Modular And Reusable Metrics All The Way
As we have run various training sessions, and we have approached modeling itself from various angles, it is time for us to consider the piece of TF functionality that ties it all together.
Ultimately, the end goal of any modeling, and of the actual training with our models, is to arrive to simply quantifiable “measures”, the losses, that we can then use in our repeated iterations of gradient calculations and subsequent tuning the weights of the model.
A loss is defined as some sort of difference, or “distance”, between the results of our model’s calculations and the given targets.
Specifically, we have been using crossentropy, (see a fun explanation here) as that “distance”.
Just as before, we need to prep our environment to run any meaningful code:
import tensorflow as tf
import dataset as qd
import custom as qc
import autograph as qa
ks = tf.keras
kl = ks.layers
In order to be able to experiment with multiple loss and metrics settings, we duplicate our tgt tensors in our newly defined adapter function of our dataset.
@tf.function
def adapter(d):
enc, dec, tgt = d['enc'], d['dec'], d['tgt']
return ((
enc.flat_values,
enc.row_splits,
dec.flat_values,
dec.row_splits,
tgt.flat_values,
tgt.row_splits,
), (
tgt.to_tensor(),
tgt.to_tensor(),
))
We also adjust our ToRagged layer.
Instead of leaving the tf.function decorator generic, which is allowing multiple version of the op to be generated based on the actual shapes of the input tensors, we restrict the generated op to only one version: the one taking triple input tensor pairs of any 1D shape.
class ToRagged(qc.ToRagged):
@tf.function(input_signature=[[
tf.TensorSpec(shape=[None], dtype=tf.int32),
tf.TensorSpec(shape=[None], dtype=tf.int64)
] * 3])
def call(self, x):
ys = []
for i in range(3):
i *= 2
fv, rs = x[i:i + 2]
ys.append(tf.RaggedTensor.from_row_splits(fv, rs))
return ys
Loss class
And now we are ready to tackle replacement loss and metric classes.
The Keras losses.Loss base class, that all the various other “losses” are derived from, has a call method with the above mentioned two arguments: the target tensor and the model’s output tensor.
Our replacement implementation of the method skips the various checks and validations from the canned version and simply flattens the two known tensors followed by calling directly the efficient graph op implementation of crossentropy:
class Loss(ks.losses.Loss):
@staticmethod
def xent(tgt, out):
tgt = tf.reshape(tf.cast(tgt, tf.int64), [-1])
s = tf.shape(out)
out = tf.reshape(out, [-1, s[-1]])
y = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=tgt,
logits=out)
return tf.reshape(y, s[:-1])
def __init__(self):
super().__init__(name='loss')
def call(self, tgt, out):
return self.xent(tgt, out)
Metric class
Our Metric class is even simpler, it adds the aggregating total and count variables and then delegates to calling our xent function (the same that our matching Loss uses).
class Metric(ks.metrics.Metric):
def __init__(self):
super().__init__(name='metric', dtype=tf.float32)
self.total = self.add_weight('total', initializer='zeros')
self.count = self.add_weight('count', initializer='zeros')
def update_state(self, tgt, out, sample_weight=None):
vs = Loss.xent(tgt, out)
self.total.assign_add(tf.math.reduce_sum(vs))
return self.count.assign_add(tf.cast(tf.size(vs), dtype=tf.float32))
def result(self):
return tf.math.divide_no_nan(self.total, self.count)
Updates to our model
Our model needs to be updated to use the newly defined components, including our new Loss and Metric classes.
As we include both the Debed and the Probe layers from our previous blogs, and they show up as a pair of output tensors respectively identifiable by their names, we can assign different losses and metrics to each.
We chose to use the same loss and metric for both. Keras, as expected, will sum both losses and metrics to calculate the end result.
To keep things simple, and since we have a pair of outputs, we also had to double the targets that the loss and metric would use:
def model_for(ps):
x = [ks.Input(shape=(), dtype='int32'), ks.Input(shape=(), dtype='int64')]
x += [ks.Input(shape=(), dtype='int32'), ks.Input(shape=(), dtype='int64')]
x += [ks.Input(shape=(), dtype='int32'), ks.Input(shape=(), dtype='int64')]
y = ToRagged()(x)
y = qc.Frames(ps)(y)
embed = qc.Embed(ps)
ye = qc.Encode(ps)(embed(y[:2]))
yd = qc.Decode(ps)(embed(y[2:]) + [ye[0]])
y = qc.Debed(ps)(yd)
ys = qa.Probe(ps)(yd)
m = ks.Model(inputs=x, outputs=[y, ys])
m.compile(
optimizer=ps.optimizer,
loss={'debed': ps.loss, 'probe': ps.loss},
metrics={'debed': [ps.metric], 'probe': [ps.metric]},
)
print(m.summary())
return m
We also need to update our parameters with instances of our specific Loss and Metric instances we want Keras to call on:
params = qc.params
params.update(
loss=Loss(),
metric=Metric(),
)
Training session
And now we are ready to start our training session.
We can confirm the model’s layers and connections and we can easily adjust the parameters to tailor the length of the sessions to our objectives.
ps = qd.Params(**params)
ps.num_epochs = 1
import masking as qm
qm.main_graph(ps, qc.dset_for(ps, adapter).take(10), model_for(ps))
Model: "model"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None,)] 0
__________________________________________________________________________________________________
input_2 (InputLayer) [(None,)] 0
__________________________________________________________________________________________________
input_3 (InputLayer) [(None,)] 0
__________________________________________________________________________________________________
input_4 (InputLayer) [(None,)] 0
__________________________________________________________________________________________________
input_5 (InputLayer) [(None,)] 0
__________________________________________________________________________________________________
input_6 (InputLayer) [(None,)] 0
__________________________________________________________________________________________________
to_ragged (ToRagged) [(None, None), (None 0 input_1[0][0]
input_2[0][0]
input_3[0][0]
input_4[0][0]
input_5[0][0]
input_6[0][0]
__________________________________________________________________________________________________
frames (Frames) [(5, 25), (None,), ( 125 to_ragged[0][0]
to_ragged[0][1]
to_ragged[0][2]
__________________________________________________________________________________________________
embed (Embed) multiple 120 frames[0][0]
frames[0][1]
frames[0][2]
frames[0][3]
__________________________________________________________________________________________________
encode (Encode) [(None, 25, 6), (Non 90516 embed[0][0]
embed[0][1]
__________________________________________________________________________________________________
decode (Decode) [(None, 15, 6), (Non 54732 embed[1][0]
embed[1][1]
encode[0][0]
__________________________________________________________________________________________________
debed (Debed) (None, None, None) 140 decode[0][0]
decode[0][1]
__________________________________________________________________________________________________
probe (Probe) (None, None, None) 140 decode[0][0]
decode[0][1]
==================================================================================================
Total params: 145,773
Trainable params: 145,648
Non-trainable params: 125
__________________________________________________________________________________________________
None
10/10 [==============================] - 11s 1s/step - loss: 5.8626 - debed_loss: 2.9345 - probe_loss: 2.9280 - probe_metric: 2.9239
With our TensorBoard callback in place, the model’s fit method will generate the standard summaries that TB can conveniently visualize.
If you haven’t run the code below, an already generated graph is here and runnable example.
#%load_ext tensorboard
#%tensorboard --logdir /tmp/q/logs
This concludes our blog, please see how to use Keras callbacks by clicking on the next blog.