Output

An OutputSeq object is the result of running a language model on some input data. It contains not only the output sequence of words generated by the model, but also other data collecting during the generation process that is useful to analyze the model.

In addition to the data, the object has methods to create plots and visualizations of that collected data. These include:

• layer_predictions()
  Which tokens did the model consider as the best outputs for a specific position in the sequence?
• rankings()
  After the model chooses an output token for a specific position, this visual looks back at the ranking of this token at each layer of the model when it was generated (layers assign scores to candidate output tokens, the higher the "probability" score, the higher the ranking of the token).
• rankings_watch()
  Shows the rankings of multiple tokens as the model scored them for a single position. For example, if the input is "The cat ___", we use this method to observe how the model ranked the words "is", "are", "was" as candidates to fill in the blank.
• saliency()
  How important was each input token in the selection of calculating the output token?

To process neuron activations, OutputSeq has methods to reduce the dimensionality and reveal underlying patterns in neuron firings. These are:

• run_nmf()

__init__(self, token_ids=None, n_input_tokens=None, tokenizer=None, output_text=None, tokens=None, encoder_hidden_states=None, decoder_hidden_states=None, embedding_states=None, attribution=None, activations=None, collect_activations_layer_nums=None, attention=None, model_outputs=None, model_type='mlm', lm_head=None, device='cpu') special

Parameters:

Name	Type	Description	Default
token_ids		The input token ids. Dimensions: (batch, position)	None
n_input_tokens		Int. The number of input tokens in the sequence.	None
tokenizer		huggingface tokenizer associated with the model generating this output	None
output_text		The output text generated by the model (if processed with generate())	None
tokens		A list of token text. Shorthand to passing the token ids by the tokenizer. dimensions are (batch, position)	None
hidden_states		A tensor of dimensions (layer, position, hidden_dimension). In layer, index 0 is for embedding hidden_state.	required
attribution		A list of attributions. One element per generated token. Each element is a list giving a value for tokens from 0 to right before the generated token.	None
activations		The activations collected from model processing. Shape is (batch, layer, neurons, position)	None
collect_activations_layer_nums			None
attention		The attention tensor retrieved from the language model	None
model_outputs		Raw return object returned by the model	None
lm_head		The trained language model head from a language model projecting a hidden state to an output vocabulary associated with teh tokenizer.	None
device		"cuda" or "cpu"	'cpu'

Source code in ecco/output.py

def __init__(self,
             token_ids=None,
             n_input_tokens=None,
             tokenizer=None,
             output_text=None,
             tokens=None,
             encoder_hidden_states=None,
             decoder_hidden_states=None,
             embedding_states=None,
             attribution=None,
             activations=None,
             collect_activations_layer_nums=None,
             attention=None,
             model_outputs=None,
             model_type: str= 'mlm',
             lm_head=None,
             device='cpu'):
    """

    Args:
        token_ids: The input token ids. Dimensions: (batch, position)
        n_input_tokens: Int. The number of input tokens in the sequence.
        tokenizer: huggingface tokenizer associated with the model generating this output
        output_text: The output text generated by the model (if processed with generate())
        tokens: A list of token text. Shorthand to passing the token ids by the tokenizer.
            dimensions are (batch, position)
        hidden_states: A tensor of  dimensions (layer, position, hidden_dimension).
            In layer, index 0 is for embedding hidden_state.
        attribution: A list of attributions. One element per generated token.
            Each element is a list giving a value for tokens from 0 to right before the generated token.
        activations: The activations collected from model processing.
            Shape is (batch, layer, neurons, position)
        collect_activations_layer_nums:
        attention: The attention tensor retrieved from the language model
        model_outputs: Raw return object returned by the model
        lm_head: The trained language model head from a language model projecting a
            hidden state to an output vocabulary associated with teh tokenizer.
        device: "cuda" or "cpu"
    """
    self.token_ids = token_ids
    self.tokenizer = tokenizer
    self.n_input_tokens = n_input_tokens
    self.output_text = output_text
    self.tokens = tokens
    self.encoder_hidden_states = encoder_hidden_states
    self.decoder_hidden_states = decoder_hidden_states
    self.embedding_states = embedding_states
    self.attribution = attribution
    self.activations = activations
    self.collect_activations_layer_nums = collect_activations_layer_nums
    self.model_outputs = model_outputs
    self.attention_values = attention
    self.lm_head = lm_head
    self.device = device
    self.model_type = model_type
    self._path = os.path.dirname(ecco.__file__)

layer_predictions(self, position=1, topk=10, layer=None, **kwargs)

Visualization plotting the topk predicted tokens after each layer (using its hidden state).

Examples:

Parameters:

Name	Type	Description	Default
position	int	The index of the output token to trace	1
topk	Optional[int]	Number of tokens to show for each layer	10
layer	Optional[int]	None shows all layers. Can also pass an int with the layer id to show only that layer	None

Source code in ecco/output.py

def layer_predictions(self, position: int = 1, topk: Optional[int] = 10, layer: Optional[int] = None, **kwargs):
    """
        Visualization plotting the topk predicted tokens after each layer (using its hidden state).

        Example:
        ![prediction scores](../../img/layer_predictions_ex_london.png)

        Args:
            position: The index of the output token to trace
            topk: Number of tokens to show for each layer
            layer: None shows all layers. Can also pass an int with the layer id to show only that layer
    """

    assert self.model_type != 'mlm', "method not supported for Masked-LMs"

    _, dec_hidden_states = self._get_hidden_states()
    assert dec_hidden_states is not None, "decoder hidden states not found"

    if position == 0:
        raise ValueError(f"'position' is set to 0. There is never a hidden state associated with this position."
                         f"Possible values are 1 and above -- the position of the token of interest in the sequence")

    if self.model_type == 'enc-dec':
        # For enc-dec LMs, the position starts at the first generated token, not as in causal LMs
        # In enc-dec LMs, the position is relative. By that means, position self.n_input_tokens is the first generated token
        new_position = position - self.n_input_tokens
        assert new_position >= 0, f"position={position} not supported, minimum is " \
                                  f"position={self.n_input_tokens} for the first generated token"
        position = new_position

    if layer is not None:
        # If a layer is specified, choose it only.
        assert dec_hidden_states is not None
        dec_hidden_states = dec_hidden_states[layer].unsqueeze(0)

    k = topk
    top_tokens = []
    probs = []
    data = []

    # loop through layer levels
    for layer_no, h in enumerate(dec_hidden_states):

        hidden_state = h[position - 1]

        # Use lm_head to project the layer's hidden state to output vocabulary
        logits = self.lm_head(self.to(hidden_state))
        softmax = F.softmax(logits, dim=-1)

        # softmax dims are (number of words in vocab) - 50257 in GPT2
        sorted_softmax = self.to(torch.argsort(softmax))

        # Not currently used. If we're "watching" a specific token, this gets its ranking
        # idx = sorted_softmax.shape[0] - torch.nonzero((sorted_softmax == watch)).flatten()

        layer_top_tokens = [self.tokenizer.decode(t) for t in sorted_softmax[-k:]][::-1]
        top_tokens.append(layer_top_tokens)
        layer_probs = softmax[sorted_softmax[-k:]].cpu().detach().numpy()[::-1]
        probs.append(layer_probs.tolist())

        # Package in output format
        layer_data = []
        for idx, (token, prob) in enumerate(zip(layer_top_tokens, layer_probs)):
            layer_num = layer if layer is not None else layer_no
            layer_data.append({'token': token,
                               'prob': str(prob),
                               'ranking': idx + 1,
                               'layer': layer_num
                               })

        data.append(layer_data)

    d.display(d.HTML(filename=os.path.join(self._path, "html", "setup.html")))
    d.display(d.HTML(filename=os.path.join(self._path, "html", "basic.html")))

    js = f"""
     requirejs(['basic', 'ecco'], function(basic, ecco){{
        const viz_id = basic.init()


        let pred = new ecco.LayerPredictions({{
            parentDiv: viz_id,
            data:{json.dumps(data)}
        }})
        pred.init()
     }}, function (err) {{
        console.log(viz_id, err);
    }})"""
    d.display(d.Javascript(js))

    if 'printJson' in kwargs and kwargs['printJson']:
        print(data)
        return data

rankings(self, **kwargs)

Plots the rankings (across layers) of the tokens the model selected. Each column is a position in the sequence. Each row is a layer.

Source code in ecco/output.py

def rankings(self, **kwargs):
    """
    Plots the rankings (across layers) of the tokens the model selected.
    Each column is a position in the sequence. Each row is a layer.

    ![Rankings watch](../../img/rankings_ex_eu_1.png)
    """

    assert self.model_type != 'mlm', "method not supported for Masked-LMs"

    _, dec_hidden_states = self._get_hidden_states()
    assert dec_hidden_states is not None, "decoder hidden states not found"

    n_layers_dec = len(dec_hidden_states)

    if self.model_type == 'causal':
        position = dec_hidden_states.shape[1] - self.n_input_tokens + 1
    elif self.model_type == 'enc-dec':
        position = dec_hidden_states.shape[1]
    else:
        raise NotImplemented(f"model_type={self.model_type} not supported")

    rankings = np.zeros((n_layers_dec, position), dtype=np.int32)
    predicted_tokens = np.empty((n_layers_dec, position), dtype='U25')
    token_found_mask = np.ones((n_layers_dec, position))

    # loop through layer levels
    for i, level in enumerate(dec_hidden_states):
        # Loop through generated/output positions
        offset = 0 if self.model_type == 'enc-dec' else self.n_input_tokens - 1
        for j, hidden_state in enumerate(level[offset:]):

            # Project hidden state to vocabulary
            # (after debugging pain: ensure input is on GPU, if appropriate)
            logits = self.lm_head(self.to(hidden_state))

            # Sort by score (ascending)
            sorted = torch.argsort(logits)

            # What token was sampled in this position?
            offset = self.n_input_tokens + 1 if self.model_type == 'enc-dec' else self.n_input_tokens
            token_id = torch.tensor(self.token_ids[0][offset + j])

            # token_id = self.token_ids.clone().detach()[self.n_input_tokens + j]
            # What's the index of the sampled token in the sorted list?
            r = torch.nonzero((sorted == token_id)).flatten()

            # subtract to get ranking (where 1 is the top scoring, because sorting was in ascending order)
            ranking = sorted.shape[0] - r
            token = self.tokenizer.decode([token_id])
            predicted_tokens[i, j] = token
            rankings[i, j] = int(ranking)
            if token_id == self.token_ids[0][j + 1]:
                token_found_mask[i, j] = 0

    input_tokens = [repr(t) for t in self.tokens[0][self.n_input_tokens - 1:-1]]
    offset = self.n_input_tokens + 1 if self.model_type == 'enc-dec' else self.n_input_tokens
    output_tokens = [repr(t) for t in self.tokens[0][offset:]]

    lm_plots.plot_inner_token_rankings(input_tokens,
                                       output_tokens,
                                       rankings,
                                       **kwargs)

    if 'printJson' in kwargs and kwargs['printJson']:
        data = {
            'input_tokens': input_tokens,
            'output_tokens': output_tokens,
            'rankings': rankings,
            'predicted_tokens': predicted_tokens,
            'token_found_mask': token_found_mask
        }
        print(data)
        return data

rankings_watch(self, watch=None, position=-1, **kwargs)

Plots the rankings of the tokens whose ids are supplied in the watch list. Only considers one position.

Source code in ecco/output.py

def rankings_watch(self, watch: List[int] = None, position: int = -1, **kwargs):
    """
    Plots the rankings of the tokens whose ids are supplied in the watch list.
    Only considers one position.

    ![Rankings plot](../../img/ranking_watch_ex_is_are_1.png)
    """

    assert self.model_type != 'mlm', "method not supported for Masked-LMs"

    _, dec_hidden_states = self._get_hidden_states()
    assert dec_hidden_states is not None, "decoder hidden states not found"

    if position != -1:
        if self.model_type == 'causal':
            position = position - 1  # e.g. position 5 corresponds to hidden state 4
        elif self.model_type == 'enc-dec':
            # In enc-dec LMs, the position is relative. By that means, position self.n_input_tokens + 1 is the first generated token
            new_position = position - 1 - self.n_input_tokens
            assert new_position >= 0, f"position={position} not supported, minimum is " \
                                      f"position={self.n_input_tokens + 1} for the first generated token"
            position = new_position
        else:
            raise NotImplemented(f"model_type={self.model_type} not supported")

    n_layers_dec = len(dec_hidden_states) if dec_hidden_states is not None else 0
    n_tokens_to_watch = len(watch)
    rankings = np.zeros((n_layers_dec, n_tokens_to_watch), dtype=np.int32)

    # loop through layer levels
    for i, level in enumerate(dec_hidden_states):
        # Loop through generated/output positions
        for j, token_id in enumerate(watch):

            hidden_state = level[position]

            # Project hidden state to vocabulary
            # (after debugging pain: ensure input is on GPU, if appropriate)
            logits = self.lm_head(self.to(hidden_state))

            # Sort by score (ascending)
            sorted = torch.argsort(logits)

            # What token was sampled in this position?
            token_id = torch.tensor(token_id)

            # What's the index of the sampled token in the sorted list?
            r = torch.nonzero((sorted == token_id)).flatten()

            # subtract to get ranking (where 1 is the top scoring, because sorting was in ascending order)
            ranking = sorted.shape[0] - r
            rankings[i, j] = int(ranking)

    input_tokens = [t for t in self.tokens[0]]
    output_tokens = [repr(self.tokenizer.decode(t)) for t in watch]

    lm_plots.plot_inner_token_rankings_watch(input_tokens, output_tokens, rankings,
                                             position + self.n_input_tokens if self.model_type == 'enc-dec' else position)

    if 'printJson' in kwargs and kwargs['printJson']:
        data = {'input_tokens': input_tokens,
                'output_tokens': output_tokens,
                'rankings': rankings}
        print(data)
        return data

run_nmf(self, **kwargs)

Run Non-negative Matrix Factorization on network activations of FFNN. Returns an NMF object which holds the factorization model and data and methods to visualize them.

Source code in ecco/output.py

def run_nmf(self, **kwargs):
    """
    Run Non-negative Matrix Factorization on network activations of FFNN. Returns an [NMF]() object which holds
    the factorization model and data and methods to visualize them.


    """
    return NMF(self.activations,
               n_input_tokens=self.n_input_tokens,
               token_ids=self.token_ids,
               _path=self._path,
               tokens=self.tokens,
               collect_activations_layer_nums=self.collect_activations_layer_nums,
               **kwargs)

saliency(self, attr_method='grad_x_input', style='minimal', **kwargs)

        Explorable showing saliency of each token generation step.
        Hovering-over or tapping an output token imposes a saliency map on other tokens
        showing their importance as features to that prediction.

        Examples:

        ```python
        import ecco
        lm = ecco.from_pretrained('distilgpt2')
        text= "The countries of the European Union are:

1. Austria
2. Belgium
3. Bulgaria 4." output = lm.generate(text, generate=20, do_sample=True)

       # Show saliency explorable
       output.saliency()
       ```
   
       Which creates the following interactive explorable:
       ![input saliency example 1](../../img/saliency_ex_1.png)
   
       If we want more details on the saliency values, we can use the detailed view:
   
       ```python
       # Show detailed explorable
       output.saliency(style="detailed")
       ```
   
       Which creates the following interactive explorable:
   
       ![input saliency example 2 - detailed](../../img/saliency_ex_2.png)
   
       Details:
       This view shows the Gradient * Inputs method of input saliency. The attribution values are calculated across the
       embedding dimensions, then we use the L2 norm to calculate a score for each token (from the values of its embeddings dimension)
       To get a percentage value, we normalize the scores by dividing by the sum of the attribution scores for all
       the tokens in the sequence.
   

Source code in ecco/output.py

def saliency(self, attr_method: Optional[str] = 'grad_x_input', style="minimal", **kwargs):
    """
        Explorable showing saliency of each token generation step.
        Hovering-over or tapping an output token imposes a saliency map on other tokens
        showing their importance as features to that prediction.

        Examples:

        ```python
        import ecco
        lm = ecco.from_pretrained('distilgpt2')
        text= "The countries of the European Union are:\n1. Austria\n2. Belgium\n3. Bulgaria\n4."
        output = lm.generate(text, generate=20, do_sample=True)

        # Show saliency explorable
        output.saliency()
        ```

        Which creates the following interactive explorable:
        ![input saliency example 1](../../img/saliency_ex_1.png)

        If we want more details on the saliency values, we can use the detailed view:

        ```python
        # Show detailed explorable
        output.saliency(style="detailed")
        ```

        Which creates the following interactive explorable:

        ![input saliency example 2 - detailed](../../img/saliency_ex_2.png)


        Details:
        This view shows the Gradient * Inputs method of input saliency. The attribution values are calculated across the
        embedding dimensions, then we use the L2 norm to calculate a score for each token (from the values of its embeddings dimension)
        To get a percentage value, we normalize the scores by dividing by the sum of the attribution scores for all
        the tokens in the sequence.
    """
    position = self.n_input_tokens

    importance_id = position - self.n_input_tokens
    tokens = []
    attribution = self.attribution[attr_method]
    for idx, token in enumerate(self.tokens[0]):
        type = "input" if idx < self.n_input_tokens else 'output'
        if idx < len(attribution[importance_id]):
            imp = attribution[importance_id][idx]
        else:
            imp = 0

        tokens.append({'token': token,
                       'token_id': int(self.token_ids[0][idx]),
                       'type': type,
                       'value': str(imp),  # because json complains of floats
                       'position': idx
                       })

    data = {
        'tokens': tokens,
        'attributions': [att.tolist() for att in attribution]
    }

    d.display(d.HTML(filename=os.path.join(self._path, "html", "setup.html")))
    d.display(d.HTML(filename=os.path.join(self._path, "html", "basic.html")))
    # viz_id = 'viz_{}'.format(round(random.random() * 1000000))

    if (style == "minimal"):
        js = f"""
         requirejs(['basic', 'ecco'], function(basic, ecco){{
            const viz_id = basic.init()
            // ecco.interactiveTokens(viz_id, {{}})
            window.ecco[viz_id] = new ecco.MinimalHighlighter({{
            parentDiv: viz_id,
            data: {data},
            preset: 'viridis'
         }})

         window.ecco[viz_id].init();
         window.ecco[viz_id].selectFirstToken();

         }}, function (err) {{
            console.log(err);
        }})"""
    elif (style == "detailed"):

        js = f"""
         requirejs(['basic', 'ecco'], function(basic, ecco){{
            const viz_id = basic.init()
            window.ecco[viz_id] = ecco.interactiveTokens(viz_id, {data})

         }}, function (err) {{
            console.log(err);
        }})"""

    d.display(d.Javascript(js))

    if 'printJson' in kwargs and kwargs['printJson']:
        print(data)
        return data