def predict_dataset(
self, dataset) -> Generator[NetworkPredictionResult, None, None]:
dataset_gen = self.create_dataset_inputs(
dataset,
self.batch_size,
self.network_proto.features,
self.network_proto.backend.shuffle_buffer_size,
mode='test')
out = self.model.predict(dataset_gen, )
for softmax, params, output_seq_len in zip(*out):
softmax = np.roll(softmax, 1, axis=1) # fix bla
# decode encoded params from json. On python<=3.5 this are bytes, else it already is a str
enc_param = params[0]
enc_param = json.loads(
enc_param.decode("utf-8") if isinstance(enc_param, bytes
) else enc_param)
decoded = self.ctc_decoder.decode(softmax[:output_seq_len[0]])
# return prediction result
yield NetworkPredictionResult(
softmax=softmax,
output_length=output_seq_len,
decoded=decoded,
params=enc_param,
ground_truth=None,
)
def predict_raw_batch(self, x: np.array, len_x: np.array) -> Generator[NetworkPredictionResult, None, None]:
out = self.session.run(
[self.softmax, self.output_seq_len, self.decoded],
feed_dict={
self.inputs: x / 255,
self.input_seq_len: len_x,
self.dropout_rate: 0,
})
out = out[0:2] + [TensorflowModel.__sparse_to_lists(out[2])]
for sm, sl, dec in zip(*out):
pred = NetworkPredictionResult(softmax=sm,
output_length=sl,
decoded=dec,
)
pred.softmax = np.roll(pred.softmax, 1, axis=1)
l, s = pred.softmax, pred.output_length
pred.decoded = self.ctc_decoder.decode(l[:s])
yield pred
def predict_raw(self, x, len_x) -> Generator[NetworkPredictionResult, None, None]:
out = self.session.run(
[self.softmax, self.output_seq_len, self.decoded],
feed_dict={
self.inputs: x,
self.input_seq_len: len_x,
self.dropout_rate: 0,
})
out = out[0:2] + [TensorflowModel.__sparse_to_lists(out[2])]
for sm, sl, dec in zip(*out):
yield NetworkPredictionResult(softmax=sm,
output_length=sl,
decoded=dec,
)
def predict_raw_batch(self, x: np.array, len_x: np.array) -> Generator[NetworkPredictionResult, None, None]:
out = self.model.predict_on_batch(
[tf.convert_to_tensor(x / 255.0, dtype=tf.float32),
tf.convert_to_tensor(len_x, dtype=tf.int32),
tf.zeros((len(x), 1), dtype=tf.string)],
)
for sm, params, sl in zip(*out):
sl = sl[0]
sm = np.roll(sm, 1, axis=1)
decoded = self.ctc_decoder.decode(sm[:sl])
pred = NetworkPredictionResult(softmax=sm,
output_length=sl,
decoded=decoded,
)
yield pred
def predict_dataset(self) -> Generator[NetworkPredictionResult, None, None]:
out = self.session.run(
[self.softmax, self.output_seq_len, self.serialized_params, self.decoded, self.targets],
feed_dict={
self.dropout_rate: 0,
})
out = out[0:3] + list(map(TensorflowModel.__sparse_to_lists, out[3:5]))
for sm, length, param, dec, gt in zip(*out):
# decode encoded params from json. On python<=3.5 this are bytes, else it already is a str
enc_param = param[0]
enc_param = json.loads(enc_param.decode("utf-8") if isinstance(enc_param, bytes) else enc_param)
# return prediction result
yield NetworkPredictionResult(softmax=sm,
output_length=length,
decoded=dec,
params=enc_param,
ground_truth=self.codec.decode(gt) if gt is not None else None,
)