Skip to content
Snippets Groups Projects
Code owners
Assign users and groups as approvers for specific file changes. Learn more.
tf.py 8.19 KiB
from contextlib import contextmanager
import numpy as np
import tensorflow as tf
from time import time
from analysis.util import atq
from io import BytesIO
import h5py
from tensorflow.python.keras.engine.saving import (
    save_weights_to_hdf5_group,
    load_weights_from_hdf5_group,
)

from evil import pin, ccall


from tensorflow.python.util import deprecation

deprecation._PRINT_DEPRECATION_WARNINGS = False


def last_map(func, obj, order=1):
    return tf.concat(
        [
            obj[..., :-1],
            tf.expand_dims(
                last_map(func=func, obj=obj[..., -1], order=order - 1)
                if 1 < order
                else func(obj[..., -1]),
                axis=-1,
            ),
        ],
        axis=-1,
    )


def tf_meanstd(val, **kwargs):
    mean = tf.reduce_mean(val, **kwargs)
    std = tf.reduce_mean(tf.square(val), **kwargs) - tf.square(mean)
    return mean, std


# this is mostly unused ...


class FD(dict):
    def add(self, data, dtype=None, shape=None, **kwargs):
        if shape is True:
            shape = data.shape
        elif shape is None:
            shape = (None,) + data.shape[1:]
        if dtype is None:
            dtype = data.dtype
        ph = tf.placeholder(dtype, shape=shape, **kwargs)
        self[ph] = data
        return ph


class Chain(object):
    def __init__(
        self,
        __name__,
        loss,
        sumnbs=(),
        step=None,
        opt=None,
        train=None,
        ema=(1, 2, 3, 4, 5),
        **kwargs
    ):
        if step is None:
            step = tf.Variable(0, trainable=False, name="%s_step" % __name__)
        if opt is None:
            opt = tf.train.AdamOptimizer()
        if train is None:
            train = opt.minimize(loss, global_step=step)
        reset_opt = tf.variables_initializer(opt.variables())

        summaries = []
        _ema_val = {}
        _ema_out = SKDict()
        kwargs["loss"] = loss
        with tf.name_scope(__name__):
            for key, val in kwargs.items():
                if not isinstance(val, tf.Tensor):
                    continue
                nd = val.shape.ndims
                if nd > 1 and sumnbs is not True and val not in sumnbs:
                    continue
                if val.dtype == tf.bool:
                    val = tf.reduce_mean(tf.cast(val, tf.float32))
                if nd:
                    # summaries.append(tf.summary.histogram(key, val))
                    mean, std = tf_meanstd(val)
                    summaries.append(tf.summary.scalar(key + "_mean", mean))
                    summaries.append(tf.summary.scalar(key + "_std", std))
                    _ema_val[key + "_mean"] = mean
                    _ema_val[key + "_std"] = std
                else:
                    summaries.append(tf.summary.scalar(key, val))
                    _ema_val[key] = val

        if ema:
            _ema_ops = []
            reset_ema_to = tf.constant(0, tf.int32)
            ema_step = tf.Variable(reset_ema_to, trainable=False, name="%s_ema_step" % __name__)
            reset_ema = ema_step.initializer
            for i in ema:
                with tf.name_scope("%s_%d" % (__name__, i)):
                    ema = tf.train.ExponentialMovingAverage(
                        decay=1.0
                        - tf.maximum(
                            0.1 ** i,
                            tf.exp((np.log(0.1 ** i) / 10.0) * tf.cast(ema_step, tf.float32)),
                        )
                    )
                    _ema_ops.append(ema.apply(_ema_val.values()))
                    for key, val in _ema_val.items():
                        val = ema.average(val)
                        _ema_out[key, i] = val
                        summaries.append(tf.summary.scalar("%s" % key, val))
            with tf.control_dependencies([train]):
                with tf.control_dependencies(_ema_ops):
                    train = ema_step.assign_add(1)

        ema = _ema_out
        summaries = tf.summary.merge(summaries)
        self.__dict__.update(kwargs)
        pin(locals(), True, kwargs)


Chain.cc = ccall(Chain, True)


class Runner(object):
    def __init__(self, sess, writer):
        self.step = 0
        pin(locals())

    def __call__(self, *args, **kwargs):
        return self.sess.run(*args, **kwargs)

    def train(self, chain, extra=None, step=1, sumiv=10, sumskip=0, **kwargs):
        if extra is None:
            extra = chain.loss
        for i in count():
            for j in xrange(step):
                self(chain.train, **kwargs)
            sum = sumiv and not (i < sumskip or i % sumiv)
            get = (chain.train, extra)
            if sum:
                get += (chain.summaries, chain.step)
            out = self(get, **kwargs)
            if sum:
                self.writer.add_summary(*out[2:])
            yield out[1]

    def conf(self, *args, **kwargs):
        return Confirmer(self.train(*args, **kwargs).next)

    @classmethod
    @contextmanager
    def make(cls, path, gpuOpts={}, flush_secs=20):
        with cls.make_session(gpuOpts=gpuOpts) as sess:
            with tf.summary.FileWriter(
                path, session=sess, flush_secs=flush_secs, graph=sess.graph
            ) as writer:
                yield cls(sess, writer)

    @classmethod
    def make_session(cls, gpuOpts={}):
        return tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(**gpuOpts)))


class Confirmer(object):
    last = np.inf
    task = None
    tdelay = 60
    tbestbump = 10

    def __init__(self, func):
        self.reset.func = func

    def __call__(self, target, its=np.inf, pre=0, **kwargs):
        if not kwargs.setdefault("disable", not kwargs):
            kwargs.setdefault("auto", self._prog)
        assert 0 < target
        tnext = 0
        with atq(xrange(its), **kwargs) as tq:
            for i in tq:
                if self.task is not None:
                    if self.task.check_stop():
                        break
                    if tnext < time():
                        tnext = time() + self.tdelay
                        pTotal = 100.0 * i / its
                        self.task.set_progress_percentage(pTotal)
                        self.task.set_status_message(
                            "total=%.1f%% conf=%.1f%% best=%.2e"
                            % (pTotal, 100.0 * self.conf / target, self.best)
                        )
                if not (self.conf < target):
                    break
                self.step += 1
                self.last = self.func()
                if self.step < pre:
                    continue
                elif self.best < self.last:
                    self.conf += 1
                else:
                    self.on_best()
                    self.conf = 0
                    self.best = self.last
                    tnext -= self.tbestbump
        return not (self.conf < target)

    def on_best(self):
        pass

    @property
    def reset(self):
        self.last = self.best = np.inf
        self.step = self.conf = 0
        return self

    @property
    def rdlb(self):
        # Relative Delta Last to Best
        return np.float_(self.last - self.best) / self.best

    def _prog(self, i, prefix=""):
        return {
            prefix + k: v
            for k, v in dict(
                step=self.step if self.step != i else None,
                last=self.last if self.rdlb else None,
                rdlb=self.rdlb or None,
                conf=self.conf or None,
                best=self.best,
            ).items()
            if v is not None
        }

    def __repr__(self):
        return "Confirmer(step=%d, best=%.3e, conf=%d)" % (self.step, self.best, self.conf)


class Bestie(object):
    def __init__(self, model, path):
        pin(locals())

    def save(self):
        self.model.save_weights(self.path)

    def load(self):
        self.model.load_weights(self.path)

    def finish(self):
        self.load()
        self.model.save(self.path)


class BestieMemory(Bestie):
    def __init__(self, *args, **kwargs):
        super(BestieMemory, self).__init__(*args, **kwargs)

    def save(self):
        self.h5 = h5py.File(BytesIO())
        save_weights_to_hdf5_group(self.h5, self.model.layers)

    def load(self):
        load_weights_from_hdf5_group(self.h5, self.model.layers)