def save_to_file(self):
save_dir = os.path.join(self.saved_models_dir, self.name, self.tag,
str(self.training_steps))
# Allow overwriting an existing save because we don't lose any data here
tools.rmdir_if_exists(save_dir)
os.makedirs(save_dir)
file_path = os.path.join(save_dir, 'data')
logging.info('Saving model to file %s' % file_path)
extra = {
'encoding_name': self.encoding_name,
'encoding_args': self.encoding_args,
'encoding_kwargs': self.encoding_kwargs,
'state_sizes': self.state_sizes,
'training_steps': self.training_steps,
'round_steps': self.round_steps
}
extra_file_path = file_path + '.pkl'
with open(extra_file_path, 'xb') as f:
pickle.dump(extra, f)
self.saver.save(self.sess, file_path)
logging.info('Saved model to file %s' % file_path)
def init_from_file(self, from_name=None, training_steps=None):
if from_name is None:
from_name = self.name
training_steps_dir = os.path.join(self.saved_models_dir, from_name,
self.tag)
if training_steps is None:
training_steps = tools.get_latest_in_dir(training_steps_dir,
key=int)
save_dir = os.path.join(training_steps_dir, str(training_steps))
file_path = os.path.join(save_dir, 'data')
logging.info('Loading model from file %s' % file_path)
extra_file_path = file_path + '.pkl'
with open(extra_file_path, 'rb') as f:
extra = pickle.load(f)
encoding_name = extra['encoding_name']
encoding_args = extra['encoding_args']
encoding_kwargs = extra['encoding_kwargs']
self._setup_encoding(encoding_name, *encoding_args, **encoding_kwargs)
self.state_sizes = extra['state_sizes']
self.training_steps = extra.get('training_steps', int(training_steps))
self.round_steps = extra.get('round_steps', 0)
self._close_sess_if_open()
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf.Session(graph=self.graph)
meta_file_path = file_path + '.meta'
self.saver = tf.train.import_meta_graph(meta_file_path)
self.saver.restore(self.sess, file_path)
self.inputs = self.graph.get_tensor_by_name('inputs:0')
self.labels = self.graph.get_tensor_by_name('labels:0')
self.logits = self.graph.get_tensor_by_name('logits:0')
self.losses = self.graph.get_tensor_by_name('losses:0')
self.loss_mean = self.graph.get_tensor_by_name('loss_mean:0')
self.loss_max = self.graph.get_tensor_by_name('loss_max:0')
self.loss_min = self.graph.get_tensor_by_name('loss_min:0')
self.loss_sum = self.graph.get_tensor_by_name('loss_sum:0')
self.optimize = self.graph.get_operation_by_name('optimize')
self.probabilities = self.graph.get_tensor_by_name(
'probabilities:0')
self.next_probabilities = tools.get_tensor_by_name_if_exists(
self.graph, 'next_probabilities:0')
# self.init_states = tuple(self.graph.get_tensor_by_name('init_state_' + str(i) + ':0')
# for i in range(len(self.state_sizes)))
# self.output_states = tuple(self.graph.get_tensor_by_name('output_state_' + str(i) + ':0')
# for i in range(len(self.state_sizes)))
self.summary = self.graph.get_tensor_by_name('summary:0')
self.graph.finalize()
logging.info('Loaded model from file %s' % file_path)
def detect():
screenshot = screen_shotter.ScreenShot()
# On parcourt la colonne du milieu de l'écran, on cherche le premier pixel
# dont la couleur correspondrait à la ligne de fishing.
coord_fish_line = find_first_color(
# TODO : le X de la colonne à analyser est complètement à l'arrache.
# il faudrait tester plusieurs colonnes, par dichotomie.
screenshot, (900, 0), (0, +1),
(0, 204, 255), "rgb", True)
if coord_fish_line is None:
info("Impossible de trouver la ligne de fishing")
return None
y_line = coord_fish_line[1]
debug("y:" + str(y_line))
# Ça va planter si y'a aucun pixel de la couleur cherchée, dans la ligne.
# Pas supposée arriver puisqu'on vient de trouver un bon pixel dans cette
# ligne là, juste avant.
(x1_line, _) = find_first_color(
screenshot, (0, y_line), (+1, 0),
(0, 204, 255), "rgb", True)
(x2_line, _) = find_first_color(
screenshot, (screen_shotter.size_screen[0]-1, y_line), (-1, 0),
(0, 204, 255), "rgb", True)
debug("x1: " + str(x1_line) + " x2: " + str(x2_line))
return (y_line, x1_line, x2_line)
def __init__(self, max_batch_size, max_batch_width, pad_item, skip_padding=False):
logging.info('Using skip padding %s' % skip_padding)
self.inputs_array = np.zeros((max_batch_size, max_batch_width), dtype=np.int32)
self.labels_array = np.zeros((max_batch_size, max_batch_width), dtype=np.int32)
self.max_batch_size = max_batch_size
self.max_batch_width = max_batch_width
self.pad_item = pad_item
self.skip_padding = skip_padding
def init_from_file_or_encoding(self, *args, **kwargs):
logging.info('Initializing %s from file or encoding...' % self.name)
try:
self.init_from_file()
except FileNotFoundError as e:
logging.info(
'File not found for %s, initializing from encoding...' %
self.name)
self.init_from_encoding(*args, **kwargs)
def test_logger(self):
# ignore warnings
warnings.simplefilter("ignore")
# log something
log_file = default_log_file + '_basic_test'
my_logging.prepare_logger(log_file)
my_logging.info("ABCD")
my_logging.shutdown()
# check logfile
success = 'ABCD' in read_file(log_file + '_info.log')
self.assertTrue(success)
def test_data(self):
log_file = default_log_file + '_data_test'
my_logging.prepare_logger(log_file)
my_logging.data('key', 2)
my_logging.info('ABCD')
my_logging.shutdown()
# check
content = read_file(log_file + '_data.log')
d = json.loads(content)
self.assertEqual(d['key'], 'key')
self.assertEqual(d['value'], 2)
self.assertTrue('ABCD' not in content)
def __init__(self,
name,
tag=None,
saved_models_dir=None,
saved_summaries_dir=None):
self.uuid = uuid.uuid4()
logging.info('Running __init__ %s' % self.uuid)
self.name = name
if tag is None:
tag = config.TAG
logging.info('Using tag %s and name %s' % (tag, name))
if saved_models_dir is None:
saved_models_dir = config.SAVED_MODELS_DIR
if saved_summaries_dir is None:
saved_summaries_dir = config.SAVED_SUMMARIES_DIR
self.tag = tag
self.saved_models_dir = saved_models_dir
self.saved_summaries_dir = saved_summaries_dir
self.training_steps = 0
self.round_steps = 0
self.sess = None
def sample(self, starting=None, max_num=50, context=200):
"""
Args:
starting: A thing that encodes to a python iterable of numbers
"""
if starting is None:
starting = self.encoding.empty()
curr = list(self.encoding.encode_input_for_testing(starting))
curr_output = list(curr)
try:
for i in range(max_num):
with tools.DelayedKeyboardInterrupt():
# Sample using the last n chars where n = context
encoded_inpt = self.encoding.make_input_for_sample(
curr[-context:])
if self.next_probabilities is not None:
probs_batch = self._run(self.next_probabilities,
[encoded_inpt])
probs = probs_batch[0]
else:
probs_batch = self._run(self.probabilities,
[encoded_inpt])
probs = probs_batch[0][-1]
next_int = np.random.choice(self.effective_alphabet_size,
1,
p=probs).item()
curr_dec = self._decode_output_to_list(curr[-5:])
next_dec = self._decode_if_ok(next_int)
probs_dec = self._make_probs_dec(probs)
logging.info('Step %s: curr: %s next: %s probs: %s' %
(i, curr_dec, repr(next_dec), probs_dec))
if next_int == self.encoding.token_item:
break
if next_int < self.alphabet_size:
curr_output.append(next_int)
curr.append(next_int)
except KeyboardInterrupt:
logging.info('Cancelling sample...')
return self.encoding.decode_output(curr_output)
def find(haystack,
needle,
all=False,
first=False,
ignore_case=True,
max_results=50):
'''Prints path to needle in iterable haystack (can be nested list or dict).
Example:
>>> haystack = {
'fruits': [
{'color': 'yellow', 'name': 'banana'},
{'color': 'red', 'name': 'strawberry'},
{'color': 'yellow', 'name': 'lemon'}],
'vegetables': [
{'color': 'green', 'name': 'green pepper'}
]}
>>> find(haystack, 'pepper')
INFO: Searching...
haystack['vegetables'][0]['name']: green pepper
INFO: Found 1 result
>>> haystack['vegetables'][0]['name']
'green pepper'
>>> find(haystack, 'cucumber')
INFO: Searching...
INFO: Not found
>>> find(haystack, 'yellow')
INFO: Searching...
haystack['fruits'][0]['color']: yellow
haystack['fruits'][2]['color']: yellow
INFO: Found 2 results
# Case-sensitive
>>> find(haystack, 'YeLLow', ignore_case=False)
INFO: Searching...
INFO: Not found
Args:
haystack (dict or list): Nested dict or list.
needle (anything): Target to find.
all (bool): Shows all results if True. Up to max_results otherwise.
first (bool): Returns first result if True. All results otherwise.
ignore_case (bool): Case-insensitive if True.
max_results (int): Max number of results to return. Unlimited if negative or zero.
'''
info('Searching...')
all_ = all
results = _find(haystack, needle, first, ignore_case)
if not results:
info('Not found')
return
for index, result in enumerate(results):
if not all_ and index >= max_results > 0:
# Max number of results reached.
print 'Displayed first %i results...' % max_results
break
# Print path to value and value itself.
path = ''
for attr in result:
if _is_text(attr):
path += "['%s']" % attr
else:
path += '[%s]' % attr
print '%s%s:' % (var_name(haystack), path),
print eval('haystack%s' % path)
num_results = len(results)
if not first:
info('Found %s result%s' %
(human_int(num_results), 's' if num_results > 1 else ''))
def handler(self, sig, frame):
self.signal_received = (sig, frame)
logging.info('KeyboardInterrupt received.')
import my_logging as logging
logging.info('Importing tensorflow...')
import tensorflow as tf
logging.info('Imported tensorflow.')
tf.get_default_graph().finalize()
import numpy as np
from datetime import datetime
import uuid
import os
import json
import pickle
import itertools
from abc import ABC, abstractmethod
import config
import tools
"""
Terminology:
input/inpt: A single string or list of numbers
inputs: An iterable where each element is an input
exp_output/exp_outpt: A list of numbers
exp_outputs: An iterable where each element is an exp_output
Directory structure:
saved_models/
shakespeare/
baseline/
3000/
# -*- coding: utf-8 -*-
from __future__ import unicode_literals
import time
from my_logging import debug, info
import screen_shotter
import fish_line_detector
from fish_line_analyzer import FishLineAnalyzer, fst
from fish_subline_analyzer import FishSublineAnalyzer
from key_presser import KeyPresser
debug("coucou éé")
info("pouet")
DICT_DELAY_FROM_FISHING_STATE = {
fst.HIGHLIGHTED : 0.5,
fst.SIGNAL_SENT : 0.1,
fst.WAIT_FISH : 0.1,
fst.FISH_NEAR : 0.000001,
fst.SEND_SIGNAL : 0.000001,
fst.CRITICAL_ZONE_NOT_FOUND : 0.1,
}
fish_line_info = fish_line_detector.detect()
if fish_line_info is None:
raise Exception("impossible de trouver la ligne de fishing à l'écran")
(y_line, x1_line, x2_line) = fish_line_info
fish_subline_analyzer = FishSublineAnalyzer(y_line+5, x1_line, x2_line)
fish_line_analyzer = FishLineAnalyzer(
def train(self,
inputs,
autosave=None,
cont=False,
count=False,
save_graph=True):
"""
Args:
inputs: A python iterable of things that encode to python iterables (if long) or lists (if short) of numbers
"""
if count:
logging.info(
'Counting enabled! This should only be used if inputs can be iterated over multiple times.'
)
logging.info('Counting number of steps...')
# Only use count if inputs and the items in it can be iterated over multiple times
total_round_steps = self.encode_and_count_batches_for_training(
inputs)
logging.info('Number of steps is %s' % total_round_steps)
batches = self.encode_and_make_batches_for_training(inputs)
if cont:
logging.info('Continuing enabled! Starting from round %s' %
self.round_steps)
# Note: round_steps is the number of rounds completed, so
# is also the index of the next round to be done
batches = itertools.islice(batches, self.round_steps, None)
else:
self.round_steps = 0
save_dir = os.path.join(self.saved_summaries_dir, self.name, self.tag)
os.makedirs(save_dir, exist_ok=True)
summaries_dir = save_dir
logging.info('Using summaries dir %s' % summaries_dir)
if save_graph and self.training_steps == 0:
logging.info('Saving graph...')
# Only write graph during the first training attempt
summary_writer = tf.summary.FileWriter(summaries_dir, self.graph)
else:
summary_writer = tf.summary.FileWriter(summaries_dir)
logging.info('Starting training...')
curr_states = None
starting_i = self.training_steps
starting_round = self.round_steps
# Need this i set for if there ends up being no batches due to continuing
i = starting_i
try:
for batch in batches:
with tools.DelayedKeyboardInterrupt():
# Use key instead of i or self.training_steps
# to prevent the unlikely race condition
# of hitting the interrupt after
# the key is updated but before entering the
# DelayedKeyboardInterrupt section
self.training_steps += 1
self.round_steps += 1
i = self.training_steps
# i, training_steps will be steps completed after the run_batch
if i - 1 == starting_i:
loss_max, loss_mean, loss_min = self._run_batch(
[self.loss_max, self.loss_mean, self.loss_min],
batch, curr_states)
logging.info(
'Starting values: steps: %s round: %s loss max: %s mean: %s min: %s'
% (starting_i, starting_round, loss_max, loss_mean,
loss_min))
if count:
time_rem = tools.TimeRemaining(
total_round_steps, starting_round)
if i % 10 != 0:
_ = self._run_batch([self.optimize], batch,
curr_states)
else:
_, summary, loss_max, loss_mean, loss_min = self._run_batch(
[
self.optimize, self.summary, self.loss_max,
self.loss_mean, self.loss_min
], batch, curr_states)
summary_writer.add_summary(summary, i)
# Note: The printed numbers are the numbers from before the optimization update happens
# Note: Step numbers start from 1
if count:
time_rem_str = time_rem.get_str(self.round_steps)
logging.info(
'Step %s round %s/%s: loss max: %s mean: %s min: %s time rem: %s'
%
(i, self.round_steps, total_round_steps,
loss_max, loss_mean, loss_min, time_rem_str))
else:
logging.info(
'Step %s round %s: loss max: %s mean: %s min: %s'
% (i, self.round_steps, loss_max, loss_mean,
loss_min))
#losses, probabilities = self.sess.run([self.losses, self.probabilities], feed_dict={self.inputs: batch})
#logging.info('Step %s: losses: %s probs: %s' % (i, losses, probabilities))
# curr_states = new_states
if autosave is not None and (autosave is not True
and i % autosave == 0):
self.save_to_file()
except KeyboardInterrupt:
logging.info('Cancelling training...')
logging.info('Saved summaries to %s' % summaries_dir)
if autosave is not None and (autosave is True
or i % autosave != 0):
# Save the last one only if it hasn't already been saved
self.save_to_file()
tools.exit(22)
else:
self.round_steps = 0
logging.info('Saved summaries to %s' % summaries_dir)
if autosave is not None:
# Always save the last one because round_steps has changed
self.save_to_file()
logging.info('Done training.')
def __del__(self):
logging.info('Running __del__ %s' % self.uuid)
self._close_sess_if_open()
def init_from_encoding(self, encoding_name, *args, **kwargs):
logging.info('Initializing from encoding...')
self._setup_encoding(encoding_name, *args, **kwargs)
self._close_sess_if_open()
self.graph = tf.Graph()
with self.graph.as_default():
self.inputs = tf.placeholder(tf.int32,
shape=(None, None),
name='inputs')
self.labels = tf.placeholder(tf.int32,
shape=(None, None),
name='labels')
# inputs is list of batches where each batch is a list of integers in [0, alphabet_size]
# inputs is tensor with dim batch_size x timesteps
# inputs_padded and labels_padded are tensors with dim batch_size x (timesteps+1)
self.inputs_one_hot = tf.one_hot(self.inputs,
self.effective_alphabet_size,
name='inputs_one_hot')
# inputs_one_hot is a list of batches where each batch is a list of one hot encoded lists
# inputs_one_hot is a tensor with dim batch_size x (timesteps+1) x (alphabet_size+1)
self.state_sizes = [4 * self.effective_alphabet_size] * 2
# def make_init_state(i, state_size):
# return tf.placeholder(tf.float32, [2, None, state_size], name='init_state_' + str(i))
# def make_lstm_state_tuple(init_state):
# c, h = tf.unstack(init_state)
# return tf.contrib.rnn.LSTMStateTuple(c, h)
# self.init_states = tuple(make_init_state(i, state_size) for i, state_size in enumerate(self.state_sizes))
# rnn_init_state = tuple(make_lstm_state_tuple(init_state) for init_state in self.init_states)
lstm_cells = list(
map(tf.contrib.rnn.BasicLSTMCell, self.state_sizes))
lstm = tf.contrib.rnn.MultiRNNCell(lstm_cells)
rnn_output, rnn_states = tf.nn.dynamic_rnn(lstm,
self.inputs_one_hot,
dtype=tf.float32)
tf.summary.histogram('rnn_output', rnn_output)
# rnn_outputs is a tensor with dim batch_size x (timesteps+1) x (alphabet_size+1)
# state is a list of tensors with dim batch_size x state_size
# def make_output_state(i, state_tuple):
# c = state_tuple.c
# h = state_tuple.h
# return tf.identity(tf.stack([c, h]), name='output_state_' + str(i))
# self.output_states = tuple(make_output_state(i, state) for i, state in enumerate(rnn_states))
logits = tf.contrib.layers.fully_connected(
rnn_output, self.effective_alphabet_size, activation_fn=None)
self.logits = tf.identity(logits, name='logits')
tf.summary.histogram('logits', self.logits)
losses = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=self.labels, logits=self.logits)
self.losses = tf.identity(losses, name='losses')
tf.summary.histogram('losses', self.losses)
# losses is a tensor with dim batch_size x (timestamps+1)
self.loss_mean = tf.reduce_mean(self.losses, name='loss_mean')
tf.summary.scalar('loss_mean', self.loss_mean)
self.loss_max = tf.reduce_max(self.losses, name='loss_max')
tf.summary.scalar('loss_max', self.loss_max)
self.loss_min = tf.reduce_min(self.losses, name='loss_min')
tf.summary.scalar('loss_min', self.loss_min)
self.loss_sum = tf.reduce_sum(self.losses, name='loss_sum')
tf.summary.scalar('loss_sum', self.loss_sum)
# loss_* is a tensor with a single float
self.optimize = tf.train.AdamOptimizer().minimize(self.loss_mean,
name='optimize')
# self.optimize = tf.train.AdadeltaOptimizer(0.03).minimize(self.loss_mean, name='optimize')
self.probabilities = tf.identity(tf.nn.softmax(self.logits),
name='probabilities')
tf.summary.histogram('probabilities', self.probabilities)
self.next_probabilities = tf.identity(self.probabilities[:, -1],
name='next_probabailities')
self.saver = tf.train.Saver()
self.summary = tf.identity(tf.summary.merge_all(), name='summary')
self.glob_var_init = tf.global_variables_initializer()
self.graph.finalize()
self.sess = tf.Session(graph=self.graph)
self.sess.run(self.glob_var_init)
logging.info('Initialized from encoding.')
def _close_sess_if_open(self):
if self.sess is not None:
self.sess.close()
self.sess = None
logging.info('Closed session %s' % self.uuid)