I have wrote a custom code to build a UNet architecture. To do so I have firstly subclassed the tf.keras.layers.Layer object to define an encoder convolutional block composed by a conv3D layer, a BatchNormalization layer and a Activation layer, similarly I defined a decoder inverse convolutional block composed by a Conv3DTranspose layer, a BatchNormalization layer, an Activation layer and a Concatenate layer. Finally I subclassed the tf.keras.Model object to define the full model, composed by 4 enconding blocks and 4 decoding blocks.
To checkpoint the model while training I have used the tf.keras.callbacks.ModelCheckpoint callback. However when a I try to load back the model (that in fact is still training) with tf.keras.models.load_model() I receive the following error: ValueError: No model found in config file.
Here the full code for the model definition, building and fitting:
import tensorflow as tf
# Encoder block
class ConvBlock(tf.keras.layers.Layer):
def __init__(self, n_filters, conv_size, conv_stride, **kwargs):
super(ConvBlock, self).__init__(**kwargs)
self.conv3D = tf.keras.layers.Conv3D(
filters=n_filters,
kernel_size=conv_size,
strides=conv_stride,
padding="same",
)
self.batch_norm = tf.keras.layers.BatchNormalization()
self.relu = tf.keras.layers.Activation("relu")
def call(self, inputs, training=None):
h = self.conv3D(inputs)
if training:
h = self.batch_norm(h)
h = self.relu(h)
return h
# Decoder block
class InvConvBlock(tf.keras.layers.Layer):
def __init__(self, n_filters, conv_size, conv_stride, activation, **kwargs):
super(InvConvBlock, self).__init__(**kwargs)
self.conv3D_T = tf.keras.layers.Conv3DTranspose(
filters=n_filters,
kernel_size=conv_size,
strides=conv_stride,
padding="same",
)
self.batch_norm = tf.keras.layers.BatchNormalization()
self.activ = tf.keras.layers.Activation(activation)
self.concat = tf.keras.layers.Concatenate(axis=-1)
def call(self, inputs, feat_concat=None, training=None):
h = self.conv3D_T(inputs)
if training:
h = self.batch_norm(h)
h = self.activ(h)
if feat_concat is not None:
h = self.concat([h, feat_concat])
return h
class UNet(tf.keras.Model):
def __init__(self, n_filters, e_size, e_stride, d_size, d_stride, **kwargs):
super(UNet, self).__init__(**kwargs)
# Encoder
self.conv_block_1 = ConvBlock(n_filters, e_size, e_stride)
self.conv_block_2 = ConvBlock(n_filters * 2, e_size, e_stride)
self.conv_block_3 = ConvBlock(n_filters * 4, e_size, (1, 1, 1))
self.conv_block_4 = ConvBlock(n_filters * 8, e_size, (1, 1, 1))
# Decoder
self.inv_conv_block_1 = InvConvBlock(n_filters * 4, d_size, (1, 1, 1), "relu")
self.inv_conv_block_2 = InvConvBlock(n_filters * 2, d_size, (1, 1, 1), "relu")
self.inv_conv_block_3 = InvConvBlock(n_filters, d_size, d_stride, "relu")
self.inv_conv_block_4 = InvConvBlock(1, d_size, d_stride, "sigmoid")
def call(self, inputs, **kwargs):
h1 = self.conv_block_1(inputs, **kwargs)
h2 = self.conv_block_2(h1, **kwargs)
h3 = self.conv_block_3(h2, **kwargs)
h = self.conv_block_4(h3, **kwargs)
h = self.inv_conv_block_1(h, feat_concat=h3, **kwargs)
h = self.inv_conv_block_2(h, feat_concat=h2, **kwargs)
h = self.inv_conv_block_3(h, feat_concat=h1, **kwargs)
h = self.inv_conv_block_4(h, **kwargs)
return h
model = UNet(
n_filters,
e_size,
e_stride,
d_size,
d_stride,
)
model.build((None, *input_shape, 1))
loss = tf.keras.losses.BinaryCrossentropy(from_logits=True)
optimizer = tf.keras.optimizers.Adam(learning_rate)
metrics = [tf.keras.metrics.Precision(), tf.keras.metrics.Recall()]
model.compile(
loss=loss,
optimizer=optimizer,
metrics=metrics,
)
CP_callback = tf.keras.callbacks.ModelCheckpoint(
f"{checkpoint_dir}/model.h5", save_freq='epoch', monitor="loss"
)
unet.fit(
data,
epochs=opts.epochs,
callbacks=[CP_callback],
)
To load the model I used the following code on another python console:
import tensorflow as tf
model = tf.keras.models.load_model(f'{checkpoint_dir}/model.h5')
but here I receive the above mentioned error. What am I missing? Or what am I doing wrong?
Thank you in advance for your help.
This is because you don't define the get_config method in your custom layers. For this check, this exited answer in SO.
Otherwise, you can save the trained weights (not the full model) and load the model as follows. In that case, you don't need to define this function. Please note, it's good practice to do, however. Here is a workaround for your problem:
# callback
tf.keras.callbacks.ModelCheckpoint('model.h5',
monitor='val_loss',
verbose= 1,
save_best_only=True,
mode= 'min',
save_weights_only=True) # <---- only save weight
# train
model = UNet(
n_filters,
e_size,
e_stride,
d_size,
d_stride,
)
model.compile(...)
model.fit(...)
# inference
model = UNet(
n_filters,
e_size,
e_stride,
d_size,
d_stride,
)
model.build((None, *input_shape, 1))
model.load_weights('model.h5')
For more details, see the documentation of Serialization and saving and also collab demonstration of François Chollet. Also, We've written an article about model subclassing and custom training stuff in tf 2.x, in the Save and Load section (at the bottom) of this article, we've demonstrated many strategies, here, hope that help.
Update
I've run your public colab notebook. Unfortunately, I am facing the same issue, and it's a bit weird and currently, I don't have the exact answer for saving the entire model in the ModelCheckpoint callback with Custom Layer even if we define the get_config() method.
However, there is another workaround that may come in handy for you. As we know there are two major ways to save tf models: (1). SaveModel and HDF5 format. The way is we choose the SaveMoedl format. Which is recommended by the way and safe to use.
The key difference between HDF5 and SavedModel is that HDF5 uses object configs to save the model architecture, while SavedModel saves the execution graph. Thus, SavedModels are able to save custom objects like subclassed models and custom layers without requiring the original code.
Now, as for your requirements, you are saving the entire model along with the best loss or val_loss in training time. For that, we can define a custom callback do save the model for lowest validation_loss (or whatever you want). As follows:
class SaveModelH5(tf.keras.callbacks.Callback):
def on_train_begin(self, logs=None):
self.val_loss = []
def on_epoch_end(self, epoch, logs=None):
current_val_loss = logs.get("val_loss")
self.val_loss.append(logs.get("val_loss"))
if current_val_loss <= min(self.val_loss):
print('Find lowest val_loss. Saving entire model.')
self.model.save('unet', save_format='tf') # < ----- Here
save_model = SaveModelH5()
unet.fit(.., callbacks=save_model)
Using
model.save('any_name', save_format=`tf`)
allows us create a any_name working directory, inside which it contains assets, saved_model.pb, and variables. The model architecture and training configuration, including the optimizer, losses, and metrics are stored in saved_model.pb. The weights are saved in the variables directory.
When saving the model and its layers, the SavedModel format stores the class name, call function, losses, and weights (and the config, if implemented). The call function defines the computation graph of the model/layer. In the absence of the model/layer config, the call function is used to create a model that exists like the original model which can be trained, evaluated, and used for inference. When we need to re-load the saved model, we can do as follows:
new_unet = tf.keras.models.load_model("unet", compile=False)
Colab.
I am building a simple Sequential model in Keras (tensorflow backend). During training I want to inspect the individual training batches and model predictions. Therefore, I am trying to create a custom Callback that saves the model predictions and targets for each training batch. However, the model is not using the current batch for prediction, but the entire training data.
How can I hand over only the current training batch to the Callback?
And how can I access the batches and targets that the Callback saves in self.predhis and self.targets?
My current version looks as follows:
callback_list = [prediction_history((self.x_train, self.y_train))]
self.model.fit(self.x_train, self.y_train, batch_size=self.batch_size, epochs=self.n_epochs, validation_data=(self.x_val, self.y_val), callbacks=callback_list)
class prediction_history(keras.callbacks.Callback):
def __init__(self, train_data):
self.train_data = train_data
self.predhis = []
self.targets = []
def on_batch_end(self, epoch, logs={}):
x_train, y_train = self.train_data
self.targets.append(y_train)
prediction = self.model.predict(x_train)
self.predhis.append(prediction)
tf.logging.info("Prediction shape: {}".format(prediction.shape))
tf.logging.info("Targets shape: {}".format(y_train.shape))
NOTE: this answer is outdated and only works with TF1. Check #bers's answer for a solution tested on TF2.
After model compilation, the placeholder tensor for y_true is in model.targets and y_pred is in model.outputs.
To save the values of these placeholders at each batch, you can:
First copy the values of these tensors into variables.
Evaluate these variables in on_batch_end, and store the resulting arrays.
Now step 1 is a bit involved because you'll have to add an tf.assign op to the training function model.train_function. Using current Keras API, this can be done by providing a fetches argument to K.function() when the training function is constructed.
In model._make_train_function(), there's a line:
self.train_function = K.function(inputs,
[self.total_loss] + self.metrics_tensors,
updates=updates,
name='train_function',
**self._function_kwargs)
The fetches argument containing the tf.assign ops can be provided via model._function_kwargs (only works after Keras 2.1.0).
As an example:
from keras.layers import Dense
from keras.models import Sequential
from keras.callbacks import Callback
from keras import backend as K
import tensorflow as tf
import numpy as np
class CollectOutputAndTarget(Callback):
def __init__(self):
super(CollectOutputAndTarget, self).__init__()
self.targets = [] # collect y_true batches
self.outputs = [] # collect y_pred batches
# the shape of these 2 variables will change according to batch shape
# to handle the "last batch", specify `validate_shape=False`
self.var_y_true = tf.Variable(0., validate_shape=False)
self.var_y_pred = tf.Variable(0., validate_shape=False)
def on_batch_end(self, batch, logs=None):
# evaluate the variables and save them into lists
self.targets.append(K.eval(self.var_y_true))
self.outputs.append(K.eval(self.var_y_pred))
# build a simple model
# have to compile first for model.targets and model.outputs to be prepared
model = Sequential([Dense(5, input_shape=(10,))])
model.compile(loss='mse', optimizer='adam')
# initialize the variables and the `tf.assign` ops
cbk = CollectOutputAndTarget()
fetches = [tf.assign(cbk.var_y_true, model.targets[0], validate_shape=False),
tf.assign(cbk.var_y_pred, model.outputs[0], validate_shape=False)]
model._function_kwargs = {'fetches': fetches} # use `model._function_kwargs` if using `Model` instead of `Sequential`
# fit the model and check results
X = np.random.rand(10, 10)
Y = np.random.rand(10, 5)
model.fit(X, Y, batch_size=8, callbacks=[cbk])
Unless the number of samples can be divided by the batch size, the final batch will have a different size than other batches. So K.variable() and K.update() can't be used in this case. You'll have to use tf.Variable(..., validate_shape=False) and tf.assign(..., validate_shape=False) instead.
To verify the correctness of the saved arrays, you can add one line in training.py to print out the shuffled index array:
if shuffle == 'batch':
index_array = _batch_shuffle(index_array, batch_size)
elif shuffle:
np.random.shuffle(index_array)
print('Index array:', repr(index_array)) # Add this line
batches = _make_batches(num_train_samples, batch_size)
The shuffled index array should be printed out during fitting:
Epoch 1/1
Index array: array([8, 9, 3, 5, 4, 7, 1, 0, 6, 2])
10/10 [==============================] - 0s 23ms/step - loss: 0.5670
And you can check if cbk.targets is the same as Y[index_array]:
index_array = np.array([8, 9, 3, 5, 4, 7, 1, 0, 6, 2])
print(Y[index_array])
[[ 0.75325592 0.64857277 0.1926653 0.7642865 0.38901153]
[ 0.77567689 0.13573623 0.4902501 0.42897559 0.55825652]
[ 0.33760938 0.68195038 0.12303088 0.83509441 0.20991668]
[ 0.98367778 0.61325065 0.28973401 0.28734073 0.93399794]
[ 0.26097574 0.88219054 0.87951941 0.64887846 0.41996446]
[ 0.97794604 0.91307569 0.93816428 0.2125808 0.94381495]
[ 0.74813435 0.08036688 0.38094272 0.83178364 0.16713736]
[ 0.52609421 0.39218962 0.21022047 0.58569125 0.08012982]
[ 0.61276627 0.20679494 0.24124858 0.01262245 0.0994412 ]
[ 0.6026137 0.25620512 0.7398164 0.52558182 0.09955769]]
print(cbk.targets)
[array([[ 0.7532559 , 0.64857274, 0.19266529, 0.76428652, 0.38901153],
[ 0.77567691, 0.13573623, 0.49025011, 0.42897558, 0.55825651],
[ 0.33760938, 0.68195039, 0.12303089, 0.83509439, 0.20991668],
[ 0.9836778 , 0.61325067, 0.28973401, 0.28734073, 0.93399793],
[ 0.26097575, 0.88219053, 0.8795194 , 0.64887846, 0.41996446],
[ 0.97794604, 0.91307569, 0.93816429, 0.2125808 , 0.94381493],
[ 0.74813437, 0.08036689, 0.38094273, 0.83178365, 0.16713737],
[ 0.5260942 , 0.39218962, 0.21022047, 0.58569127, 0.08012982]], dtype=float32),
array([[ 0.61276627, 0.20679495, 0.24124858, 0.01262245, 0.0994412 ],
[ 0.60261369, 0.25620511, 0.73981643, 0.52558184, 0.09955769]], dtype=float32)]
As you can see, there are two batches in cbk.targets (one "full batch" of size 8 and the final batch of size 2), and the row order is the same as Y[index_array].
Long edit (almost a new answer) for the following reasons:
Yu-Yang's 2017 answer relies on the private _make_train_function and _function_kwargs APIs, which work only in TF1 (and maybe in TF1 compatibility, so-called non-eager mode).
Similarly, Binyan Hu's 2020 answer relies on _make_test_function and does not work in TF2 by default (requiring non-eager mode as well).
My own Jan 2020 answer, which was already subject to several required configuration settings, seems to have stopped working with (or before) TF 2.5, and I was not able to make model.inputs or model.outputs work any longer.
Finally, the earlier version of this answer requires potentially expensive model evaluation to obtain the predictions for each batch. A similar solution to obtain activation histograms even led to OOM issues with repeated training of different models.
So I set out find a way to obtain all possible quantities (inputs, targets, predictions, activations), batch-wise, without using any private APIs. The aim was to be able to call .numpy() on the intended quantities, so Keras callbacks can run ordinary Python code to ease debugging (I suppose that is what this question is mainly about - for maximum performance, one would probably try to integrate as many computations as possible into TensorFlow's graph operations anyway).
This is the common base model for all solutions:
"""Demonstrate batch data access."""
import tensorflow as tf
from tensorflow import keras
class DataCallback(keras.callbacks.Callback):
"""This class is where all implementations differ."""
def tf_nan(dtype):
"""Create NaN variable of proper dtype and variable shape for assign()."""
return tf.Variable(float("nan"), dtype=dtype, shape=tf.TensorShape(None))
def main():
"""Run main."""
model = keras.Sequential([keras.layers.Dense(1, input_shape=(2,))])
callback = DataCallback()
model.compile(loss="mse", optimizer="adam")
model.fit(
x=tf.transpose(tf.range(7.0) + [[0.2], [0.4]]),
y=tf.transpose(tf.range(7.0) + 10 + [[0.5]]),
validation_data=(
tf.transpose(tf.range(11.0) + 30 + [[0.6], [0.7]]),
tf.transpose(tf.range(11.0) + 40 + [[0.9]]),
),
shuffle=False,
batch_size=3,
epochs=2,
verbose=0,
callbacks=[callback],
)
model.save("tmp.tf")
if __name__ == "__main__":
main()
The following three snippets show one possible solution each, each with their own pros and cons. The core trick is always the same: allocate a tf.Variable and use tf.Variable.assign to export the intended quantity, from some Keras code run in graph mode, into the callback. The methods differ slightly in callback initialization and (in one case) model compilation, and most importantly, in the quantities they can access, which is why I summarize them above each snippet.
Custom metric
Using a custom (fake) metric (similar to my Jan 2020 answer), while we cannot seem to access model.inputs nor model.outputs any more (and model.(_)targets does not even exist any longer), we can access y_true and y_pred, which represent the model targets and outputs:
[ ] Inputs/Samples (x)
[ ] Weights (w)
[+] Targets/Labels (y_true)
[+] Outputs/Predictions (y_pred)
[ ] All layers (or only final input/output layers)
"""Demonstrate batch data access using a custom metric."""
import tensorflow as tf
from tensorflow import keras
class DataCallback(keras.callbacks.Callback): # diff
"""Callback to operate on batch data from metric."""
def __init__(self):
"""Offer a metric to access batch data."""
super().__init__()
self.y_true = None
self.y_pred = None
def set_model(self, model):
"""Initialize variables when model is set."""
self.y_true = tf_nan(model.output.dtype)
self.y_pred = tf_nan(model.output.dtype)
def metric(self, y_true, y_pred):
"""Fake metric."""
self.y_true.assign(y_true)
self.y_pred.assign(y_pred)
return 0
def on_train_batch_end(self, _batch, _logs=None):
"""See keras.callbacks.Callback.on_train_batch_end."""
print("y_true =", self.y_true.numpy())
print("y_pred =", self.y_pred.numpy())
def on_train_end(self, _logs=None):
"""Clean up."""
del self.y_true, self.y_pred
def tf_nan(dtype):
"""Create NaN variable of proper dtype and variable shape for assign()."""
return tf.Variable(float("nan"), dtype=dtype, shape=tf.TensorShape(None))
def main():
"""Run main."""
model = keras.Sequential([keras.layers.Dense(1, input_shape=(2,))])
callback = DataCallback()
model.compile(loss="mse", optimizer="adam", metrics=[callback.metric]) # diff
model.fit(
x=tf.transpose(tf.range(7.0) + [[0.2], [0.4]]),
y=tf.transpose(tf.range(7.0) + 10 + [[0.5]]),
validation_data=(
tf.transpose(tf.range(11.0) + 30 + [[0.6], [0.7]]),
tf.transpose(tf.range(11.0) + 40 + [[0.9]]),
),
shuffle=False,
batch_size=3,
epochs=2,
verbose=0,
callbacks=[callback],
)
model.save("tmp.tf")
if __name__ == "__main__":
main()
Custom training step
A custom training step is what I used in an earlier version of this answer. The idea still works in principle, but y_pred can be expensive and it might make sense to use a custom metric (see above) if that is required.
[+] Inputs/Samples (x)
[+] Weights (w)
[+] Targets/Labels (y_true)
[~] Outputs/Predictions (y_pred) [expensive!]
[ ] All layers (or only final input/output layers)
"""Demonstrate batch data access using a custom training step."""
import tensorflow as tf
from tensorflow import keras
class DataCallback(keras.callbacks.Callback): # diff
"""Callback to operate on batch data from training step."""
def __init__(self):
"""Initialize tf.Variables."""
super().__init__()
self.x = None
self.w = None
self.y_true = None
self.y_pred = None
def set_model(self, model):
"""Wrap the model.train_step function to access training batch data."""
self.x = tf_nan(model.input.dtype)
# pylint:disable=protected-access (replace by proper dtype if you know it)
if model.compiled_loss._user_loss_weights is not None:
self.w = tf_nan(model.compiled_loss._user_loss_weights.dtype)
self.y_true = tf_nan(model.output.dtype)
self.y_pred = tf_nan(model.output.dtype)
model_train_step = model.train_step
def outer_train_step(data):
# https://github.com/keras-team/keras/blob/v2.7.0/keras/engine/training.py
x, y_true, w = keras.utils.unpack_x_y_sample_weight(data)
self.x.assign(x)
if w is not None:
self.w.assign(w)
self.y_true.assign(y_true)
result = model_train_step(data)
y_pred = model(x)
self.y_pred.assign(y_pred)
return result
model.train_step = outer_train_step
def on_train_batch_end(self, _batch, _logs=None):
"""See keras.callbacks.Callback.on_train_batch_end."""
print("x =", self.x.numpy())
if self.w is not None:
print("w =", self.w.numpy())
print("y_true =", self.y_true.numpy())
print("y_pred =", self.y_pred.numpy())
def on_train_end(self, _logs=None):
"""Clean up."""
del self.x, self.w, self.y_true, self.y_pred
def tf_nan(dtype):
"""Create NaN variable of proper dtype and variable shape for assign()."""
return tf.Variable(float("nan"), dtype=dtype, shape=tf.TensorShape(None))
def main():
"""Run main."""
model = keras.Sequential([keras.layers.Dense(1, input_shape=(2,))])
callback = DataCallback()
model.compile(loss="mse", optimizer="adam")
model.fit(
x=tf.transpose(tf.range(7.0) + [[0.2], [0.4]]),
y=tf.transpose(tf.range(7.0) + 10 + [[0.5]]),
validation_data=(
tf.transpose(tf.range(11.0) + 30 + [[0.6], [0.7]]),
tf.transpose(tf.range(11.0) + 40 + [[0.9]]),
),
shuffle=False,
batch_size=3,
epochs=2,
verbose=0,
callbacks=[callback],
)
model.save("tmp.tf")
if __name__ == "__main__":
main()
Custom layer call
A custom layer call is a super-flexible way of accessing each layer's inputs and outputs. The callback handles patching of the call functions for a list of layers. While we cannot access weights and targets (as these quantitities do not make sense at the level of individual layers), it allows us to access individual layer activations, which can be handy for questions such as How does one log activations using `tf.keras.callbacks.TensorBoard`?.
[+] Inputs/Samples (x)
[ ] Weights (w)
[ ] Targets/Labels (y_true)
[+] Outputs/Predictions (y_pred)
[+] All layers (or only final input/output layers)
"""Demonstrate batch data access using custom layer calls."""
import tensorflow as tf
from tensorflow import keras
class DataCallback(keras.callbacks.Callback): # diff
"""Callback to operate on batch data from selected (to be wrapped) layers."""
def __init__(self, layers):
"""Wrap the calls of an iterable of model layers to access layer batch data."""
super().__init__()
self.data = {}
self.inner_calls = {}
self.outer_calls = {}
for layer in layers:
self.data[layer] = {
"inputs": tf_nan(layer.input.dtype),
"outputs": tf_nan(layer.output.dtype),
}
self.inner_calls[layer] = layer.call
def outer_call(inputs, layer=layer, layer_call=layer.call):
self.data[layer]["inputs"].assign(inputs)
outputs = layer_call(inputs)
self.data[layer]["outputs"].assign(outputs)
return outputs
self.outer_calls[layer] = outer_call
def on_train_batch_begin(self, _epoch, _logs=None):
"""Wrap layer calls during each batch."""
for layer, call in self.outer_calls.items():
layer.call = call
def on_train_batch_end(self, _epoch, _logs=None):
"""Restore original layer calls for ModelCheckpoint, model.save, ..."""
for layer, call in self.inner_calls.items():
layer.call = call
for layer, data in self.data.items():
print("Layer =", layer)
print("Inputs =", data["inputs"].numpy())
print("Outputs =", data["outputs"].numpy())
def tf_nan(dtype):
"""Create NaN variable of proper dtype and variable shape for assign()."""
return tf.Variable(float("nan"), dtype=dtype, shape=tf.TensorShape(None))
def main():
"""Run main."""
model = keras.Sequential([keras.layers.Dense(1, input_shape=(2,))])
callback = DataCallback(model.layers) # diff
model.compile(loss="mse", optimizer="adam")
model.fit(
x=tf.transpose(tf.range(7.0) + [[0.2], [0.4]]),
y=tf.transpose(tf.range(7.0) + 10 + [[0.5]]),
validation_data=(
tf.transpose(tf.range(11.0) + 30 + [[0.6], [0.7]]),
tf.transpose(tf.range(11.0) + 40 + [[0.9]]),
),
shuffle=False,
batch_size=3,
epochs=2,
verbose=0,
callbacks=[callback],
)
model.save("tmp.tf")
if __name__ == "__main__":
main()
When to use which and open to-dos
I think the snippets above each solution nicely summarize what each approach is capable of. Generally,
a custom training step will be ideal to access the model input, such as batched dataset generators, effects of shuffling, etc;
a custom layer call is ideal to access the in-betweens of the model; and
a custom metric is ideal to access the outputs of the model.
I am fairly certain (but have not tried) that one can combine all approaches to be able to access all batch quantities simultaneously. I have not tested anything but training mode - each method can have further pros and cons relating to their usefulness in testing or prediction mode. Finally, I assume, but have not tested either, that their should be only minor differences between tf.keras and keras. Having tested this code on TF2.8.rc1 and Keras 2.8.0, which has moved the tf.keras code back into the keras pip package, and not using any private APIs, I believe this assumption is justified.
It would be great if this approach could be extended to access model.inputs and model.outputs again. Currently, I am getting errors such as this one:
TypeError: You are passing KerasTensor(...), an intermediate Keras symbolic input/output, to a TF API that does not allow registering custom dispatchers, such as tf.cond, tf.function, gradient tapes, or tf.map_fn. Keras Functional model construction only supports TF API calls that do support dispatching, such as tf.math.add or tf.reshape. Other APIs cannot be called directly on symbolic Kerasinputs/outputs. You can work around this limitation by putting the operation in a custom Keras layer call and calling that layer on this symbolic input/output.
Previous answer
From TF 2.2 on, you can use custom training steps rather than callbacks to achieve what you want. Here's a demo that works with tensorflow==2.2.0rc1, using inheritance to improve the keras.Sequential model. Performance-wise, this is not ideal as predictions are made twice, once in self(x, training=True) and once in super().train_step(data). But you get the idea.
This works in eager mode and does not use private APIs, so it should be pretty stable. One caveat is that you have to use tf.keras (standalone keras does not support Model.train_step), but I feel standalone keras is becoming more and more deprecated anyway. (In fact, tf.keras migrates to keras in TF2.8.)
"""Demonstrate access to Keras batch tensors in a tf.keras custom training step."""
import numpy as np
from tensorflow import keras
from tensorflow.keras import backend as K
from tensorflow.python.keras.engine import data_adapter
in_shape = (2,)
out_shape = (1,)
batch_size = 3
n_samples = 7
class SequentialWithPrint(keras.Sequential):
def train_step(self, original_data):
# Basically copied one-to-one from https://git.io/JvDTv
data = data_adapter.expand_1d(original_data)
x, y_true, w = data_adapter.unpack_x_y_sample_weight(data)
y_pred = self(x, training=True)
# this is pretty much like on_train_batch_begin
K.print_tensor(w, "Sample weight (w) =")
K.print_tensor(x, "Batch input (x) =")
K.print_tensor(y_true, "Batch output (y_true) =")
K.print_tensor(y_pred, "Prediction (y_pred) =")
result = super().train_step(original_data)
# add anything here for on_train_batch_end-like behavior
return result
# Model
model = SequentialWithPrint([keras.layers.Dense(out_shape[0], input_shape=in_shape)])
model.compile(loss="mse", optimizer="adam")
# Example data
X = np.random.rand(n_samples, *in_shape)
Y = np.random.rand(n_samples, *out_shape)
model.fit(X, Y, batch_size=batch_size)
print("X: ", X)
print("Y: ", Y)
Finally, here is a simpler example without inheritance:
"""Demonstrate access to Keras batch tensors in a tf.keras custom training step."""
import tensorflow as tf
IN_SHAPE = (2,)
OUT_SHAPE = (1,)
BATCH_SIZE = 3
N_SAMPLES = 7
def make_print_data_and_train_step(keras_model):
"""Return a train_step function that prints data batches."""
original_train_step = keras_model.train_step
def print_data_and_train_step(data):
# Adapted from https://git.io/JvDTv, skipping data_adapter.expand_1d
x, y_true, w = tf.keras.utils.unpack_x_y_sample_weight(data)
y_pred = keras_model(x, training=True)
# this is pretty much like on_train_batch_begin
tf.keras.backend.print_tensor(w, "Sample weight (w) =")
tf.keras.backend.print_tensor(x, "Batch input (x) =")
tf.keras.backend.print_tensor(y_true, "Batch output (y_true) =")
tf.keras.backend.print_tensor(y_pred, "Prediction (y_pred) =")
result = original_train_step(data)
# add anything here for on_train_batch_end-like behavior
return result
return print_data_and_train_step
# Model
model = tf.keras.Sequential([tf.keras.layers.Dense(OUT_SHAPE[0], input_shape=IN_SHAPE)])
model.train_step = make_print_data_and_train_step(model)
model.compile(loss="mse", optimizer="adam")
# Example data
X = tf.random.normal((N_SAMPLES, *IN_SHAPE))
Y = tf.random.normal((N_SAMPLES, *OUT_SHAPE))
model.fit(X, Y, batch_size=BATCH_SIZE)
print("X: ", X)
print("Y: ", Y)
Update: This approach has stopped working. See my other answer a number of solutions compatible with TF2.8 (and hopefully beyond).
One problem with #Yu-Yang's solution is that it relies on model._function_kwargs, which is not guaranteed to work as it is not part of the API. In particular, in TF2 with eager execution, session kwargs seem to be either not accepted at all or run preemptively due to eager mode.
Therefore, here is my solution tested on tensorflow==2.1.0. The trick is to replace fetches by a Keras metric, in which the assignment operations from fetches are made during training.
This even enables a Keras-only solution if the batch size divides the number of samples; otherwise, another trick has to be applied when initializing TensorFlow variables with a None shape, similar to validate_shape=False in earlier solutions (compare https://github.com/tensorflow/tensorflow/issues/35667).
Importantly, tf.keras behaves differently from keras (sometimes just ignoring assignments, or seeing variables as Keras symbolic tensors), so this updated solution takes care of both implementations (Keras==2.3.1 and tensorflow==2.1.0).
"""Demonstrate access to Keras symbolic tensors in a (tf.)keras.Callback."""
import numpy as np
import tensorflow as tf
use_tf_keras = True
if use_tf_keras:
from tensorflow import keras
from tensorflow.keras import backend as K
tf.config.experimental_run_functions_eagerly(False)
compile_kwargs = {"run_eagerly": False, "experimental_run_tf_function": False}
else:
import keras
from keras import backend as K
compile_kwargs = {}
in_shape = (2,)
out_shape = (1,)
batch_size = 3
n_samples = 7
class CollectKerasSymbolicTensorsCallback(keras.callbacks.Callback):
"""Collect Keras symbolic tensors."""
def __init__(self):
"""Initialize intermediate variables for batches and lists."""
super().__init__()
# Collect batches here
self.inputs = []
self.targets = []
self.outputs = []
# # For a pure Keras solution, we need to know the shapes beforehand;
# # in particular, batch_size must divide n_samples:
# self.input = K.variable(np.empty((batch_size, *in_shape)))
# self.target = K.variable(np.empty((batch_size, *out_shape)))
# self.output = K.variable(np.empty((batch_size, *out_shape)))
# If the shape of these variables will change (e.g., last batch), initialize
# arbitrarily and specify `shape=tf.TensorShape(None)`:
self.input = tf.Variable(0.0, shape=tf.TensorShape(None))
self.target = tf.Variable(0.0, shape=tf.TensorShape(None))
self.output = tf.Variable(0.0, shape=tf.TensorShape(None))
def on_batch_end(self, batch, logs=None):
"""Evaluate the variables and save them into lists."""
self.inputs.append(K.eval(self.input))
self.targets.append(K.eval(self.target))
self.outputs.append(K.eval(self.output))
def on_train_end(self, logs=None):
"""Print all variables."""
print("Inputs: ", *self.inputs)
print("Targets: ", *self.targets)
print("Outputs: ", *self.outputs)
#tf.function
def assign_keras_symbolic_tensors_metric(_foo, _bar):
"""
Return the assignment operations as a metric to have them evaluated by Keras.
This replaces `fetches` from the TF1/non-eager-execution solution.
"""
# Collect assignments as list of (dest, src)
assignments = (
(callback.input, model.inputs[0]),
(callback.target, model._targets[0] if use_tf_keras else model.targets[0]),
(callback.output, model.outputs[0]),
)
for (dest, src) in assignments:
dest.assign(src)
return 0
callback = CollectKerasSymbolicTensorsCallback()
metrics = [assign_keras_symbolic_tensors_metric]
# Example model
model = keras.Sequential([keras.layers.Dense(out_shape[0], input_shape=in_shape)])
model.compile(loss="mse", optimizer="adam", metrics=metrics, **compile_kwargs)
# Example data
X = np.random.rand(n_samples, *in_shape)
Y = np.random.rand(n_samples, *out_shape)
model.fit(X, Y, batch_size=batch_size, callbacks=[callback])
print("X: ", X)
print("Y: ", Y)
Inspired by the way tf.keras.callbacks.TesnsorBoard saves v1 (graph) summaries.
No variable assignments and no redundant metrics.
For use with tensorflow>=2.0.0, graph (disable eager) mode during evaluating.
Extensive operations on the numpy predictions can be implemented by overriding SavePrediction._pred_callback.
import numpy as np
import tensorflow as tf
from tensorflow import keras
tf.compat.v1.disable_eager_execution()
in_shape = (2,)
out_shape = (1,)
batch_size = 2
n_samples = 32
class SavePrediction(keras.callbacks.Callback):
def __init__(self):
super().__init__()
self._get_pred = None
self.preds = []
def _pred_callback(self, preds):
self.preds.append(preds)
def set_model(self, model):
super().set_model(model)
if self._get_pred is None:
self._get_pred = self.model.outputs[0]
def on_test_begin(self, logs):
# pylint: disable=protected-access
self.model._make_test_function()
# pylint: enable=protected-access
if self._get_pred not in self.model.test_function.fetches:
self.model.test_function.fetches.append(self._get_pred)
self.model.test_function.fetch_callbacks[self._get_pred] = self._pred_callback
def on_test_end(self, logs):
if self._get_pred in self.model.test_function.fetches:
self.model.test_function.fetches.remove(self._get_pred)
if self._get_pred in self.model.test_function.fetch_callbacks:
self.model.test_function.fetch_callbacks.pop(self._get_pred)
print(self.preds)
model = keras.Sequential([
keras.layers.Dense(out_shape[0], input_shape=in_shape)
])
model.compile(loss="mse", optimizer="adam")
X = np.random.rand(n_samples, *in_shape)
Y = np.random.rand(n_samples, *out_shape)
model.evaluate(X, Y,
batch_size=batch_size,
callbacks=[SavePrediction()])