Related
Doing things on Google Colab.
transformers: 4.10.2
pytorch-lightning: 1.2.7
import torch
from torch.utils.data import DataLoader
from transformers import BertJapaneseTokenizer, BertForSequenceClassification
import pytorch_lightning as pl
dataset_for_loader = [
{'data':torch.tensor([0,1]), 'labels':torch.tensor(0)},
{'data':torch.tensor([2,3]), 'labels':torch.tensor(1)},
{'data':torch.tensor([4,5]), 'labels':torch.tensor(2)},
{'data':torch.tensor([6,7]), 'labels':torch.tensor(3)},
]
loader = DataLoader(dataset_for_loader, batch_size=2)
for idx, batch in enumerate(loader):
print(f'# batch {idx}')
print(batch)
category_list = [
'dokujo-tsushin',
'it-life-hack',
'kaden-channel',
'livedoor-homme',
'movie-enter',
'peachy',
'smax',
'sports-watch',
'topic-news'
]
tokenizer = BertJapaneseTokenizer.from_pretrained(MODEL_NAME)
max_length = 128
dataset_for_loader = []
for label, category in enumerate(tqdm(category_list)):
# file ./text has lots of articles, categorized by category
# and they are just plain texts, whose content begins from forth line
for file in glob.glob(f'./text/{category}/{category}*'):
lines = open(file).read().splitlines()
text = '\n'.join(lines[3:])
encoding = tokenizer(
text,
max_length=max_length,
padding='max_length',
truncation=True
)
encoding['labels'] = label
encoding = { k: torch.tensor(v) for k, v in encoding.items() }
dataset_for_loader.append(encoding)
SEED=lambda:0.0
# random.shuffle(dataset_for_loader) # ランダムにシャッフル
random.shuffle(dataset_for_loader,SEED)
n = len(dataset_for_loader)
n_train = int(0.6*n)
n_val = int(0.2*n)
dataset_train = dataset_for_loader[:n_train]
dataset_val = dataset_for_loader[n_train:n_train+n_val]
dataset_test = dataset_for_loader[n_train+n_val:]
dataloader_train = DataLoader(
dataset_train, batch_size=32, shuffle=True
)
dataloader_val = DataLoader(dataset_val, batch_size=256)
dataloader_test = DataLoader(dataset_test, batch_size=256)
class BertForSequenceClassification_pl(pl.LightningModule):
def __init__(self, model_name, num_labels, lr):
super().__init__()
self.save_hyperparameters()
self.bert_sc = BertForSequenceClassification.from_pretrained(
model_name,
num_labels=num_labels
)
def training_step(self, batch, batch_idx):
output = self.bert_sc(**batch)
loss = output.loss
self.log('train_loss', loss)
return loss
def validation_step(self, batch, batch_idx):
output = self.bert_sc(**batch)
val_loss = output.loss
self.log('val_loss', val_loss)
def test_step(self, batch, batch_idx):
labels = batch.pop('labels')
output = self.bert_sc(**batch)
labels_predicted = output.logits.argmax(-1)
num_correct = ( labels_predicted == labels ).sum().item()
accuracy = num_correct/labels.size(0)
self.log('accuracy', accuracy)
def configure_optimizers(self):
return torch.optim.Adam(self.parameters(), lr=self.hparams.lr)
checkpoint = pl.callbacks.ModelCheckpoint(
monitor='val_loss',
mode='min',
save_top_k=1,
save_weights_only=True,
dirpath='model/',
)
trainer = pl.Trainer(
gpus=1,
max_epochs=10,
callbacks = [checkpoint]
)
model = BertForSequenceClassification_pl(
MODEL_NAME, num_labels=9, lr=1e-5
)
### (a) ###
# I think this is where I am doing fine-tuning
trainer.fit(model, dataloader_train, dataloader_val)
# this is to score after fine-tuning
test = trainer.test(test_dataloaders=dataloader_test)
print(f'Accuracy: {test[0]["accuracy"]:.2f}')
But I am not really sure how to do a test before fine-tuning, in order to compare two models before and after fine-tuning, in order to show how effective fine-tuning is.
Inserting the following two lines to ### (a) ###:
test = trainer.test(test_dataloaders=dataloader_test)
print(f'Accuracy: {test[0]["accuracy"]:.2f}')
I got this result:
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
<ipython-input-13-c8b2c67f2d5c> in <module>()
9
10 # 6-19
---> 11 test = trainer.test(test_dataloaders=dataloader_test)
12 print(f'Accuracy: {test[0]["accuracy"]:.2f}')
13
/usr/local/lib/python3.7/dist-packages/pytorch_lightning/trainer/trainer.py in test(self, model, test_dataloaders, ckpt_path, verbose, datamodule)
896 self.verbose_test = verbose
897
--> 898 self._set_running_stage(RunningStage.TESTING, model or self.lightning_module)
899
900 # If you supply a datamodule you can't supply train_dataloader or val_dataloaders
/usr/local/lib/python3.7/dist-packages/pytorch_lightning/trainer/trainer.py in _set_running_stage(self, stage, model_ref)
563 the trainer and the model
564 """
--> 565 model_ref.running_stage = stage
566 self._running_stage = stage
567
AttributeError: 'NoneType' object has no attribute 'running_stage'
I noticed that Trainer.fit() can take None as arguments other than model, so I tried this:
trainer.fit(model)
test=trainer.test(test_dataloaders=dataloader_test)
print(f'Accuracy: {test[0]["accuracy"]:.2f}')
The result:
MisconfigurationException: No `train_dataloader()` method defined. Lightning `Trainer` expects as minimum a `training_step()`, `train_dataloader()` and `configure_optimizers()` to be defined.
Thanks.
The Trainer needs to call its .fit() in order to set up a lot of things and then only you can do .test() or other methods.
You are right about putting a .fit() just before .test() but the fit call needs to a valid one. You have to feed a dataloader/datamodule to it. But since you don't want to do a training/validation in this fit call, just pass limit_[train/val]_batches=0 while Trainer construction.
trainer = Trainer(gpus=..., ..., limit_train_batches=0, limit_val_batches=0)
trainer.fit(model, dataloader_train, dataloader_val)
trainer.test(model, dataloader_test) # without fine-tuning
The fit call here will just set things up for you and skip training/validation. And then the testing follows. Next time run the same code but without the limit_[train/val]_batches, this will do the pretraining for you
trainer = Trainer(gpus=..., ...)
trainer.fit(model, dataloader_train, dataloader_val)
trainer.test(model, dataloader_test) # with fine-tuning
Clarifying a bit about .fit() taking None for all but model: Its not quite true - you must provide either a DataLoader or a DataModule.
I'm creating a custom yaml tag MyTag. It can contain any given valid yaml - map, scalar, anchor, sequence etc.
How do I implement class MyTag to model this tag so that ruamel parses the contents of a !mytag in exactly the same way as it would parse any given yaml? The MyTag instance just stores whatever the parsed result of the yaml contents is.
The following code works, and the asserts should should demonstrate exactly what it should do and they all pass.
But I'm not sure if it's working for the right reasons. . . Specifically in the from_yaml class method, is using commented_obj = constructor.construct_undefined(node) a recommended way of achieving this, and is consuming 1 and only 1 from the yielded generator correct? It's not just working by accident?
Should I instead be using something like construct_object, or construct_map or. . .? The examples I've been able to find tend to know what type it is constructing, so would either use construct_map or construct_sequence to pick which type of object to construct. In this case I effectively want to piggy-back of the usual/standard ruamel parsing for whatever unknown type there might be in there, and just store it in its own type.
import ruamel.yaml
from ruamel.yaml.comments import CommentedMap, CommentedSeq, TaggedScalar
class MyTag():
yaml_tag = '!mytag'
def __init__(self, value):
self.value = value
#classmethod
def from_yaml(cls, constructor, node):
commented_obj = constructor.construct_undefined(node)
flag = False
for data in commented_obj:
if flag:
raise AssertionError('should only be 1 thing in generator??')
flag = True
return cls(data)
with open('mytag-sample.yaml') as yaml_file:
yaml_parser = ruamel.yaml.YAML()
yaml_parser.register_class(MyTag)
yaml = yaml_parser.load(yaml_file)
custom_tag_with_list = yaml['root'][0]['arb']['k2']
assert type(custom_tag_with_list) is MyTag
assert type(custom_tag_with_list.value) is CommentedSeq
print(custom_tag_with_list.value)
standard_list = yaml['root'][0]['arb']['k3']
assert type(standard_list) is CommentedSeq
assert standard_list == custom_tag_with_list.value
custom_tag_with_map = yaml['root'][1]['arb']
assert type(custom_tag_with_map) is MyTag
assert type(custom_tag_with_map.value) is CommentedMap
print(custom_tag_with_map.value)
standard_map = yaml['root'][1]['arb_no_tag']
assert type(standard_map) is CommentedMap
assert standard_map == custom_tag_with_map.value
custom_tag_scalar = yaml['root'][2]
assert type(custom_tag_scalar) is MyTag
assert type(custom_tag_scalar.value) is TaggedScalar
standard_tag_scalar = yaml['root'][3]
assert type(standard_tag_scalar) is str
assert standard_tag_scalar == str(custom_tag_scalar.value)
And some sample yaml:
root:
- item: blah
arb:
k1: v1
k2: !mytag
- one
- two
- three-k1: three-v1
three-k2: three-v2
three-k3: 123 # arb comment
three-k4:
- a
- b
- True
k3:
- one
- two
- three-k1: three-v1
three-k2: three-v2
three-k3: 123 # arb comment
three-k4:
- a
- b
- True
- item: argh
arb: !mytag
k1: v1
k2: 123
# blah line 1
# blah line 2
k3:
k31: v31
k32:
- False
- string here
- 321
arb_no_tag:
k1: v1
k2: 123
# blah line 1
# blah line 2
k3:
k31: v31
k32:
- False
- string here
- 321
- !mytag plain scalar
- plain scalar
- item: no comment
arb:
- one1
- two2
In YAML you can have anchors and aliases, and it is perfectly fine to have an object be a child of itself (using an alias). If you want to dump the Python data structure data:
data = [1, 2, 4, dict(a=42)]
data[3]['b'] = data
it dumps to:
&id001
- 1
- 2
- 4
- a: 42
b: *id001
and for that anchors and aliases are necessary.
When loading such a construct, ruamel.yaml recurses into the nested data structures, but if the toplevel node has not caused a real object to be constructed to which the anchor can be made a reference, the recursive leaf cannot resolve the alias.
To solve that, a generator is used, except for scalar values. It first creates an empty object, then recurses and updates it values. In code calling the constructor a check is made to see if a generator is returned, and in that case next() is done on the data, and potential self-recursion "resolved".
Because you call construct_undefined(), you always get a generator. Practically that method could return a value if it detects a scalar node (which of course cannot recurse), but it doesn't. If it would, your code could then not load the following YAML document:
!mytag 1
without modifications that test if you get a generator or not, as is done in the code in ruamel.yaml calling the various constructors so it can handle both construct_undefined and e.g. construct_yaml_int (which is not a generator).
I am training a model with keras using the model.fit() method.
I would like to use multiple validation sets that should be validated on separately after each training epoch so that i get one loss value for each validation set. If possible they should be both displayed during training and as well be returned by the keras.callbacks.History() callback.
I am thinking of something like this:
history = model.fit(train_data, train_targets,
epochs=epochs,
batch_size=batch_size,
validation_data=[
(validation_data1, validation_targets1),
(validation_data2, validation_targets2)],
shuffle=True)
I currently have no idea how to implement this. Is it possible to achieve this by writing my own Callback? Or how else would you approach this problem?
I ended up writing my own Callback based on the History callback to solve the problem. I'm not sure if this is the best approach but the following Callback records losses and metrics for the training and validation set like the History callback as well as losses and metrics for additional validation sets passed to the constructor.
class AdditionalValidationSets(Callback):
def __init__(self, validation_sets, verbose=0, batch_size=None):
"""
:param validation_sets:
a list of 3-tuples (validation_data, validation_targets, validation_set_name)
or 4-tuples (validation_data, validation_targets, sample_weights, validation_set_name)
:param verbose:
verbosity mode, 1 or 0
:param batch_size:
batch size to be used when evaluating on the additional datasets
"""
super(AdditionalValidationSets, self).__init__()
self.validation_sets = validation_sets
for validation_set in self.validation_sets:
if len(validation_set) not in [3, 4]:
raise ValueError()
self.epoch = []
self.history = {}
self.verbose = verbose
self.batch_size = batch_size
def on_train_begin(self, logs=None):
self.epoch = []
self.history = {}
def on_epoch_end(self, epoch, logs=None):
logs = logs or {}
self.epoch.append(epoch)
# record the same values as History() as well
for k, v in logs.items():
self.history.setdefault(k, []).append(v)
# evaluate on the additional validation sets
for validation_set in self.validation_sets:
if len(validation_set) == 3:
validation_data, validation_targets, validation_set_name = validation_set
sample_weights = None
elif len(validation_set) == 4:
validation_data, validation_targets, sample_weights, validation_set_name = validation_set
else:
raise ValueError()
results = self.model.evaluate(x=validation_data,
y=validation_targets,
verbose=self.verbose,
sample_weight=sample_weights,
batch_size=self.batch_size)
for metric, result in zip(self.model.metrics_names,results):
valuename = validation_set_name + '_' + metric
self.history.setdefault(valuename, []).append(result)
which i am then using like this:
history = AdditionalValidationSets([(validation_data2, validation_targets2, 'val2')])
model.fit(train_data, train_targets,
epochs=epochs,
batch_size=batch_size,
validation_data=(validation_data1, validation_targets1),
callbacks=[history]
shuffle=True)
I tested this on TensorFlow 2 and it worked. You can evaluate on as many validation sets as you want at the end of each epoch:
class MyCustomCallback(tf.keras.callbacks.Callback):
def on_epoch_end(self, epoch, logs=None):
res_eval_1 = self.model.evaluate(X_test_1, y_test_1, verbose = 0)
res_eval_2 = self.model.evaluate(X_test_2, y_test_2, verbose = 0)
print(res_eval_1)
print(res_eval_2)
And later:
my_val_callback = MyCustomCallback()
# Your model creation code
model.fit(..., callbacks=[my_val_callback])
Considering the current keras docs, you can pass callbacks to evaluate and evaluate_generator. So you can call evaluate multiple times with different datasets.
I have not tested it, so I am happy if you comment your experiences with it below.
Does anyone know of a YAML deserializer that can provide position information for the constructed objects?
I know how to deserialize a YAML file into a Java object. Simple instructions on http://yamlbeans.sourceforge.net/.
However, I want to do some algorithmic validation on the deserialized object and report error back to the user pointing to the position in the YAML that cause the error.
Example:
=========YAML file==========
name: Nathan Sweet
age: 28
address: 4011 16th Ave S
=======JAVA class======
public class Contact {
public String name;
public int age;
public String address;
}
Imagine if I want to first load the yaml into Contact class and then validate the address against some repository and error back if its invalid. Something like:
'Line 3 Column 9: The address does not match valid entry in the database'
The problem is, currently there is no way to get the position inside a deserialized object from YAML.
Anyone know a solution to this issue?
Most YAML parsers, if they keep any information about positions around they drop it while constructing the language native objects.
In ruamel.yaml ¹, I keep more information around because I want to be able to round-trip with minimal loss of original layout (e.g. keeping comments and key order in mappings).
I don't keep information on individual key-value pairs, but I do on the "upper-left" position of a mapping². Because of the kept order of the mapping items you can give some rather nice feedback. Given an input file:
- name: anthon
age: 53
adres: Rijn en Schiekade 105
- name: Nathan Sweet
age: 28
address: 4011 16th Ave S
And a program that you call with the input file as argument:
#! /usr/bin/env python2.7
# coding: utf-8
# http://stackoverflow.com/questions/30677517/yaml-deserializer-with-position-information?noredirect=1#comment49491314_30677517
import sys
import ruamel.yaml
up_arrow = '↑'
def key_error(key, value, line, col, error, e='E'):
print('E[{}:{}]: {}'.format(line, col, error))
print('{}{}: {}'.format(' '*col, key, value))
print('{}{}'.format(' '*(col), up_arrow))
print('---')
def value_error(key, value, line, col, error, e='E'):
val_col = col + len(key) + 2
print('{}[{}:{}]: {}'.format(e, line, val_col, error))
print('{}{}: {}'.format(' '*col, key, value))
print('{}{}'.format(' '*(val_col), up_arrow))
print('---')
def value_warning(key, value, line, col, error):
value_error(key, value, line, col, error, e='W')
class Contact(object):
def __init__(self, vals):
for offset, k in enumerate(vals):
self.check(k, vals[k], vals.lc.line+offset, vals.lc.col)
for k in ['name', 'address', 'age']:
if k not in vals:
print('K[{}:{}]: {}'.format(
vals.lc.line+offset, vals.lc.col, "missing key: "+k
))
print('---')
def check(self, key, value, line, col):
if key == 'name':
if value[0].lower() == value[0]:
value_error(key, value, line, col,
'value should start with uppercase')
elif key == 'age':
if value < 50:
value_warning(key, value, line, col,
'probably too young for knowing ALGOL 60')
elif key == 'address':
pass
else:
key_error(key, value, line, col,
"unexpected key")
data = ruamel.yaml.load(open(sys.argv[1]), Loader=ruamel.yaml.RoundTripLoader)
for x in data:
contact = Contact(x)
giving you E(rrors), W(arnings) and K(eys missing):
E[0:8]: value should start with uppercase
name: anthon
↑
---
E[2:2]: unexpected key
adres: Rijn en Schiekade 105
↑
---
K[2:2]: missing key: address
---
W[4:7]: probably too young for knowing ALGOL 60
age: 28
↑
---
Which you should be able to parser in a calling program in any language to give feedback. The check method of course need adjusting to your requirements. This is not as nice as being to do that in the language the rest of your application is in, but it might be better than nothing.
In my experience handling the above format is certainly simpler than extending an existing (open source) YAML parser.
¹ Disclaimer: I am the author of that package
² I want to use that kind of information at some point to preserve spurious newlines, inserted for readability
In python, you can readily write custom Dumper/Loader objects and use them to load (or dump) your yaml code. You can have these objects track the file/line info:
import yaml
from collections import OrderedDict
class YamlOrderedDict(OrderedDict):
"""
An OrderedDict that was loaded from a yaml file, and is annotated
with file/line info for reporting about errors in the source file
"""
def _annotate(self, node):
self._key_locs = {}
self._value_locs = {}
nodeiter = node.value.__iter__()
for key in self:
subnode = nodeiter.next()
self._key_locs[key] = subnode[0].start_mark.name + ':' + \
str(subnode[0].start_mark.line+1)
self._value_locs[key] = subnode[1].start_mark.name + ':' + \
str(subnode[1].start_mark.line+1)
def key_loc(self, key):
try:
return self._key_locs[key]
except AttributeError, KeyError:
return ''
def value_loc(self, key):
try:
return self._value_locs[key]
except AttributeError, KeyError:
return ''
# Use YamlOrderedDict objects for yaml maps instead of normal dict
yaml.add_representer(OrderedDict, lambda dumper, data:
dumper.represent_dict(data.iteritems()))
yaml.add_representer(YamlOrderedDict, lambda dumper, data:
dumper.represent_dict(data.iteritems()))
def _load_YamlOrderedDict(loader, node):
rv = YamlOrderedDict(loader.construct_pairs(node))
rv._annotate(node)
return rv
yaml.add_constructor(yaml.resolver.BaseResolver.DEFAULT_MAPPING_TAG, _load_YamlOrderedDict)
Now when you read a yaml file, any mapping objects will be read as a YamlOrderedDict, which allows looking up the file location of keys in the mapping object. You can also add an iterator method like:
def iter_with_lines(self):
for key, val in self.items():
yield (key, val, self.key_loc(key))
...and now you can write a loop like:
for key,value,location in obj.iter_with_lines():
# iterate through the key/value pairs in a YamlOrderedDict, with
# the source file location
I have this YAML:
- company:
- id: toyota
- fullname: トヨタ自動車株式会社
- company:
- id: konami
- fullname: Konami Corporation
And I want to get the fullname of the company whose id is konami.
Using Ruby 1.9.2, what is the simplest/usual way to get it?
Note: In the rest of my code, I have been using require "yaml" so I would prefer to use the same library.
This works too and does not use iteration:
y = YAML.load_file('japanese_companies.yml')
result = y.select{ |x| x['company'].first['id'] == 'konami' }
result.first['company'].last['fullname'] # => "Konami Corporation"
Or if you have other attributes and you can't be sure fullname is the last one:
result.first['company'].select{ |x| x['fullname'] }.first['fullname']
I agree with Ray Toal, if you change your yml it becomes much easier. E.g.:
toyota:
fullname: トヨタ自動車株式会社
konami:
fullname: Konami Corporation
With the above yaml, fetching the fullname of konami becomes much easier:
y = YAML.load_file('test.yml')
y.fetch('konami')['fullname']
Your YAML is a little unconventional but we can compensate.
A brute force approach is (I'm not sure if this can be done without parsing the YAML):
require 'yaml'
YAML.parse_file(ARGV[0]).transform.each do |company|
properties = {}
company['company'].each {|h| properties = properties.merge(h)}
puts properties['fullname'] if properties['id'] == 'konami'
end
Pass your YAML file in as the first argument to this script.
Feel free to adapt into a method that takes the YAML as a string and returns the desired fullname. (A return is useful because it directly answers the OP's question of obtaining the first such company.)