You cannot select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
338 lines
9.7 KiB
Plaintext
338 lines
9.7 KiB
Plaintext
{
|
|
"cells": [
|
|
{
|
|
"cell_type": "markdown",
|
|
"id": "prepared-nitrogen",
|
|
"metadata": {},
|
|
"source": [
|
|
"Note on pytorch. NN optimization acts imperitively/by side effect as follows.\n",
|
|
" - Define model\n",
|
|
" - loop\n",
|
|
" - Calculate loss\n",
|
|
" - Zero gradients\n",
|
|
" - backprop to model\n",
|
|
" - check conditions for exit\n",
|
|
" - report diagnostics\n",
|
|
" - disect results\n",
|
|
" \n",
|
|
" \n",
|
|
"## Split result from NN\n",
|
|
"Goal is to train the NN and then get a couple of outputs at the end that can be used to split between value function partials and launch functions."
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 1,
|
|
"id": "grateful-conviction",
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"import torch"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 2,
|
|
"id": "incorrect-animal",
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"class DoubleNetwork(torch.nn.Module):\n",
|
|
" def __init__(self, input_size,output_size,layers_size):\n",
|
|
" super().__init__()\n",
|
|
" \n",
|
|
" #So, this next section constructs different layers within the NN\n",
|
|
" #sinlge linear section\n",
|
|
" self.linear_step_1a = torch.nn.Linear(input_size,layers_size)\n",
|
|
" \n",
|
|
" #single linear section\n",
|
|
" self.linear_step_2a = torch.nn.Linear(layers_size,output_size)\n",
|
|
" self.linear_step_2b = torch.nn.Linear(layers_size,output_size)\n",
|
|
" \n",
|
|
" def forward(self, input_values):\n",
|
|
" \n",
|
|
" intermediate_values_a = self.linear_step_1a(input_values)\n",
|
|
" \n",
|
|
" out_values_a = self.linear_step_2a(intermediate_values_a)\n",
|
|
" out_values_b = self.linear_step_2b(intermediate_values_a)\n",
|
|
" \n",
|
|
" return out_values_a,out_values_b"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 3,
|
|
"id": "ruled-letter",
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"\n",
|
|
" tensor(3.5646, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(11.7849, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(24.8772, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(5.4752, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.4457, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0925, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0490, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0290, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0178, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0111, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0070, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0045, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0029, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0019, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0012, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0008, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0005, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0003, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0002, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0001, grad_fn=<AddBackward0>)\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"model = DoubleNetwork(input_size = 5, output_size=5, layers_size=15)\n",
|
|
"\n",
|
|
"data_in = torch.tensor([1.5,2,3,4,5])\n",
|
|
"\n",
|
|
"data_in\n",
|
|
"\n",
|
|
"target = torch.zeros(5)\n",
|
|
"\n",
|
|
"def loss_fn2(output,target):\n",
|
|
" return sum((output[1] +output[0] - target)**2)\n",
|
|
" #could add a simplicity assumption i.e. l1 on parameters.\n",
|
|
"\n",
|
|
"#Prep Optimizer\n",
|
|
"optimizer = torch.optim.SGD(model.parameters(),lr=0.01)\n",
|
|
"\n",
|
|
"for i in range(20):\n",
|
|
" #training loop\n",
|
|
" optimizer.zero_grad()\n",
|
|
"\n",
|
|
" output = model.forward(data_in)\n",
|
|
" output\n",
|
|
"\n",
|
|
" l = loss_fn2(output, target)\n",
|
|
"\n",
|
|
" l.backward()\n",
|
|
"\n",
|
|
" optimizer.step()\n",
|
|
"\n",
|
|
" print(\"\\n\",l)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 4,
|
|
"id": "quantitative-keeping",
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"class SplitNetwork(torch.nn.Module):\n",
|
|
" def __init__(self, input_size,output_size_a,output_size_b,layers_size):\n",
|
|
" super().__init__()\n",
|
|
" \n",
|
|
" #So, this next section constructs different layers within the NN\n",
|
|
" #sinlge linear section\n",
|
|
" self.linear_step_1 = torch.nn.Linear(input_size,layers_size)\n",
|
|
" self.linear_step_2 = torch.nn.Linear(layers_size,layers_size)\n",
|
|
" self.linear_step_3 = torch.nn.Linear(layers_size,layers_size)\n",
|
|
" self.linear_step_4 = torch.nn.Linear(layers_size,layers_size)\n",
|
|
" \n",
|
|
" #single linear section\n",
|
|
" self.linear_step_split_a = torch.nn.Linear(layers_size,output_size_a)\n",
|
|
" self.linear_step_split_b = torch.nn.Linear(layers_size,output_size_b)\n",
|
|
" \n",
|
|
" def forward(self, input_values):\n",
|
|
" \n",
|
|
" intermediate_values = self.linear_step_1(input_values)\n",
|
|
" intermediate_values = self.linear_step_2(intermediate_values)\n",
|
|
" intermediate_values = self.linear_step_3(intermediate_values)\n",
|
|
" intermediate_values = self.linear_step_4(intermediate_values)\n",
|
|
" \n",
|
|
" out_values_a = self.linear_step_split_a(intermediate_values)\n",
|
|
" out_values_b = self.linear_step_split_b(intermediate_values)\n",
|
|
" \n",
|
|
" return out_values_a,out_values_b"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 5,
|
|
"id": "vietnamese-prophet",
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"model = SplitNetwork(input_size = 6, output_size_a=5, output_size_b=7, layers_size=15)\n",
|
|
"\n",
|
|
"data_in = torch.tensor([1.5,2,3,4,5,6])\n",
|
|
"\n",
|
|
"\n",
|
|
"target_a = torch.zeros(5)\n",
|
|
"target_b = torch.ones(7)\n",
|
|
"\n",
|
|
"def loss_fn3(output,target_a, target_b):\n",
|
|
" return sum((output[0] - target_a)**2) + sum((output[1] - target_b)**2)\n",
|
|
" #could add a simplicity assumption i.e. l1 on parameters."
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 6,
|
|
"id": "limiting-slide",
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"\n",
|
|
" tensor(9.6420, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(4.1914, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(5.1337, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(1.4943, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.5210, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.1217, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0605, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0256, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0126, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0057, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0028, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0013, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0006, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0003, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(0.0001, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(7.2050e-05, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(3.5139e-05, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(1.7068e-05, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(8.3342e-06, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(4.0624e-06, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(1.9857e-06, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(9.7029e-07, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(4.7492e-07, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(2.3232e-07, grad_fn=<AddBackward0>)\n",
|
|
"\n",
|
|
" tensor(1.1381e-07, grad_fn=<AddBackward0>)\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"#Prep Optimizer\n",
|
|
"optimizer = torch.optim.SGD(model.parameters(),lr=0.01)\n",
|
|
"\n",
|
|
"for i in range(25):\n",
|
|
" #training loop\n",
|
|
" optimizer.zero_grad()\n",
|
|
"\n",
|
|
" output = model.forward(data_in)\n",
|
|
" output\n",
|
|
"\n",
|
|
" l = loss_fn3(output, target_a, target_b)\n",
|
|
"\n",
|
|
" l.backward()\n",
|
|
"\n",
|
|
" optimizer.step()\n",
|
|
"\n",
|
|
" print(\"\\n\",l)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 7,
|
|
"id": "elder-karen",
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"data": {
|
|
"text/plain": [
|
|
"(tensor([ 3.4232e-05, 3.7350e-05, 5.3748e-05, -2.7344e-05, -1.0052e-04],\n",
|
|
" grad_fn=<AddBackward0>),\n",
|
|
" tensor([1.0001, 1.0001, 1.0000, 1.0000, 1.0001, 1.0000, 1.0001],\n",
|
|
" grad_fn=<AddBackward0>))"
|
|
]
|
|
},
|
|
"execution_count": 7,
|
|
"metadata": {},
|
|
"output_type": "execute_result"
|
|
}
|
|
],
|
|
"source": []
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"id": "agreed-community",
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": []
|
|
}
|
|
],
|
|
"metadata": {
|
|
"kernelspec": {
|
|
"display_name": "Python 3",
|
|
"language": "python",
|
|
"name": "python3"
|
|
},
|
|
"language_info": {
|
|
"codemirror_mode": {
|
|
"name": "ipython",
|
|
"version": 3
|
|
},
|
|
"file_extension": ".py",
|
|
"mimetype": "text/x-python",
|
|
"name": "python",
|
|
"nbconvert_exporter": "python",
|
|
"pygments_lexer": "ipython3",
|
|
"version": "3.8.8"
|
|
}
|
|
},
|
|
"nbformat": 4,
|
|
"nbformat_minor": 5
|
|
}
|