I will describe my intention here. I want to import BERT pretrained model via tf-hub function hub.module(bert_url, trainable = True) and utilize it for text classification task. I plan to use a large corpus to fine-tune weights of BERT as well as a few dense layers whose inputs are the BERT outputs. I would then like to freeze layers of BERT and train only the dense layers following BERT. How can I do this efficiently?
You mention Hub's TF1 API hub.Module, so I suppose you are writing TF1 code and using the TF1-compatible Hub assets google/bert/..., such as https://tfhub.dev/google/bert_cased_L-12_H-768_A-12/1
Are you going to have separate run of your program for the two phases of training? If so, maybe you can just drop trainable=True from the hub.Module call in the second run. This doesn't affect variable names, so you can restore the training result from the first run, including BERT's adjusted weights. (To be clear: the pre-trained weights shipped with the hub.Module are only used for initialization at the very start of training; restoring a checkpoint overrides them.)
Related
I trained a resnet model whose final layer has 3 outputs (multiclass classification). I want to use these model weights to pretrain a regression model which has the exact same architecture except the last layer, which has 1 output.
This seems like a very basic use case, but I do not see how to do this. Restoring a checkpoint gives an error since the architectures are not the same (mismatched shape). All other solutions I have found are either for TF1 (eg https://innerpeace-wu.github.io/2017/12/13/Tensorflow-Restore-partial-weights/) or using Keras .h5 restore.
How can I do this in TF2?
A layer (....) which is an input to the Conv operator producing the output array model/re_lu_1/Relu, is lacking min/max data, which is necessary for quantization. If accuracy matters, either target a non-quantized output format, or run quantized training with your model from a floating point checkpoint to change the input graph to contain min/max information. If you don't care about accuracy, you can pass --default_ranges_min= and --default_ranges_max= for easy experimentation.
For tensorflow 1.x, if you want to quantize, you have to place it with fake quantization nodes to activate the quantization of the model.
There are 3 phases of quantization:
Training part: load your model to graph => create training graph by contrib => train and store weights ckpt
Eval part: load your model to graph without weights => create eval graph => restore graph => export to frozen model
Toco/tflite convert frozen model to quantized model
However, the most important factor is the configuration of batch_normalization in the model. After trying multiple configuration, the best one is using batch_normalization without fused option from tensorflow.keras.layers.
The reason is because Tensorflow want to avoid the folding result to be quantized. Therefore, activation behind batchnorm wont work. Details in [here][1]
In short, this layer should be attached only under tensorflow.keras.layers.Conv2D with parsed activation param, which is Relu/Relu6/Identity
If you conduct the above process: Conv2d=>Activation=>BatchNorm
the layer will not yield errors does not have MinMax information
Is it possible to define a graph in native TensorFlow and then convert this graph to a Keras model?
My intention is simply combining (for me) the best of the two worlds.
I really like the Keras model API for prototyping and new experiments, i.e. using the awesome multi_gpu_model(model, gpus=4) for training with multiple GPUs, saving/loading weights or whole models with oneliners, all the convenience functions like .fit(), .predict(), and others.
However, I prefer to define my model in native TensorFlow. Context managers in TF are awesome and, in my opinion, it is much easier to implement stuff like GANs with them:
with tf.variable_scope("Generator"):
# define some layers
with tf.variable_scope("Discriminator"):
# define some layers
# model losses
G_train_op = ...AdamOptimizer(...)
.minimize(gloss,
var_list=tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope="Generator")
D_train_op = ...AdamOptimizer(...)
.minimize(dloss,
var_list=tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope="Discriminator")
Another bonus is structuring the graph this way. In TensorBoard debugging complicated native Keras models are hell since they are not structured at all. With heavy use of variable scopes in native TF you can "disentangle" the graph and look at a very structured version of a complicated model for debugging.
By utilizing this I can directly setup custom loss function and do not have to freeze anything in every training iteration since TF will only update the weights in the correct scope, which is (at least in my opinion) far easier than the Keras solution to loop over all the existing layers and set .trainable = False.
TL;DR:
Long story short: I like the direct access to everything in TF, but most of the time a simple Keras model is sufficient for training, inference, ... later on. The model API is much easier and more convenient in Keras.
Hence, I would prefer to set up a graph in native TF and convert it to Keras for training, evaluation, and so on. Is there any way to do this?
I don't think it is possible to create a generic automated converter for any TF graph, that will come up with a meaningful set of layers, with proper namings etc. Just because graphs are more flexible than a sequence of Keras layers.
However, you can wrap your model with the Lambda layer. Build your model inside a function, wrap it with Lambda and you have it in Keras:
def model_fn(x):
layer_1 = tf.layers.dense(x, 100)
layer_2 = tf.layers.dense(layer_1, 100)
out_layer = tf.layers.dense(layer_2, num_classes)
return out_layer
model.add(Lambda(model_fn))
That is what sometimes happens when you use multi_gpu_model: You come up with three layers: Input, model, and Output.
Keras Apologetics
However, integration between TensorFlow and Keras can be much more tighter and meaningful. See this tutorial for use cases.
For instance, variable scopes can be used pretty much like in TensorFlow:
x = tf.placeholder(tf.float32, shape=(None, 20, 64))
with tf.name_scope('block1'):
y = LSTM(32, name='mylstm')(x)
The same for manual device placement:
with tf.device('/gpu:0'):
x = tf.placeholder(tf.float32, shape=(None, 20, 64))
y = LSTM(32)(x) # all ops / variables in the LSTM layer will live on GPU:0
Custom losses are discussed here: Keras: clean implementation for multiple outputs and custom loss functions?
This is how my model defined in Keras looks in Tensorboard:
So, Keras is indeed only a simplified frontend to TensorFlow so you can mix them quite flexibly. I would recommend you to inspect source code of Keras model zoo for clever solutions and patterns that allows you to build complex models using clean API of Keras.
You can insert TensorFlow code directly into your Keras model or training pipeline! Since mid-2017, Keras has fully adopted and integrated into TensorFlow. This article goes into more detail.
This means that your TensorFlow model is already a Keras model and vice versa. You can develop in Keras and switch to TensorFlow whenever you need to. TensorFlow code will work with Keras APIs, including Keras APIs for training, inference and saving your model.
Is there a way to load a pretrained model in Tensorflow and remove the top layers in the network? I am looking at Tensorflow release r1.10
The only documentation I could find is with tf.keras.Sequential.pop
https://www.tensorflow.org/versions/r1.10/api_docs/python/tf/keras/Sequential#pop
I want to manually prune a pretrained network by removing bunch of top convolution layers and add a custom fully convoluted layer.
EDIT:
The model is ssd_mobilenet_v1_coco downloaded from Tensorflow Model Zoo. I have access to both the frozen_inference_graph.pb model file and checkpoint file.
I donot have access to the python code which is used to construct the model.
Thanks.
From inspecting the code, SSDMobileNetV1FeatureExtractor.extract_features redirects research.slim.nets:
from nets import mobilenet_v1 # nets will have to be on your PYTHONPATH
with tf.variable_scope('MobilenetV1',
reuse=self._reuse_weights) as scope:
with slim.arg_scope(
mobilenet_v1.mobilenet_v1_arg_scope(
is_training=None, regularize_depthwise=True)):
with (slim.arg_scope(self._conv_hyperparams_fn())
if self._override_base_feature_extractor_hyperparams
else context_manager.IdentityContextManager()):
_, image_features = mobilenet_v1.mobilenet_v1_base(
ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
final_endpoint='Conv2d_13_pointwise',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
use_explicit_padding=self._use_explicit_padding,
scope=scope)
The mobilenet_v1_base function takes a final_endpoint argument. Rather than prune the constructed graph, just construct the graph up until the endpoint you want.
In classify_image.py, the input image is fed with a loaded image in
predictions = sess.run(softmax_tensor,{'DecodeJpeg/contents:0': image_data})
What if I want to add new layers to the inception model and train the whole model again? Are the variables loaded from classify_image_graph_def.pb trainable? I saw that freeze_graph.py used convert_variables_to_constants to produce freezed graph. So can those loaded weights be trained again, are they constants? And how can I connect the input('shuffle_batch:0') to the inception model to the output of tf.train.shuffle_batch?
The model used in classify_image.py has its variables frozen into constants, and doesn't have any gradient ops, so it's not easy to turn it back into something trainable. You can see how we remove one layer and replace it with something trainable here:
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/examples/image_retraining/retrain.py
It's hard to generalize though. You'd be better off looking at some examples of fine-tuning here:
https://github.com/tensorflow/models/tree/master/inception#how-to-fine-tune-a-pre-trained-model-on-a-new-task