costs)
print(f"Validation score: {np.round(mean_cost, 4)}")
return mean_cost
class RolloutBaseline:
def __init__(self, model, filename,
from_checkpoint=False,
path_to_checkpoint=None,
wp_n_epochs=1,
epoch=0,
num_samples=10000,
warmup_exp_beta=0.8,
embedding_dim=128,
graph_size=20
):
"""
Args:
model: current model
filename: suffix for baseline checkpoint filename
from_checkpoint: start from checkpoint flag
path_to_checkpoint: path to baseline model weights
wp_n_epochs: number of warm-up epochs
epoch: current epoch number
num_samples: number of samples to be generated for baseline dataset
warmup_exp_beta: warmup mixing parameter (exp. moving average parameter)
"""
self.num_samples = num_samples
self.cur_epoch = epoch
self.wp_n_epochs = wp_n_epochs
self.beta = warmup_exp_beta
# controls the amount of warmup
self.alpha = 0.0
self.running_average_cost = None
# Checkpoint params
self.filename = filename
self.from_checkpoint = from_checkpoint
self.path_to_checkpoint = path_to_checkpoint
# Problem params
self.embedding_dim = embedding_dim
self.graph_size = graph_size
# create and evaluate initial baseline
self._update_baseline(model, epoch)
def _update_baseline(self, model, epoch):
# Load or copy baseline model based on self.from_checkpoint condition
if self.from_checkpoint and self.alpha == 0:
print('Baseline model loaded')
self.model = load_tf_model(self.path_to_checkpoint,
embedding_dim=self.embedding_dim,
graph_size=self.graph_size)
else:
self.model = copy_of_tf_model(model,
embedding_dim=self.embedding_dim,
graph_size=self.graph_size)
# For checkpoint
self.model.save_weights('baseline_checkpoint_epoch_{}_{}.h5'.format(epoch, self.filename), save_format='h5')
# We generate a new dataset for baseline model on each baseline update to prevent possible overfitting
self.dataset = generate_data_onfly(num_samples=self.num_samples, graph_size=self.graph_size)
print(f"Evaluating baseline model on baseline dataset (epoch = {epoch})")
self.bl_vals = rollout(self.model, self.dataset)
self.mean = tf.reduce_mean(self.bl_vals)
self.cur_epoch = epoch
def ema_eval(self, cost):
"""This is running average of cost through previous batches (only for warm-up epochs)
"""
if self.running_average_cost is None:
self.running_average_cost = tf.reduce_mean(cost)
else:
self.running_average_cost = self.beta * self.running_average_cost + (1. - self.beta) * tf.reduce_mean(cost)
return self.running_average_cost
def eval(self, batch, cost):
"""Evaluates current baseline model on single training batch
"""
if self.alpha == 0:
return self.ema_eval(cost)
if self.alpha < 1:
v_ema = self.ema_eval(cost)
else:
v_ema = 0.0
v_b, _ = self.model(batch)
v_b = tf.stop_gradient(v_b)
v_ema = tf.stop_gradient(v_ema)
# Combination of baseline cost and exp. moving average cost
return self.alpha * v_b + (1 - self.alpha) * v_ema
def eval_all(self, dataset):
"""Evaluates current baseline model on the whole dataset only for non warm-up epochs
"""
if self.alpha < 1:
return None
val_costs = rollout(self.model, dataset, batch_size=2048)
return val_costs
def epoch_callback(self, model, epoch):
"""Compares current baseline model with the training model and updates baseline if it is improved
"""
self.cur_epoch = epoch
print(f"Evaluating candidate model on baseline dataset (callback epoch = {self.cur_epoch})")
candidate_vals = rollout(model, self.dataset) # costs for training model on baseline dataset
candidate_mean = tf.reduce_mean(candidate_vals)
diff = candidate_mean - self.mean
print(f"Epoch {self.cur_epoch} candidate mean {candidate_mean}, baseline epoch {self.cur_epoch} mean {self.mean}, difference {diff}")
if diff < 0:
# statistic + p-value
t, p = ttest_rel(candidate_vals, self.bl_vals)
p_val = p / 2
print(f"p-value: {p_val}")
if p_val < 0.05:
print('Update baseline')
self._update_baseline(model, self.cur_epoch)
# alpha controls the amount of warmup
if self.alpha < 1.0:
self.alpha = (self.cur_epoch + 1) / float(self.wp_n_epochs)
print(f"alpha was updated to {self.alpha}")
def load_tf_model(path, embedding_dim=128, graph_size=20, n_encode_layers=2):
"""Load model weights from hd5 file
"""
# https://stackoverflow.com/questions/51806852/cant-save-custom-subclassed-model
CAPACITIES = {10: 20.,
20: 30.,
50: 40.,
100: 50.
}
data_random = [tf.random.uniform((2, 2,), minval=0, maxval=1, dtype=tf.dtypes.float32),
tf.random.uniform((2, graph_size, 2), minval=0, maxval=1, dtype=tf.dtypes.float32),
tf.cast(tf.random.uniform(minval=1, maxval=10, shape=(2, graph_size),
dtype=tf.int32), tf.float32) / tf.cast(CAPACITIES[graph_size], tf.float32)]
model_loaded = AttentionDynamicModel(embedding_dim,n_encode_layers=n_encode_layers)
set_decode_type(model_loaded, "greedy")
_, _ = model_loaded(data_random)
model_loaded.load_weights(path)
return model_loaded
|