clearml/trains/binding/artifacts.py

import hashlib
from copy import deepcopy
from multiprocessing.pool import ThreadPool
from tempfile import mkdtemp
from threading import Thread, Event

import numpy as np
from pathlib2 import Path

from ..debugging.log import LoggerRoot

try:
    import pandas as pd
except ImportError:
    pd = None


class Artifacts(object):
    _flush_frequency_sec = 300.
    # notice these two should match
    _save_format = '.csv.gz'
    _compression = 'gzip'
    # hashing constants
    _hash_block_size = 65536

    class _ProxyDictWrite(dict):
        """ Dictionary wrapper that updates an arguments instance on any item set in the dictionary """
        def __init__(self, artifacts_manager, *args, **kwargs):
            super(Artifacts._ProxyDictWrite, self).__init__(*args, **kwargs)
            self._artifacts_manager = artifacts_manager
            # list of artifacts we should not upload (by name & weak-reference)
            self.artifact_metadata = {}

        def __setitem__(self, key, value):
            # check that value is of type pandas
            if isinstance(value, np.ndarray) or (pd and isinstance(value, pd.DataFrame)):
                super(Artifacts._ProxyDictWrite, self).__setitem__(key, value)

                if self._artifacts_manager:
                    self._artifacts_manager.flush()
            else:
                raise ValueError('Artifacts currently support pandas.DataFrame objects only')

        def unregister_artifact(self, name):
            self.artifact_metadata.pop(name, None)
            self.pop(name, None)

        def add_metadata(self, name, metadata):
            self.artifact_metadata[name] = deepcopy(metadata)

        def get_metadata(self, name):
            return self.artifact_metadata.get(name)

    @property
    def artifacts(self):
        return self._artifacts_dict

    @property
    def summary(self):
        return self._summary

    def __init__(self, task):
        self._task = task
        # notice the double link, this important since the Artifact
        # dictionary needs to signal the Artifacts base on changes
        self._artifacts_dict = self._ProxyDictWrite(self)
        self._last_artifacts_upload = {}
        self._unregister_request = set()
        self._thread = None
        self._flush_event = Event()
        self._exit_flag = False
        self._thread_pool = ThreadPool()
        self._summary = ''
        self._temp_folder = []

    def register_artifact(self, name, artifact, metadata=None):
        # currently we support pandas.DataFrame (which we will upload as csv.gz)
        if name in self._artifacts_dict:
            LoggerRoot.get_base_logger().info('Register artifact, overwriting existing artifact \"{}\"'.format(name))
        self._artifacts_dict[name] = artifact
        if metadata:
            self._artifacts_dict.add_metadata(name, metadata)

    def unregister_artifact(self, name):
        # Remove artifact from the watch list
        self._unregister_request.add(name)
        self.flush()

    def flush(self):
        # start the thread if it hasn't already:
        self._start()
        # flush the current state of all artifacts
        self._flush_event.set()

    def stop(self, wait=True):
        # stop the daemon thread and quit
        # wait until thread exists
        self._exit_flag = True
        self._flush_event.set()
        if wait:
            if self._thread:
                self._thread.join()
            # remove all temp folders
            for f in self._temp_folder:
                try:
                    Path(f).rmdir()
                except Exception:
                    pass

    def _start(self):
        if not self._thread:
            # start the daemon thread
            self._flush_event.clear()
            self._thread = Thread(target=self._daemon)
            self._thread.daemon = True
            self._thread.start()

    def _daemon(self):
        while not self._exit_flag:
            self._flush_event.wait(self._flush_frequency_sec)
            self._flush_event.clear()
            try:
                artifact_keys = list(self._artifacts_dict.keys())
                for name in artifact_keys:
                    self._upload_artifacts(name)
            except Exception as e:
                LoggerRoot.get_base_logger().warning(str(e))

        # create summary
        self._summary = self._get_statistics()

    def _upload_artifacts(self, name):
        logger = self._task.get_logger()
        artifact = self._artifacts_dict.get(name)

        # remove from artifacts watch list
        if name in self._unregister_request:
            try:
                self._unregister_request.remove(name)
            except KeyError:
                pass
            self._artifacts_dict.unregister_artifact(name)

        if artifact is None:
            return

        local_csv = (Path(self._get_temp_folder()) / (name + self._save_format)).absolute()
        if local_csv.exists():
            # we are still uploading... get another temp folder
            local_csv = (Path(self._get_temp_folder(force_new=True)) / (name + self._save_format)).absolute()
        artifact.to_csv(local_csv.as_posix(), index=False, compression=self._compression)
        current_sha2 = self.sha256sum(local_csv.as_posix(), skip_header=32)
        if name in self._last_artifacts_upload:
            previous_sha2 = self._last_artifacts_upload[name]
            if previous_sha2 == current_sha2:
                # nothing to do, we can skip the upload
                local_csv.unlink()
                return
        self._last_artifacts_upload[name] = current_sha2
        # now upload and delete at the end.
        logger.report_image_and_upload(title='artifacts', series=name, path=local_csv.as_posix(),
                                       delete_after_upload=True, iteration=self._task.get_last_iteration(),
                                       max_image_history=2)

    def _get_statistics(self):
        summary = ''
        thread_pool = ThreadPool()

        try:
            # build hash row sets
            artifacts_summary = []
            for a_name, a_df in self._artifacts_dict.items():
                if not pd or not isinstance(a_df, pd.DataFrame):
                    continue

                a_unique_hash = set()

                def hash_row(r):
                    a_unique_hash.add(hash(bytes(r)))

                a_shape = a_df.shape
                # parallelize
                thread_pool.map(hash_row, a_df.values)
                # add result
                artifacts_summary.append((a_name, a_shape, a_unique_hash,))

            # build intersection summary
            for i, (name, shape, unique_hash) in enumerate(artifacts_summary):
                summary += '[{name}]: shape={shape}, {unique} unique rows, {percentage:.1f}% uniqueness\n'.format(
                    name=name, shape=shape, unique=len(unique_hash), percentage=100*len(unique_hash)/float(shape[0]))
                for name2, shape2, unique_hash2 in artifacts_summary[i+1:]:
                    intersection = len(unique_hash & unique_hash2)
                    summary += '\tIntersection with [{name2}] {intersection} rows: {percentage:.1f}%\n'.format(
                        name2=name2, intersection=intersection, percentage=100*intersection/float(len(unique_hash2)))
        except Exception as e:
            LoggerRoot.get_base_logger().warning(str(e))
        finally:
            thread_pool.close()
            thread_pool.terminate()
        return summary

    def _get_temp_folder(self, force_new=False):
        if force_new or not self._temp_folder:
            new_temp = mkdtemp(prefix='artifacts_')
            self._temp_folder.append(new_temp)
            return new_temp
        return self._temp_folder[0]

    @staticmethod
    def sha256sum(filename, skip_header=0):
        # create sha2 of the file, notice we skip the header of the file (32 bytes)
        # because sometimes that is the only change
        h = hashlib.sha256()
        b = bytearray(Artifacts._hash_block_size)
        mv = memoryview(b)
        with open(filename, 'rb', buffering=0) as f:
            # skip header
            f.read(skip_header)
            for n in iter(lambda: f.readinto(mv), 0):
                h.update(mv[:n])
        return h.hexdigest()