Tian et al. 2011 Model 5 Conv. 2d Hilbert Curve

This model uses hilbert curves

# Hacky way to schedule. Here I'm setting these to sleep until the gpus should be free.
# At the end of the notebooks  os._exit(00) will kill the kernel freeing the gpu. 
#                          Hours to wait
# import time; time.sleep( 24 * (60*60))

# Run Settings:
nb_name = '15_TianEtAl2011'# Set manually! -----------------------------------

downsample_obs = False
train_n = 90
test_n = 10

dataloader_batch_size = 8 #16 #64
run_epochs = 200

use_gpu_num = 0

# Imports --------------------------------------------------------------------
import os
import pandas as pd
import numpy as np
import re

import torch
from torch.utils.data import Dataset
from torch.utils.data import DataLoader
from torch import nn

import tqdm
from tqdm import tqdm

import plotly.graph_objects as go
import plotly.express as px
import plotly.io as pio
pio.templates.default = "plotly_white"


import dlgwas
from dlgwas.kegg import ensure_dir_path_exists
from dlgwas.kegg import get_cached_result
from dlgwas.kegg import put_cached_result

from dlgwas.dlfn import calc_cs
from dlgwas.dlfn import apply_cs
from dlgwas.dlfn import reverse_cs

from dlgwas.dlfn import TianEtAl2011Dataset
from dlgwas.dlfn import train_loop
from dlgwas.dlfn import train_error
from dlgwas.dlfn import test_loop
from dlgwas.dlfn import train_nn
from dlgwas.dlfn import yhat_loop


device = "cuda" if torch.cuda.is_available() else "cpu"
if use_gpu_num in [0, 1]: 
    torch.cuda.set_device(use_gpu_num)
print(f"Using {device} device")


ensure_dir_path_exists(dir_path = '../models/'+nb_name)
ensure_dir_path_exists(dir_path = '../reports/'+nb_name)

ensure_dir_path_exists(dir_path = '../models/'+nb_name)
ensure_dir_path_exists(dir_path = '../reports/'+nb_name)

ModuleNotFoundError: No module named 'dlgwas'

Load Cleaned Data

# Read in cleaned data
taxa_groupings = pd.read_csv('../models/10_TianEtAl2011/taxa_groupings.csv')
data           = pd.read_csv('../models/10_TianEtAl2011/clean_data.csv')

# Define holdout sets (Populations)
uniq_pop = list(set(taxa_groupings['Population']))
print(str(len(uniq_pop))+" Unique Holdout Groups.")
taxa_groupings['Holdout'] = None
for i in range(len(uniq_pop)):
    mask = (taxa_groupings['Population'] == uniq_pop[i])
    taxa_groupings.loc[mask, 'Holdout'] = i

taxa_groupings

Setup Holdouts

#randomly holdout a population if there is not a file with the population held out.
# Holdout_Int = 0
Holdout_Int_path = '../models/'+nb_name+'/holdout_pop_int.pkl'
if None != get_cached_result(Holdout_Int_path):
    Holdout_Int = get_cached_result(Holdout_Int_path)
else:
    Holdout_Int = int(np.random.choice([i for i in range(len(uniq_pop))], 1))
    put_cached_result(Holdout_Int_path, Holdout_Int)

    
print("Holding out i="+str(Holdout_Int)+": "+uniq_pop[Holdout_Int])

mask = (taxa_groupings['Holdout'] == Holdout_Int)
train_idxs = list(taxa_groupings.loc[~mask, ].index)
test_idxs = list(taxa_groupings.loc[mask, ].index)

# downsample_obs = True
# train_n = 900
# test_n = 100

if downsample_obs == True:
    train_idxs = np.random.choice(train_idxs, train_n)
    test_idxs = np.random.choice(test_idxs, test_n)
    print([len(e) for e in [test_idxs, train_idxs]])
    
# used to go from index in tensor to index in data so that the right xs tensor can be loaded in
idx_original = np.array(data.index)

y1 = data['leaf_length']
y2 = data['leaf_width']
y3 = data['upper_leaf_angle']
y1 = np.array(y1)
y2 = np.array(y2)
y3 = np.array(y3)

Scale data

scale_dict_path = '../models/'+nb_name+'/scale_dict.pkl'
if None != get_cached_result(scale_dict_path):
    scale_dict = get_cached_result(scale_dict_path)
else:
    scale_dict = {
        'y1':calc_cs(y1[train_idxs]),
        'y2':calc_cs(y2[train_idxs]),
        'y3':calc_cs(y3[train_idxs])
    }
    put_cached_result(scale_dict_path, scale_dict)

y1 = apply_cs(y1, scale_dict['y1'])
y2 = apply_cs(y2, scale_dict['y2'])
y3 = apply_cs(y3, scale_dict['y3'])

Allow for cycling data onto and off of GPU

# loading this into memory causes the session to crash

y1_train = torch.from_numpy(y1[train_idxs])[:, None]
y2_train = torch.from_numpy(y2[train_idxs])[:, None]
y3_train = torch.from_numpy(y3[train_idxs])[:, None]

idx_original_train = torch.from_numpy(idx_original[train_idxs])

y1_test = torch.from_numpy(y1[test_idxs])[:, None]
y2_test = torch.from_numpy(y2[test_idxs])[:, None]
y3_test = torch.from_numpy(y3[test_idxs])[:, None]

idx_original_test = torch.from_numpy(idx_original[test_idxs])


# dataloader_batch_size = 64

training_dataloader = DataLoader(
    TianEtAl2011Dataset(
        y1 = y1_train,
        y2 = y2_train,
        y3 = y3_train,
        marker_type = 'hilbert',
        idx_original = idx_original_train,
        use_gpu_num = use_gpu_num,
#         device = 'cpu'
    ), 
    batch_size = dataloader_batch_size, 
    shuffle = True)

testing_dataloader = DataLoader(
    TianEtAl2011Dataset(
        y1 = y1_test,
        y2 = y2_test,
        y3 = y3_test,
        marker_type = 'hilbert',
        idx_original = idx_original_test,
        use_gpu_num = use_gpu_num,
#         device = 'cpu'
    ), 
    batch_size = dataloader_batch_size, 
    shuffle = True)

Non-Boilerplate

class NeuralNetwork(nn.Module):
    def __init__(self):
        super(NeuralNetwork, self).__init__()    

        # Block 1 ------------------------------------------------------------
        self.long_way_0 = nn.Sequential(
            nn.Conv2d(
                    in_channels= 4, # second channel
                    out_channels= 4,
                    kernel_size= (3, 3),
                    stride= 2,
                    bias = True
                ),
#             nn.BatchNorm1d(4),
            nn.Conv2d(
                    in_channels= 4, 
                    out_channels= 4,
                    kernel_size= (3, 3),
                    stride= 1,
                    padding = 1,
                    bias = True
                ),
#             nn.BatchNorm1d(4),
            nn.Dropout(p=0.75)
            )
        
        self.shortcut_0 = nn.Sequential(
            nn.Conv2d(
                    in_channels= 4, 
                    out_channels= 4,
                    kernel_size= (3, 3),
                    stride= 2,
                    bias = True
                )
        )
        # Block 2 ------------------------------------------------------------
        self.long_way_1 = nn.Sequential(
            nn.Conv2d(
                    in_channels= 4, # second channel
                    out_channels= 4,
                    kernel_size= (3, 3),
                    stride= 2,
                    bias = True
                ),
#             nn.BatchNorm1d(4),
            nn.Conv2d(
                    in_channels= 4, 
                    out_channels= 4,
                    kernel_size= (3, 3),
                    stride= 1,
                    padding = 1,
                    bias = True
                ),
#             nn.BatchNorm1d(4),
            nn.Dropout(p=0.75)
            )
        
        self.shortcut_1 = nn.Sequential(
            nn.Conv2d(
                    in_channels= 4, 
                    out_channels= 4,
                    kernel_size= (3, 3),
                    stride= 2,
                    bias = True
                )
        )
        # Block 3 ------------------------------------------------------------
        self.long_way_2 = nn.Sequential(
            nn.Conv2d(
                    in_channels= 4, # second channel
                    out_channels= 4,
                    kernel_size= (3, 3),
                    stride= 2,
                    bias = True
                ),
#             nn.BatchNorm1d(4),
            nn.Conv2d(
                    in_channels= 4, 
                    out_channels= 4,
                    kernel_size= (3, 3),
                    stride= 1,
                    padding = 1,
                    bias = True
                ),
#             nn.BatchNorm1d(4),
            nn.Dropout(p=0.75)
            )
        
        self.shortcut_2 = nn.Sequential(
            nn.Conv2d(
                    in_channels= 4, 
                    out_channels= 4,
                    kernel_size= (3, 3),
                    stride= 2,
                    bias = True
                )
        )        
        
        
        self.feature_processing = nn.Sequential(
            nn.Conv2d(
                    in_channels= 4, 
                    out_channels= 4,
                    kernel_size= (3, 3),
                    stride = 2,
                    bias = True
                ),
            nn.Dropout(p=0.75)
        )

        self.output_processing = nn.Sequential(
            nn.Flatten(),
            nn.ReLU(), # They used inverse square root activation $y = \frac{x}{\sqrt{1+ax^2}}$
            nn.Dropout(p=0.75),
            nn.Linear(15876, 1)
        )            
        
    def forward(self, x):
        x_out = self.long_way_0(x)
        x_shortcut = self.shortcut_0(x)
        x_out += x_shortcut
        
        x = x_out
        x_out = self.long_way_1(x)
        x_shortcut = self.shortcut_1(x)
        x_out += x_shortcut
        
        x = x_out
        x_out = self.long_way_2(x)
        x_shortcut = self.shortcut_2(x)        
        x_out += x_shortcut
        
        x_out = self.feature_processing(x_out)
        x_out = self.output_processing(x_out)
        
        return x_out

# xs_i, y1_i, y2_i, y3_i = next(iter(training_dataloader))

# model = NeuralNetwork().to(device)

# res = model(xs_i) # try prediction on one batch
# res.shape

def train_nn(
    nb_name,
    training_dataloader,
    testing_dataloader,
    model,
    learning_rate = 1e-3,
    batch_size = 64,
    epochs = 500
):
    # Initialize the loss function
    loss_fn = nn.MSELoss()
#     optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)

    # Optimizer with L2 normalization
    optimizer = torch.optim.Adam([
        {'params':model.long_way_0.parameters(), 'weight_decay': 0.1},
        {'params':model.shortcut_0.parameters(), 'weight_decay': 0.1},
        {'params':model.long_way_1.parameters(), 'weight_decay': 0.1},
        {'params':model.shortcut_1.parameters(), 'weight_decay': 0.1},
        {'params':model.long_way_2.parameters(), 'weight_decay': 0.1},
        {'params':model.shortcut_2.parameters(), 'weight_decay': 0.1},
        {'params':model.feature_processing.parameters(),        'weight_decay': 0.1},
        {'params':model.output_processing.parameters(),         'weight_decay': 0.01},
    ], lr=learning_rate)

    loss_df = pd.DataFrame([i for i in range(epochs)], columns = ['Epoch'])
    loss_df['TrainMSE'] = np.nan
    loss_df['TestMSE']  = np.nan

    for t in tqdm(range(epochs)):        
#         print(f"Epoch {t+1}\n-------------------------------")
        train_loop(training_dataloader, model, loss_fn, optimizer, silent = True)

        loss_df.loc[loss_df.index == t, 'TrainMSE'
                   ] = train_error(training_dataloader, model, loss_fn, silent = True)
        
        loss_df.loc[loss_df.index == t, 'TestMSE'
                   ] = test_loop(testing_dataloader, model, loss_fn, silent = True)
        
        if (t+1)%10: # Cache in case training is interupted
#             print(loss_df.loc[loss_df.index == t, ['TrainMSE', 'TestMSE']])
            torch.save(model.state_dict(), 
                       '../models/'+nb_name+'/model_'+str(t)+'_'+str(epochs)+'.pt') # convention is to use .pt or .pth
            loss_df.to_csv('../reports/'+nb_name+'/loss_df'+str(t)+'_'+str(epochs)+'.csv', index=False) 
        
    return([model, loss_df])

# don't run if either of these exist because there may be cases where we want the results but not the model

if not os.path.exists('../models/'+nb_name+'/model.pt'): 
    # Shared setup (train from scratch and load latest)
    model = NeuralNetwork()

    # find the biggest model to save
    saved_models = os.listdir('../models/'+nb_name+'/')
    saved_models = [e for e in saved_models if re.match('model*', e)]

    if saved_models == []:
        epochs_run = 0
    else:
        # if there are saved models reload and resume training
        saved_models_numbers = [int(e.replace('model_', ''
                                    ).replace('.pt', ''
                                    ).split('_')[0]) for e in saved_models]
        # saved_models
        epochs_run = max(saved_models_numbers)+1 # add 1 to account for 0 index
        latest_model = [e for e in saved_models if re.match(
            '^model_'+str(epochs_run-1)+'_.*\.pt$', e)][0] # subtract 1 to convert back
        model.load_state_dict(torch.load('../models/'+nb_name+'/'+latest_model))
        print('Resuming Training: '+str(epochs_run)+'/'+str(run_epochs)+' epochs run.')
    
    model.to(device)   

    model, loss_df = train_nn(
        nb_name,
        training_dataloader,
        testing_dataloader,
        model,
        learning_rate = 1e-3,
        batch_size = dataloader_batch_size,
        epochs = (run_epochs - epochs_run)
    )
    
    # experimental outputs:
    # 1. Model
    torch.save(model.state_dict(), '../models/'+nb_name+'/model.pt') # convention is to use .pt or .pth

    # 2. loss_df
    loss_df.to_csv('../reports/'+nb_name+'/loss_df.csv', index=False)  
    
    
    # 3. predictions 
    yhats = pd.concat([
        yhat_loop(testing_dataloader, model).assign(Split = 'Test'),
        yhat_loop(training_dataloader, model).assign(Split = 'Train')], axis = 0)

    yhats.to_csv('../reports/'+nb_name+'/yhats.csv', index=False)

NeuralNetwork()

Standard Visualizations

loss_df = pd.read_csv('../reports/'+nb_name+'/loss_df.csv')

loss_df.TrainMSE = reverse_cs(loss_df.TrainMSE, scale_dict['y1'])
loss_df.TestMSE  = reverse_cs(loss_df.TestMSE , scale_dict['y1'])


fig = go.Figure()
fig.add_trace(go.Scatter(x=loss_df.Epoch, y=loss_df.TestMSE,
                    mode='lines', name='Test'))
fig.add_trace(go.Scatter(x=loss_df.Epoch, y=loss_df.TrainMSE,
                    mode='lines', name='Train'))
fig.show()

yhats = pd.read_csv('../reports/'+nb_name+'/yhats.csv')

yhats.y_true = reverse_cs(yhats.y_true, scale_dict['y1'])
yhats.y_pred = reverse_cs(yhats.y_pred, scale_dict['y1'])

px.scatter(yhats, x = 'y_true', y = 'y_pred', color = 'Split', trendline="ols")

yhats['Error'] = yhats.y_true - yhats.y_pred

px.histogram(yhats, x = 'Error', color = 'Split',
             marginal="box", # can be `rug`, `violin`
             nbins= 50)

# automatically kill kernel after running. 
# This is a hacky way to free up _all_ space on the gpus
# os._exit(00)