def forward(self, xNum, xCat):
# embed the auxiliary variables
embedConcat = nd.concat(
self.stationEmbedding(xCat[:,:,0]),
self.nYearEmbedding(xCat[:,:,1]),
self.nMonthEmbedding(xCat[:,:,2]),
self.mDayEmbedding(xCat[:,:,3]),
self.wdayEmbedding(xCat[:,:,4]),
self.nHourEmbedding(xCat[:,:,5]),
dim=2)
# The training and testing
embedTrain = embedConcat[:,0:168,:]
embedTest = embedConcat[:,168:,:]
# The input series for encoding
xNum = xNum.reshape((xNum.shape[0],xNum.shape[1],1))
inputSeries = nd.concat(xNum, embedTrain, dim=2)
inputSeries = nd.transpose(inputSeries, axes=(0,2,1))
for subTCN in self.encoder:
inputSeries = subTCN(inputSeries)
# The output
output = inputSeries
output = nd.transpose(output, axes=(0,2,1))
output = nd.reshape(output,(output.shape[0], 1,-1))
output = nd.broadcast_axis(output, axis=1, size=self.outputSize)
# the decoder
output=self.outputLayer(self.decoder(output, embedTest))
#output = nd.sum_axis(output, axis=2)
mu = nd.sum_axis(self.mu(output), axis=2)
sigma = nd.sum_axis(self.sigma(output), axis=2)
return mu, sigma
def forward(self, current, previous, doc_encode):
"""[summary]
Args:
current ([type]): h_j (batch_size, sentence_hidden_size * 2)
previous ([type]): s_j (batch_size, sentence_hidden_size * 2)
doc_encode ([type]): d (batch_size, ndoc_dims)
"""
# content: (batch_size, 1)
content = self.content_encoder(current)
# salience: (batch_size, sentence_hidden_size * 2)
salience = self.salience_encoder(doc_encode)
salience = current * salience
# salience: (batch_size,)
salience = nd.sum_axis(salience, -1)
# salience: (batch_size, 1)
salience = nd.expand_dims(salience, -1)
# novelty: (bathc_size, sentence_hidden_size * 2)
novelty = self.novelty_encoder(nd.tanh(previous))
novelty = current * novelty
# salience: (batch_size,)
novelty = nd.sum_axis(novelty, -1)
# salience: (batch_size, 1)
novelty = nd.expand_dims(novelty, -1)
# P: (batch_size, 1)
P = nd.sigmoid(content + salience - novelty)
return P
def forward(self, xNum, xCat):
# embed the auxiliary variables
embedConcat = nd.concat(self.stationEmbedding(xCat[:, :, 0]),
self.nYearEmbedding(xCat[:, :, 1]),
self.nMonthEmbedding(xCat[:, :, 2]),
self.mDayEmbedding(xCat[:, :, 3]),
self.wdayEmbedding(xCat[:, :, 4]),
self.nHourEmbedding(xCat[:, :, 5]),
dim=2)
# The training and testing
embedTrain = embedConcat[:, 0:self.inputSize,
] # only consider the id for the input series
embedTest = embedConcat[:, self.inputSize:, :]
# The input series for encoding
xNum = xNum.reshape((xNum.shape[0], xNum.shape[1], 1))
#inputSeries = nd.concat(xNum, embedTrain, dim=2)
inputSeries = xNum
inputSeries = nd.transpose(inputSeries, axes=(0, 2, 1))
for subTCN in self.encoder:
inputSeries = subTCN(inputSeries)
# The output
output = inputSeries
output = nd.transpose(output, axes=(0, 2, 1))
output = nd.reshape(output, (output.shape[0], 1, -1))
output = nd.broadcast_axis(output, axis=1, size=self.outputSize)
# the decoder
output = self.outputLayer(self.decoder(output, embedTest))
#output = nd.sum_axis(output, axis=2)
# The quantile outputs
outputQ10 = nd.sum_axis(self.Q10(output), axis=2)
outputQ50 = nd.sum_axis(self.Q50(output), axis=2)
outputQ90 = nd.sum_axis(self.Q90(output), axis=2)
return outputQ10, outputQ50, outputQ90
def preprocess(self, data, label):
gray_data = nd.sum_axis(nd.array([[[[0.3]], [[0.59]], [[0.11]]]]) *
data,
1,
keepdims=True)
gray_label = nd.sum_axis(nd.array([[[[1]], [[1]], [[1]]]]) * label, 1)
one_hot_label = self.one_hot(gray_label)
return gray_data, one_hot_label
def transform_fn(net, data, input_content_type, output_content_type):
try:
inp = json.loads(json.loads(data)[0])
bucket = inp['bucket']
prefix = inp['prefix']
s3_response = download_from_s3(bucket, prefix)
img = decode_response(s3_response)
img = nd.expand_dims(nd.transpose(img, (2, 0, 1)), 0)
img = nd.sum_axis(nd.array([[[[0.3]], [[0.59]], [[0.11]]]]) * img,
1,
keepdims=True)
batch = mx.io.DataBatch([img])
net.forward(batch)
raw_output = net.get_outputs()[0].asnumpy()
mask = np.argmax(raw_output, axis=(1))[0].astype(np.uint8)
output_prefix = os.path.join(
'output', '/'.join(prefix.split('/')[1:]).split('.')[0] +
'_MASK_PREDICTION.png')
push_to_s3(mask, bucket, output_prefix)
response = {'bucket': bucket, 'prefix': output_prefix}
except Exception as e:
response = {'Error': str(e)}
return json.dumps(response), output_content_type