Deploying to gh-pages from @ 7732bb4 🚀

.buildinfo

@@ -1,4 +1,4 @@
 # Sphinx build info version 1
 # This file hashes the configuration used when building these files. When it is not found, a full rebuild will be done.
-config: 91dfa39bc45fef944cf143f45c8d2be4
+config: 6f0c17a356e9b39c9908867ce6b2ddbf
 tags: 645f666f9bcd5a90fca523b33c5a78b7

_downloads/0749894295c4ed0d731492395c04895e/train_hmm.py

@@ -1,8 +1,12 @@
 """
 Example
 =======
+
+dddd
 """
 
+# sphinx_gallery_thumbnail_path = 'imgs/juice.png'
+
 import pyjuice as juice
 import torch
 import torchvision

_downloads/2e145951c19f5006f1bdad9dcdd41f8d/train_mnist_pdhclt.ipynb

@@ -4,7 +4,7 @@
       "cell_type": "markdown",
       "metadata": {},
       "source": [
-        "\n# Example\n"
+        "\n# Example\n\naaaa\n"
       ]
     },
     {
@@ -15,7 +15,7 @@
       },
       "outputs": [],
       "source": [
-        "import pyjuice as juice\nimport torch\nimport torchvision\nimport time\nfrom torch.utils.data import TensorDataset, DataLoader\n\n\ndef evaluate(pc, loader):\n    lls_total = 0.0\n    for batch in loader:\n        x = batch[0].to(pc.device)\n        lls = pc(x)\n        lls_total += lls.mean().detach().cpu().numpy().item()\n    \n    lls_total /= len(loader)\n    return lls_total\n\n\ndef mini_batch_em_epoch(num_epochs, pc, optimizer, scheduler, train_loader, test_loader, device):\n    for epoch in range(num_epochs):\n        t0 = time.time()\n        train_ll = 0.0\n        for batch in train_loader:\n            x = batch[0].to(device)\n\n            optimizer.zero_grad()\n\n            lls = pc(x)\n            lls.mean().backward()\n\n            train_ll += lls.mean().detach().cpu().numpy().item()\n\n            optimizer.step()\n            if scheduler is not None:\n                scheduler.step()\n\n        train_ll /= len(train_loader)\n\n        t1 = time.time()\n        test_ll = evaluate(pc, loader=test_loader)\n        t2 = time.time()\n\n        print(f\"[Epoch {epoch}/{num_epochs}][train LL: {train_ll:.2f}; test LL: {test_ll:.2f}].....[train forward+backward+step {t1-t0:.2f}; test forward {t2-t1:.2f}] \")\n\n\ndef full_batch_em_epoch(pc, train_loader, test_loader, device):\n    with torch.no_grad():\n        t0 = time.time()\n        train_ll = 0.0\n        for batch in train_loader:\n            x = batch[0].to(device)\n\n            lls = pc(x)\n            pc.backward(x, flows_memory = 1.0)\n\n            train_ll += lls.mean().detach().cpu().numpy().item()\n\n        pc.mini_batch_em(step_size = 1.0, pseudocount = 0.1)\n\n        train_ll /= len(train_loader)\n\n        t1 = time.time()\n        test_ll = evaluate(pc, loader=test_loader)\n        t2 = time.time()\n        print(f\"[train LL: {train_ll:.2f}; test LL: {test_ll:.2f}].....[train forward+backward+step {t1-t0:.2f}; test forward {t2-t1:.2f}] \")\n\n\ndef pd_hclt_test():\n\n    device = torch.device(\"cuda:0\")\n\n    train_dataset = torchvision.datasets.MNIST(root = \"./examples/data\", train = True, download = True)\n    test_dataset = torchvision.datasets.MNIST(root = \"./examples/data\", train = False, download = True)\n\n    train_data = train_dataset.data.reshape(60000, 28*28)\n    test_data = test_dataset.data.reshape(10000, 28*28)\n\n    num_features = train_data.size(1)\n\n    train_loader = DataLoader(\n        dataset = TensorDataset(train_data),\n        batch_size = 512,\n        shuffle = True,\n        drop_last = True\n    )\n    test_loader = DataLoader(\n        dataset = TensorDataset(test_data),\n        batch_size = 512,\n        shuffle = False,\n        drop_last = True\n    )\n\n    ns = juice.structures.PDHCLT(\n        train_data.cuda(),\n        data_shape = (28, 28),\n        num_latents = 128,\n        split_intervals = (4, 4),\n        structure_type = \"sum_dominated\"\n    )\n    pc = juice.TensorCircuit(ns)\n\n    pc.to(device)\n\n    optimizer = juice.optim.CircuitOptimizer(pc, lr = 0.1, pseudocount = 0.1)\n    scheduler = juice.optim.CircuitScheduler(\n        optimizer, \n        method = \"multi_linear\", \n        lrs = [0.9, 0.1, 0.05], \n        milestone_steps = [0, len(train_loader) * 100, len(train_loader) * 350]\n    )\n\n    pc.print_statistics()\n\n    mini_batch_em_epoch(350, pc, optimizer, None, train_loader, test_loader, device)\n    full_batch_em_epoch(pc, train_loader, test_loader, device)\n\n\nif __name__ == \"__main__\":\n    torch.manual_seed(2391)\n    pd_hclt_test()"
+        "# sphinx_gallery_thumbnail_path = 'imgs/juice.png'\n\nimport pyjuice as juice\nimport torch\nimport torchvision\nimport time\nfrom torch.utils.data import TensorDataset, DataLoader\n\n\ndef evaluate(pc, loader):\n    lls_total = 0.0\n    for batch in loader:\n        x = batch[0].to(pc.device)\n        lls = pc(x)\n        lls_total += lls.mean().detach().cpu().numpy().item()\n    \n    lls_total /= len(loader)\n    return lls_total\n\n\ndef mini_batch_em_epoch(num_epochs, pc, optimizer, scheduler, train_loader, test_loader, device):\n    for epoch in range(num_epochs):\n        t0 = time.time()\n        train_ll = 0.0\n        for batch in train_loader:\n            x = batch[0].to(device)\n\n            optimizer.zero_grad()\n\n            lls = pc(x)\n            lls.mean().backward()\n\n            train_ll += lls.mean().detach().cpu().numpy().item()\n\n            optimizer.step()\n            if scheduler is not None:\n                scheduler.step()\n\n        train_ll /= len(train_loader)\n\n        t1 = time.time()\n        test_ll = evaluate(pc, loader=test_loader)\n        t2 = time.time()\n\n        print(f\"[Epoch {epoch}/{num_epochs}][train LL: {train_ll:.2f}; test LL: {test_ll:.2f}].....[train forward+backward+step {t1-t0:.2f}; test forward {t2-t1:.2f}] \")\n\n\ndef full_batch_em_epoch(pc, train_loader, test_loader, device):\n    with torch.no_grad():\n        t0 = time.time()\n        train_ll = 0.0\n        for batch in train_loader:\n            x = batch[0].to(device)\n\n            lls = pc(x)\n            pc.backward(x, flows_memory = 1.0)\n\n            train_ll += lls.mean().detach().cpu().numpy().item()\n\n        pc.mini_batch_em(step_size = 1.0, pseudocount = 0.1)\n\n        train_ll /= len(train_loader)\n\n        t1 = time.time()\n        test_ll = evaluate(pc, loader=test_loader)\n        t2 = time.time()\n        print(f\"[train LL: {train_ll:.2f}; test LL: {test_ll:.2f}].....[train forward+backward+step {t1-t0:.2f}; test forward {t2-t1:.2f}] \")\n\n\ndef pd_hclt_test():\n\n    device = torch.device(\"cuda:0\")\n\n    train_dataset = torchvision.datasets.MNIST(root = \"./examples/data\", train = True, download = True)\n    test_dataset = torchvision.datasets.MNIST(root = \"./examples/data\", train = False, download = True)\n\n    train_data = train_dataset.data.reshape(60000, 28*28)\n    test_data = test_dataset.data.reshape(10000, 28*28)\n\n    num_features = train_data.size(1)\n\n    train_loader = DataLoader(\n        dataset = TensorDataset(train_data),\n        batch_size = 512,\n        shuffle = True,\n        drop_last = True\n    )\n    test_loader = DataLoader(\n        dataset = TensorDataset(test_data),\n        batch_size = 512,\n        shuffle = False,\n        drop_last = True\n    )\n\n    ns = juice.structures.PDHCLT(\n        train_data.cuda(),\n        data_shape = (28, 28),\n        num_latents = 128,\n        split_intervals = (4, 4),\n        structure_type = \"sum_dominated\"\n    )\n    pc = juice.TensorCircuit(ns)\n\n    pc.to(device)\n\n    optimizer = juice.optim.CircuitOptimizer(pc, lr = 0.1, pseudocount = 0.1)\n    scheduler = juice.optim.CircuitScheduler(\n        optimizer, \n        method = \"multi_linear\", \n        lrs = [0.9, 0.1, 0.05], \n        milestone_steps = [0, len(train_loader) * 100, len(train_loader) * 350]\n    )\n\n    pc.print_statistics()\n\n    mini_batch_em_epoch(350, pc, optimizer, None, train_loader, test_loader, device)\n    full_batch_em_epoch(pc, train_loader, test_loader, device)\n\n\nif __name__ == \"__main__\":\n    torch.manual_seed(2391)\n    pd_hclt_test()"
       ]
     }
   ],

_downloads/3dc6334404121f1adc55f4e2108f25f1/train_mnist_pd.ipynb

@@ -4,7 +4,7 @@
       "cell_type": "markdown",
       "metadata": {},
       "source": [
-        "\n# Example\n"
+        "\n# Example\n\nbbbb\n"
       ]
     },
     {
@@ -15,7 +15,7 @@
       },
       "outputs": [],
       "source": [
-        "import pyjuice as juice\nimport torch\nimport torchvision\nimport time\nfrom torch.utils.data import TensorDataset, DataLoader\n\n\ndef evaluate(pc, loader):\n    lls_total = 0.0\n    for batch in loader:\n        x = batch[0].to(pc.device)\n        lls = pc(x)\n        lls_total += lls.mean().detach().cpu().numpy().item()\n    \n    lls_total /= len(loader)\n    return lls_total\n\n\ndef mini_batch_em_epoch(num_epochs, pc, optimizer, scheduler, train_loader, test_loader, device):\n    for epoch in range(num_epochs):\n        t0 = time.time()\n        train_ll = 0.0\n        for batch in train_loader:\n            x = batch[0].to(device)\n\n            optimizer.zero_grad()\n\n            lls = pc(x)\n            lls.mean().backward()\n\n            train_ll += lls.mean().detach().cpu().numpy().item()\n\n            optimizer.step()\n            if scheduler is not None:\n                scheduler.step()\n\n        train_ll /= len(train_loader)\n\n        t1 = time.time()\n        test_ll = evaluate(pc, loader=test_loader)\n        t2 = time.time()\n\n        print(f\"[Epoch {epoch}/{num_epochs}][train LL: {train_ll:.2f}; test LL: {test_ll:.2f}].....[train forward+backward+step {t1-t0:.2f}; test forward {t2-t1:.2f}] \")\n\n\ndef full_batch_em_epoch(pc, train_loader, test_loader, device):\n    with torch.no_grad():\n        t0 = time.time()\n        train_ll = 0.0\n        for batch in train_loader:\n            x = batch[0].to(device)\n\n            lls = pc(x)\n            pc.backward(x, flows_memory = 1.0)\n\n            train_ll += lls.mean().detach().cpu().numpy().item()\n\n        pc.mini_batch_em(step_size = 1.0, pseudocount = 0.1)\n\n        train_ll /= len(train_loader)\n\n        t1 = time.time()\n        test_ll = evaluate(pc, loader=test_loader)\n        t2 = time.time()\n        print(f\"[train LL: {train_ll:.2f}; test LL: {test_ll:.2f}].....[train forward+backward+step {t1-t0:.2f}; test forward {t2-t1:.2f}] \")\n\n\ndef train_mnist_pd():\n\n    device = torch.device(\"cuda:0\")\n\n    train_dataset = torchvision.datasets.MNIST(root = \"./examples/data\", train = True, download = True)\n    test_dataset = torchvision.datasets.MNIST(root = \"./examples/data\", train = False, download = True)\n\n    train_data = train_dataset.data.reshape(60000, 28*28)\n    test_data = test_dataset.data.reshape(10000, 28*28)\n\n    num_features = train_data.size(1)\n\n    train_loader = DataLoader(\n        dataset = TensorDataset(train_data),\n        batch_size = 512,\n        shuffle = True,\n        drop_last = True\n    )\n    test_loader = DataLoader(\n        dataset = TensorDataset(test_data),\n        batch_size = 512,\n        shuffle = False,\n        drop_last = True\n    )\n\n    ns = juice.structures.PD(\n        data_shape = (28, 28),\n        num_latents = 512,\n        split_intervals = (4, 4),\n        structure_type = \"sum_dominated\"\n    )\n    pc = juice.TensorCircuit(ns)\n    pc.print_statistics()\n\n    pc.to(device)\n\n    optimizer = juice.optim.CircuitOptimizer(pc, lr = 0.1, pseudocount = 0.0001)\n\n    mini_batch_em_epoch(350, pc, optimizer, None, train_loader, test_loader, device)\n    full_batch_em_epoch(pc, train_loader, test_loader, device)\n\n\nif __name__ == \"__main__\":\n    train_mnist_pd()"
+        "# sphinx_gallery_thumbnail_path = 'imgs/juice.png'\n\nimport pyjuice as juice\nimport torch\nimport torchvision\nimport time\nfrom torch.utils.data import TensorDataset, DataLoader\n\n\ndef evaluate(pc, loader):\n    lls_total = 0.0\n    for batch in loader:\n        x = batch[0].to(pc.device)\n        lls = pc(x)\n        lls_total += lls.mean().detach().cpu().numpy().item()\n    \n    lls_total /= len(loader)\n    return lls_total\n\n\ndef mini_batch_em_epoch(num_epochs, pc, optimizer, scheduler, train_loader, test_loader, device):\n    for epoch in range(num_epochs):\n        t0 = time.time()\n        train_ll = 0.0\n        for batch in train_loader:\n            x = batch[0].to(device)\n\n            optimizer.zero_grad()\n\n            lls = pc(x)\n            lls.mean().backward()\n\n            train_ll += lls.mean().detach().cpu().numpy().item()\n\n            optimizer.step()\n            if scheduler is not None:\n                scheduler.step()\n\n        train_ll /= len(train_loader)\n\n        t1 = time.time()\n        test_ll = evaluate(pc, loader=test_loader)\n        t2 = time.time()\n\n        print(f\"[Epoch {epoch}/{num_epochs}][train LL: {train_ll:.2f}; test LL: {test_ll:.2f}].....[train forward+backward+step {t1-t0:.2f}; test forward {t2-t1:.2f}] \")\n\n\ndef full_batch_em_epoch(pc, train_loader, test_loader, device):\n    with torch.no_grad():\n        t0 = time.time()\n        train_ll = 0.0\n        for batch in train_loader:\n            x = batch[0].to(device)\n\n            lls = pc(x)\n            pc.backward(x, flows_memory = 1.0)\n\n            train_ll += lls.mean().detach().cpu().numpy().item()\n\n        pc.mini_batch_em(step_size = 1.0, pseudocount = 0.1)\n\n        train_ll /= len(train_loader)\n\n        t1 = time.time()\n        test_ll = evaluate(pc, loader=test_loader)\n        t2 = time.time()\n        print(f\"[train LL: {train_ll:.2f}; test LL: {test_ll:.2f}].....[train forward+backward+step {t1-t0:.2f}; test forward {t2-t1:.2f}] \")\n\n\ndef train_mnist_pd():\n\n    device = torch.device(\"cuda:0\")\n\n    train_dataset = torchvision.datasets.MNIST(root = \"./examples/data\", train = True, download = True)\n    test_dataset = torchvision.datasets.MNIST(root = \"./examples/data\", train = False, download = True)\n\n    train_data = train_dataset.data.reshape(60000, 28*28)\n    test_data = test_dataset.data.reshape(10000, 28*28)\n\n    num_features = train_data.size(1)\n\n    train_loader = DataLoader(\n        dataset = TensorDataset(train_data),\n        batch_size = 512,\n        shuffle = True,\n        drop_last = True\n    )\n    test_loader = DataLoader(\n        dataset = TensorDataset(test_data),\n        batch_size = 512,\n        shuffle = False,\n        drop_last = True\n    )\n\n    ns = juice.structures.PD(\n        data_shape = (28, 28),\n        num_latents = 512,\n        split_intervals = (4, 4),\n        structure_type = \"sum_dominated\"\n    )\n    pc = juice.TensorCircuit(ns)\n    pc.print_statistics()\n\n    pc.to(device)\n\n    optimizer = juice.optim.CircuitOptimizer(pc, lr = 0.1, pseudocount = 0.0001)\n\n    mini_batch_em_epoch(350, pc, optimizer, None, train_loader, test_loader, device)\n    full_batch_em_epoch(pc, train_loader, test_loader, device)\n\n\nif __name__ == \"__main__\":\n    train_mnist_pd()"
       ]
     }
   ],