diff --git a/python/Demonstration.ipynb b/python/Demonstration.ipynb
index 52ac7c12..84e9cd18 100644
--- a/python/Demonstration.ipynb
+++ b/python/Demonstration.ipynb
@@ -4,7 +4,7 @@
"cell_type": "markdown",
"metadata": {},
"source": [
- "# CircuitsVis Demonstration"
+ "# CircuitsVis Demonstration\n"
]
},
{
@@ -15,7 +15,7 @@
"source": [
"## Setup/Imports\n",
"\n",
- "__Note:__ To run Jupyter directly within this repo, you may need to run `poetry run pip install jupyter`."
+ "**Note:** To run Jupyter directly within this repo, you may need to run `poetry run pip install jupyter`.\n"
]
},
{
@@ -27,8 +27,7 @@
"name": "stdout",
"output_type": "stream",
"text": [
- "The autoreload extension is already loaded. To reload it, use:\n",
- " %reload_ext autoreload\n"
+ "In dev mode: True\n"
]
}
],
@@ -39,12 +38,14 @@
"\n",
"# Imports\n",
"import numpy as np\n",
- "from circuitsvis.attention import attention_patterns, attention_pattern\n",
+ "from circuitsvis.attention import attention_patterns, attention_pattern, attention_heads\n",
"from circuitsvis.activations import text_neuron_activations\n",
- "from circuitsvis.examples import hello\n",
"from circuitsvis.tokens import colored_tokens\n",
"from circuitsvis.topk_tokens import topk_tokens\n",
- "from circuitsvis.topk_samples import topk_samples"
+ "from circuitsvis.topk_samples import topk_samples\n",
+ "\n",
+ "from circuitsvis.utils.render import is_in_dev_mode\n",
+ "print(\"In dev mode:\", is_in_dev_mode())"
]
},
{
@@ -53,82 +54,9906 @@
"tags": []
},
"source": [
- "## Built In Visualizations"
+ "## Built In Visualizations\n"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "### Activations\n"
+ ]
+ },
+ {
+ "attachments": {},
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "#### Text Neuron Activations (single sample)\n"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 2,
+ "metadata": {},
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ " "
+ ],
+ "text/plain": [
+ ""
+ ]
+ },
+ "execution_count": 2,
+ "metadata": {},
+ "output_type": "execute_result"
+ }
+ ],
+ "source": [
+ "tokens = ['Hi', ' and', ' welcome', ' to', ' the', ' Attention', ' Patterns', ' example']\n",
+ "n_layers = 3\n",
+ "n_neurons_per_layer = 4\n",
+ "activations = np.random.normal(size=(len(tokens), n_layers, n_neurons_per_layer))\n",
+ "activations = np.exp(activations) / np.exp(activations).sum(axis=0, keepdims=True)\n",
+ "text_neuron_activations(tokens=tokens, activations=activations)"
+ ]
+ },
+ {
+ "attachments": {},
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "#### Text Neuron Activations (multiple samples)\n"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 3,
+ "metadata": {},
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ " "
+ ],
+ "text/plain": [
+ ""
+ ]
+ },
+ "execution_count": 3,
+ "metadata": {},
+ "output_type": "execute_result"
+ }
+ ],
+ "source": [
+ "tokens = [['Hi', ' and', ' welcome', ' to', ' the', ' Attention', ' Patterns', ' example'], ['This', ' is', ' another', ' example', ' of', ' colored', ' tokens'], ['And', ' here', ' another', ' example', ' of', ' colored', ' tokens', ' with', ' more', ' words.'], ['This', ' is', ' another', ' example', ' of', ' tokens.']]\n",
+ "n_layers = 3\n",
+ "n_neurons_per_layer = 4\n",
+ "activations = []\n",
+ "for sample in tokens:\n",
+ " sample_activations = np.random.normal(size=(len(sample), n_layers, n_neurons_per_layer)) * 5\n",
+ " activations.append(sample_activations)\n",
+ "text_neuron_activations(tokens=tokens, activations=activations)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
- "### Activations"
+ "### Attention\n"
]
},
{
- "attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
- "#### Text Neuron Activations (single sample)"
+ "#### Attention Heads\n"
]
},
{
"cell_type": "code",
- "execution_count": 2,
+ "execution_count": 4,
"metadata": {},
"outputs": [
{
"data": {
"text/html": [
- "\n",
+ "\n",
" "
],
"text/plain": [
- ""
+ ""
]
},
- "execution_count": 2,
+ "execution_count": 4,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"tokens = ['Hi', ' and', ' welcome', ' to', ' the', ' Attention', ' Patterns', ' example']\n",
- "n_layers = 3\n",
- "n_neurons_per_layer = 4\n",
- "activations = np.random.normal(size=(len(tokens), n_layers, n_neurons_per_layer))\n",
- "activations = np.exp(activations) / np.exp(activations).sum(axis=0, keepdims=True) \n",
- "text_neuron_activations(tokens=tokens, activations=activations)"
- ]
- },
- {
- "attachments": {},
- "cell_type": "markdown",
- "metadata": {},
- "source": [
- "#### Text Neuron Activations (multiple samples)"
+ "n_tokens = len(tokens)\n",
+ "single_head_attention = np.tril(np.random.normal(loc=0.3, scale=0.2, size=(n_tokens, n_tokens)))\n",
+ "attention_heads(tokens=tokens, attention=single_head_attention, show_tokens=True)"
]
},
{
"cell_type": "code",
- "execution_count": 3,
+ "execution_count": 5,
"metadata": {},
"outputs": [
{
"data": {
"text/html": [
- "\n",
+ "\n",
" "
],
"text/plain": [
- ""
+ ""
]
},
- "execution_count": 3,
+ "execution_count": 5,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
- "tokens = [['Hi', ' and', ' welcome', ' to', ' the', ' Attention', ' Patterns', ' example'], ['This', ' is', ' another', ' example', ' of', ' colored', ' tokens'], ['And', ' here', ' another', ' example', ' of', ' colored', ' tokens', ' with', ' more', ' words.'], ['This', ' is', ' another', ' example', ' of', ' tokens.']]\n",
"n_layers = 3\n",
- "n_neurons_per_layer = 4\n",
- "activations = []\n",
- "for sample in tokens:\n",
- " sample_activations = np.random.normal(size=(len(sample), n_layers, n_neurons_per_layer)) * 5\n",
- " activations.append(sample_activations)\n",
- "text_neuron_activations(tokens=tokens, activations=activations)\n",
- "\n"
- ]
- },
- {
- "cell_type": "markdown",
- "metadata": {},
- "source": [
- "### Attention"
+ "n_heads = 4\n",
+ "\n",
+ "multi_head_attention = np.tril(np.random.normal(loc=0.3, scale=0.2, size=(n_layers * n_heads, n_tokens, n_tokens)))\n",
+ "head_names = [f\"L{layer}H{head}\" for layer in range(n_layers) for head in range(n_heads)]\n",
+ "attention_heads(tokens=tokens, attention=multi_head_attention, attention_head_names=head_names, show_tokens=True, match_color=True)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
- "#### Attention Pattern (single head)"
+ "#### Attention Pattern (single head)\n"
]
},
{
"cell_type": "code",
- "execution_count": 4,
+ "execution_count": 6,
"metadata": {},
"outputs": [
{
"data": {
"text/html": [
- "\n",
+ "\n",
" "
],
"text/plain": [
- ""
+ ""
]
},
- "execution_count": 4,
+ "execution_count": 6,
"metadata": {},
"output_type": "execute_result"
}
@@ -14797,67 +24602,67 @@
"cell_type": "markdown",
"metadata": {},
"source": [
- "#### Attention Patterns"
+ "#### Attention Patterns [deprecated]\n"
]
},
{
"cell_type": "code",
- "execution_count": 5,
+ "execution_count": 7,
"metadata": {},
"outputs": [
{
"data": {
"text/html": [
- "\n",
+ "\n",
"