x0 cut prediction

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
plt.rcParams['figure.dpi'] = 120

Loading the data

doe = pd.read_csv('../data/doe.csv')

data = pd.read_csv('../data/cut_x0_all.csv')
data.drop(data[data.doe_id == 1000].index, inplace=True)
data.drop(data[data.doe_id == 247].index, inplace=True)

Creating the regressor

from mesh_predictor import CutPredictor

reg = CutPredictor()
reg.load_data(
    doe = doe,
    data = data,
    index='doe_id',
    process_parameters = [
        'Blechdicke', 
        'Niederhalterkraft', 
        'Ziehspalt', 
        'Einlegeposition', 
        'Ziehtiefe',
        'Rp0',
    ],
    categorical = [
        'Ziehspalt', 
        'Ziehtiefe',
    ],
    position = 'tp',
    output = ['deviationc'],#, 'y', 'z'],
    validation_split=0.1,
    validation_method='leaveoneout',
    position_scaler='minmax'
)
reg.save_config("../models/cut_x0_deviation.pkl")

reg.data_summary()

Data summary
------------------------------------------------------------

Process parameters:
    - Blechdicke : numerical [ 0.99  ...  1.48 ]
    - Niederhalterkraft : numerical [ 10  ...  500 ]
    - Ziehspalt : categorical [1.6, 2.4]
    - Einlegeposition : numerical [ -5  ...  5 ]
    - Ziehtiefe : categorical [30, 50, 70]
    - Rp0 : numerical [ 133.18263199999998  ...  296.5565 ]
Input variables:
    - tp : numerical, [ 0.0 / 5.0 ] 
Output variable(s):
    - deviationc : numerical, [ -3.16506211149574 / 7.228601057768613 ]

Inputs (879000, 10)
Outputs (879000, 1)
Total number of experiments: 879
Total number of samples: 879000
Number of training samples: 792000
Number of test samples: 87000
Number of experiments in the test set: 87

Training methods

Autotuning

best_config = reg.autotune(
    save_path='../models/best_x0_model',
    trials=100,
    max_epochs=20, 
    layers=[4, 6],
    neurons=[64, 256, 64],
    dropout=[0.0, 0.5, 0.1],
    learning_rate=[1e-5, 1e-3]
)
print(best_config)
reg.training_summary()

Alternative: define a custom network and do the optimization yourself

One can also run the autotuning for a limited number of epochs and then fine-tune the best configuration by training it longer.

config = {
    'batch_size': 2048,
    'max_epochs': 50,
    'layers': [256, 256, 256, 256, 256],
    'dropout': 0.0,
    'learning_rate': 0.01
}

# or best_config from autotune if you already did it once

reg.custom_model(save_path='../models/current_x0_model', config=config, verbose=True)
reg.training_summary()

Metal device set to: Apple M1 Pro

systemMemory: 16.00 GB
maxCacheSize: 5.33 GB

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense (Dense)               (None, 256)               2816      

 re_lu (ReLU)                (None, 256)               0         

 dense_1 (Dense)             (None, 256)               65792     

 re_lu_1 (ReLU)              (None, 256)               0         

 dense_2 (Dense)             (None, 256)               65792     

 re_lu_2 (ReLU)              (None, 256)               0         

 dense_3 (Dense)             (None, 256)               65792     

 re_lu_3 (ReLU)              (None, 256)               0         

 dense_4 (Dense)             (None, 256)               65792     

 re_lu_4 (ReLU)              (None, 256)               0         

 dense_5 (Dense)             (None, 1)                 257       

=================================================================
Total params: 266,241
Trainable params: 266,241
Non-trainable params: 0
_________________________________________________________________
Epoch 1/50
387/387 [==============================] - 7s 12ms/step - loss: 0.0951 - val_loss: 0.0015
Epoch 2/50
387/387 [==============================] - 5s 12ms/step - loss: 0.0010 - val_loss: 9.7784e-04
Epoch 3/50
387/387 [==============================] - 5s 12ms/step - loss: 7.8229e-04 - val_loss: 8.4430e-04
Epoch 4/50
387/387 [==============================] - 5s 12ms/step - loss: 6.6143e-04 - val_loss: 6.9625e-04
Epoch 5/50
387/387 [==============================] - 5s 12ms/step - loss: 5.6085e-04 - val_loss: 7.4612e-04
Epoch 6/50
387/387 [==============================] - 5s 12ms/step - loss: 4.8998e-04 - val_loss: 5.6353e-04
Epoch 7/50
387/387 [==============================] - 5s 12ms/step - loss: 4.8325e-04 - val_loss: 4.4705e-04
Epoch 8/50
387/387 [==============================] - 5s 12ms/step - loss: 3.9933e-04 - val_loss: 6.2635e-04
Epoch 9/50
387/387 [==============================] - 5s 12ms/step - loss: 3.6695e-04 - val_loss: 5.3525e-04
Epoch 10/50
387/387 [==============================] - 5s 12ms/step - loss: 3.6690e-04 - val_loss: 3.7967e-04
Epoch 11/50
387/387 [==============================] - 5s 13ms/step - loss: 3.4395e-04 - val_loss: 3.8449e-04
Epoch 12/50
387/387 [==============================] - 5s 13ms/step - loss: 3.2885e-04 - val_loss: 6.1913e-04
Epoch 13/50
387/387 [==============================] - 5s 13ms/step - loss: 2.7704e-04 - val_loss: 4.1812e-04
Epoch 14/50
387/387 [==============================] - 5s 12ms/step - loss: 2.7017e-04 - val_loss: 2.5869e-04
Epoch 15/50
387/387 [==============================] - 5s 13ms/step - loss: 2.4560e-04 - val_loss: 5.4358e-04
Epoch 16/50
387/387 [==============================] - 5s 13ms/step - loss: 2.7837e-04 - val_loss: 3.0109e-04
Epoch 17/50
387/387 [==============================] - 5s 13ms/step - loss: 2.4484e-04 - val_loss: 3.3601e-04
Epoch 18/50
387/387 [==============================] - 5s 13ms/step - loss: 2.0835e-04 - val_loss: 3.8868e-04
Epoch 19/50
387/387 [==============================] - 6s 15ms/step - loss: 2.4355e-04 - val_loss: 2.4871e-04
Epoch 20/50
387/387 [==============================] - 5s 13ms/step - loss: 1.9843e-04 - val_loss: 3.7203e-04
Epoch 21/50
387/387 [==============================] - 5s 12ms/step - loss: 2.1181e-04 - val_loss: 2.4572e-04
Epoch 22/50
387/387 [==============================] - 5s 12ms/step - loss: 1.8778e-04 - val_loss: 2.5494e-04
Epoch 23/50
387/387 [==============================] - 5s 13ms/step - loss: 1.8066e-04 - val_loss: 2.5111e-04
Epoch 24/50
387/387 [==============================] - 5s 12ms/step - loss: 1.8112e-04 - val_loss: 2.9211e-04
Epoch 25/50
387/387 [==============================] - 5s 12ms/step - loss: 2.0284e-04 - val_loss: 2.5421e-04
Epoch 26/50
387/387 [==============================] - 4s 12ms/step - loss: 1.6379e-04 - val_loss: 2.2906e-04
Epoch 27/50
387/387 [==============================] - 5s 12ms/step - loss: 1.7656e-04 - val_loss: 1.9107e-04
Epoch 28/50
387/387 [==============================] - 5s 12ms/step - loss: 1.6522e-04 - val_loss: 1.8863e-04
Epoch 29/50
387/387 [==============================] - 5s 13ms/step - loss: 1.5896e-04 - val_loss: 2.4476e-04
Epoch 30/50
387/387 [==============================] - 6s 15ms/step - loss: 1.6748e-04 - val_loss: 3.0141e-04
Epoch 31/50
387/387 [==============================] - 5s 13ms/step - loss: 1.6237e-04 - val_loss: 2.0276e-04
Epoch 32/50
387/387 [==============================] - 5s 12ms/step - loss: 1.7063e-04 - val_loss: 2.3029e-04
Epoch 33/50
387/387 [==============================] - 6s 15ms/step - loss: 1.4083e-04 - val_loss: 2.2445e-04
Epoch 34/50
387/387 [==============================] - 5s 12ms/step - loss: 1.3416e-04 - val_loss: 2.4996e-04
Epoch 35/50
387/387 [==============================] - 4s 12ms/step - loss: 1.6694e-04 - val_loss: 2.5844e-04
Epoch 36/50
387/387 [==============================] - 4s 11ms/step - loss: 1.4822e-04 - val_loss: 1.9492e-04
Epoch 37/50
387/387 [==============================] - 5s 12ms/step - loss: 1.2537e-04 - val_loss: 2.6402e-04
Epoch 38/50
387/387 [==============================] - 6s 15ms/step - loss: 1.4451e-04 - val_loss: 2.0392e-04
Epoch 39/50
387/387 [==============================] - 5s 13ms/step - loss: 1.3550e-04 - val_loss: 2.0369e-04
Epoch 40/50
387/387 [==============================] - 6s 14ms/step - loss: 1.3397e-04 - val_loss: 1.7202e-04
Epoch 41/50
387/387 [==============================] - 5s 13ms/step - loss: 1.2029e-04 - val_loss: 2.1723e-04
Epoch 42/50
387/387 [==============================] - 5s 13ms/step - loss: 1.2936e-04 - val_loss: 2.4467e-04
Epoch 43/50
387/387 [==============================] - 5s 12ms/step - loss: 1.3571e-04 - val_loss: 1.7959e-04
Epoch 44/50
387/387 [==============================] - 6s 15ms/step - loss: 1.0874e-04 - val_loss: 1.8706e-04
Epoch 45/50
387/387 [==============================] - 5s 12ms/step - loss: 1.1912e-04 - val_loss: 1.8208e-04
Epoch 46/50
387/387 [==============================] - 5s 12ms/step - loss: 1.2890e-04 - val_loss: 1.8400e-04
Epoch 47/50
387/387 [==============================] - 5s 13ms/step - loss: 1.2369e-04 - val_loss: 2.1889e-04
Epoch 48/50
387/387 [==============================] - 5s 13ms/step - loss: 1.0986e-04 - val_loss: 2.1243e-04
Epoch 49/50
387/387 [==============================] - 5s 13ms/step - loss: 1.1663e-04 - val_loss: 1.6970e-04
Epoch 50/50
387/387 [==============================] - 6s 16ms/step - loss: 1.0531e-04 - val_loss: 1.9543e-04
43/43 [==============================] - 0s 10ms/step - loss: 1.6970e-04
INFO:tensorflow:Assets written to: ../models/current_x0_model/assets
Validation mse: 0.00016970359138213098

Other alternative: the model has already been trained

We just need to reload it to make predictions.

reg.load_network(load_path='../models/best_x0_model')

Visualization

Prediction for single process parameter values

x, y = reg.predict({
        'Blechdicke': 1.01, 
        'Niederhalterkraft': 410.0, 
        'Ziehspalt': 2.4, 
        'Einlegeposition': -5, 
        'Ziehtiefe': 30,
        'Stempel_ID': 3,
        'E': 191.37245,
        'Rp0': 138.22696,
        'Rp50': 449.528189,
    }, 
    positions=1000)

plt.figure()
plt.plot(x, y[:, 0])
plt.xlabel('tp')
plt.ylabel('deviationc')
plt.show()

Comparison with the ground truth on the training set

Randomly choose an id and compare the prediction to the ground truth. If the experiment does not exist in the data, an error will be thrown.

idx = np.random.choice(data['doe_id'].unique())
print("Doe_ID", idx)
data = reg.compare(idx)

Doe_ID 106

%matplotlib inline
plt.rcParams['figure.dpi'] = 150

def viz(x, y):
    plt.figure()
    plt.plot(x, y[:, 0])
    plt.xlabel('tp')
    plt.ylabel('deviationc')


reg.interactive(function=viz, positions=100)