I am pretty new to allennlp and I am struggling with building a model that does not seem to fit perfectly in the standard way of building model in allennlp.
I want to build a pipeline model using NLP. The pipeline consists basically of two models, let's call them A and B. First A is trained and based on the prediction of the full train A, B trained afterwards.
What I have seen is that people define two separate models, train both using the command line interface allennlp train ... in a shell script that looks like
# set a bunch of environment variables
...
allennlp train -s $OUTPUT_BASE_PATH_A --include-package MyModel --force $CONFIG_MODEL_A
# prepare environment variables for model b
...
allennlp train -s $OUTPUT_BASE_PATH_B --include-package MyModel --force $CONFIG_MODEL_B
I have two concerns about that:
This code is hard to debug
It's not very flexible. When I want to do a forward pass of the fully trained model I have write another script that bash script that does that.
Any ideas on how to do that in a better way?
I thought about using a python script instead of a shell script and invoke allennlp.commands.main(..) directly. Doing so at least you have a joint python module you can run using a debugger.
There are two possibilities.
If you're really just plugging the output of one model into the input of another, you could merge them together into one model and run it that way. You can do this with two already-trained models if you initialize the combined model with the two trained models using a from_file model. To do it at training time is a little harder, but not impossible. You would train the first model like you do now. For the second step, you train the combined model directly, with the inner first model's weights frozen.
The other thing you can do is use AllenNLP as a library, without the config files. We have a template up on GitHub that shows you how to do this. The basic insight is that everything you configure in one of the Jsonnet configuration files corresponds 1:1 to a Python class that you can use directly from Python. There is no requirement to use the configuration files. If you use AllenNLP this way, have much more flexibility, including chaining things together.
Related
I am using pytorch lightning to finetune t5 transformer on a specific task. However, I was not able to understand how the finetuning works. I always see this code :
tokenizer = AutoTokenizer.from_pretrained(hparams.model_name_or_path) model = AutoModelForSeq2SeqLM.from_pretrained(hparams.model_name_or_path)
I don't get how the finetuning is done, are they freezing the whole model and training the head only, (if so how can I change the head) or are they using the pre-trained model as a weight initializing? I have been looking for an answer for couple days already. Any links or help are appreciated.
If you are using PyTorch Lightning, then it won't freeze the head until you specify it do so. Lightning has a callback which you can use to freeze your backbone and training only the head module. See Backbone Finetuning
Also checkout Ligthning-Flash, it allows you to quickly build model for various text tasks and uses Transformers library for backbone. You can use the Trainer to specify which kind of finetuning you want to apply for your training.
Thanks
The official document of AllenNLP suggests specifying "validation_data_path" in the configuration file, but what if one wants to construct a dataset from a single source and then randomly split it into train and validation datasets with a given ratio?
Does AllenNLP support this? I would greatly appreciate your comments.
AllenNLP does not have this functionality yet, but we are working on some stuff to get there.
In the meantime, here is how I did it for the VQAv2 reader: https://github.com/allenai/allennlp-models/blob/main/allennlp_models/vision/dataset_readers/vqav2.py#L354
This reader supports Python slicing syntax where you, for example, specify a data_path as "my_source_file[:1000]" to take the first 1000 instances from my_source_file. You can also supply multiple paths by setting data_path: ["file1", "file2[:1000]", "file3[1000-"]]. You can probably steal the top two blocks in that file (line 354 to 369) and put them into your own dataset reader to achieve the same result.
I have a PPO policy based model that I train with RLLib using the Ray Tune API on some standard gym environments (with no fancy preprocessing). I have model checkpoints saved which I can load from and restore for further training.
Now, I want to export my model for production onto a system that should ideally have no dependencies on Ray or RLLib. Is there a simple way to do this?
I know that there is an interface export_model in the rllib.policy.tf_policy class, but it doesn't seem particularly easy to use. For instance, after calling export_model('savedir') in my training script, and in another context loading via model = tf.saved_model.load('savedir'), the resulting model object is troublesome (something like model.signatures['serving_default'](gym_observation) doesn't work) to feed the correct inputs into for evaluation. I'm ideally looking for a method that would allow for easy out of the box model loading and evaluation on observation objects
Once you have restored from checkpoint with agent.restore(**checkpoint_path**), you can use agent.export_policy_model(**output_dir**) to export the model as a .pb file and variables folder.
There are some things that I would like to modify in the OpenAI environments. If we use the Cartpole example then we can edit things that are in the class init function but with environments that use Box2D it doesn't seem to be as straightforward.
For example, consider the BipedalWalker environment.
In this case, how would I edit things like the SPEED_HIP or SPEED_KNEE variables?
Yes, you can modify or create new environments in gym. The simplest (but not recommended) way is to modify the constants in your local gym installation directly, but of course that's not really nice.
A nicer way is to download the bipedal walker environment file (from here) and save it to a file (say, my_bipedal_walker.py)
Then you modify the constants in the my_bipedal_walker.py file, and then just import it in your code (assuming you put the file in a path that is importable, or the same folder as your other code files):
import gym
from my_bipedal_walker import BipedalWalker
env = BipedalWalker()
Then you have the env variable being an instance of the environment, with your defined constants for the physics computation, which you can use with any RL algorithm.
An even nicer way would be making your custom environment available in the OpenAI gym registry, which you can do by following the instructions here
You can edit the bipedal walker environment just like you can modify the cartpole environment.
All you have to do is modify the constants for SPEED_HIP and SPEED_KNEE.
If you want to change how those constants are used in the locomotion of the agent, you might also want to tweak the step method.
After making changes to the code, when you instantiate the environment, the new instance will be using the modifications you made.
I was consulting several references to discover how I may output trained Weka models into Java source code so that I may use the classifiers I am training in actual code for research applications I have been developing.
As I was playing with Weka 3.7, I noticed that while it does output Java code to its main text buffer when use simpler classification (supervised in my case this time) methods such as J48 decision tree, it removes the option (rather, it voids it by removing the ability to checkmark it and fades the text) to output Java code for RandomTree and RandomForest (which are the ones that give me the best performance in my situation).
Note: I am clicking on the "More Options" button and checking "Output source code:".
Does Weka not allow you to output RandomTree or RandomForest as Java code? If so, why? Or if it does and just doesn't put it in the output buffer (since RF is multiple decision trees which I imagine it doesn't want to waste buffer space), how does one go digging up where in the file system Weka outputs java code by default?
Are there any tricks to get Weka to give me my trained RandomForest as Java code? Or is Serialization of the output *.model files my only hope when it comes to RF and RandomTree?
Thanks in advance to those who provide help.
NOTE: (As an addendum to the answer provided below) If you run across a similar situation (requiring you to use your trained classifier/ML model in your code), I recommend following the links posted in the answer that was provided in response to my question. If you do not specifically need the Java code for the RandomForest, as an example, de-serializing the model works quite nicely and fits into Java application code, fulfilling its task as a trained model/hardened algorithm meant to predict future unlabelled instances.
RandomTree and RandomForest can't be output as Java code. I'm not sure for the reasoning why, but they don't implement the "Sourceable" interface.
This explains a little about outputting a classifier as Java code: Link 1
This shows which classifiers can be output as Java code: Link 2
Unfortunately I think the easiest route will be Serialization, although, you could maybe try implementing "Sourceable" for other classifiers on your own.
Another, but perhaps inconvenient solution, would be to use Weka to build the classifier every time you use it. You wouldn't need to load the ".model" file, but you would need to load your training data and relearn the model. Here is a starters guide to building classifiers in your own java code http://weka.wikispaces.com/Use+WEKA+in+your+Java+code.
Solved the problem for myself by turning the output of WEKA's -printTrees option of the RandomForest classifier into Java source code.
http://pielot.org/2015/06/exporting-randomforest-models-to-java-source-code/
Since I am using classifiers with Android, all of the existing options had disadvantages:
shipping Android apps with serialized models didn't reliably work across devices
computing the model on the phone took too much resources
The final code will consist of three classes only: the class with the generated model + two classes to make the classification work.