clean up

README.md

@@ -46,7 +46,6 @@ To set up the environment locally and run the code:
 	```bash
 	jupyter notebook
 	```
-	Open either `VVN.ipynb` or `kMVN.ipynb`.
 
 ---
 

cif_MVN.ipynb

@@ -264,7 +264,7 @@
       "cell_type": "markdown",
       "metadata": {},
       "source": [
-        "### load pre-trained model\n",
+        "### Load pre-trained model\n",
         "Load the pre-trained model from the archive. Our code automatically load the hyperparameters and the learned weights from the .torch file. "
       ]
     },

cif_VVN.ipynb

@@ -266,7 +266,7 @@
       "cell_type": "markdown",
       "metadata": {},
       "source": [
-        "### load pre-trained model\n",
+        "### Load pre-trained model\n",
         "Load the pre-trained model from the archive. Our code automatically load the hyperparameters and the learned weights from the .torch file. "
       ]
     },

cif_kMVN.ipynb

@@ -91,7 +91,7 @@
             "  warnings.warn(\n",
             "/home/rokabe/anaconda3/envs/phonon/lib/python3.9/site-packages/seekpath/hpkot/__init__.py:362: EdgeCaseWarning: aP lattice, but the k_gamma3 angle is almost equal to 90 degrees\n",
             "  warnings.warn(\n",
-            "100%|██████████| 5/5 [00:00<00:00, 56.20it/s]\n"
+            "100%|██████████| 5/5 [00:00<00:00, 55.82it/s]\n"
           ]
         }
       ],
@@ -264,7 +264,7 @@
       "cell_type": "markdown",
       "metadata": {},
       "source": [
-        "### load pre-trained model\n",
+        "### Load pre-trained model\n",
         "Load the pre-trained model from the archive. Our code automatically load the hyperparameters and the learned weights from the .torch file. "
       ]
     },
@@ -328,7 +328,7 @@
           "text": [
             "  0%|          | 0/5 [00:00<?, ?it/s]/data1/rokabe/phonon/phonon_prediction/utils/utils_plot.py:209: FutureWarning: The behavior of DataFrame concatenation with empty or all-NA entries is deprecated. In a future version, this will no longer exclude empty or all-NA columns when determining the result dtypes. To retain the old behavior, exclude the relevant entries before the concat operation.\n",
             "  df = pd.concat([df, df0], ignore_index=True)\n",
-            "100%|██████████| 5/5 [00:04<00:00,  1.07it/s]\n"
+            "100%|██████████| 5/5 [00:05<00:00,  1.00s/it]\n"
           ]
         }
       ],
@@ -339,7 +339,7 @@
     },
     {
       "cell_type": "code",
-      "execution_count": 10,
+      "execution_count": 8,
       "metadata": {},
       "outputs": [
         {
@@ -368,7 +368,7 @@
               "Text(0.5, 1.0, '[mp-149] Si₈: phonon band structure')"
             ]
           },
-          "execution_count": 10,
+          "execution_count": 8,
           "metadata": {},
           "output_type": "execute_result"
         },

tutorial_MVN_previous.ipynb → previous_codes/tutorial_MVN_previous.ipynb

@@ -219,7 +219,7 @@
       },
       "source": [
         "## Data Preparation\n",
-        "### Data provenance\n",
+        "### Download dataset\n",
         "We train our model using the database of Density Functional Perturbation Theory (DFPT)-calculated phonon bands, containing approximately 1,500 crystalline solids [[Petretto et al. 2018]](https://doi.org/10.1038/sdata.2018.65)."
       ]
     },
@@ -478,8 +478,8 @@
         "id": "LZtSOwB9Cols"
       },
       "source": [
-        "### Network architecture\n",
-        "We build a model based on the `Network` described in the `e3nn` [Documentation](https://docs.e3nn.org/en/latest/api/nn/models/gate_points_2101.html), modified to incorporate the periodic boundaries we imposed on the crystal graphs. The network applies equivariant convolutions to each atomic node and finally takes an average over all nodes, normalizing the output."
+        "## Network architecture\n",
+        "We build a model based on the `Network` described in the `e3nn` [Documentation](https://docs.e3nn.org/en/latest/api/nn/models/gate_points_2101.html), modified to incorporate our Virtual node Graph Neural Network (VGNN) to deal with variable output dimension. "
       ]
     },
     {
@@ -580,7 +580,7 @@
         "id": "-5cuaOFOC_D_"
       },
       "source": [
-        "### Training\n",
+        "## Train the model\n",
         "The model is trained using a mean-squared error loss function with an Adam optimizer.\n",
         "Plot the prediction results (train/test data) after 'max_iter' epochs. \n",
         "Here we demonstrate training only with zero (or one) epoch. If you just want to use pre-trained model for phonon prediction you can skip this tab. "
@@ -608,8 +608,8 @@
         "        batch_size,\n",
         "        k_fold,\n",
         "        option=option,\n",
-        "        factor=factor,    #!\n",
-        "        conf_dict=None)  #!"
+        "        factor=factor,    \n",
+        "        conf_dict=None)  "
       ]
     },
     {
@@ -651,7 +651,7 @@
         "id": "5kpYlnpwEFW3"
       },
       "source": [
-        "### Results\n",
+        "## Results\n",
         "We evaluate our model by visualizing the predicted and true phonon band in each error tertile. "
       ]
     },

tutorial_VVN_previous.ipynb → previous_codes/tutorial_VVN_previous.ipynb

@@ -95,7 +95,7 @@
         "print('training ratio: ', tr_ratio)\n",
         "print('batch size: ', batch_size)\n",
         "print('irreduceble output representation: ', irreps_out)\n",
-        "print('factor to multiply/divide label data: ', factor) #!"
+        "print('factor to multiply/divide label data: ', factor) "
       ]
     },
     {
@@ -221,7 +221,7 @@
       },
       "source": [
         "## Data Preparation\n",
-        "### Data provenance\n",
+        "### Download dataset\n",
         "We train our model using the database of Density Functional Perturbation Theory (DFPT)-calculated phonon bands, containing approximately 1,500 crystalline solids [[Petretto et al. 2018]](https://doi.org/10.1038/sdata.2018.65)."
       ]
     },
@@ -489,8 +489,8 @@
         "id": "LZtSOwB9Cols"
       },
       "source": [
-        "### Network architecture\n",
-        "We build a model based on the `Network` described in the `e3nn` [Documentation](https://docs.e3nn.org/en/latest/api/nn/models/gate_points_2101.html), modified to incorporate the periodic boundaries we imposed on the crystal graphs. The network applies equivariant convolutions to each atomic node and finally takes an average over all nodes, normalizing the output."
+        "## Network architecture\n",
+        "We build a model based on the `Network` described in the `e3nn` [Documentation](https://docs.e3nn.org/en/latest/api/nn/models/gate_points_2101.html), modified to incorporate our Virtual node Graph Neural Network (VGNN) to deal with variable output dimension. "
       ]
     },
     {
@@ -591,7 +591,7 @@
         "id": "-5cuaOFOC_D_"
       },
       "source": [
-        "### Training\n",
+        "## Train the model\n",
         "The model is trained using a mean-squared error loss function with an Adam optimizer.\n",
         "Plot the prediction results (train/test data) after 'max_iter' epochs. \n",
         "Here we demonstrate training only with zero (or one) epoch. If you just want to use pre-trained model for phonon prediction you can skip this tab. "
@@ -619,8 +619,8 @@
         "            batch_size,\n",
         "            k_fold,\n",
         "            option=option,\n",
-        "            factor=factor,    #!\n",
-        "            conf_dict=None)  #!"
+        "            factor=factor,    \n",
+        "            conf_dict=None)  "
       ]
     },
     {
@@ -662,7 +662,7 @@
         "id": "5kpYlnpwEFW3"
       },
       "source": [
-        "### Results\n",
+        "## Results\n",
         "We evaluate our model by visualizing the predicted and true phonon band in each error tertile. "
       ]
     },
@@ -704,8 +704,8 @@
       ],
       "source": [
         "# Generate Data Frame\n",
-        "df_tr = generate_dataframe(model, tr_loader, loss_fn, device, option, factor)   #!\n",
-        "df_te = generate_dataframe(model, te_loader, loss_fn, device, option, factor)  #!"
+        "df_tr = generate_dataframe(model, tr_loader, loss_fn, device, option, factor)   \n",
+        "df_te = generate_dataframe(model, te_loader, loss_fn, device, option, factor)  "
       ]
     },
     {

tutorial_kMVN_previous.ipynb → previous_codes/tutorial_kMVN_previous.ipynb

@@ -219,7 +219,7 @@
       },
       "source": [
         "## Data Preparation\n",
-        "### Data provenance\n",
+        "### Download dataset\n",
         "We train our model using the database of Density Functional Perturbation Theory (DFPT)-calculated phonon bands, containing approximately 1,500 crystalline solids [[Petretto et al. 2018]](https://doi.org/10.1038/sdata.2018.65)."
       ]
     },
@@ -485,8 +485,8 @@
         "id": "LZtSOwB9Cols"
       },
       "source": [
-        "### Network architecture\n",
-        "We build a model based on the `Network` described in the `e3nn` [Documentation](https://docs.e3nn.org/en/latest/api/nn/models/gate_points_2101.html), modified to incorporate the periodic boundaries we imposed on the crystal graphs. The network applies equivariant convolutions to each atomic node and finally takes an average over all nodes, normalizing the output."
+        "## Network architecture\n",
+        "We build a model based on the `Network` described in the `e3nn` [Documentation](https://docs.e3nn.org/en/latest/api/nn/models/gate_points_2101.html), modified to incorporate our Virtual node Graph Neural Network (VGNN) to deal with variable output dimension. "
       ]
     },
     {
@@ -587,7 +587,7 @@
         "id": "-5cuaOFOC_D_"
       },
       "source": [
-        "### Training\n",
+        "## Train the model\n",
         "The model is trained using a mean-squared error loss function with an Adam optimizer.\n",
         "Plot the prediction results (train/test data) after 'max_iter' epochs. \n",
         "Here we demonstrate training only with zero (or one) epoch. If you just want to use pre-trained model for phonon prediction you can skip this tab. "
@@ -641,7 +641,7 @@
       ],
       "source": [
         "model_name = pretrained_name if use_pretrained else run_name  # model name (load from pretrained model if exists)\n",
-        "save_name = f'./models/{model_name}'    #!\n",
+        "save_name = f'./models/{model_name}'    \n",
         "model_file = save_name + '.torch'\n",
         "model.load_state_dict(torch.load(model_file)['state'])\n",
         "model = model.to(device)\n",
@@ -656,7 +656,7 @@
         "id": "5kpYlnpwEFW3"
       },
       "source": [
-        "### Results\n",
+        "## Results\n",
         "We evaluate our model by visualizing the predicted and true phonon band in each error tertile. "
       ]
     },

tutorial_MVN.ipynb

@@ -219,7 +219,7 @@
       },
       "source": [
         "## Data Preparation\n",
-        "### Data provenance\n",
+        "### Download dataset\n",
         "We train our model using the database of Density Functional Perturbation Theory (DFPT)-calculated phonon bands, containing approximately 1,500 crystalline solids [[Petretto et al. 2018]](https://doi.org/10.1038/sdata.2018.65)."
       ]
     },
@@ -478,8 +478,8 @@
         "id": "LZtSOwB9Cols"
       },
       "source": [
-        "### Network architecture\n",
-        "We build a model based on the `Network` described in the `e3nn` [Documentation](https://docs.e3nn.org/en/latest/api/nn/models/gate_points_2101.html), modified to incorporate the periodic boundaries we imposed on the crystal graphs. The network applies equivariant convolutions to each atomic node and finally takes an average over all nodes, normalizing the output."
+        "## Network architecture\n",
+        "We build a model based on the `Network` described in the `e3nn` [Documentation](https://docs.e3nn.org/en/latest/api/nn/models/gate_points_2101.html), modified to incorporate our Virtual node Graph Neural Network (VGNN) to deal with variable output dimension. "
       ]
     },
     {
@@ -580,7 +580,7 @@
         "id": "-5cuaOFOC_D_"
       },
       "source": [
-        "### Training\n",
+        "## Train the model\n",
         "The model is trained using a mean-squared error loss function with an Adam optimizer.\n",
         "Plot the prediction results (train/test data) after 'max_iter' epochs. \n",
         "Here we demonstrate training only with zero (or one) epoch. If you just want to use pre-trained model for phonon prediction you can skip this tab. "
@@ -608,8 +608,8 @@
         "        batch_size,\n",
         "        k_fold,\n",
         "        option=option,\n",
-        "        factor=factor,    #!\n",
-        "        conf_dict=None)  #!"
+        "        factor=factor,    \n",
+        "        conf_dict=None)  "
       ]
     },
     {
@@ -678,8 +678,8 @@
         "id": "5kpYlnpwEFW3"
       },
       "source": [
-        "### Results\n",
-        "We evaluate our model by visualizing the predicted and true phonon band in each error tertile. "
+        "## Results\n",
+        "We evaluate our model by visualizing the predicted and true phonon in each error tertile. "
       ]
     },
     {

tutorial_VVN.ipynb

@@ -30,7 +30,7 @@
         "from utils.utils_load import load_band_structure_data   \n",
         "from utils.utils_data import generate_data_dict, get_lattice_parameters\n",
         "from utils.utils_model import BandLoss, GraphNetwork_VVN, train, load_model\n",
-        "from utils.utils_plot import plot_element_count_stack, generate_dataframe, plot_gphonons, plot_loss, compare_models #!\n",
+        "from utils.utils_plot import plot_element_count_stack, generate_dataframe, plot_gphonons, plot_loss, compare_models \n",
         "# from utils.helpers import make_dict\n",
         "from config_file import pretrained_model_dict\n",
         "torch.set_default_dtype(torch.float64)\n",
@@ -88,14 +88,14 @@
         "batch_size = 1\n",
         "k_fold = 5\n",
         "irreps_out = '1x0e'\n",
-        "factor = 1000   #!\n",
+        "factor = 1000   \n",
         "\n",
         "print('\\ndata parameters')\n",
         "print('method: ', k_fold, '-fold cross validation')\n",
         "print('training ratio: ', tr_ratio)\n",
         "print('batch size: ', batch_size)\n",
         "print('irreduceble output representation: ', irreps_out)\n",
-        "print('factor to multiply/divide label data: ', factor) #!"
+        "print('factor to multiply/divide label data: ', factor) "
       ]
     },
     {
@@ -221,7 +221,7 @@
       },
       "source": [
         "## Data Preparation\n",
-        "### Data provenance\n",
+        "### Download dataset\n",
         "We train our model using the database of Density Functional Perturbation Theory (DFPT)-calculated phonon bands, containing approximately 1,500 crystalline solids [[Petretto et al. 2018]](https://doi.org/10.1038/sdata.2018.65)."
       ]
     },
@@ -482,8 +482,8 @@
         "id": "LZtSOwB9Cols"
       },
       "source": [
-        "### Network architecture\n",
-        "We build a model based on the `Network` described in the `e3nn` [Documentation](https://docs.e3nn.org/en/latest/api/nn/models/gate_points_2101.html), modified to incorporate the periodic boundaries we imposed on the crystal graphs. The network applies equivariant convolutions to each atomic node and finally takes an average over all nodes, normalizing the output."
+        "## Network architecture\n",
+        "We build a model based on the `Network` described in the `e3nn` [Documentation](https://docs.e3nn.org/en/latest/api/nn/models/gate_points_2101.html), modified to incorporate our Virtual node Graph Neural Network (VGNN) to deal with variable output dimension. "
       ]
     },
     {
@@ -584,7 +584,7 @@
         "id": "-5cuaOFOC_D_"
       },
       "source": [
-        "### Training\n",
+        "## Train the model\n",
         "The model is trained using a mean-squared error loss function with an Adam optimizer.\n",
         "Plot the prediction results (train/test data) after 'max_iter' epochs. \n",
         "Here we demonstrate training only with zero (or one) epoch. If you just want to use pre-trained model for phonon prediction you can skip this tab. "
@@ -612,8 +612,8 @@
         "            batch_size,\n",
         "            k_fold,\n",
         "            option=option,\n",
-        "            factor=factor,    #!\n",
-        "            conf_dict=None)  #!"
+        "            factor=factor,    \n",
+        "            conf_dict=None)  "
       ]
     },
     {
@@ -642,7 +642,7 @@
       "source": [
         "\n",
         "model_name = pretrained_name if use_pretrained else run_name  # model name (load from pretrained model if exists)\n",
-        "save_name = f'./models/{model_name}'    #!\n",
+        "save_name = f'./models/{model_name}'    \n",
         "model_file = save_name + '.torch'\n",
         "model, conf_dict, history, s0 = load_model(GraphNetwork_VVN, model_file, device)\n",
         "num_params = sum(p.numel() for p in model.parameters() if p.requires_grad)\n",
@@ -653,7 +653,8 @@
       "cell_type": "markdown",
       "metadata": {},
       "source": [
-        "### Plot the loss function"
+        "### Plot the loss function\n",
+        "Plot the loss history (train, valid)."
       ]
     },
     {
@@ -684,7 +685,7 @@
       },
       "source": [
         "### Results\n",
-        "We evaluate our model by visualizing the predicted and true phonon band in each error tertile. "
+        "We evaluate our model by visualizing the predicted and true phonon in each error tertile. "
       ]
     },
     {
@@ -718,8 +719,8 @@
       ],
       "source": [
         "# Generate Data Frame\n",
-        "df_tr = generate_dataframe(model, tr_loader, loss_fn, device, option, factor)   #!\n",
-        "df_te = generate_dataframe(model, te_loader, loss_fn, device, option, factor)  #!"
+        "df_tr = generate_dataframe(model, tr_loader, loss_fn, device, option, factor)   \n",
+        "df_te = generate_dataframe(model, te_loader, loss_fn, device, option, factor)  "
       ]
     },
     {