tools: add some data probe (#152)

tools/data_probe/collector.py

@@ -0,0 +1,67 @@
+from itertools import chain
+
+import torch
+from torch.distributions import Categorical
+from transformers import AutoTokenizer
+
+from data_juicer.format import load_formatter
+
+
+class TextTokenDistCollector(object):
+    """Tokenize and collect distribution of tokens for given
+    dataset with a specified tokenizer.
+    """
+
+    def __init__(self, tokenizer):
+        """
+        Initialization method.
+
+        :param tokenizer: tokenizer name on huggingface
+        """
+        self.tokenizer = AutoTokenizer.from_pretrained(tokenizer,
+                                                       trust_remote_code=True)
+        self.vocab_size = len(self.tokenizer)
+
+    def collect(self, data_path, text_key, num_proc=1) -> 'Categorical':
+        """
+        Tokenize and collect tokens distribution of input dataset
+        :param data_path: path to input dataset.
+        :param text_key: field keys that will be considered into token counts.
+        :param num_proc: number of processes to count tokens.
+        :return: token distribution.
+        """
+
+        formatter = load_formatter(data_path)
+        dataset = formatter.load_dataset(num_proc=num_proc)
+        assert text_key in dataset.features, f'[{text_key} not find in dataset'
+
+        def prepare_tokenizer(
+            tokenizer,
+            text_key,
+        ):
+            """
+            Prepare a tokenizer function for dataset.
+            :param tokenizer: a tokenizer to tokenize sample.
+            :param text_key: field keys that will be
+                considered into token counts.
+            """
+
+            def _tokenize_fn(example, ):
+                example = tokenizer(example[text_key],
+                                    add_special_tokens=False)
+                return example
+
+            return _tokenize_fn
+
+        tokenize_proc = prepare_tokenizer(self.tokenizer, text_key)
+        dataset = dataset.map(tokenize_proc,
+                              num_proc=num_proc,
+                              desc=f'tokenize {data_path.split("/")[-1]}')
+
+        token_count = torch.zeros(self.vocab_size, dtype=torch.int64)
+        token_ids = torch.tensor(
+            list(chain.from_iterable(dataset['input_ids'])))
+        indices, counts = token_ids.unique(return_counts=True)
+        token_count.scatter_(0, indices, counts.to(token_count.dtype))
+        dist = Categorical(token_count)
+        return dist

tools/data_probe/draw.py

@@ -0,0 +1,41 @@
+import matplotlib.pyplot as plt
+import numpy as np
+import seaborn as sns
+
+
+def draw_heatmap(data, xlabels, ylables=None, figsize=None, triangle=False):
+    """
+    Draw heatmap of input data with special lables.
+    :param data: input data, now support
+        [`list`, `tuple`, `numpy array`, 'torch tensor']
+    :param xlabels: x axis labels.
+    :param ylabels: y axis labels, if None, use xlabels.
+    :param figsize: figure size.
+    :param triangle: only display triangle.
+    :return: a plot figure.
+    """
+    figsize = figsize if figsize else (8 * 2.5, 6 * 2.5)
+    _, ax = plt.subplots(figsize=figsize)
+    mask = None
+    if triangle:
+        mask = np.triu(np.ones_like(data))
+    ax.tick_params(
+        right=True,
+        top=True,
+        labelright=True,
+        labeltop=True,
+    )
+    sns.heatmap(data,
+                ax=ax,
+                cmap='Oranges',
+                annot=True,
+                mask=mask,
+                linewidths=.05,
+                square=True,
+                xticklabels=xlabels,
+                yticklabels=ylables,
+                annot_kws={'size': 8})
+    plt.subplots_adjust(left=.1, right=0.95, bottom=0.22, top=0.95)
+    fig = plt.gcf()
+    plt.show()
+    return fig

tools/data_probe/measure.py

@@ -0,0 +1,109 @@
+import torch
+import torch.nn.functional as F
+from torch import Tensor
+from torch.distributions import Categorical
+
+
+class Measure(object):
+    """Base class for Measure distribution.
+    """
+    name = 'base'
+
+    def measure(self, *args, **kwargs):
+        pass
+
+    def __call__(self, *args, **kwargs):
+        return self.measure(*args, **kwargs)
+
+    def _convert_to_tensor(self, p):
+        """
+        Convert input data to torch tensor.
+        :param p: input data, now support
+            [`scalar`,`list`, `tuple`, `torch binary file`, and `Categorical`].
+        :return: torch tensor
+        """
+        if isinstance(p, Tensor):
+            return p
+        elif isinstance(p, Categorical):
+            return p.probs
+        elif isinstance(p, str):
+            return torch.load(p)
+        else:
+            return torch.tensor(p)
+
+    def _convert_to_categorical(self, p):
+        """
+        Convert input data to torch Categorical.
+        :param p: input data, now support
+            [`scalar`,`list`, `tuple`, `torch binary file`, and `Categorical`].
+        :return: torch Categorical
+        """
+        if isinstance(p, Categorical):
+            return p
+        elif isinstance(p, Tensor):
+            return Categorical(p)
+        elif isinstance(p, str):
+            return Categorical(torch.load(p))
+        else:
+            return Categorical(torch.tensor(p))
+
+
+class KLDivMeasure(Measure):
+    """
+    Measure Kullback-Leibler divergence.
+    """
+    name = 'kl_divergence'
+
+    def measure(self, p, q):
+        p = self._convert_to_categorical(p)
+        q = self._convert_to_categorical(q)
+        assert p.probs.shape == q.probs.shape, \
+            'The two inputs have different shape:' \
+            f'{p.probs.shape} != {q.probs.shape} in {self.name}'
+        return F.kl_div(q.logits, p.probs, log_target=False, reduction='sum')
+
+
+class JSDivMeasure(Measure):
+    """
+    Measure Jensen-Shannon divergence.
+    """
+    name = 'js_divergence'
+
+    def measure(self, p, q):
+        p = self._convert_to_tensor(p)
+        q = self._convert_to_tensor(q)
+        assert p.shape == q.shape,  \
+            'The two inputs have different shape:' \
+            f'{p.shape} != {q.shape} in {self.name}'
+
+        m = 0.5 * (p + q)
+        kl_p = KLDivMeasure()(p, m)
+        kl_q = KLDivMeasure()(q, m)
+        js = 0.5 * (kl_p + kl_q)
+        return js
+
+
+class CrossEntropyMeasure(Measure):
+    """
+    Measure Cross-Entropy.
+    """
+    name = 'cross_entropy'
+
+    def measure(self, p, q):
+        p = self._convert_to_categorical(p)
+        q = self._convert_to_categorical(q)
+        assert p.probs.shape == q.probs.shape, \
+            'The two inputs have different shape: '\
+            f'{p.probs.shape} != {q.probs.shape} in {self.name}'
+        return F.cross_entropy(q.logits, p.probs, reduction='sum')
+
+
+class EntropyMeasure(Measure):
+    """
+    Measure Entropy.
+    """
+    name = 'entropy'
+
+    def measure(self, p):
+        p = self._convert_to_categorical(p)
+        return p.entropy()