ium_426206/train_mlflow.py

import torch
import numpy as np

import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset, TensorDataset, DataLoader
import argparse
import mlflow
import mlflow.pytorch
from urllib.parse import urlparse
from mlflow.models.signature import infer_signature

class LayerLinearRegression(nn.Module):
    def __init__(self):
        super().__init__()
        # Instead of our custom parameters, we use a Linear layer with single input and single output
        self.linear = nn.Linear(1, 1)
                
    def forward(self, x):
        # Now it only takes a call to the layer to make predictions
        return self.linear(x)

if __name__ == "__main__":

    parser = argparse.ArgumentParser(description='Program do uczenia modelu')
    parser.add_argument('-l', '--lr', type=float, default=1e-3, help="Współczynik uczenia (lr)", required=False)
    parser.add_argument('-e', '--epochs', type=int, default=100, help="Liczba epok", required=False)
    args = parser.parse_args()

    lr = args.lr
    n_epochs = args.epochs
    mlflow.set_experiment("s426206")
    with mlflow.start_run():
        mlflow.log_param("lr", lr)
        mlflow.log_param("epochs", n_epochs)

        train_dataset = torch.load('train_dataset.pt')
        #val_dataset = torch.load('val_dataset.pt')

        train_loader = DataLoader(dataset=train_dataset)
        #val_loader = DataLoader(dataset=val_dataset)
        
        model = LayerLinearRegression()
        # Checks model's parameters
        #print(model.state_dict())   

        loss_fn = nn.MSELoss(reduction='mean')
        optimizer = optim.SGD(model.parameters(), lr=lr)

        def make_train_step(model, loss_fn, optimizer):
            # Builds function that performs a step in the train loop
            def train_step(x, y):
                # Sets model to TRAIN mode
                model.train()
                # Makes predictions
                yhat = model(x)
                # Computes loss
                loss = loss_fn(y, yhat)
                # Computes gradients
                loss.backward()
                # Updates parameters and zeroes gradients
                optimizer.step()
                optimizer.zero_grad()
                # Returns the loss
                return loss.item()
            
            # Returns the function that will be called inside the train loop
            return train_step

        # Creates the train_step function for our model, loss function and optimizer
        train_step = make_train_step(model, loss_fn, optimizer)
        training_losses = []
        validation_losses = []
        #print(model.state_dict())   
        # For each epoch...
        for epoch in range(n_epochs):

            losses = []
            # Uses loader to fetch one mini-batch for training
            for x_batch, y_batch in train_loader:
                # NOW, sends the mini-batch data to the device
                # so it matches location of the MODEL
                # x_batch = x_batch.to(device)
                # y_batch = y_batch.to(device)
                # One stpe of training
                loss = train_step(x_batch, y_batch)
                losses.append(loss)
            training_loss = np.mean(losses)
            training_losses.append(training_loss)

            mlflow.log_metric("MSE", training_loss)
                
            # After finishing training steps for all mini-batches,
            # it is time for evaluation!
            # Ewaluacja jest już tutaj nie potrzebna bo odbywa sie w evaluation.py. Można jednak włączyć podgląd ewaluacji dla poszczególnych epok.    
            # # We tell PyTorch to NOT use autograd...
            # # Do you remember why?
            # with torch.no_grad():
            #     val_losses = []
            #     # Uses loader to fetch one mini-batch for validation
            #     for x_val, y_val in val_loader:
            #         # Again, sends data to same device as model
            #         # x_val = x_val.to(device)
            #         # y_val = y_val.to(device)
                    
            #         model.eval()
            #         # Makes predictions
            #         yhat = model(x_val)
            #         # Computes validation loss
            #         val_loss = loss_fn(y_val, yhat)
            #         val_losses.append(val_loss.item())
            #     validation_loss = np.mean(val_losses)
            #     validation_losses.append(validation_loss)

            # print(f"[{epoch+1}] Training loss: {training_loss:.3f}\t Validation loss: {validation_loss:.3f}")
            print(f"[{epoch+1}] Training loss: {training_loss:.3f}\t")

        torch.save({
                    'model_state_dict': model.state_dict(),
                    'optimizer_state_dict': optimizer.state_dict(),
                        'loss': lr,
                    }, 'model.pt')

        x_train = np.array(train_dataset)[:,0] #(Sales Sum row)
        input_example = np.reshape(x_train, (-1,1))

        with torch.no_grad():        
            model.eval()
            siganture = infer_signature(x_train, model(torch.tensor(np.reshape(x_train, (-1,1))).float()).numpy())

        #mlflow.set_experiment("s426206")
        #mlflow.set_tracking_uri("http://172.17.0.1:5000")
        tracking_url_type_store = urlparse(mlflow.get_tracking_uri()).scheme
        # print(tracking_url_type_store)
        # Model registry does not work with file store

        if tracking_url_type_store != "file":
            mlflow.pytorch.log_model(model, "model", registered_model_name="s426206", signature=siganture, input_example=input_example)
        else:
            mlflow.pytorch.log_model(model, "model", signature=siganture, input_example=input_example)
            mlflow.pytorch.save_model(model, "my_model", signature=siganture, input_example=input_example)

        #export MLFLOW_CONDA_HOME=/home/jan/miniconda3/
        #mlflow models serve -m my_model/