class LocalEmbedder:
embedder = None
def __init__(self, model, model_settings, layer):
self.model = model
self.layer = layer
self._load_model()
self._target_image_size = model_settings["target_image_size"]
self._session = cachecontrol.CacheControl(
requests.session(),
cache=cachecontrol.caches.FileCache(
join(cache_dir(), __name__ + ".ImageEmbedder.httpcache"))
)
self.cancelled = False
self._image_loader = ImageLoader()
self._cache = EmbedderCache(model, layer)
def _load_model(self):
self.embedder = squeezenet(include_softmax=False)
def from_file_paths(self, file_paths, image_processed_callback=None):
all_embeddings = []
for image in file_paths:
all_embeddings.append(self._embed(image))
if image_processed_callback:
image_processed_callback(success=True)
self._cache.persist_cache()
return np.array(all_embeddings)
def _embed(self, file_path):
""" Load images and compute cache keys and send requests to
an http2 server for valid ones.
"""
if self.cancelled:
raise EmbeddingCancelledException()
image = self._image_loader.load_image_or_none(
file_path, self._target_image_size)
if image is None:
return None
image = self._image_loader.preprocess_squeezenet(image)
cache_key = self._cache.md5_hash(image)
cached_im = self._cache.get_cached_result_or_none(cache_key)
if cached_im is not None:
return cached_im
embedded_image = self.embedder.predict([image])
embedded_image = embedded_image[0][0]
self._cache.add(cache_key, embedded_image)
return embedded_image
class LocalEmbedder:
embedder = None
def __init__(self, model, model_settings, layer):
self.model = model
self.layer = layer
self._load_model()
self._target_image_size = model_settings["target_image_size"]
self._session = cachecontrol.CacheControl(
requests.session(),
cache=cachecontrol.caches.FileCache(
join(cache_dir(), __name__ + ".ImageEmbedder.httpcache")))
self.cancelled = False
self._image_loader = ImageLoader()
self._cache = EmbedderCache(model, layer)
def _load_model(self):
self.embedder = squeezenet(include_softmax=False)
def from_file_paths(self, file_paths, image_processed_callback=None):
all_embeddings = []
for image in file_paths:
all_embeddings.append(self._embed(image))
if image_processed_callback:
image_processed_callback(success=True)
self._cache.persist_cache()
return np.array(all_embeddings)
def _embed(self, file_path):
""" Load images and compute cache keys and send requests to
an http2 server for valid ones.
"""
if self.cancelled:
raise EmbeddingCancelledException()
image = self._image_loader.load_image_or_none(file_path,
self._target_image_size)
if image is None:
return None
image = self._image_loader.preprocess_squeezenet(image)
cache_key = self._cache.md5_hash(image)
cached_im = self._cache.get_cached_result_or_none(cache_key)
if cached_im is not None:
return cached_im
embedded_image = self.embedder.predict([image])
embedded_image = embedded_image[0][0]
self._cache.add(cache_key, embedded_image)
return embedded_image
class ServerEmbedder(Http2Client):
MAX_REPEATS = 4
CANNOT_LOAD = "cannot load"
def __init__(self, model, model_settings, layer, server_url):
super().__init__(server_url)
self._model = model
self._layer = layer
self._target_image_size = model_settings['target_image_size']
# attribute that offers support for cancelling the embedding
# if ran in another thread
self.cancelled = False
self.machine_id = \
QSettings().value('error-reporting/machine-id', '', type=str) \
or str(uuid.getnode())
self.session_id = None
self._image_loader = ImageLoader()
self._cache = EmbedderCache(model, layer)
def from_file_paths(self, file_paths, image_processed_callback=None):
"""Send the images to the remote server in batches. The batch size
parameter is set by the http2 remote peer (i.e. the server).
Parameters
----------
file_paths: list
A list of file paths for images to be embedded.
image_processed_callback: callable (default=None)
A function that is called after each image is fully processed
by either getting a successful response from the server,
getting the result from cache or skipping the image.
Returns
-------
embeddings: array-like
Array-like of float16 arrays (embeddings) for
successfully embedded images and Nones for skipped images.
Raises
------
ConnectionError:
If disconnected or connection with the server is lost
during the embedding process.
EmbeddingCancelledException:
If cancelled attribute is set to True (default=False).
"""
if not self.is_connected_to_server():
self.reconnect_to_server()
self.session_id = str(random.randint(1, 1e10))
all_embeddings = [None] * len(file_paths)
repeats_counter = 0
# repeat while all images has embeddings or
# while counter counts out (prevents cycling)
while len([el for el in all_embeddings if el is None]) > 0 and \
repeats_counter < self.MAX_REPEATS:
# take all images without embeddings yet
selected_indices = [i for i, v in enumerate(all_embeddings)
if v is None]
file_paths_wo_emb = [(file_paths[i], i) for i in selected_indices]
for batch in self._yield_in_batches(file_paths_wo_emb):
b_images, b_indices = zip(*batch)
try:
embeddings = self._send_to_server(
b_images, image_processed_callback, repeats_counter
)
except (MaxNumberOfRequestsError, BrokenPipeError):
# maximum number of http2 requests through a single
# connection is exceeded and a remote peer has closed
# the connection so establish a new connection and retry
# with the same batch (should happen rarely as the setting
# is usually set to >= 1000 requests in http2)
self.reconnect_to_server()
embeddings = [None] * len(batch)
# insert embeddings into the list
for i, emb in zip(b_indices, embeddings):
all_embeddings[i] = emb
self._cache.persist_cache()
repeats_counter += 1
# change images that were not loaded from 'cannot loaded' to None
all_embeddings = \
[None if not isinstance(el, np.ndarray) and el == self.CANNOT_LOAD
else el for el in all_embeddings]
return np.array(all_embeddings)
def _yield_in_batches(self, list_):
gen_ = (path for path in list_)
batch_size = self._max_concurrent_streams
num_yielded = 0
while True:
batch = list(islice(gen_, batch_size))
num_yielded += len(batch)
yield batch
if num_yielded == len(list_):
return
def _send_to_server(self, file_paths, image_processed_callback, retry_n):
""" Load images and compute cache keys and send requests to
an http2 server for valid ones.
"""
cache_keys = []
http_streams = []
for file_path in file_paths:
if self.cancelled:
raise EmbeddingCancelledException()
image = self._image_loader.load_image_bytes(
file_path, self._target_image_size)
if not image:
# skip the sending because image was skipped at loading
http_streams.append(None)
cache_keys.append(None)
continue
cache_key = self._cache.md5_hash(image)
cache_keys.append(cache_key)
if self._cache.exist_in_cache(cache_key):
# skip the sending because image is present in the
# local cache
http_streams.append(None)
continue
try:
headers = {
'Content-Type': 'image/jpeg',
'Content-Length': str(len(image))
}
stream_id = self._send_request(
method='POST',
url='/image/' + self._model +
'?machine={}&session={}&retry={}'
.format(self.machine_id, self.session_id, retry_n),
headers=headers,
body_bytes=image
)
http_streams.append(stream_id)
except (ConnectionError, BrokenPipeError):
self._cache.persist_cache()
raise
# wait for the responses in a blocking manner
return self._get_responses_from_server(
http_streams,
cache_keys,
image_processed_callback
)
def _get_responses_from_server(self, http_streams, cache_keys,
image_processed_callback):
"""Wait for responses from an http2 server in a blocking manner."""
embeddings = []
for stream_id, cache_key in zip(http_streams, cache_keys):
if self.cancelled:
raise EmbeddingCancelledException()
if not stream_id and not cache_key:
# when image cannot be loaded
embeddings.append(self.CANNOT_LOAD)
if image_processed_callback:
image_processed_callback(success=False)
continue
if not stream_id:
# skip rest of the waiting because image was either
# skipped at loading or is present in the local cache
embedding = self._cache.get_cached_result_or_none(cache_key)
embeddings.append(embedding)
if image_processed_callback:
image_processed_callback(success=embedding is not None)
continue
try:
response = self._get_json_response_or_none(stream_id)
except (ConnectionError, MaxNumberOfRequestsError):
self._cache.persist_cache()
self.reconnect_to_server()
return embeddings
if not response or 'embedding' not in response:
# returned response is not a valid json response
# or the embedding key not present in the json
embeddings.append(None)
else:
# successful response
embedding = np.array(response['embedding'], dtype=np.float16)
embeddings.append(embedding)
self._cache.add(cache_key, embedding)
if image_processed_callback:
image_processed_callback(embeddings[-1] is not None)
return embeddings
class ServerEmbedder(Http2Client):
MAX_REPEATS = 4
CANNOT_LOAD = "cannot load"
def __init__(self, model, model_settings, layer, server_url):
super().__init__(server_url)
self._model = model
self._layer = layer
self._target_image_size = model_settings['target_image_size']
# attribute that offers support for cancelling the embedding
# if ran in another thread
self.cancelled = False
self.machine_id = \
QSettings().value('error-reporting/machine-id', '', type=str) \
or str(uuid.getnode())
self.session_id = None
self.num_images_sent = 0
self._image_loader = ImageLoader()
self._cache = EmbedderCache(model, layer)
def from_file_paths(self, file_paths, image_processed_callback=None):
"""
This function provides a fix for ProtocolError that happens on every
1000 image. Now only 999 images are send through one connection.
Parameters
----------
file_paths: list
A list of file paths for images to be embedded.
image_processed_callback: callable (default=None)
A function that is called after each image is fully processed
by either getting a successful response from the server,
getting the result from cache or skipping the image.
Returns
-------
embeddings: array-like
Array-like of float16 arrays (embeddings) for
successfully embedded images and Nones for skipped images.
"""
embeddings = []
for batch in self._yield_in_batches(file_paths, 999):
if self.num_images_sent > 0: # prevents reconnection for first batch
self.reconnect_to_server()
embeddings += self.embedd_batch(batch, image_processed_callback)
self.num_images_sent += len(batch)
return np.array(embeddings)
def embedd_batch(self, file_paths, image_processed_callback=None):
"""Send the images to the remote server in batches. The batch size
parameter is set by the http2 remote peer (i.e. the server).
Parameters
----------
file_paths: list
A list of file paths for images to be embedded.
image_processed_callback: callable (default=None)
A function that is called after each image is fully processed
by either getting a successful response from the server,
getting the result from cache or skipping the image.
Returns
-------
embeddings: array-like
Array-like of float16 arrays (embeddings) for
successfully embedded images and Nones for skipped images.
Raises
------
ConnectionError:
If disconnected or connection with the server is lost
during the embedding process.
EmbeddingCancelledException:
If cancelled attribute is set to True (default=False).
"""
if not self.is_connected_to_server():
self.reconnect_to_server()
self.session_id = str(random.randint(1, 1e10))
all_embeddings = [None] * len(file_paths)
repeats_counter = 0
# repeat while all images has embeddings or
# while counter counts out (prevents cycling)
while len([el for el in all_embeddings if el is None]) > 0 and \
repeats_counter < self.MAX_REPEATS:
# take all images without embeddings yet
selected_indices = [
i for i, v in enumerate(all_embeddings) if v is None
]
file_paths_wo_emb = [(file_paths[i], i) for i in selected_indices]
for batch in self._yield_in_batches(file_paths_wo_emb):
b_images, b_indices = zip(*batch)
try:
embeddings = self._send_to_server(
b_images, image_processed_callback, repeats_counter)
except (MaxNumberOfRequestsError, BrokenPipeError):
# maximum number of http2 requests through a single
# connection is exceeded and a remote peer has closed
# the connection so establish a new connection and retry
# with the same batch (should happen rarely as the setting
# is usually set to >= 1000 requests in http2)
self.reconnect_to_server()
embeddings = [None] * len(batch)
# insert embeddings into the list
for i, emb in zip(b_indices, embeddings):
all_embeddings[i] = emb
self._cache.persist_cache()
repeats_counter += 1
# change images that were not loaded from 'cannot loaded' to None
all_embeddings = \
[None if not isinstance(el, np.ndarray) and el == self.CANNOT_LOAD
else el for el in all_embeddings]
return all_embeddings
def _yield_in_batches(self, list_, batch_size=None):
gen_ = (path for path in list_)
if batch_size is None:
batch_size = self._max_concurrent_streams
num_yielded = 0
while True:
batch = list(islice(gen_, batch_size))
num_yielded += len(batch)
yield batch
if num_yielded == len(list_):
return
def _send_to_server(self, file_paths, image_processed_callback, retry_n):
""" Load images and compute cache keys and send requests to
an http2 server for valid ones.
"""
cache_keys = []
http_streams = []
for file_path in file_paths:
if self.cancelled:
raise EmbeddingCancelledException()
image = self._image_loader.load_image_bytes(
file_path, self._target_image_size)
if not image:
# skip the sending because image was skipped at loading
http_streams.append(None)
cache_keys.append(None)
continue
cache_key = self._cache.md5_hash(image)
cache_keys.append(cache_key)
if self._cache.exist_in_cache(cache_key):
# skip the sending because image is present in the
# local cache
http_streams.append(None)
continue
try:
headers = {
'Content-Type': 'image/jpeg',
'Content-Length': str(len(image))
}
stream_id = self._send_request(
method='POST',
url='/image/' + self._model +
'?machine={}&session={}&retry={}'.format(
self.machine_id, self.session_id, retry_n),
headers=headers,
body_bytes=image)
http_streams.append(stream_id)
except (ConnectionError, BrokenPipeError):
self._cache.persist_cache()
raise
# wait for the responses in a blocking manner
return self._get_responses_from_server(http_streams, cache_keys,
image_processed_callback)
def _get_responses_from_server(self, http_streams, cache_keys,
image_processed_callback):
"""Wait for responses from an http2 server in a blocking manner."""
embeddings = []
for stream_id, cache_key in zip(http_streams, cache_keys):
if self.cancelled:
raise EmbeddingCancelledException()
if not stream_id and not cache_key:
# when image cannot be loaded
embeddings.append(self.CANNOT_LOAD)
if image_processed_callback:
image_processed_callback(success=False)
continue
if not stream_id:
# skip rest of the waiting because image was either
# skipped at loading or is present in the local cache
embedding = self._cache.get_cached_result_or_none(cache_key)
embeddings.append(embedding)
if image_processed_callback:
image_processed_callback(success=embedding is not None)
continue
try:
response = self._get_json_response_or_none(stream_id)
except (ConnectionError, MaxNumberOfRequestsError):
self._cache.persist_cache()
self.reconnect_to_server()
return embeddings
if not response or 'embedding' not in response:
# returned response is not a valid json response
# or the embedding key not present in the json
embeddings.append(None)
else:
# successful response
embedding = np.array(response['embedding'], dtype=np.float16)
embeddings.append(embedding)
self._cache.add(cache_key, embedding)
if image_processed_callback:
image_processed_callback(embeddings[-1] is not None)
return embeddings
class ServerEmbedder:
MAX_REPEATS = 3
count_errors = 10
initial_count_errors = 10
def __init__(self, model: str, model_settings: Dict[str, Any],
server_url: str) -> None:
self.server_url = getenv('ORANGE_EMBEDDING_API_URL', server_url)
self._model = model
self._im_size = model_settings['target_image_size']
# attribute that offers support for cancelling the embedding
# if ran in another thread
self.cancelled = False
self.machine_id = \
QSettings().value('error-reporting/machine-id', '', type=str) \
or str(uuid.getnode())
self.session_id = str(random.randint(1, 1e10))
self._image_loader = ImageLoader()
self._cache = EmbedderCache(model)
self.batch_size = 10000
# default embedding timeouts are too small we need to increase them
self.timeouts = httpx.TimeoutConfig(timeout=60)
self.num_parallel_requests = 0
self.max_parallel = model_settings['batch_size']
def from_file_paths(
self,
file_paths: List[str],
image_processed_callback: Callable[[bool], None] = None
) -> List[Optional[List[float]]]:
"""
This function repeats calling embedding function until all images
are embedded. It prevents skipped images due to network issues.
The process is repeated for each image maximally MAX_REPEATS times.
Parameters
----------
file_paths
A list of images paths to be embedded.
image_processed_callback
A function that is called after each image is embedded
by either getting a successful response from the server,
getting the result from cache or skipping the image.
Returns
-------
embeddings
Array-like of float arrays (embeddings) for successfully embedded
images and Nones for skipped images.
Raises
------
EmbeddingConnectionError
Error which indicate that the embedding is not possible due to
connection error.
EmbeddingCancelledException:
If cancelled attribute is set to True (default=False).
"""
# if there is less images than 10 connection error should be raised
# earlier
self.count_errors = len(file_paths) * 3 \
if len(file_paths) * 3 < 10 else 10
self.initial_count_errors = self.count_errors
loop = asyncio.new_event_loop()
asyncio.set_event_loop(loop)
embeddings = [None] * len(file_paths)
repeats_count = 0
to_embedd = list(enumerate(file_paths))
# repeats until all embedded or max_repeats reached
while len(to_embedd) > 0 and repeats_count < self.MAX_REPEATS:
repeats_count += 1
# get paths that are not embedded yet
idx, paths = list(zip(*to_embedd))
# call embedding
new_embeddings = asyncio.get_event_loop().run_until_complete(
self.embedd_batch(paths, repeats_count,
image_processed_callback))
to_embedd = []
# save embeddings to list
for i, e, p in zip(idx, new_embeddings, paths):
if e is None:
# return it back to the embedding process embedding none
to_embedd.append((i, p))
else:
embeddings[i] = e
loop.close()
return embeddings
async def embedd_batch(
self,
file_paths: List[str],
n_repeats: int,
proc_callback: Callable[[bool], None] = None
) -> List[Optional[List[float]]]:
"""
Function perform embedding of a batch of images.
Parameters
----------
file_paths
A list of file paths for images to be embedded.
n_repeats
The index of retry. It is zero when batch is sent to server
for the first time. In case when first trial was not successful
we will send images again.
proc_callback
A function that is called after each image is fully processed
by either getting a successful response from the server,
getting the result from cache or skipping the image.
Returns
-------
embeddings
Array-like of float arrays (embeddings) for successfully embedded
images and Nones for skipped images.
Raises
------
EmbeddingCancelledException:
If cancelled attribute is set to True (default=False).
"""
requests = []
async with httpx.Client(timeout=self.timeouts,
base_url=self.server_url) as client:
for p in file_paths:
if self.cancelled:
raise EmbeddingCancelledException()
requests.append(
self._send_to_server(p, n_repeats, client, proc_callback))
embeddings = await asyncio.gather(*requests)
self._cache.persist_cache()
assert self.num_parallel_requests == 0
return embeddings
async def __wait_until_released(self) -> None:
while self.num_parallel_requests >= self.max_parallel:
await asyncio.sleep(0.1)
async def _send_to_server(
self,
image: str,
n_repeats: int,
client: httpx.Client,
proc_callback: Callable[[bool],
None] = None) -> Optional[List[float]]:
"""
Function get list of images objects. It send them to server and
retrieve responses.
Parameters
----------
image
Single image path.
n_repeats
The index of retry. It is zero when batch is sent to server
for the first time. In case when first trial was not successful
we will send images again.
client
HTTPX client that communicates with the server
proc_callback
A function that is called after each image is fully processed
by either getting a successful response from the server,
getting the result from cache or skipping the image.
Returns
-------
emb
Embedding. For images that are not successfully embedded returns
None.
"""
await self.__wait_until_released()
if self.cancelled:
raise EmbeddingCancelledException()
self.num_parallel_requests += 1
# load image
im = self._image_loader.load_image_bytes(image, self._im_size)
if im is None:
self.num_parallel_requests -= 1
return None
# if image in cache return it
cache_key = self._cache.md5_hash(im)
emb = self._cache.get_cached_result_or_none(cache_key)
if emb is None:
# gather responses
url = f"/image/{self._model}?machine={self.machine_id}" \
f"&session={self.session_id}&retry={n_repeats}"
emb = await self._send_request(client, im, url)
if emb is not None:
self._cache.add(cache_key, emb)
if proc_callback:
proc_callback(emb is not None)
self.num_parallel_requests -= 1
return emb
async def _send_request(self, client: httpx.Client, image: bytes,
url: str) -> Optional[List[float]]:
"""
This function sends a single request to the server.
Parameters
----------
client
HTTPX client that communicates with the server
image
Single image packed in sequence of bytes.
url
Rest of the url string.
Returns
-------
embedding
Embedding. For images that are not successfully embedded returns
None.
"""
headers = {
'Content-Type': 'image/jpeg',
'Content-Length': str(len(image))
}
try:
response = await client.post(url, headers=headers, data=image)
except ReadTimeout:
# it happens when server do not respond in 60 seconds, in
# this case we return None and images will be resent later
return None
except OSError:
# it happens when no connection and images cannot be sent to the
# server
# we count number of consecutive errors
self.count_errors -= 1
# if there is more than 10 consecutive errors it means that
# there is p:qly no connection so we stop with embedding
# with EmbeddingConnectionError
if self.count_errors <= 0:
self.num_parallel_requests = 0 # for safety reasons
raise EmbeddingConnectionError
return None
# we reset the counter at successful embedding
self.count_errors = self.initial_count_errors
return ServerEmbedder._parse_response(response)
@staticmethod
def _parse_response(response: httpx.Response) -> Optional[List[float]]:
"""
This function get response and extract embeddings out of them.
Parameters
----------
response
Response by the server
Returns
-------
embedding
Embedding. For images that are not successfully embedded returns
None.
"""
if response.content:
try:
cont = json.loads(response.content.decode('utf-8'))
return cont.get('embedding', None)
except JSONDecodeError:
# in case that embedding was not successful response is not
# valid JSON
return None
else:
return None