Related
I have my encode function that looks like this:
from transformers import BertTokenizer, BertModel
MODEL = 'bert-base-multilingual-uncased'
tokenizer = BertTokenizer.from_pretrained(MODEL)
def encode(texts, tokenizer=tokenizer, maxlen=10):
# import pdb; pdb.set_trace()
inputs = tokenizer.encode_plus(
texts,
return_tensors='tf',
return_attention_masks=True,
return_token_type_ids=True,
pad_to_max_length=True,
max_length=maxlen
)
return inputs['input_ids'], inputs["token_type_ids"], inputs["attention_mask"]
I want to get my data encoded on the fly by doing this:
x_train = (tf.data.Dataset.from_tensor_slices(df_train.comment_text.astype(str).values)
.map(encode))
However, this chucks the error:
ValueError: Input is not valid. Should be a string, a list/tuple of strings or a list/tuple of integers.
Now from my understanding when I set a breakpoint inside encode it was because I was sending a non-numpy array. How do I get huggingface transformers to play nice with tensorflow strings as inputs?
If you need a dummy dataframe here it is:
df_train = pd.DataFrame({'comment_text': ['Today was a good day']*5})
What I tried
So I tried to use from_generator so that I can parse in the strings to the encode_plus function. However, this does not work with TPUs.
AUTO = tf.data.experimental.AUTOTUNE
def get_gen(df):
def gen():
for i in range(len(df)):
yield encode(df.loc[i, 'comment_text']) , df.loc[i, 'toxic']
return gen
shapes = ((tf.TensorShape([maxlen]), tf.TensorShape([maxlen]), tf.TensorShape([maxlen])), tf.TensorShape([]))
train_dataset = tf.data.Dataset.from_generator(
get_gen(df_train),
((tf.int32, tf.int32, tf.int32), tf.int32),
shapes
)
train_dataset = train_dataset.batch(BATCH_SIZE).prefetch(AUTO)
Version Info:
transformers.__version__, tf.__version__ => ('2.7.0', '2.1.0')
the tokenizer of bert works on a string, a list/tuple of strings or a list/tuple of integers. So, check is your data getting converted to string or not. To apply tokenizer on whole dataset I used Dataset.map, but this runs on graph mode. So, I need to wrap it in a tf.py_function. The tf.py_function will pass regular tensors (with a value and a .numpy() method to access it), to the wrapped python function. My data was getting converted to bytes after using py_function hence I applied tf.compat.as_str to convert bytes to string.
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
def encode(lang1, lang2):
lang1 = tokenizer.encode(tf.compat.as_str(lang1.numpy()), add_special_tokens=True)
lang2 = tokenizer.encode(tf.compat.as_str(lang2.numpy()), add_special_tokens=True)
return lang1, lang2
def tf_encode(pt, en):
result_pt, result_en = tf.py_function(func = encode, inp = [pt, en], Tout=[tf.int64, tf.int64])
result_pt.set_shape([None])
result_en.set_shape([None])
return result_pt, result_en
train_dataset = dataset3.map(tf_encode)
BUFFER_SIZE = 200
BATCH_SIZE = 64
train_dataset = train_dataset.shuffle(BUFFER_SIZE).padded_batch(BATCH_SIZE,
padded_shapes=(60, 60))
a,p = next(iter(train_dataset))
When you create the tensorflow dataset with: tf.data.Dataset.from_tensor_slices(df_train.comment_text.astype(str).values)
tensorflow converts your strings into tensors of string type which is not an accepted input of of tokenizer.encode_plus. Like the error message says it only accepts a string, a list/tuple of strings or a list/tuple of integers. You can verify this by adding a print(type(texts)) inside your encode function (Output:<class 'tensorflow.python.framework.ops.Tensor'>).
I'm not sure what your follow up plan is and why you need a tf.data.Dataset, but you have to encode your input before you turn it into a tf.data.Dataset:
import tensorflow as tf
from transformers import BertTokenizer, BertModel
MODEL = 'bert-base-multilingual-uncased'
tokenizer = BertTokenizer.from_pretrained(MODEL)
texts = ['Today was a good day', 'Today was a bad day',
'Today was a rainy day', 'Today was a sunny day',
'Today was a cloudy day']
#inputs['input_ids'], inputs["token_type_ids"], inputs["attention_mask"]
inputs = tokenizer.batch_encode_plus(
texts,
return_tensors='tf',
return_attention_masks=True,
return_token_type_ids=True,
pad_to_max_length=True,
max_length=10
)
dataset = tf.data.Dataset.from_tensor_slices((inputs['input_ids'],
inputs['attention_mask'],
inputs['token_type_ids']))
print(type(dataset))
Output:
<class 'tensorflow.python.data.ops.dataset_ops.TensorSliceDataset'>
I had this exact error but my mistake was simple, I had a few NaNs in my texts.
So make sure to check if there are NaNs in your texts dataframe.
I am adding this summarization of my issue to make it easier to understand:
I want to do exactly what is done in the following tensorflow example:
https://www.tensorflow.org/guide/datasets
# Reads an image from a file, decodes it into a dense tensor, and resizes it
# to a fixed shape.
def _parse_function(filename, label):
image_string = tf.read_file(filename)
image_decoded = tf.image.decode_jpeg(image_string)
image_resized = tf.image.resize_images(image_decoded, [28, 28])
return image_resized, label
# A vector of filenames.
filenames = tf.constant(["/var/data/image1.jpg", "/var/data/image2.jpg", ...])
# `labels[i]` is the label for the image in `filenames[i].
labels = tf.constant([0, 37, ...])
dataset = tf.data.Dataset.from_tensor_slices((filenames, labels))
dataset = dataset.map(_parse_function)
The only differences are: I read the data from CSV that has many more features and then I call the map method:
dataset = tf.data.experimental.make_csv_dataset(file_pattern=CSV_PATH_TRAIN,
batch_size=2,
header=True,
label_name = 'label').map(_parse_function)
How does my _parse_function need to look like? How do I access the image path features, updates it to be an image presentation and return a modified numeric matrix feature of the image without changing anything at the other features?
thanks,
eilalan
==================Here are my code tries:==================
My code reads a CSV with feature columns and label. One of the features is image path, the others are strings.
The image path need to be processed into image numbers matrix.
I have tried doing so with the following options. In both ways tf.read_file fails with the input dimension error.
My question is how to pass one image at a time into the map methods
def read_image_png_option_1(image_path, depth=3, scale=False):
"""Reads the image from image_path (tf.string tensor) [jpg image].
Cast the result to float32 and if scale=True scale it in [-1,1]
using scale_image. Otherwise the values are in [0,1]
Reuturn:
the decoded jpeg image, casted to float32
"""
image = tf.image.convert_image_dtype(
tf.image.decode_png(tf.read_file(image_path), channels=depth),
dtype=tf.float32)
if scale:
image = scale_image(image)
return image
def read_image_png_option_2(features, depth=3, scale=False):
"""Reads the image from image_path (tf.string tensor) [jpg image].
Cast the result to float32 and if scale=True scale it in [-1,1]
using scale_image. Otherwise the values are in [0,1]
Reuturn:
the decoded jpeg image, casted to float32
"""
image = tf.image.convert_image_dtype(
tf.image.decode_png(tf.read_file(features['image']), channels=depth),
dtype=tf.float32)
if scale:
image = scale_image(image)
features['image'] = image
return features
def make_input_fn(fileName,batch_size=8, perform_shuffle=True):
"""An input function for training """
def _input_fn():
def decode_csv(line):
print('line is ',line)
filename_col,label_col,gender_col,ethinicity = tf.decode_csv(line,
[[""]]*amount_of_columns_csv,
field_delim=",",
na_value='NA',
select_cols=None)
image_col = read_image_png_option_1(filename_col)
d = dict(zip(['image','label','gender','ethinicity'], [image_col,label_col,gender_col,ethinicity])), label
return d
## OPTION 1:
# filenames could be more than one
# dataset = tf.data.TextLineDataset(filenames=fileName).skip(1).batch(batch_size).map(decode_csv)
## OPTION 2:
dataset = tf.data.experimental.make_csv_dataset(file_pattern=CSV_PATH_TRAIN,
batch_size=2,
header=True,
label_name = 'label').map(read_image_png_option_2)
#select_columns=[0,1]) #[tf.string,tf.string,tf.string,tf.string])
if perform_shuffle:
dataset = dataset.shuffle(buffer_size=256)
return dataset
return _input_fn()
train_input_fn = lambda: make_input_fn(CSV_PATH_TRAIN)
train_spec = tf.estimator.TrainSpec(input_fn=train_input_fn, max_steps=50)
eval_input_fn = lambda: make_input_fn(CSV_PATH_VAL)
eval_spec = tf.estimator.EvalSpec(eval_input_fn)
feature_columns = [tf.feature_column.numeric_column("image",shape=(224,224)), # here i need a pyhton method to transform
tf.feature_column.categorical_column_with_vocabulary_list("gender", ["ww","ee"]),
tf.feature_column.categorical_column_with_vocabulary_list("ethinicity",["xx","yy"])]
estimator = tf.estimator.DNNClassifier(feature_columns=feature_columns,hidden_units=[1024, 512, 256],warm_start_from=ws)
tf.estimator.train_and_evaluate(estimator, train_spec=train_spec, eval_spec=eval_spec)
Error for option 2:
ValueError: Shape must be rank 0 but is rank 1 for 'ReadFile' (op: 'ReadFile') with input shapes: [2].
Error for option 1:
ValueError: Shape must be rank 0 but is rank 1 for 'ReadFile' (op: 'ReadFile') with input shapes: [?].
Any help is appreciated.
Thanks
First you need to read the CSV file into dataset.
Then for each row in your CSV you can call your parse function.
def getInput(fileList):
# returns a dataset containing list of filenames
files = tf.data.Dataset.from_tensor_slices(fileList)
# Returs a dataset containing list of rows taken from all the files in file list.
# dataset is filled dynamically and not all entries are read at once
dataset = files.interleave(tf.data.TextLineDataset)
# call parse function for each row
# returned dataset will contain list of whatever the parse function is returning for the row
# we want the image path to be converted to decoded image in parse function
dataset = dataset.map(_parse_function, num_parallel_calls=8)
# return an iterator for the dataset which will be used to get elements.
return dataset.make_one_shot_iterator().get_next()
The parse function will be passed only one parameter that will be a single row from the CSV file. You need to decode the CSV and do further processing on each value.
Let's say you have 3 columns in your CSV each being a string.
def _parse_function(value):
columns_default = [[""], [""], [""]]
# this will be a tensor of columns in the row
columns = tf.decode_csv(value, record_defaults=columns_default,
field_delim=',')
col_names = ["label", "imagepath", "c3"]
features = dict(zip(col_names, columns))
for f, tensor in features.items():
# process imagepath to decoded image
if f == "imagepath":
image_string = tf.read_file(tensor)
image_decoded = tf.image.decode_jpeg(image_string)
image_resized = tf.image.resize_images(image_decoded, [28, 28])
features[f] = image_resized
labels = tf.equal(features.pop('label'), "1")
labels = tf.expand_dims(labels, 0)
return features, labels
Edit:
Explanation for comment:
Dataset object simply contains a list of elements. The elements can be tensors or a tuple of tensors etc. Tensor object can contain anything. It could represent a single feature, a single record or a batch of record. Further dataset API provide handy methods to manipulate the elements within.
If you are using dataset with another API like estimator then they expect the dataset elements to be in specific format which is what need to return from our input function for eg.
https://www.tensorflow.org/api_docs/python/tf/estimator/Estimator#train
I have edited my code block above to describe what dataset object at each step will contain.
From what I understand is that you have image path as one of the field in your CSV and you want to convert that path into an actual decoded image which you will use as one of the feature.
Since the image is going to be just one of the feature, you should not try to create a dataset using image files alone. Dataset object will include all your features at once.
So doing this would be incorrect:
files = tf.data.Dataset.from_tensor_slices(ds['imagepath'])
dataset = files.interleave(tf.data.TextLineDataset)
If you are using make_csv() function to read your csv then it will convert each row of your csv into one record where one record will contain list of all features, same as columns of csv.
So each element in the returned dataset should contain a single tensor containing all your features.
Here your image path will be one of the features. now you want to transform that image path to decoded image.
I suppose you can do it by applying a parse function to elements of dataset using map() function but it will be slightly tricky as all your features are already packed inside a single tensor.
I currently am using the Tensorflow dataset api to perform some augmentations to images at a specified path. The filename itself contains information that states whether to augment the file or not. So what I want to do is read in the files from the dataset and for each file, perform a find within the filename and if I find a specific substring, then set a bool flag and replace the substring with "".
The error I get is:
AttributeError: 'Tensor' object has no attribute 'find'
I can't perform a "find" on the tensor with dtype string entries because find is not a part of the Tensor, so I am trying to figure out how I can go about performing the above action. I have shared some code below that I think demonstrates what I am trying to do. Performance is important, so I would prefer to do this the correct way if anyone sees that I am going about doing this via the Dataset API incorrectly.
def preproc_img(filenames):
def parse_fn(filename):
augment_inst = False
if cfg.SPLIT_INTO_INST:
#*****************************************************
#*** THIS IS WHERE THE LOGIC IS CURRENTLY BREAKING ***
#*****************************************************
if filename.find('_data_augmentation') != -1:
augment_inst = True
filename = filename.replace('_data_augmentation', '')
image_string = tf.read_file(filename)
img = tf.image.decode_image(image_string, channels=3)
return dict(zip([filename], [img]))
dataset = tf.data.Dataset.from_tensor_slices(filenames)
dataset = dataset.map(parse_fn)
iterator = dataset.make_one_shot_iterator()
return iterator.get_next()
def perform_train():
if __name__ == '__main__':
filenames = helper.get_image_paths()
next_batch = preproc_img(filenames)
with tf.Session() as sess:
with sess .graph.as_default():
sess.run(tf.local_variables_initializer())
sess.run(tf.global_variables_initializer())
dat = sess.run(next_batch)
# I would now go about calling any of my tf op code below
You can use tf.regex_replace for replacing text in a tf.string tensor.
filename = tf.regex_replace(filename, "_data_augmentation", "")
For TF 2.0
filename = tf.strings.regex_replace(filename, "_data_augmentation", "")
I am trying to convert an encoded byte string back into the original array in the tensorflow graph (using tensorflow operations) in order to make a prediction in a tensorflow model. The array to byte conversion is based on this answer and it is the suggested input to tensorflow model prediction on google cloud's ml-engine.
def array_request_example(input_array):
input_array = input_array.astype(np.float32)
byte_string = input_array.tostring()
string_encoded_contents = base64.b64encode(byte_string)
return string_encoded_contents.decode('utf-8')}
Tensorflow code
byte_string = tf.placeholder(dtype=tf.string)
audio_samples = tf.decode_raw(byte_string, tf.float32)
audio_array = np.array([1, 2, 3, 4])
bstring = array_request_example(audio_array)
fdict = {byte_string: bstring}
with tf.Session() as sess:
[tf_samples] = sess.run([audio_samples], feed_dict=fdict)
I have tried using decode_raw and decode_base64 but neither return the original values.
I have tried setting the the out_type of decode raw to the different possible datatypes and tried altering what data type I am converting the original array to.
So, how would I read the byte array in tensorflow? Thanks :)
Extra Info
The aim behind this is to create the serving input function for a custom Estimator to make predictions using gcloud ml-engine local predict (for testing) and using the REST API for the model stored on the cloud.
The serving input function for the Estimator is
def serving_input_fn():
feature_placeholders = {'b64': tf.placeholder(dtype=tf.string,
shape=[None],
name='source')}
audio_samples = tf.decode_raw(feature_placeholders['b64'], tf.float32)
# Dummy function to save space
power_spectrogram = create_spectrogram_from_audio(audio_samples)
inputs = {'spectrogram': power_spectrogram}
return tf.estimator.export.ServingInputReceiver(inputs, feature_placeholders)
Json request
I use .decode('utf-8') because when attempting to json dump the base64 encoded byte strings I receive this error
raise TypeError(repr(o) + " is not JSON serializable")
TypeError: b'longbytestring'
Prediction Errors
When passing the json request {'audio_bytes': 'b64': bytestring} with gcloud local I get the error
PredictionError: Invalid inputs: Expected tensor name: b64, got tensor name: [u'audio_bytes']
So perhaps google cloud local predict does not automatically handle the audio bytes and base64 conversion? Or likely somethings wrong with my Estimator setup.
And the request {'instances': [{'audio_bytes': 'b64': bytestring}]} to REST API gives
{'error': 'Prediction failed: Error during model execution: AbortionError(code=StatusCode.INVALID_ARGUMENT, details="Input to DecodeRaw has length 793713 that is not a multiple of 4, the size of float\n\t [[Node: DecodeRaw = DecodeRaw[_output_shapes=[[?,?]], little_endian=true, out_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:CPU:0"](_arg_source_0_0)]]")'}
which confuses me as I explicitly define the request to be a float and do the same in the serving input receiver.
Removing audio_bytes from the request and utf-8 encoding the byte strings allows me to get predictions, though in testing the decoding locally, I think the audio is being incorrectly converted from the byte string.
The answer that you referenced, is written assuming you are running the model on CloudML Engine's service. The service actually takes care of the JSON (including UTF-8) and base64 encoding.
To get your code working locally or in another environment, you'll need the following changes:
def array_request_example(input_array):
input_array = input_array.astype(np.float32)
return input_array.tostring()
byte_string = tf.placeholder(dtype=tf.string)
audio_samples = tf.decode_raw(byte_string, tf.float32)
audio_array = np.array([1, 2, 3, 4])
bstring = array_request_example(audio_array)
fdict = {byte_string: bstring}
with tf.Session() as sess:
tf_samples = sess.run([audio_samples], feed_dict=fdict)
That said, based on your code, I suspect you are looking to send data as JSON; you can use gcloud local predict to simulate CloudML Engine's service. Or, if you prefer to write your own code, perhaps something like this:
def array_request_examples,(input_arrays):
"""input_arrays is a list (batch) of np_arrays)"""
input_arrays = (a.astype(np.float32) for a in input_arrays)
# Convert each image to byte strings
bytes_strings = (a.tostring() for a in input_arrays)
# Base64 encode the data
encoded = (base64.b64encode(b) for b in bytes_strings)
# Create a list of images suitable to send to the service as JSON:
instances = [{'audio_bytes': {'b64': e}} for e in encoded]
# Create a JSON request
return json.dumps({'instances': instances})
def parse_request(request):
# non-TF to simulate the CloudML Service which does not expect
# this to be in the submitted graphs.
instances = json.loads(request)['instances']
return [base64.b64decode(i['audio_bytes']['b64']) for i in instances]
byte_strings = tf.placeholder(dtype=tf.string, shape=[None])
decode = lambda raw_byte_str: tf.decode_raw(raw_byte_str, tf.float32)
audio_samples = tf.map_fn(decode, byte_strings, dtype=tf.float32)
audio_array = np.array([1, 2, 3, 4])
request = array_request_examples([audio_array])
fdict = {byte_strings: parse_request(request)}
with tf.Session() as sess:
tf_samples = sess.run([audio_samples], feed_dict=fdict)
I am running the sample iris program in TensorFlow Serving. Since it is a TF.Learn model, I am exporting the model using the following classifier.export(export_dir=model_dir,signature_fn=my_classification_signature_fn) and the signature_fn is defined as shown below:
def my_classification_signature_fn(examples, unused_features, predictions):
"""Creates classification signature from given examples and predictions.
Args:
examples: `Tensor`.
unused_features: `dict` of `Tensor`s.
predictions: `Tensor` or dict of tensors that contains the classes tensor
as in {'classes': `Tensor`}.
Returns:
Tuple of default classification signature and empty named signatures.
Raises:
ValueError: If examples is `None`.
"""
if examples is None:
raise ValueError('examples cannot be None when using this signature fn.')
if isinstance(predictions, dict):
default_signature = exporter.classification_signature(
examples, classes_tensor=predictions['classes'])
else:
default_signature = exporter.classification_signature(
examples, classes_tensor=predictions)
named_graph_signatures={
'inputs': exporter.generic_signature({'x_values': examples}),
'outputs': exporter.generic_signature({'preds': predictions})}
return default_signature, named_graph_signatures
The model gets successfully exported using the following piece of code.
I have created a client which makes real-time predictions using TensorFlow Serving.
The following is the code for the client:
flags.DEFINE_string("model_dir", "/tmp/iris_model_dir", "Base directory for output models.")
tf.app.flags.DEFINE_integer('concurrency', 1,
'maximum number of concurrent inference requests')
tf.app.flags.DEFINE_string('server', '', 'PredictionService host:port')
#connection
host, port = FLAGS.server.split(':')
channel = implementations.insecure_channel(host, int(port))
stub = prediction_service_pb2.beta_create_PredictionService_stub(channel)
# Classify two new flower samples.
new_samples = np.array([5.8, 3.1, 5.0, 1.7], dtype=float)
request = predict_pb2.PredictRequest()
request.model_spec.name = 'iris'
request.inputs["x_values"].CopyFrom(
tf.contrib.util.make_tensor_proto(new_samples))
result = stub.Predict(request, 10.0) # 10 secs timeout
However, on making the predictions, the following error is displayed:
grpc.framework.interfaces.face.face.AbortionError: AbortionError(code=StatusCode.INTERNAL, details="Output 0 of type double does not match declared output type string for node _recv_input_example_tensor_0 = _Recv[client_terminated=true, recv_device="/job:localhost/replica:0/task:0/cpu:0", send_device="/job:localhost/replica:0/task:0/cpu:0", send_device_incarnation=2016246895612781641, tensor_name="input_example_tensor:0", tensor_type=DT_STRING, _device="/job:localhost/replica:0/task:0/cpu:0"]()")
Here is the entire stack trace.
enter image description here
The iris model is defined in the following manner:
# Specify that all features have real-value data
feature_columns = [tf.contrib.layers.real_valued_column("", dimension=4)]
# Build 3 layer DNN with 10, 20, 10 units respectively.
classifier = tf.contrib.learn.DNNClassifier(feature_columns=feature_columns,
hidden_units=[10, 20, 10],
n_classes=3, model_dir=model_dir)
# Fit model.
classifier.fit(x=training_set.data,
y=training_set.target,
steps=2000)
Kindly guide a solution for this error.
I think the problem is that your signature_fn is going on the else branch and passing predictions as the output to the classification signature, which expects a string output and not a double output. Either use a regression signature function or add something to the graph to get the output in the form of a string.