Create a pytorch model from my H2O model

[matt7salomon]

My organization only allows pytroch models to go into production. I wrote the following code to convert my trained h2o model into pytorch but it errors out.

import torch
import torch.nn as nn

# Define a PyTorch model with the same architecture as the H2O model
class ConvertedH2OModel(nn.Module):
    def __init__(self, input_dim):
        super(ConvertedH2OModel, self).__init__()
        self.fc1 = nn.Linear(input_dim, 256)
        self.bn1 = nn.BatchNorm1d(256)
        self.fc2 = nn.Linear(256, 128)
        self.bn2 = nn.BatchNorm1d(128)
        self.fc3 = nn.Linear(128, 64)
        self.fc4 = nn.Linear(64, 1)
        self.dropout = nn.Dropout(p=0.3)

    def forward(self, x):
        x = torch.relu(self.bn1(self.fc1(x)))
        x = self.dropout(x)
        x = torch.relu(self.bn2(self.fc2(x)))
        x = self.dropout(x)
        x = torch.relu(self.fc3(x))
        x = self.fc4(x)
        return x

# Initialize the PyTorch model
input_dim = num_features  # Assuming num_features is defined
torch_model = ConvertedH2OModel(input_dim)

# Check if the model has weights and biases
if specific_model._model_json["output"]["weights"] is not None and specific_model._model_json["output"]["biases"] is not None:
    # Extract weights and biases from the H2O model
    h2o_weights = [specific_model.weights(i) for i in range(len(specific_model.params['hidden']))]
    h2o_biases = [specific_model.biases(i) for i in range(len(specific_model.params['hidden']))]

    # Assign the extracted weights and biases to the PyTorch model
    with torch.no_grad():
        torch_model.fc1.weight = nn.Parameter(torch.tensor(h2o_weights[0].T, dtype=torch.float32))
        torch_model.fc1.bias = nn.Parameter(torch.tensor(h2o_biases[0], dtype=torch.float32))
        torch_model.fc2.weight = nn.Parameter(torch.tensor(h2o_weights[1].T, dtype=torch.float32))
        torch_model.fc2.bias = nn.Parameter(torch.tensor(h2o_biases[1], dtype=torch.float32))
        torch_model.fc3.weight = nn.Parameter(torch.tensor(h2o_weights[2].T, dtype=torch.float32))
        torch_model.fc3.bias = nn.Parameter(torch.tensor(h2o_biases[2], dtype=torch.float32))
        torch_model.fc4.weight = nn.Parameter(torch.tensor(h2o_weights[3].T, dtype=torch.float32))
        torch_model.fc4.bias = nn.Parameter(torch.tensor(h2o_biases[3], dtype=torch.float32))

    # Now the torch_model is initialized with the weights and biases from the H2O model
else:
    print("The model does not have weights and biases available for extraction.")

Here is my h2o mode summary:

DeepLearning_1_AutoML_1_20240923_211230
Status of Neuron Layers: predicting target, regression, gaussian distribution, Quadratic loss, 3,101 weights/biases, 77.1 KB, 104,018 training samples, mini-batch size 1
    layer    units    type       dropout    l1    l2    mean_rate               rate_rms                momentum    mean_weight             weight_rms           mean_bias             bias_rms
--  -------  -------  ---------  ---------  ----  ----  ----------------------  ----------------------  ----------  ----------------------  -------------------  --------------------  -----------------------
    1        286      Input      0.0
    2        10       Rectifier  0.0        0.0   0.0   0.025230179532672037    0.009651198983192444    0.0         0.00023786274331370066  0.0954655110836029   0.7788656104215739    0.5845489501953125
    3        10       Rectifier  0.0        0.0   0.0   0.008876945667725521    0.008416924625635147    0.0         -0.004443328110501171   0.28273701667785645  0.9809969561491746    0.06552445888519287
    4        10       Rectifier  0.0        0.0   0.0   0.011014057626889553    0.023908764123916626    0.0         0.02570307268295437     0.3091564178466797   1.022937802614036     0.04745239019393921
    5        1        Linear                0.0   0.0   0.00039743693923810497  0.00022761948639526963  0.0         0.034631247818470004    0.31644153594970703  0.017902538961798036  1.0971281125650402e-154
{'default': 'Automatic', 'actual': 'Automatic', 'input': 'Automatic'}

[wendycwong]

@matt7salomon : You can use our DAI tools which will allow you to generate pytorch models: https://www.google.com/search?client=safari&rls=en&q=h2o+Driverless+AI&ie=UTF-8&oe=UTF-8 . Currently, we don't support pytorch on our opensource platform.

[matt7salomon]

Thanks but my code seem to be able to get the structure transferred from h2o to pytroch and not the weights and biases but i dont need those. I am trying to train in torch again using the same structure and same parameters. I should be able to get all the model info from h2o connection.

[wendycwong]

@matt7salomon : Sorry I misunderstood your question.

To get the weight and biases exposed to the returned model, you need to set the parameter export_weights_and_biases=True like this:

dlmodel = H2ODeepLearningEstimator(hidden=[17,191], epochs=1, balance_classes=False, reproducible=True, seed=1234, export_weights_and_biases=True)

The weights and biases will come back as H2O frames. Here is an example of this:

==================================================================== covtype = h2o.upload_file(pyunit_utils.locate("smalldata/covtype/covtype.20k.data")) covtype[54] = covtype[54].asfactor()

dlmodel = H2ODeepLearningEstimator(hidden=[17,191], epochs=1, balance_classes=False, reproducible=True, seed=1234, export_weights_and_biases=True) dlmodel.train(x=list(range(54)),y=54,training_frame=covtype) print(dlmodel)

weights1 = dlmodel.weights(0) weights2 = dlmodel.weights(1) weights3 = dlmodel.weights(2)

biases1 = dlmodel.biases(0) biases2 = dlmodel.biases(1) biases3 = dlmodel.biases(2)

w1c = weights1.ncol w1r = weights1.nrow assert w1c == 52, "wrong dimensionality! expected {0}, but got {1}.".format(52, w1c) assert w1r == 17, "wrong dimensionality! expected {0}, but got {1}.".format(17, w1r)

w2c = weights2.ncol w2r = weights2.nrow assert w2c == 17, "wrong dimensionality! expected {0}, but got {1}.".format(17, w2c) assert w2r == 191, "wrong dimensionality! expected {0}, but got {1}.".format(191, w2r)

w3c = weights3.ncol w3r = weights3.nrow assert w3c == 191, "wrong dimensionality! expected {0}, but got {1}.".format(191, w3c) assert w3r == 7, "wrong dimensionality! expected {0}, but got {1}.".format(7, w3r)

b1c = biases1.ncol b1r = biases1.nrow assert b1c == 1, "wrong dimensionality! expected {0}, but got {1}.".format(1, b1c) assert b1r == 17, "wrong dimensionality! expected {0}, but got {1}.".format(17, b1r)

b2c = biases2.ncol b2r = biases2.nrow assert b2c == 1, "wrong dimensionality! expected {0}, but got {1}.".format(1, b2c) assert b2r == 191, "wrong dimensionality! expected {0}, but got {1}.".format(191, b2r)

b3c = biases3.ncol b3r = biases3.nrow assert b3c == 1, "wrong dimensionality! expected {0}, but got {1}.".format(1, b3c) assert b3r == 7, "wrong dimensionality! expected {0}, but got {1}.".format(7, b3r)

[wendycwong]

If you already have your model trained without setting that parameter, you can still get the weights and biases via mojo. Here is an example of how to do that:

assumed that you have model:

=========================== model_mojo = model.download_mojo("/some/directory/forMojo")

The model_mojo will be a zip file. Unzip the file and you will find the weights and biases information in the file model.ini.

Here is an example of my model.ini