refactor(Docs): Add table of all actions, and tiny cleanup

Fix end token calculation for long prompts

README.md

@@ -29,33 +29,10 @@ See example workflow in examples folder.
     - Example: `"Hello World"`, `'It\'s a sunny day'`
     - Represents string literals.
 
-## Functions
-
-Here are the available functions and their usage:
-
-1. **Sum Function**:
-    - Syntax: `sum(arg1 | arg2 | ... | argN)`
-    - Adds together multiple embeddings.
-    - Example: `sum(embedding:face1 | dog)`
-
-2. **Negation Function**:
-    - Syntax: `neg(arg)`
-    - Negates the output.
-    - Example: `neg(A embedding:happycats outside)`
-
-3. **Normalization Function**:
-    - Syntax: `norm(arg)`
-    - Normalizes the given vector embedding.
-    - Example: `norm(sum(embedding:face1 | embedding:face2))`
-
-4. **Difference Function**:
-    - Syntax: `diff(arg1 | arg2 | ... | argN)`
-    - Computes the difference between multiple vector embeddings.
-    - Example: `diff(embedding:face1 | embedding:face2)`
 
 ### Notes on Arguments:
 - Each function takes one or more arguments.
-- An argument (`arg`) can be an embedding, a word, another function, or a quoted string.
+- An argument (`arg`) can be an embedding, multiple words, another function, or a quoted string.
 - For functions that accept multiple arguments, they are separated by the `|` symbol.
 
 ## Examples
@@ -78,4 +55,25 @@ Here are the available functions and their usage:
    ```
    sum(king|neg(man)|woman)
    ```
-   ```
+
+## Functions
+
+Here are the available functions and their usage:
+
+| Display Name | Action Name | Description | Usage Examples |
+| --- | --- | --- | --- |
+| Multiply | mult | Multiplies the provided segments or actions by the multiplier. | <ul><li>mult(The cat is\|2.5)</li><li>mult(Cat\|-1)</li></ul> |
+| Set Dimensions | setDims | Sets the specified dimensions of the input embeddings to the specified value | <ul><li>The setDims(cat\|4, -0.01253\|76, 1.2) is happy</li></ul> |
+| Negate | neg | Negates the provided segments or actions. | <ul><li>neg(cat)</li><li>sum(king\|neg(man)\|women)</li></ul> |
+| Normalize | norm | Normalizes the provided segments or actions. | <ul><li>norm(cat)</li><li>sum(cat\|norm(sum(tiger\|fish)))</li></ul> |
+| Positional Embedding Scale | posScale | Scales(Multiplies) the positional embeddings of the provided segments or actions by the multiplier. | <ul><li>A posScale(cat\|1.5) on a rainy day</li></ul> |
+| Random Embedding | rand | Returns a random embedding of the specified token length, with the values optionally bounded by the second and third arguments. | <ul><li>A rand(1) cat</li><li>A rand(1\|-1\|1) cat</li></ul> |
+| Difference | diff | Subtracts the segments in the order they are given. The first segment is subtracted from the second, then the third from the result, and so on. | <ul><li>diff(The cat is\|The dog is)</li><li>diff(Cat\|Dog)</li><li>sum(diff(king\|man)\|woman)</li></ul> |
+| Slerp | slerp | Performs a slerp(Interpolation) between two segments or actions, with the given weight. The recommended weight is 0 - 1 | <ul><li>The slerp(cat\|dog\|0.5) is happy</li></ul> |
+| Pooled Average(Experimental) | _exp-pooledAvg | Processes the provided segments or actions fully through CLIP and creates a pooled average of the last hidden state by averaging the last hidden state of each token. | <ul><li>A cat on a _exp-pooledAvg(beautiful sunny day)</li><li>A _exp-pooledAvg(broken glass) bottle</li></ul> |
+| Sum | sum | Adds the embeddings of the provided segments or actions. | <ul><li>A happy sum(cat\|dog\|shark)</li></ul> |
+| Pooler Output(Experimental) | _exp-pooler | Processes the provided segments or actions fully through CLIP and returns the pooler_output from the transformer | <ul><li>A cat on a _exp-pooler(beautiful sunny day)</li><li>A _exp-pooler(broken glass) bottle</li></ul> |
+| Scale Dimensions | scaleDims | Scales the specified dimensions of the input embeddings by the specified amount | <ul><li>The scaleDims(cat\|4,1.5\|76,1.2) is happy</li></ul> |
+| Ignore Positional Embeddings | postPos | Prevents positional embeddings from being applied to the provided segments or actions. | <ul><li>A postPos(cat) on a rainy day</li></ul> |
+| Average | avg | Performs a weighted average between two segments or actions. The recommended weight is 0 - 1. | <ul><li>avg(The cat is\|The dog is\|0.5)</li><li>avg(Cat\|Dog\|0.5)</li></ul> |
+

lib/__init__.py

@@ -1,8 +1,16 @@
+from custom_nodes.KepPromptLang.lib.actions.avg import AverageAction
 from custom_nodes.KepPromptLang.lib.actions.diff import DiffAction
 from custom_nodes.KepPromptLang.lib.actions.mult import MultiplyAction
 from custom_nodes.KepPromptLang.lib.actions.neg import NegAction
 from custom_nodes.KepPromptLang.lib.actions.norm import NormAction
+
+from custom_nodes.KepPromptLang.lib.actions.pooled_avg import PooledAvgAction
+from custom_nodes.KepPromptLang.lib.actions.pooler import PoolerAction
+from custom_nodes.KepPromptLang.lib.actions.pos_scale import PosScaleAction
+from custom_nodes.KepPromptLang.lib.actions.post_pos import PostPosAction
 from custom_nodes.KepPromptLang.lib.actions.rand import RandAction
+from custom_nodes.KepPromptLang.lib.actions.scale_dims import ScaleDims
+from custom_nodes.KepPromptLang.lib.actions.set_dims import SetDims
 from custom_nodes.KepPromptLang.lib.actions.slerp import SlerpAction
 from custom_nodes.KepPromptLang.lib.actions.sum import SumAction
 from custom_nodes.KepPromptLang.lib.parser.registration import register_action
@@ -14,3 +22,10 @@ register_action(NormAction)
 register_action(RandAction)
 register_action(SumAction)
 register_action(SlerpAction)
+register_action(AverageAction)
+register_action(ScaleDims)
+register_action(SetDims)
+register_action(PosScaleAction)
+register_action(PoolerAction)
+register_action(PostPosAction)
+register_action(PooledAvgAction)

lib/action/base.py

@@ -1,9 +1,10 @@
 from abc import ABC, abstractmethod
 from enum import Enum
-from typing import Union, List
+from typing import Union, List, TypedDict, Tuple
 
 from torch import Tensor
 from torch.nn import Embedding
+from transformers.models.clip.modeling_clip import CLIPTextTransformer
 
 from custom_nodes.KepPromptLang.lib.parser.prompt_segment import PromptSegment
 
@@ -13,6 +14,14 @@ class ActionArity(Enum):
     SINGLE = 1
     MULTI = 2
 
+class PostModifiers(TypedDict):
+    """
+    A dictionary of post modifiers for an action result.
+    """
+    position_embed_scale: Union[float, None]
+    bypass_pos_embed: Union[bool, None]
+
+
 class Action(ABC):
     @property
     @abstractmethod
@@ -30,9 +39,23 @@ class Action(ABC):
 
     @property
     @abstractmethod
-    def name(self) -> str:
+    def display_name(self) -> str:
         pass
 
+    @property
+    @abstractmethod
+    def action_name(self) -> str:
+        pass
+
+    @property
+    @abstractmethod
+    def description(self) -> str:
+        pass
+
+    @property
+    def usage_examples(self) -> List[str]:
+        return []
+
     @property
     @abstractmethod
     def grammar(self) -> str:
@@ -67,7 +90,7 @@ class Action(ABC):
         pass
 
     @abstractmethod
-    def get_result(self, embedding_module: Embedding) -> Tensor:
+    def get_result(self, embedding_module: Embedding) -> Union[Tensor, Tuple[Tensor, PostModifiers]]:
         """
         Get the result of this action. This is called when the embeddings are being calculated.
         :param embedding_module: The embedding module to use to get the base embeddings for tokens.
@@ -75,6 +98,13 @@ class Action(ABC):
         """
         pass
 
+    def process_with_transformer(self, transformer: CLIPTextTransformer, embedding_module: Embedding) -> None:
+        """
+        For actions that need access to the TextTransformer, this method is called. Results are expected to be returned via get_result still.
+        :param transformer: An instance of CLIPTextTransformer
+        """
+        pass
+
     def depth_repr(self, depth: int = 1) -> str:
         raise NotImplementedError()
 
@@ -90,12 +120,17 @@ class SingleArgAction(Action, ABC):
                 segments.append(seg_or_action)
         return segments
 
+    def process_with_transformer(self, transformer: CLIPTextTransformer, embedding_module: Embedding) -> None:
+        for seg_or_action in self.arg:
+            if isinstance(seg_or_action, Action):
+                seg_or_action.process_with_transformer(transformer, embedding_module)
+
     def __init__(self, arg: List[Union[PromptSegment, Action]]):
         # TODO: Target is a list now... what does this mean for us..
         self.arg = arg
 
     def __repr__(self) -> str:
-        return f"{self.name}({self.arg})"
+        return f"{self.display_name}({self.arg})"
 
 class MultiArgAction(Action, ABC):
     arity = ActionArity.MULTI
@@ -110,6 +145,12 @@ class MultiArgAction(Action, ABC):
 
         return segments
 
+    def process_with_transformer(self, transformer: CLIPTextTransformer, embedding_module: Embedding) -> None:
+        for arg in self.all_args:
+            for seg_or_action in arg:
+                if isinstance(seg_or_action, Action):
+                    seg_or_action.process_with_transformer(transformer, embedding_module)
+
     def __init__(
             self,
             args: List[List[Union[PromptSegment, Action]]],

lib/actions/action_utils.py

@@ -1,3 +1,5 @@
+from typing import List
+
 from torch import Tensor
 from torch.nn import Embedding
 
@@ -9,3 +11,10 @@ def get_embedding(seg_or_action: SegOrAction, embedding_module: Embedding) -> Te
     if isinstance(seg_or_action, Action):
         return seg_or_action.get_result(embedding_module)
     return seg_or_action.get_embeddings(embedding_module)
+
+def get_total_length(args: List[SegOrAction]) -> int:
+    total_length = 0
+    for seg_or_action in args:
+        total_length += seg_or_action.token_length()
+
+    return total_length

lib/actions/avg.py

@@ -0,0 +1,108 @@
+from typing import List, Union
+
+import torch
+from torch.nn import Embedding
+
+from custom_nodes.KepPromptLang.lib.action.base import MultiArgAction, Action
+from custom_nodes.KepPromptLang.lib.actions.action_utils import get_embedding
+from custom_nodes.KepPromptLang.lib.actions.types import SegOrAction
+from custom_nodes.KepPromptLang.lib.actions.utils import slerp
+from custom_nodes.KepPromptLang.lib.parser.prompt_segment import PromptSegment
+
+
+class AverageAction(MultiArgAction):
+    grammar = 'avg(" arg "|" arg "|" arg ")"'
+    chars = ["+", "+"]
+
+    display_name = "Average"
+    action_name = "avg"
+    description = "Performs a weighted average between two segments or actions. The recommended weight is 0 - 1."
+    usage_examples = [
+        "avg(The cat is|The dog is|0.5)",
+        "avg(Cat|Dog|0.5)",
+    ]
+
+
+    def __init__(self, args: List[List[Union[PromptSegment, Action]]]) -> None:
+        super().__init__(args)
+
+        if len(args) != 3:
+            raise ValueError("Average action should have exactly three arguments(2 vectors and a weight)")
+
+        self.first_arg = args[0]
+        self.second_arg = args[1]
+        self._parse_weight(args[2])
+
+        self._validate_args()
+
+
+    def _parse_weight(self, arg: List[SegOrAction]) -> None:
+        if len(arg) != 1:
+            raise ValueError("Average weight should have exactly one segment")
+
+        weight_seg_or_action = arg[0]
+
+        if isinstance(weight_seg_or_action, Action):
+            raise ValueError("Average weight should not have an action as an argument")
+
+        try:
+            self.parsed_weight = float(weight_seg_or_action.text)
+        except ValueError:
+            raise ValueError("Average should have an integer/float as the weight")
+
+    def _validate_args(self) -> None:
+        first_arg_token_length = sum(seg_or_action.token_length() for seg_or_action in self.first_arg)
+        second_arg_token_length = sum(seg_or_action.token_length() for seg_or_action in self.second_arg)
+        if first_arg_token_length != second_arg_token_length:
+            raise ValueError(f"Average start and end arguments should have the same length. Got {start_arg_token_length} and {end_arg_token_length}")
+
+        if self.parsed_weight < 0 or self.parsed_weight > 1:
+            print(f"WARNING: Average weight should be between 0 and 1. Got {self.parsed_weight}")
+
+    def token_length(self) -> int:
+        # Average interpolates between the embeddings of the start and end segments, so the length is the length of the start segment
+        return sum(seg_or_action.token_length() for seg_or_action in self.first_arg)
+
+    def get_result(self, embedding_module: Embedding) -> torch.Tensor:
+        # Calculate the embeddings for the start segment
+        all_start_embeddings = [
+            get_embedding(seg_or_action, embedding_module)
+            for seg_or_action in self.first_arg
+        ]
+
+        start_embedding = torch.cat(all_start_embeddings, dim=1)
+
+        # Calculate the embeddings for the end segment
+        all_end_embeddings = [
+            get_embedding(seg_or_action, embedding_module)
+            for seg_or_action in self.second_arg
+        ]
+
+        end_embedding = torch.cat(all_end_embeddings, dim=1)
+
+        # Perform the weighted average
+        result = start_embedding * (1 - self.parsed_weight) + end_embedding * self.parsed_weight
+
+        return result
+
+    # def __repr__(self):
+    #     return f"sum(\n\tbase_segment={self.base_segment},\n\targs={self.args}\n)"
+    def __repr__(self) -> str:
+        return f"sum({', '.join(map(str, self.additional_args))})"
+
+    def depth_repr(self, depth=1):
+        out = "NudgeAction(\n"
+        if isinstance(self.base_arg, Action):
+            base_segment_repr = self.base_arg.depth_repr(depth + 1)
+            out += "\t" * depth + f"base_segment={base_segment_repr}\n"
+        else:
+            out += "\t" * depth + f"base_segment={self.base_arg.depth_repr()},\n"
+
+        if isinstance(self.additional_args, Action):
+            target_repr = self.additional_args.depth_repr(depth + 1)
+            out += "\t" * depth + f"target={target_repr},\n"
+        else:
+            out += "\t" * depth + f"target={self.additional_args.depth_repr()},\n"
+        out += "\t" * depth + f"weight={self.weight},\n"
+        out += "\t" * (depth - 1) + ")"
+        return out

lib/actions/diff.py

@@ -11,9 +11,17 @@ from custom_nodes.KepPromptLang.lib.parser.registration import register_action
 
 class DiffAction(MultiArgAction):
     grammar = 'diff(" arg ("|" arg)* ")"'
-    name = "diff"
     chars = ["-", "-"]
 
+    display_name = "Difference"
+    action_name = "diff"
+    description = "Subtracts the segments in the order they are given. The first segment is subtracted from the second, then the third from the result, and so on."
+    usage_examples = [
+        "diff(The cat is|The dog is)",
+        "diff(Cat|Dog)",
+        "sum(diff(king|man)|woman)",
+    ]
+
     def __init__(self, args: List[List[Union[PromptSegment, Action]]]):
         super().__init__(args)
         self.base_arg = args[0]

lib/actions/mult.py

@@ -13,9 +13,16 @@ from custom_nodes.KepPromptLang.lib.parser.registration import register_action
 
 class MultiplyAction(MultiArgAction):
     grammar = 'mult(" arg+ ")"'
-    name = "mult"
     chars = ["[", "]"]
 
+    display_name = "Multiply"
+    action_name = "mult"
+    description = "Multiplies the provided segments or actions by the multiplier."
+    usage_examples = [
+        "mult(The cat is|2.5)",
+        "mult(Cat|-1)",
+    ]
+
     def __init__(self, args: List[List[SegOrAction]]) -> None:
         super().__init__(args)
         if len(args) != 2:
@@ -26,17 +33,17 @@ class MultiplyAction(MultiArgAction):
 
     def _parse_multiplier(self, arg: List[SegOrAction]) -> None:
         if len(arg) != 1:
-            raise ValueError("Multiply action first argument should have exactly one segment")
+            raise ValueError("Multiply actions multiplier should have exactly one segment")
 
         multiplier_seg_or_action = arg[0]
 
         if isinstance(multiplier_seg_or_action, Action):
-            raise ValueError("Multiply action should not have an action as an argument")
+            raise ValueError("Multiply actions multiplier must be a number")
 
         try:
             self.parsed_multiplier = float(multiplier_seg_or_action.text)
         except ValueError:
-            raise ValueError("Multiply action should have an integer/float as the first argument")
+            raise ValueError("Multiply action should have an integer/float as the multiplier")
 
     def token_length(self) -> int:
         """

lib/actions/neg.py

@@ -7,9 +7,16 @@ from custom_nodes.KepPromptLang.lib.parser.registration import register_action
 
 class NegAction(SingleArgAction):
     grammar = 'neg(" arg+ ")"'
-    name = "neg"
     chars = ["[", "]"]
 
+    display_name = "Negate"
+    action_name = "neg"
+    description = "Negates the provided segments or actions."
+    usage_examples = [
+        "neg(cat)",
+        "sum(king|neg(man)|women)",
+    ]
+
     def token_length(self) -> int:
         """
         Neg negates the embeddings of the base segment, so the length is the length of the base segment

lib/actions/norm.py

@@ -10,9 +10,16 @@ from custom_nodes.KepPromptLang.lib.parser.registration import register_action
 
 class NormAction(SingleArgAction):
     grammar = 'norm(" arg+ ")"'
-    name = "norm"
     chars = None
 
+    display_name = "Normalize"
+    action_name = "norm"
+    description = "Normalizes the provided segments or actions."
+    usage_examples = [
+        "norm(cat)",
+        "sum(cat|norm(sum(tiger|fish)))",
+    ]
+
     def token_length(self) -> int:
         """
         Norm normalizes the embeddings of the base segment, so the length is the length of the base segment

lib/actions/pooled_avg.py

@@ -0,0 +1,68 @@
+import torch
+from torch.nn import Embedding
+from transformers.modeling_outputs import BaseModelOutputWithPooling
+from transformers.models.clip.modeling_clip import CLIPTextTransformer
+
+from custom_nodes.KepPromptLang.lib.action.base import Action, SingleArgAction
+from custom_nodes.KepPromptLang.lib.actions.action_utils import get_total_length
+from custom_nodes.KepPromptLang.lib.fun_clip_stuff import (
+    PromptLangCLIPTextEmbeddings,
+    PrompLangCLIPTextTransformer,
+)
+from custom_nodes.KepPromptLang.lib.parser.prompt_segment import PromptSegment
+
+
+class PooledAvgAction(SingleArgAction):
+    grammar = 'pooledAvg(" arg+ ")"'
+    chars = ["[", "]"]
+
+    display_name = "Pooled Average(Experimental)"
+    action_name = "_exp-pooledAvg"
+    description = "Processes the provided segments or actions fully through CLIP and creates a pooled average of the last hidden state by averaging the last hidden state of each token."
+    usage_examples = [
+        "A cat on a _exp-pooledAvg(beautiful sunny day)",
+        "A _exp-pooledAvg(broken glass) bottle",
+    ]
+
+    def __init__(self, args):
+        super().__init__(args)
+        self.result = None
+
+    def token_length(self) -> int:
+        """
+        PooledAvg returns the average of the last hidden state, so the length is 1
+        """
+        return 1
+
+    def process_with_transformer(
+        self, transformer: CLIPTextTransformer, embedding_module: Embedding
+    ) -> None:
+        """ """
+        # SOT + tokens + EOT
+        eot_token = embedding_module.num_embeddings - 1
+        print("Using EOT token", eot_token)
+        #TODO: Play with impact of padding on pooled output
+        arg_length = get_total_length(self.arg)
+
+        # SOT + arg length + EOT
+        empty_tokens = [[49406] + [eot_token] * (arg_length + 1)]
+
+        transformer_results: BaseModelOutputWithPooling = transformer(
+            [
+                [PromptSegment(text="_Empty Batch_", tokens=empty_tokens[0])],
+                [PromptSegment(text="[SOT]", tokens=[49406])]
+                + self.arg
+                + [PromptSegment(text="[EOT]", tokens=[eot_token])],
+            ]
+        )
+        self.result = (
+            transformer_results.last_hidden_state[1, 1:-1, :].mean(dim=0).unsqueeze(0).unsqueeze(0)
+        )
+
+    def get_result(self, embedding_module: Embedding) -> torch.Tensor:
+        if self.result is not None:
+            return self.result
+
+        raise Exception(
+            "PooledAvg action result is not set. Did you forget to call process_with_transformer?"
+        )

lib/actions/pooler.py

@@ -0,0 +1,67 @@
+import torch
+from torch.nn import Embedding
+from transformers.modeling_outputs import BaseModelOutputWithPooling
+from transformers.models.clip.modeling_clip import CLIPTextTransformer
+
+from custom_nodes.KepPromptLang.lib.action.base import Action, SingleArgAction
+from custom_nodes.KepPromptLang.lib.actions.action_utils import get_total_length
+from custom_nodes.KepPromptLang.lib.fun_clip_stuff import (
+    PromptLangCLIPTextEmbeddings,
+    PrompLangCLIPTextTransformer,
+)
+from custom_nodes.KepPromptLang.lib.parser.prompt_segment import PromptSegment
+
+
+class PoolerAction(SingleArgAction):
+    grammar = 'pooler(" arg+ ")"'
+    chars = ["[", "]"]
+
+    display_name = "Pooler Output(Experimental)"
+    action_name = "_exp-pooler"
+    description = "Processes the provided segments or actions fully through CLIP and returns the pooler_output from the transformer"
+    usage_examples = [
+        "A cat on a _exp-pooler(beautiful sunny day)",
+        "A _exp-pooler(broken glass) bottle",
+    ]
+
+
+    def __init__(self, args):
+        super().__init__(args)
+        self.result = None
+
+    def token_length(self) -> int:
+        """
+        Pooler returns the embedding of the EOT token, so the length is 1
+        """
+        return 1
+
+    def process_with_transformer(
+        self, transformer: CLIPTextTransformer, embedding_module: Embedding
+    ) -> None:
+        """ """
+        # SOT + tokens + EOT
+        eot_token = embedding_module.num_embeddings - 1
+        print("Using EOT token", eot_token)
+
+        arg_length = get_total_length(self.arg)
+
+        # SOT + arg length + EOT
+        empty_tokens = [[49406] + [eot_token] * (arg_length + 1)]
+
+        transformer_results: BaseModelOutputWithPooling = transformer(
+            [
+                [PromptSegment(text="_Empty Batch_", tokens=empty_tokens[0])],
+                [PromptSegment(text="[SOT]", tokens=[49406])]
+                + self.arg
+                + [PromptSegment(text="[EOT]", tokens=[eot_token])],
+            ]
+        )
+        self.result = transformer_results.pooler_output[1].unsqueeze(0).unsqueeze(0)
+
+    def get_result(self, embedding_module: Embedding) -> torch.Tensor:
+        if self.result is not None:
+            return self.result
+
+        raise Exception(
+            "Pooled action result is not set. Did you forget to call process_with_transformer?"
+        )

lib/actions/pos_scale.py

@@ -0,0 +1,71 @@
+from typing import Tuple, List
+
+import torch
+from torch.nn import Embedding
+
+from custom_nodes.KepPromptLang.lib.action.base import (
+    Action,
+    PostModifiers,
+    MultiArgAction,
+)
+from custom_nodes.KepPromptLang.lib.actions.types import SegOrAction
+
+
+class PosScaleAction(MultiArgAction):
+    grammar = 'posScale(" arg+ ")"'
+    chars = ["[", "]"]
+
+    display_name = "Positional Embedding Scale"
+    action_name = "posScale"
+    description = "Scales(Multiplies) the positional embeddings of the provided segments or actions by the multiplier."
+    usage_examples = [
+        "A posScale(cat|1.5) on a rainy day",
+    ]
+
+    def __init__(self, args: List[List[SegOrAction]]) -> None:
+        super().__init__(args)
+        if len(args) != 2:
+            raise ValueError("PosScale action should have exactly two arguments")
+
+        self.target_arg = args[0]
+        self._parse_multiplier(args[1])
+
+    def _parse_multiplier(self, arg: List[SegOrAction]) -> None:
+        if len(arg) != 1:
+            raise ValueError(
+                "PosScale actions multiplier should have exactly one segment"
+            )
+
+        multiplier_seg_or_action = arg[0]
+
+        if isinstance(multiplier_seg_or_action, Action):
+            raise ValueError("PosScale actions multiplier must be a number")
+
+        try:
+            self.parsed_multiplier = float(multiplier_seg_or_action.text)
+        except ValueError:
+            raise ValueError(
+                "PosScale action should have an integer/float as the multiplier"
+            )
+
+    def token_length(self) -> int:
+        """
+        PosScale modifies the posional embeddings of the base segment, so the length is the length of the base segment
+        :return:
+        """
+        total_length = 0
+        for seg_or_action in self.target_arg:
+            total_length += seg_or_action.token_length()
+
+        return total_length
+
+    def get_result(self, embedding_module: Embedding) -> Tuple[torch.Tensor, PostModifiers]:
+        all_embeddings = []
+        for seg_or_action in self.target_arg:
+            if isinstance(seg_or_action, Action):
+                all_embeddings.append(seg_or_action.get_result(embedding_module))
+            else:
+                all_embeddings.append(seg_or_action.get_embeddings(embedding_module))
+
+        target_embeddings = torch.cat(all_embeddings, dim=1)
+        return target_embeddings, {"position_embed_scale": self.parsed_multiplier}

lib/actions/post_pos.py

@@ -0,0 +1,47 @@
+from typing import Tuple
+
+import torch
+from torch import Tensor
+from torch.nn import Embedding
+
+from custom_nodes.KepPromptLang.lib.action.base import (
+    SingleArgAction,
+    Action,
+    PostModifiers,
+)
+
+
+class PostPosAction(SingleArgAction):
+    grammar = 'postPos(" arg+ ")"'
+    chars = ["[", "]"]
+
+    display_name = "Ignore Positional Embeddings"
+    action_name = "postPos"
+    description = "Prevents positional embeddings from being applied to the provided segments or actions."
+    usage_examples = [
+        "A postPos(cat) on a rainy day",
+    ]
+
+    def __init__(self, args):
+        super().__init__(args)
+        self.result = None
+
+    def token_length(self) -> int:
+        """
+        PostPos returns the results of the wrapped action, so the length is the length of the wrapped action
+        """
+        total_length = 0
+        for seg_or_action in self.arg:
+            total_length += seg_or_action.token_length()
+
+        return total_length
+
+    def get_result(self, embedding_module: Embedding) -> Tuple[Tensor, PostModifiers]:
+        all_embeddings = []
+        for seg_or_action in self.arg:
+            if isinstance(seg_or_action, Action):
+                all_embeddings.append(seg_or_action.get_result(embedding_module))
+            else:
+                all_embeddings.append(seg_or_action.get_embeddings(embedding_module))
+
+        return torch.cat(all_embeddings, dim=1), PostModifiers(position_embed_scale=None, bypass_pos_embed=True)

lib/actions/rand.py

@@ -17,6 +17,14 @@ class RandAction(MultiArgAction):
     name = "rand"
     chars = None
 
+    display_name = "Random Embedding"
+    action_name = "rand"
+    description = "Returns a random embedding of the specified token length, with the values optionally bounded by the second and third arguments."
+    usage_examples = [
+        "A rand(1) cat",
+        "A rand(1|-1|1) cat",
+    ]
+
     parsed_token_length = 0
     range_min = 0
     range_max = 1

lib/actions/scale_dims.py

@@ -0,0 +1,76 @@
+from typing import List, Union
+
+import torch
+from torch.nn import Embedding
+
+from custom_nodes.KepPromptLang.lib.action.base import MultiArgAction, Action
+from custom_nodes.KepPromptLang.lib.actions.action_utils import get_embedding
+from custom_nodes.KepPromptLang.lib.parser.prompt_segment import PromptSegment
+
+
+class ScaleDims(MultiArgAction):
+    grammar = 'scaleDims(" arg ("|" arg)* ")"'
+    chars = ["-", "-"]
+
+    display_name = "Scale Dimensions"
+    action_name = "scaleDims"
+    description = "Scales the specified dimensions of the input embeddings by the specified amount"
+    usage_examples = [
+        "The scaleDims(cat|4,1.5|76,1.2) is happy",
+    ]
+
+    def __init__(self, args: List[List[Union[PromptSegment, Action]]]):
+        super().__init__(args)
+
+        self.base_arg = args[0]
+        self._parse_scale_args(args[1:])
+
+
+    def _parse_scale_args(self, args: List[List[Union[PromptSegment, Action]]]) -> None:
+        # scaleDims scale args should have format of "<dim>,<scale>" where dim is the dimension to scale and scale is the amount to scale it by
+        # scaleDims(some words|4,1.5|76,1.2)
+        self.scale_args = []
+        for arg in args:
+            if isinstance(arg, Action):
+                raise ValueError("ScaleDims scale args must be in the format of <dim>,<scale>(e.g. 4,1.5) but got an action")
+
+            if len(arg) != 1:
+                raise ValueError("ScaleDims scale args must be in the format of <dim>,<scale>(e.g. 4,1.5) but got multiple segments")
+            extracted_arg = arg[0]
+            assert isinstance(extracted_arg, PromptSegment)
+
+            if "," not in extracted_arg.text:
+                raise ValueError("ScaleDims scale args must be in the format of <dim>,<scale>(e.g. 4,1.5) but got a segment with no comma: " + extracted_arg.text)
+
+            # Split prompt segment into text and scale args
+            dim, scale = extracted_arg.text.split(",")
+            try:
+                # TODO: Check that dim is within the bounds of the embedding
+                parsed_dim = int(dim)
+            except ValueError:
+                raise ValueError("ScaleDims scale args must be in the format of <dim>,<scale>(e.g. 4,1.5) but got a segment with a non-integer dim: " + str(dim))
+
+            try:
+                parsed_scale = float(scale)
+            except ValueError:
+                raise ValueError("ScaleDims scale args must be in the format of <dim>,<scale>(e.g. 4,1.5) but got a segment with a non-float scale: " + str(scale))
+
+            self.scale_args.append((parsed_dim, parsed_scale))
+
+
+    def token_length(self) -> int:
+        # scaleDims modifies the embeddings of the base segment, so the length is the length of the base segment
+        return sum(seg_or_action.token_length() for seg_or_action in self.base_arg)
+
+    def get_result(self, embedding_module: Embedding) -> torch.Tensor:
+        # Calculate the embeddings for the base segment
+        all_base_embeddings = [
+            get_embedding(seg_or_action, embedding_module)
+            for seg_or_action in self.base_arg
+        ]
+
+        base_embeddings = torch.cat(all_base_embeddings, dim=1)
+        for dim, scale in self.scale_args:
+            base_embeddings[0, :, dim] *= scale
+
+        return base_embeddings

lib/actions/set_dims.py

@@ -0,0 +1,76 @@
+from typing import List, Union
+
+import torch
+from torch.nn import Embedding
+
+from custom_nodes.KepPromptLang.lib.action.base import MultiArgAction, Action
+from custom_nodes.KepPromptLang.lib.actions.action_utils import get_embedding
+from custom_nodes.KepPromptLang.lib.parser.prompt_segment import PromptSegment
+
+
+class SetDims(MultiArgAction):
+    grammar = 'setDims(" arg ("|" arg)* ")"'
+    chars = ["-", "-"]
+
+    display_name = "Set Dimensions"
+    action_name = "setDims"
+    description = "Sets the specified dimensions of the input embeddings to the specified value"
+    usage_examples = [
+        "The setDims(cat|4, -0.01253|76, 1.2) is happy"
+    ]
+
+    def __init__(self, args: List[List[Union[PromptSegment, Action]]]):
+        super().__init__(args)
+
+        self.base_arg = args[0]
+        self._parse_value_args(args[1:])
+
+
+    def _parse_value_args(self, args: List[List[Union[PromptSegment, Action]]]) -> None:
+        # setDims args should have format of "<dim>,<value>" where dim is the dimension to set and value is the value to set it to
+        # setDims(some words|4,-0.01254|76,1.2)
+        self.value_args = []
+        for arg in args:
+            if isinstance(arg, Action):
+                raise ValueError("SetDims value args must be in the format of <dim>,<value>(e.g. 4,1.5) but got an action")
+
+            if len(arg) != 1:
+                raise ValueError("SetDims value args must be in the format of <dim>,<value>(e.g. 4,1.5) but got multiple segments")
+            extracted_arg = arg[0]
+            assert isinstance(extracted_arg, PromptSegment)
+
+            if "," not in extracted_arg.text:
+                raise ValueError("SetDims value args must be in the format of <dim>,<value>(e.g. 4,1.5) but got a segment with no comma: " + extracted_arg.text)
+
+            # Split prompt segment into text and value args
+            dim, value = extracted_arg.text.split(",")
+            try:
+                # TODO: Check that dim is within the bounds of the embedding
+                parsed_dim = int(dim)
+            except ValueError:
+                raise ValueError("SetDims value args must be in the format of <dim>,<value>(e.g. 4,1.5) but got a segment with a non-integer dim: " + str(dim))
+
+            try:
+                parsed_value = float(value)
+            except ValueError:
+                raise ValueError("SetDims value args must be in the format of <dim>,<value>(e.g. 4,1.5) but got a segment with a non-float scale: " + str(value))
+
+            self.value_args.append((parsed_dim, parsed_value))
+
+
+    def token_length(self) -> int:
+        # setDims modifies the embeddings of the base segment, so the length is the length of the base segment
+        return sum(seg_or_action.token_length() for seg_or_action in self.base_arg)
+
+    def get_result(self, embedding_module: Embedding) -> torch.Tensor:
+        # Calculate the embeddings for the base segment
+        all_base_embeddings = [
+            get_embedding(seg_or_action, embedding_module)
+            for seg_or_action in self.base_arg
+        ]
+
+        base_embeddings = torch.cat(all_base_embeddings, dim=1)
+        for dim, value in self.value_args:
+            base_embeddings[0, :, dim] = value
+
+        return base_embeddings

lib/actions/slerp.py

@@ -12,9 +12,15 @@ from custom_nodes.KepPromptLang.lib.parser.prompt_segment import PromptSegment
 
 class SlerpAction(MultiArgAction):
     grammar = 'slerp(" arg "|" arg "|" arg ")"'
-    name = "slerp"
     chars = ["+", "+"]
 
+    display_name = "Slerp"
+    action_name = "slerp"
+    description = "Performs a slerp(Interpolation) between two segments or actions, with the given weight. The recommended weight is 0 - 1"
+    usage_examples = [
+        "The slerp(cat|dog|0.5) is happy",
+    ]
+
     def __init__(self, args: List[List[Union[PromptSegment, Action]]]) -> None:
         super().__init__(args)
 

lib/actions/sum.py

@@ -11,9 +11,15 @@ from custom_nodes.KepPromptLang.lib.parser.registration import register_action
 
 class SumAction(MultiArgAction):
     grammar = 'sum(" arg ("|" arg)+ ")"'
-    name = "sum"
     chars = ["+", "+"]
 
+    display_name = "Sum"
+    action_name = "sum"
+    description = "Adds the embeddings of the provided segments or actions."
+    usage_examples = [
+        "A happy sum(cat|dog|shark)",
+    ]
+
     def __init__(self, args: List[List[Union[PromptSegment, Action]]]) -> None:
         super().__init__(args)
 

lib/clip_config_bigg.json

@@ -0,0 +1,23 @@
+{
+  "architectures": [
+    "CLIPTextModel"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "dropout": 0.0,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_size": 1280,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 5120,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 77,
+  "model_type": "clip_text_model",
+  "num_attention_heads": 20,
+  "num_hidden_layers": 32,
+  "pad_token_id": 1,
+  "projection_dim": 1280,
+  "torch_dtype": "float32",
+  "vocab_size": 49408
+}

lib/clip_model.py

@@ -7,6 +7,8 @@ from transformers import CLIPTextConfig, modeling_utils
 
 from comfy import model_management
 import comfy.ops
+from comfy.sd1_clip import SD1ClipModel
+from comfy.sdxl_clip import SDXLClipModel
 from custom_nodes.KepPromptLang.lib.action.base import Action
 from custom_nodes.KepPromptLang.lib.actions.types import SegOrAction
 from custom_nodes.KepPromptLang.lib.fun_clip_stuff import PromptLangTextModel
@@ -14,7 +16,7 @@ from custom_nodes.KepPromptLang.lib.parser.prompt_segment import PromptSegment
 
 
 # Methods with no comment can be assumed to be the same as comfy.sd1_clip.SD1ClipModel
-class PromptLangClipModel(torch.nn.Module):
+class PromptLangSDClipModel(torch.nn.Module):
     """Uses the CLIP transformer encoder for text (from huggingface)"""
     LAYERS = [
         "last",
@@ -106,12 +108,11 @@ class PromptLangClipModel(torch.nn.Module):
                                 tokens_temp += [next_new_token]
                                 next_new_token += 1
                             else:
-                                print("WARNING: shape mismatch when trying to apply embedding, embedding will be ignored",
+                                raise Exception("WARNING: shape mismatch when trying to apply embedding. Should have been caught during tokenization.",
                                       tid_or_tensor.shape[0], current_embeds.weight.shape[1])
                     if len(tokens_temp) < segment_length:
-                        # Pretty sure this is only needed if the embedding is not the same size as the CLIP embedding
-                        print("WARNING: segment length mismatch, padding with EOS token")
-                        tokens_temp.extend([self.empty_tokens[0][-1] * (segment_length - len(tokens_temp))])
+                        # This should never happen...
+                        raise Exception("Segment size mismatch. Please submit an issue on Github.")
                     segment.tokens = tokens_temp
 
         n = token_dict_size
@@ -209,3 +210,33 @@ class PromptLangClipModel(torch.nn.Module):
         if (len(output) == 0):
             return z_empty.cpu(), first_pooled.cpu()
         return torch.cat(output, dim=-2).cpu(), first_pooled.cpu()
+
+class PromptLangSD1ClipModel(SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, clip_name="l", clip_model=PromptLangSDClipModel):
+        super().__init__()
+        self.clip_name = clip_name
+        self.clip = "clip_{}".format(self.clip_name)
+        setattr(self, self.clip, clip_model(device=device, dtype=dtype))
+
+
+class PromptLangSDXLClipModel(SDXLClipModel):
+    def __init__(self, device="cpu", dtype=None) -> None:
+        # Skip SDXLClipModel's init
+        super(SDXLClipModel, self).__init__()
+        self.clip_l = PromptLangSDClipModel(layer="hidden", layer_idx=11, device=device, dtype=dtype)
+        self.clip_l.layer_norm_hidden_state = False
+        self.clip_g = PromptLangSDXLClipG(device, dtype)
+
+class PromptLangSDXLClipG(PromptLangSDClipModel):
+    def __init__(self, device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, textmodel_path=None, dtype=None):
+        if layer == "penultimate":
+            layer="hidden"
+            layer_idx=-2
+
+        textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json")
+        super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, textmodel_path=textmodel_path, dtype=dtype)
+        self.empty_tokens = [[49406] + [49407] + [0] * 75]
+        self.layer_norm_hidden_state = False
+
+    def load_sd(self, sd):
+        return super().load_sd(sd)

lib/fun_clip_stuff.py

@@ -1,10 +1,11 @@
-from typing import Optional, Tuple, Union, List
+from typing import Optional, Tuple, Union, List, TypedDict, TYPE_CHECKING
+from importlib.metadata import version as import_version
+from packaging import version
 
 import torch
 from transformers import CLIPTextConfig
 from transformers.modeling_outputs import BaseModelOutputWithPooling
 from transformers.models.clip.modeling_clip import (
-    _expand_mask,
     CLIPTextEmbeddings,
     CLIPTextTransformer,
     CLIPTextModel,
@@ -13,6 +14,11 @@ from transformers.models.clip.modeling_clip import (
 from custom_nodes.KepPromptLang.lib.action.base import Action
 from custom_nodes.KepPromptLang.lib.actions.types import SegOrAction
 
+if TYPE_CHECKING:
+    from torch import Tensor
+    from custom_nodes.KepPromptLang.lib.action.base import PostModifiers
+
+
 def slerp(val, low, high):
     low = low.unsqueeze(0)
     high = high.unsqueeze(0)
@@ -23,6 +29,17 @@ def slerp(val, low, high):
     res = (torch.sin((1.0-val)*omega)/so).unsqueeze(1)*low + (torch.sin(val*omega)/so).unsqueeze(1) * high
     return res
 
+
+class PosModifier(TypedDict):
+    """
+    A dictionary of post modifiers for an action result.
+    """
+
+    position_embed_scale: Union[float]
+    start_idx: Union[int]
+    end_idx: Union[int]
+
+
 class PromptLangCLIPTextEmbeddings(CLIPTextEmbeddings):
     def __init__(self, config: CLIPTextConfig):
         super().__init__(config)
@@ -34,19 +51,42 @@ class PromptLangCLIPTextEmbeddings(CLIPTextEmbeddings):
             position_ids: Optional[torch.LongTensor] = None,
             inputs_embeds: Optional[torch.FloatTensor] = None,
     ) -> torch.Tensor:
-
         if input_dicts is None:
             raise ValueError("You have to specify input_dicts")
 
-        batches = []
+        batches: List[List[Tensor | Tuple[Tensor, PostModifiers] | Action]] = []
+        pos_modifiers: List[List[PosModifier]] = []
         for batch_idx, batch in enumerate(input_dicts):
             results = []
+            batch_pos_modifiers = []
+            token_idx = 0
             for seg_or_action in batch:
                 if isinstance(seg_or_action, Action):
-                    results.append(seg_or_action.get_result(self.token_embedding))
+                    action_result: Union[
+                        Tensor, Tuple[Tensor, PostModifiers]
+                    ] = seg_or_action.get_result(self.token_embedding)
+                    if isinstance(action_result, tuple):
+                        result, post_modifiers = action_result
+                        if post_modifiers.get("position_embed_scale", None) is not None:
+                            post_modifiers["start_idx"] = token_idx
+                            post_modifiers["end_idx"] = (
+                                token_idx + seg_or_action.token_length()
+                            )
+
+                        if post_modifiers.get("bypass_pos_embed", False):
+                            post_modifiers["start_idx"] = token_idx
+                            post_modifiers["end_idx"] = (
+                                token_idx + seg_or_action.token_length()
+                            )
+                        batch_pos_modifiers.append(post_modifiers)
+                    else:
+                        result = action_result
                 else:
-                    results.append(seg_or_action.get_embeddings(self.token_embedding))
+                    result = seg_or_action.get_embeddings(self.token_embedding)
+                results.append(result)
+                token_idx += seg_or_action.token_length()
             batches.append(results)
+            pos_modifiers.append(batch_pos_modifiers)
 
         seq_length = batches[0][0].shape[-2]
 
@@ -60,8 +100,28 @@ class PromptLangCLIPTextEmbeddings(CLIPTextEmbeddings):
             else:
                 embeds.append(torch.cat(batch, dim=-2))
 
-        position_embeddings = self.position_embedding(position_ids)
-        embeddings = torch.cat(embeds, dim=0) + position_embeddings
+        # Iterate over the batches and apply the pos modifiers to the position embeddings then add them to the embeddings
+        for idx, batch_pos_modifiers in enumerate(pos_modifiers):
+            position_embeddings = self.position_embedding(position_ids)
+            if len(batch_pos_modifiers) > 0:
+                print(f"Found {len(batch_pos_modifiers)} pos modifiers for batch {idx}")
+                # Apply each pos modifier to the position embeddings at the specified indices
+                for post_modifier in batch_pos_modifiers:
+                    if post_modifier.get("bypass_pos_embed", False):
+                        position_embeddings[
+                            0, post_modifier["start_idx"] : post_modifier["end_idx"]
+                        ] = 0
+                    elif post_modifier["position_embed_scale"] is not None:
+                        position_embeddings[
+                            0, post_modifier["start_idx"] : post_modifier["end_idx"]
+                        ] *= post_modifier["position_embed_scale"]
+                    else:
+                        raise ValueError(
+                            "Pos modifier must have a scale or bypass_pos_embed"
+                        )
+            # Add the possibly modified position embeddings to the embeddings
+            embeds[idx] = embeds[idx] + position_embeddings
+        embeddings = torch.cat(embeds, dim=0)
 
         return embeddings
 
@@ -70,6 +130,40 @@ class PrompLangCLIPTextTransformer(CLIPTextTransformer):
     def __init__(self, config: CLIPTextConfig):
         super().__init__(config)
         self.embeddings = PromptLangCLIPTextEmbeddings(config)
+        self.transformers_version = version.parse(import_version('transformers'))
+
+    def process_attention_mask(self, hidden_states, attention_mask, bsz, seq_len):
+        # Parse the transformer version
+        input_shape = torch.Size([bsz, seq_len])
+
+        v4_30 = version.parse('4.30.0')
+        v4_35 = version.parse('4.35')
+        if self.transformers_version < v4_30:
+            print("Using transformers < 4.30.0")
+            causal_attention_mask = self._build_causal_attention_mask(bsz, seq_len, hidden_states.dtype).to(
+                hidden_states.device)
+        elif v4_30 <= self.transformers_version < v4_35:
+            print("Using transformers >= 4.30.0 and <= 4.34.*")
+            from transformers.models.clip.modeling_clip import _make_causal_mask
+            causal_attention_mask = _make_causal_mask(input_shape, hidden_states.dtype, device=hidden_states.device)
+        else:
+            print("Using transformers >= 4.35")
+            from transformers.modeling_attn_mask_utils import _create_4d_causal_attention_mask
+            causal_attention_mask = _create_4d_causal_attention_mask(
+                input_shape, hidden_states.dtype, device=hidden_states.device
+            )
+
+        # Expand attention_mask if it exists
+        if attention_mask is not None:
+            # Import _expand_mask or _prepare_4d_attention_mask based on version
+            if self.transformers_version < v4_35:
+                from transformers.models.clip.modeling_clip import _expand_mask
+                attention_mask = _expand_mask(attention_mask, hidden_states.dtype)
+            else:
+                from transformers.modeling_attn_mask_utils import _prepare_4d_attention_mask
+                attention_mask = _prepare_4d_attention_mask(attention_mask, hidden_states.dtype)
+
+        return causal_attention_mask, attention_mask
 
     def forward(
             self,
@@ -96,29 +190,19 @@ class PrompLangCLIPTextTransformer(CLIPTextTransformer):
         # input_shape = input_ids.size()
         # input_ids = input_ids.view(-1, input_shape[-1])
 
+        for batch_idx, batch in enumerate(input_ids):
+            for seg_or_action in batch:
+                if isinstance(seg_or_action, Action):
+                    seg_or_action.process_with_transformer(
+                        self, self.embeddings.token_embedding
+                    )
+
         hidden_states = self.embeddings(input_dicts=input_ids)
 
         bsz = len(input_ids)
-        # TODO: Properly gather this
-        seq_len = 77
-        input_shape = torch.Size([bsz, seq_len])
-        # CLIP's text model uses causal mask, prepare it here.
-        # https://github.com/openai/CLIP/blob/cfcffb90e69f37bf2ff1e988237a0fbe41f33c04/clip/model.py#L324
-        ## VERSION DIFF ##
-        # transformers < 4.30.0
-        if hasattr(self, "_build_causal_attention_mask"):
-            print("Using transformers < 4.30.0")
-            causal_attention_mask = self._build_causal_attention_mask(bsz, seq_len, hidden_states.dtype).to(hidden_states.device)
-        else:
-            # transformers >= 4.30.0
-            print("Using transformers >= 4.30.0")
-            from transformers.models.clip.modeling_clip import _make_causal_mask
-            causal_attention_mask = _make_causal_mask(input_shape, hidden_states.dtype, device=hidden_states.device)
+        seq_len = hidden_states.shape[1]
 
-        # expand attention_mask
-        if attention_mask is not None:
-            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
-            attention_mask = _expand_mask(attention_mask, hidden_states.dtype)
+        causal_attention_mask, attention_mask = self.process_attention_mask(hidden_states, attention_mask, bsz, seq_len)
 
         encoder_outputs = self.encoder(
             inputs_embeds=hidden_states,
@@ -141,8 +225,11 @@ class PrompLangCLIPTextTransformer(CLIPTextTransformer):
                 if isinstance(seg_or_action, Action):
                     idx += seg_or_action.token_length()
                 else:
-                    if seg_or_action.text == '__PAD__':
+                    if seg_or_action.text == "__PAD__" or seg_or_action.text == "[EOT]":
                         break
+
+                    # Is a segment, and isn't the pad segment
+                    idx += seg_or_action.token_length()
             eot_idx.append(idx)
         # text_embeds.shape = [batch_size, sequence_length, transformer.width]
         # take features from the eot embedding (eot_token is the highest number in each sequence)

lib/parser/registration.py

@@ -5,7 +5,13 @@ from custom_nodes.KepPromptLang.lib.action.base import Action
 action_registry: Dict[str, Type[Action]] = {}
 
 def register_action(action: Type[Action]) -> None:
-    action_registry[str(action.name)] = action
+    """
+
+    :rtype: object
+    """
+    if action.action_name in action_registry:
+        raise ValueError(f"Action {action.action_name} already registered")
+    action_registry[str(action.action_name)] = action
 
 def get_action_by_name(name: str) -> Type[Action]:
     if name not in action_registry:

lib/parser/transformer.py

@@ -2,32 +2,28 @@ from typing import List
 
 from lark import Transformer, Token
 
-from comfy.sd1_clip import SD1Tokenizer
+from comfy.sd1_clip import SDTokenizer
 from custom_nodes.KepPromptLang.lib.action.base import Action, ActionArity
-from custom_nodes.KepPromptLang.lib.actions.diff import DiffAction
-from custom_nodes.KepPromptLang.lib.actions.rand import RandAction
 from custom_nodes.KepPromptLang.lib.parser.registration import get_action_by_name
 from custom_nodes.KepPromptLang.lib.parser.utils import build_prompt_segment
-from custom_nodes.KepPromptLang.lib.actions.neg import NegAction
-from custom_nodes.KepPromptLang.lib.actions.norm import NormAction
-from custom_nodes.KepPromptLang.lib.actions.sum import SumAction
 from custom_nodes.KepPromptLang.lib.parser.prompt_segment import PromptSegment
 
 
 class PromptTransformer(Transformer):
+    """
+    Transforms the parsed prompt into a list of segments and actions
+    Types from the grammar are mapped to the methods in this class
+    """
     # def WORD(self, items):
     #     return items
 
-    def __init__(self, tokenizer: SD1Tokenizer):
+    def __init__(self, tokenizer: SDTokenizer):
         super().__init__()
         self.tokenizer = tokenizer
 
     def item(self, items: List[Token]):
         for item in items:
-            if isinstance(item, Action):
-                return item
-
-            if isinstance(item, PromptSegment):
+            if isinstance(item, (Action, PromptSegment)):
                 return item
 
             if item.type == "WORD":
@@ -41,7 +37,7 @@ class PromptTransformer(Transformer):
             elif item.type == "embedding":
                 return build_prompt_segment(item, self.tokenizer)
             elif item.type == "function":
-                return item
+                raise Exception("Unexpected type in prompt transformer: function. Please report this issue on GitHub.")
             else:
                 raise Exception("Unknown item type: " + str(item.type))
 

lib/parser/utils.py

@@ -1,6 +1,6 @@
 from lark import Token
 
-from comfy.sd1_clip import SD1Tokenizer
+from comfy.sd1_clip import SDTokenizer
 from custom_nodes.KepPromptLang.lib.parser.prompt_segment import PromptSegment
 
 
@@ -11,7 +11,7 @@ def flatten_tree(tree):
         return [str(tree.data)] + sum([flatten_tree(child) for child in tree.children], [])
 
 
-def build_prompt_segment(text: str, tokenizer: SD1Tokenizer) -> PromptSegment:
+def build_prompt_segment(text: str, tokenizer: SDTokenizer) -> PromptSegment:
     split_text = text.split(" ")
     tokens = []
     for word in split_text:
@@ -24,10 +24,13 @@ def build_prompt_segment(text: str, tokenizer: SD1Tokenizer) -> PromptSegment:
             if embedding is None:
                 print(f"warning, embedding:{embedding_name} does not exist, ignoring")
             else:
-                if len(embedding.shape) == 1:
-                    tokens.append(embedding)
+                if embedding.shape[1] != tokenizer.embedding_size:
+                    print(f"warning, embedding:{embedding_name} has size {embedding.shape[1]}, expected {tokenizer.embedding_size}, ignoring")
                 else:
-                    tokens.extend(embedding)
+                    if len(embedding.shape) == 1:
+                        tokens.append(embedding)
+                    else:
+                        tokens.extend(embedding)
 
             if leftover != "":
                 word = leftover

lib/tokenizer.py

@@ -1,18 +1,18 @@
-from typing import List
+from typing import List, Dict
 
 from lark import Tree
 
-from comfy.sd1_clip import SD1Tokenizer
+from comfy.sd1_clip import SD1Tokenizer, SDTokenizer
 from custom_nodes.KepPromptLang.lib.actions.types import SegOrAction
 
 from custom_nodes.KepPromptLang.lib.parser import PromptParser
 from custom_nodes.KepPromptLang.lib.parser.transformer import PromptTransformer
 from custom_nodes.KepPromptLang.lib.parser.prompt_segment import PromptSegment
 
-class PromptLangTokenizer(SD1Tokenizer):
-    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', special_tokens=None):
-        super().__init__(tokenizer_path, max_length, pad_with_end, embedding_directory, embedding_size, embedding_key)
 
+class PromptLangSDTokenizer(SDTokenizer):
+    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l'):
+        super().__init__(tokenizer_path, max_length, pad_with_end, embedding_directory, embedding_size, embedding_key)
     """
     Doesn't actually tokenize...
     Returns batches of segments and actions
@@ -43,9 +43,9 @@ class PromptLangTokenizer(SD1Tokenizer):
 
             # If the segment is too large to fit in a single batch, pad the current batch and start a new one
             if num_tokens + batch_size > self.max_length - 1:
-                remaining_length = self.max_length - batch_size - 1 # -1 for end token
+                remaining_length = self.max_length - batch_size
                 # Pad batch
-                batch.append(PromptSegment("__PAD__", [self.end_token] + [pad_token] * remaining_length - 1))
+                batch.append(PromptSegment("__PAD__", [self.end_token] + [pad_token] * (remaining_length - 1))) # -1 for end token
                 batched_segments.append(batch)
 
                 # start new batch
@@ -66,3 +66,23 @@ class PromptLangTokenizer(SD1Tokenizer):
         #     batch_size_info(batch)
 
         return batched_segments
+
+class PromptLangSD1Tokenizer(SD1Tokenizer):
+    def __init__(self, embedding_directory=None, clip_name='l', tokenizer=PromptLangSDTokenizer) -> None:
+        super().__init__(embedding_directory, clip_name, tokenizer)
+
+
+class PromptLangSDXLClipGTokenizer(PromptLangSDTokenizer):
+    def __init__(self, tokenizer_path=None, embedding_directory=None):
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=1280, embedding_key='clip_g')
+
+class PromptLangSDXLTokenizer(SD1Tokenizer):
+    def __init__(self, embedding_directory=None) -> None:
+        self.clip_l = PromptLangSDTokenizer(embedding_directory=embedding_directory)
+        self.clip_g = PromptLangSDXLClipGTokenizer(embedding_directory=embedding_directory)
+
+    def tokenize_with_weights(self, text:str, return_word_ids=False) -> Dict[str, List[List[SegOrAction]]]:
+        out = {}
+        out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids)
+        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids)
+        return out

nodes.py

@@ -1,5 +1,5 @@
-import random
-from typing import List, Tuple
+import os
+from typing import List, Tuple, Any
 
 import numpy as np
 from PIL import Image
@@ -7,9 +7,18 @@ from PIL import Image
 import folder_paths
 import comfy.sd
 import comfy.ops
-from custom_nodes.KepPromptLang.lib.clip_model import PromptLangClipModel
+from comfy.sd2_clip import SD2ClipModel
+from comfy.sdxl_clip import SDXLClipModel
+from comfy.supported_models_base import ClipTarget
+from custom_nodes.KepPromptLang.lib.clip_model import (
+    PromptLangSDXLClipModel,
+    PromptLangSD1ClipModel,
+)
 
-from custom_nodes.KepPromptLang.lib.tokenizer import PromptLangTokenizer
+from custom_nodes.KepPromptLang.lib.tokenizer import (
+    PromptLangSDXLTokenizer,
+    PromptLangSD1Tokenizer,
+)
 
 
 class EmptyClass:
@@ -32,15 +41,22 @@ class SpecialClipLoader:
 
     @staticmethod
     def load_clip(source_clip: comfy.sd.CLIP) -> Tuple[comfy.sd.CLIP]:
-        clip_target = EmptyClass()
-        clip_target.params = {}
-        clip_target.clip = PromptLangClipModel
-        clip_target.tokenizer = PromptLangTokenizer
 
-        clip = comfy.sd.CLIP(clip_target, embedding_directory=source_clip.tokenizer.embedding_directory)
-        comfy.sd.load_clip_weights(
-            clip.cond_stage_model, source_clip.cond_stage_model.state_dict()
-        )
+        if isinstance(source_clip.cond_stage_model, SDXLClipModel):
+            clip_target = ClipTarget(PromptLangSDXLTokenizer, PromptLangSDXLClipModel)
+            clip = comfy.sd.CLIP(clip_target, embedding_directory=source_clip.tokenizer.clip_g.embedding_directory)
+            comfy.sd.load_clip_weights(clip.cond_stage_model.clip_g,source_clip.cond_stage_model.clip_g.state_dict())
+            comfy.sd.load_clip_weights(
+                clip.cond_stage_model.clip_l, source_clip.cond_stage_model.clip_l.state_dict()
+            )
+        elif isinstance(source_clip, SD2ClipModel):
+            raise ValueError("SD2 Clip model is not supported.")
+        else:
+            clip_target = ClipTarget(PromptLangSD1Tokenizer, PromptLangSD1ClipModel)
+            clip = comfy.sd.CLIP(clip_target, embedding_directory=source_clip.tokenizer.clip_l.embedding_directory)
+            comfy.sd.load_clip_weights(
+                clip.cond_stage_model, source_clip.cond_stage_model.state_dict()
+            )
         return (clip,)
 
 
@@ -49,9 +65,9 @@ def tensor2img(tensor_img) -> Image.Image:
     i_np_arr = np.clip(i, 0, 255, out=i).astype(np.uint8, copy=False)
     return Image.fromarray(i_np_arr)
 
-
 class BuildGif:
     def __init__(self) -> None:
+        self.output_dir = folder_paths.get_output_directory()
         pass
 
     @classmethod
@@ -60,6 +76,7 @@ class BuildGif:
             "required": {
                 "images": ("IMAGE",),
                 "split_every": ("INT", {"default": -1}),
+                "frame_duration": ("INT", {"default": 125}),
                 "output_mode": (
                     ["One Per Split", "Big Grid"],
                     {"default": "Big Grid"},
@@ -68,23 +85,31 @@ class BuildGif:
         }
 
     RELOAD_INST = True
-    RETURN_TYPES = ("IMAGE",)
-    RETURN_NAMES = ("Gifs",)
+    RETURN_TYPES = ()
+    # RETURN_NAMES = ("Gifs",)
     INPUT_IS_LIST = True
     FUNCTION = "build_gif"
-    OUTPUT_IS_LIST = (True,)
-    # OUTPUT_NODE = False
+    # OUTPUT_IS_LIST = (True,)
+    OUTPUT_NODE = True
 
     CATEGORY = "List Stuff"
 
-    @staticmethod
-    def build_gif(images: list, split_every: List[int], output_mode: str):
+    def build_gif(self, images: List[Any], split_every: List[int], frame_duration: List[int], output_mode: List[str]):
         print("Build GIF called!")
         print(f"{type(images)}")
 
         if len(split_every) > 1:
             raise Exception("List input for split every is not supported.")
 
+        if len(output_mode) > 1:
+            raise Exception("List input for output_mode is not supported.")
+        output_mode = output_mode[0]
+
+        if len(frame_duration) > 1:
+            raise Exception("List input for frame_duration is not supported.")
+        frame_duration = frame_duration[0]
+
+        full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix="Gif", output_dir=self.output_dir, image_width=0, image_height=0)
         split_every_val = split_every[0]
         batch_size = images[0].size()[0]
         if split_every_val == -1:
@@ -93,8 +118,6 @@ class BuildGif:
         else:
             split_chunks = int(len(images) / split_every_val)
 
-        out = []
-
         num_wide = batch_size
         num_tall = split_chunks
 
@@ -104,6 +127,7 @@ class BuildGif:
         ]
 
         frames = []
+        results = list()
 
         if output_mode == "Big Grid":
             # For every image in gif
@@ -122,8 +146,9 @@ class BuildGif:
                         )
                 frames.append(img_frame)
 
+            file = f"{filename}_{counter:05}_"
             save_path = (
-                f"{folder_paths.get_output_directory()}/{random.randint(1, 100)}"
+                f"{os.path.join(full_output_folder, file)}"
             )
             frames[0].save(
                 f"{save_path}.webp",
@@ -133,15 +158,24 @@ class BuildGif:
                 save_all=True,
                 append_images=frames[1:],
                 optimize=False,
-                duration=125,
+                duration=frame_duration,
                 loop=0,
             )
+            results.append({
+                "filename": f"{file}.webp",
+                "subfolder": subfolder,
+                "type": "output"
+            })
         elif output_mode == "One Per Split":
             for split_idx in range(int(split_chunks)):
                 split_start = split_every_val * split_idx
                 split_end = split_every_val * (split_idx + 1)
                 for batch_idx in range(batch_size):
-                    save_path = f"{folder_paths.get_output_directory()}/-{batch_idx}-{random.randint(1, 100)}"
+                    file = f"{filename}_{counter:05}_"
+                    save_path = (
+                        f"{os.path.join(full_output_folder, file)}"
+                    )
+                    counter += 1
                     print(save_path)
                     tensor2img(images[split_start][batch_idx]).save(
                         f"{save_path}.webp",
@@ -151,7 +185,12 @@ class BuildGif:
                             for nested_batch in images[split_start + 1 : split_end]
                         ],
                         optimize=False,
-                        duration=125,
+                        duration=frame_duration,
                         loop=0,
                     )
-        return (out,)
+                    results.append({
+                        "filename": f"{file}.webp",
+                        "subfolder": subfolder,
+                        "type": "output"
+                    })
+        return { "ui": { "images": results } }

requirements.txt

@@ -1 +1,2 @@
 lark
+packaging

tools/build_docs.py

@@ -0,0 +1,83 @@
+import importlib
+import inspect
+import os
+from typing import List, Type
+
+from custom_nodes.KepPromptLang.lib.action.base import Action
+
+
+EXCLUDED_MODULES = ["utils.py", "action_utils.py", "types.py"]
+def import_module_from_path(path: str):
+    module_name = path.replace("/", ".")[:-3]
+    return importlib.import_module(module_name, package="custom_nodes.KepPromptLang.lib.actions")
+
+
+# Function to find and import action classes
+def find_action_classes(directory: str) -> List[Type[Action]]:
+    action_classes = []
+    for filename in os.listdir(directory):
+        if filename.endswith(".py") and not filename.startswith("__") and filename not in EXCLUDED_MODULES:
+            module_path = os.path.join(directory, filename)
+            module = import_module_from_path(module_path)
+            for name, obj in inspect.getmembers(module, inspect.isclass):
+                if issubclass(obj, Action) and obj is not Action and obj.__name__ != "MultiArgAction" and obj.__name__ != "SingleArgAction":
+                    action_classes.append(obj)
+    return action_classes
+
+
+# Function to extract info from an action class
+def extract_class_info(cls: Type[Action]) -> dict:
+    class_info = {
+        'class_name': cls.__name__,
+        'properties': {
+            'display_name': getattr(cls, 'display_name', None),
+            'action_name': getattr(cls, 'action_name', None),
+            'description': getattr(cls, 'description', None),
+            'usage_examples': getattr(cls, 'usage_examples', None)
+        }
+    }
+    return class_info
+
+def escape_pipes(text: str) -> str:
+    return text.replace('|', '\\|')
+
+def generate_markdown_documentation(classes_info: List[dict]) -> str:
+    documentation = "# Actions Documentation\n\n"
+
+    # Define table columns
+    # columns = ["Class", "Display Name", "Action Name", "Description", "Usage Examples"]
+    columns = ["Display Name", "Action Name", "Description", "Usage Examples"]
+    documentation += "| " + " | ".join(columns) + " |\n"
+    documentation += "| --- " * len(columns) + "|\n"
+
+    for cls_info in classes_info:
+        # row = [cls_info['class_name']]
+        row = []
+        # Iterate over properties in a predefined order
+        for prop in ["display_name", "action_name", "description", "usage_examples"]:
+            prop_doc = cls_info['properties'].get(prop, 'N/A')
+
+            # Format and escape usage examples
+            if isinstance(prop_doc, list):
+                escaped_examples = [escape_pipes(example) for example in prop_doc]
+                prop_doc = "<ul>" + "".join([f"<li>{example}</li>" for example in escaped_examples]) + "</ul>"
+            else:
+                prop_doc = escape_pipes(prop_doc)
+
+            row.append(prop_doc)
+
+        documentation += "| " + " | ".join(row) + " |\n"
+
+    return documentation
+
+
+
+# Main execution
+if __name__ == "__main__":
+    actions_directory = (
+        "../lib/actions"  # Update this path as per your project structure
+    )
+    action_classes = find_action_classes(actions_directory)
+    class_infos = [extract_class_info(cls) for cls in action_classes]
+    docs = generate_markdown_documentation(class_infos)
+    print(docs)  # Or write to a file