{
"cells": [
{
"cell_type": "markdown",
"source": [
"Building on [PonderNet](https://arxiv.org/abs/2107.05407), this notebook implements a neural alternative of the [Variable Rate Coding](https://doi.org/10.32470/CCN.2019.1397-0) model to produce human-like responses.\n",
"\n",
"Given stimulus symbols as inputs, the model produces two outputs:\n",
"\n",
"- Response symbol, which, in comparison with the input stimuli, can be used to measure accuracy).\n",
"- Remaining entropy (to be contrasted against a decision threshold and ultimateely halt the process).\n",
"\n",
"Under the hood, the model uses a RNN along with multiple Poisson processes to...\n",
"\n",
"\n",
"## Resources\n",
"\n",
"- [Network model](https://drive.google.com/file/d/16eiUUwKGWfh9pu9VUxzlx046hQNHV0Qe/view?usp=sharinghttps://drive.google.com/file/d/16eiUUwKGWfh9pu9VUxzlx046hQNHV0Qe/view?usp=sharing)\n"
],
"metadata": {}
},
{
"cell_type": "markdown",
"source": [
"## Problem setting\n",
"\n",
"### Model\n",
"Given input and output data, we want to learn a supervised model of the function $X \\to y$ as follows:\n",
"\n",
"$\n",
"f: X,h_n \\mapsto \\tilde{y},h_{n+1}, \\lambda_n\n",
"$\n",
"\n",
"where $X$ and $y$ denote stimulus and response symbols, $\\lambda_n$ denotes halting probability at time $n$, and $h_{n}$ is the latent state of the model. The learninig continious up to the time point $N$.\n",
"\n",
"For the brevity and compatibility, both data are one-hot encoded.\n",
"\n",
"\n",
"### Input\n",
"\n",
"One-hot encoded symbols.\n",
"\n",
"### Output\n",
"\n",
"One-hot encoded symbols.\n",
"\n",
"### Criterion\n",
"\n",
"L = L_cross_entropy + L_halting"
],
"metadata": {}
},
{
"cell_type": "code",
"execution_count": 161,
"source": [
"# Setup and imports\n",
"import torch\n",
"from torch import nn\n",
"from torch.utils.tensorboard import SummaryWriter\n",
"\n",
"from sklearn.metrics import accuracy_score\n",
"\n",
"import numpy as np\n",
"from scipy import stats\n",
"import pandas as pd\n",
"\n",
"import matplotlib.pyplot as plt\n",
"import seaborn as sns; sns.set()\n",
"\n",
"import tensorflow as tf\n",
"import tensorboard as tb\n",
"tf.io.gfile = tb.compat.tensorflow_stub.io.gfile"
],
"outputs": [],
"metadata": {}
},
{
"cell_type": "code",
"execution_count": 2,
"source": [
"# produce a tarin of spikes and store timestamps of each spike in `spike_timestamps`.\n",
"\n",
"signal_rate = 2\n",
"noise_rate = 1\n",
"rate = signal_rate + noise_rate\n",
"max_duration_in_sec = 10.\n",
"resolution_in_sec = .1\n",
"\n",
"n_total_timesteps = int(max_duration_in_sec / resolution_in_sec)\n",
"n_spikes = np.random.poisson(rate * max_duration_in_sec)\n",
"\n",
"# method 1: shuffle timesteps\n",
"spike_timesteps = np.sort(np.random.choice(n_total_timesteps, size=n_spikes, replace=False))\n",
"\n",
"# method 2: exponential isi -> timestamps\n",
"# isi = np.random.exponential(1 / rate, n_spikes)\n",
"# spike_timestamps = np.cumsum(isi)\n",
"\n",
"# method 3: homogenous spikes -> timestamps\n",
"# spike_timestamps = stats.uniform.rvs(loc=0, scale=max_duration_in_sec, size=n_spikes)"
],
"outputs": [],
"metadata": {}
},
{
"cell_type": "markdown",
"source": [
"## PonderRNN"
],
"metadata": {}
},
{
"cell_type": "markdown",
"source": [
"## Mock data"
],
"metadata": {}
},
{
"cell_type": "code",
"execution_count": 3,
"source": [
"\n",
"\n",
"def generate_mock_data(n_subjects, n_trials, n_stimuli):\n",
" \"\"\"[summary]\n",
"\n",
" # TODO required data columns: subject_index, trial_index, stimulus_index, accuracy, response_time\n",
"\n",
" Args:\n",
" n_subjects (int): [description]\n",
" n_trials (int): [description]\n",
" n_stimuli (int): [description]\n",
"\n",
" Returns:\n",
" (X, accuracies, response_times): A tuple containing generated mock X, accuracies, and response_times (in sec).\n",
" \"\"\"\n",
" # stimuli\n",
" X = np.random.randint(low=1, high=n_stimuli+1, size=(n_subjects, n_trials))\n",
"\n",
" # response accuracy\n",
" subject_accuracies = np.random.uniform(low=0.2, high=1.0, size=n_subjects)\n",
" subject_accuracies = np.round(subject_accuracies * n_trials) / n_trials\n",
" accuracies = np.empty(shape=(n_subjects, n_trials))\n",
" for subj in range(n_subjects):\n",
" accuracies[subj,:] = np.random.choice(\n",
" [0,1],\n",
" p=[1-subject_accuracies[subj],subject_accuracies[subj]],\n",
" size=n_trials)\n",
"\n",
" # generate output w.r.t the accuracy (and fill incorrect trials with invalid response)\n",
" y = np.where(accuracies == 1., X, X+1 % (n_stimuli+1))\n",
"\n",
" # response time\n",
" response_times = np.random.exponential(.5, size=accuracies.shape)\n",
"\n",
" return X, y, accuracies, response_times"
],
"outputs": [],
"metadata": {}
},
{
"cell_type": "code",
"execution_count": 105,
"source": [
"# mock data parameters\n",
"n_subjects = 1\n",
"n_trials = 20\n",
"n_stimuli = 6\n",
"\n",
"X, y, accuracies, response_times = generate_mock_data(n_subjects, n_trials, n_stimuli)"
],
"outputs": [],
"metadata": {}
},
{
"cell_type": "code",
"execution_count": 172,
"source": [
"logs = SummaryWriter()\n",
"\n",
"class PonderRNN(nn.Module):\n",
" def __init__(self, n_inputs, n_channels, n_outputs):\n",
" super(PonderRNN, self).__init__()\n",
" self.fc1 = nn.Linear(n_inputs, n_channels, bias=False)\n",
" self.rnn = nn.RNN(n_channels, n_channels)\n",
" self.fc2 = nn.Linear(n_channels,n_outputs, bias=False)\n",
" self.softmax = nn.Softmax(dim=2)\n",
"\n",
" def forward(self, x):\n",
" sent_msg = self.fc1(x)\n",
" rcvd_msg, embedding = self.rnn(sent_msg)\n",
" logits = self.fc2(rcvd_msg)\n",
"\n",
" out = self.softmax(logits).squeeze()\n",
"\n",
" halting_prob = 0. # TODO\n",
"\n",
" return out, embedding, halting_prob\n",
"\n",
"def criterion(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:\n",
" return torch.functional.binary_cross_entropy_with_logits()\n",
"\n",
"n_epoches = 1500\n",
"\n",
"model = PonderRNN(n_stimuli+1, n_stimuli, n_stimuli+1)\n",
"optimizer = torch.optim.Adam(model.parameters(), lr=0.001)\n",
"criterion = torch.nn.CrossEntropyLoss()\n",
"\n",
"X_train = nn.functional.one_hot(torch.tensor(X)).type(torch.float)\n",
"y_train = torch.tensor(y) - 1\n",
"\n",
"for epoch in range(n_epoches):\n",
" model.train()\n",
" optimizer.zero_grad()\n",
" y_pred, h, _ = model(X_train)\n",
"\n",
" # logs.add_embedding(h.reshape(n_trials,n_stimuli), global_step=epoch, tag='embedding')\n",
"\n",
" model_accuracy = accuracy_score(y_train.squeeze(), torch.argmax(y_pred.detach(),dim=1))\n",
" logs.add_scalar('accurracy/train', model_accuracy, epoch) \n",
"\n",
" loss = criterion(y_pred, y_train.squeeze())\n",
" logs.add_scalar('loss/train', loss, epoch)\n",
" loss.backward()\n",
" optimizer.step()\n",
"\n",
"# tensorboard --logdir==runs/"
],
"outputs": [
{
"output_type": "stream",
"name": "stdout",
"text": [
"torch.Size([1, 20, 6])\n"
]
},
{
"output_type": "error",
"ename": "NameError",
"evalue": "name 'xx' is not defined",
"traceback": [
"\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
"\u001b[0;31mNameError\u001b[0m Traceback (most recent call last)",
"\u001b[0;32m<ipython-input-172-3dd48c33ed48>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[1;32m 39\u001b[0m \u001b[0mprint\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mh\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0msize\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 40\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 41\u001b[0;31m \u001b[0mxx\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 42\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 43\u001b[0m \u001b[0mlogs\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0madd_embedding\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mh\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mreshape\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mn_trials\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0mn_stimuli\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mglobal_step\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mepoch\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mtag\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;34m'embedding'\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
"\u001b[0;31mNameError\u001b[0m: name 'xx' is not defined"
]
}
],
"metadata": {}
},
{
"cell_type": "code",
"execution_count": 137,
"source": [
"model.eval()\n",
"y_pred, _ = model(X_train)\n",
"y_pred = np.argmax(y_pred.detach().numpy(), axis=1) + 1\n",
"y_pred, y"
],
"outputs": [
{
"output_type": "execute_result",
"data": {
"text/plain": [
"(array([2, 6, 5, 2, 6, 4, 4, 7, 3, 7, 5, 3, 4, 3, 3, 6, 3, 5, 7, 2]),\n",
" array([[2, 6, 5, 1, 6, 4, 4, 6, 3, 7, 5, 3, 4, 3, 3, 6, 3, 5, 7, 2]]))"
]
},
"metadata": {},
"execution_count": 137
}
],
"metadata": {}
}
],
"metadata": {
"orig_nbformat": 4,
"language_info": {
"name": "python",
"version": "3.9.4",
"mimetype": "text/x-python",
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"pygments_lexer": "ipython3",
"nbconvert_exporter": "python",
"file_extension": ".py"
},
"kernelspec": {
"name": "python3",
"display_name": "Python 3.9.4 64-bit ('py3': conda)"
},
"interpreter": {
"hash": "5ddcf14c786c671500c086f61f0b66d0417d6c58ff12753e346e191a84f72b84"
}
},
"nbformat": 4,
"nbformat_minor": 2
}
