def open_as_needed(filename, mode='rb'):
"""Return a file-object given either a filename or an object.
Handles opening with the right class based on the file extension.
"""
# Handle file-like objects
if hasattr(filename, 'read'):
# See if the file object is really gzipped or bzipped.
lead = filename.read(4)
# If we can seek, seek back to start, otherwise read all the data into an
# in-memory file-like object.
if hasattr(filename, 'seek'):
filename.seek(0)
else:
filename = BytesIO(lead + filename.read())
# If the leading bytes match one of the signatures, pass into the appropriate class.
try:
lead = lead.encode('ascii')
except AttributeError:
pass
if lead.startswith(b'\x1f\x8b'):
filename = gzip.GzipFile(fileobj=filename)
elif lead.startswith(b'BZh'):
filename = bz2.BZ2File(filename)
return filename
# This will convert pathlib.Path instances to strings
filename = str(filename)
if filename.endswith('.bz2'):
return bz2.BZ2File(filename, mode)
elif filename.endswith('.gz'):
return gzip.GzipFile(filename, mode)
else:
kwargs = {'errors': 'surrogateescape'} if mode != 'rb' else {}
return open(filename, mode, **kwargs)
def build_model(**params):
# List of parameters
if 'density' not in params: params['density'] = int((params['lstmsize']//1.5)*2)
if 'activation' not in params: params['activation'] = 'relu'
if 'twice' not in params: params['twice'] = False
if 'optimizer' not in params: params['optimizer'] = 'adam'
if 'shuffle' not in params: params['shuffle'] = False
# Model definition
model = Sequential()
model.add(LSTM(params['lstmsize'], input_shape=(92,5,), return_sequences=params['twice']))
if 'dropout' in params:
model.add(Dropout(params['dropout']))
if params['twice']:
model.add(LSTM(params['lstmsize']))
if 'dropout' in params:
model.add(Dropout(params['dropout']))
model.add(Dense(params['density'], activation=params['activation']))
if 'full_density' in params and params['full_density']:
density = params['density']//2
while density >= 12:
model.add(Dense(density, activation=params['activation']))
density //= 2
model.add(Dense(5, activation='linear'))
model.compile(loss='mse', optimizer=params['optimizer'])
storage_client = storage.Client()
bucket = storage_client.get_bucket(env["BUCKET"])
blob = bucket.blob(f"models/weights/{params['symbol']}_models/{params['tag']}.hdf5")
temp_file = BytesIO()
# Este pequeño hack se me hizo necesario para sortear el check que "load_weights" le hace al nombre del archivo.
def endswith(suffix, start=0, end=-1):
return True
temp_file.endswith = endswith
blob.download_to_file(temp_file)
model.load_weights(temp_file)
return model
