Model

class poutyne.framework.Model(model, optimizer, loss_function, *, metrics=None, batch_metrics=None, epoch_metrics=None)[source]

The Model class encapsulates a PyTorch module/network, a PyTorch optimizer, a loss function and metric functions. It allows the user to train a neural network without hand-coding the epoch/step logic.

Parameters
  • model (torch.nn.Module) – A PyTorch module.

  • optimizer (Union[torch.optim.Optimizer, str]) – If torch.optim.Optimier, an initialized PyTorch. If str, should be the optimizer’s name in Pytorch (i.e. ‘Adam’ for torch.optim.Adam). (Default value = ‘sgd’)

  • loss_function (Union[Callable, str]) – can also be a string with the same name as a PyTorch loss function (either the functional or object name). The loss function must have the signature loss_function(input, target) where input is the prediction of the network and target is the ground truth. (Default value = None)

  • metrics (list) – *metrics is deprecated as of version 0.5.1. Use batch_metrics instead.* List of functions with the same signature as the loss function. Each metric can be any PyTorch loss function. It can also be a string with the same name as a PyTorch loss function (either the functional or object name). ‘accuracy’ (or just ‘acc’) is also a valid metric. Each metric function is called on each batch of the optimization and on the validation batches at the end of the epoch. (Default value = None)

  • batch_metrics (list) – List of functions with the same signature as the loss function. Each metric can be any PyTorch loss function. It can also be a string with the same name as a PyTorch loss function (either the functional or object name). ‘accuracy’ (or just ‘acc’) is also a valid metric. Each metric function is called on each batch of the optimization and on the validation batches at the end of the epoch. (Default value = None)

  • epoch_metrics (list) – List of functions with the same signature as EpochMetric (Default value = None)

model

The associated PyTorch module.

Type

torch.nn.Module

optimizer

The associated PyTorch optimizer.

Type

torch.optim.Optimizer

loss_function

The associated loss function.

metrics

*metrics is deprecated as of version 0.5.1. Use batch_metrics instead.* The associated metric functions.

Type

list

batch_metrics

The associated metric functions for every batch.

Type

list

epoch_metrics

The associated metric functions for every epoch.

Type

list

Example

Using Numpy arrays (or tensors) dataset:

from poutyne.framework import Model
import torch
import numpy as np

num_features = 20
num_classes = 5

# Our training dataset with 800 samples.
num_train_samples = 800
train_x = np.random.randn(num_train_samples, num_features).astype('float32')
train_y = np.random.randint(num_classes, size=num_train_samples).astype('int64')

# Our validation dataset with 200 samples.
num_valid_samples = 200
valid_x = np.random.randn(num_valid_samples, num_features).astype('float32')
valid_y = np.random.randint(num_classes, size=num_valid_samples).astype('int64')

pytorch_module = torch.nn.Linear(num_features, num_classes) # Our network

# We create and optimize our model
model = Model(pytorch_module, 'sgd', 'cross_entropy', batch_metrics=['accuracy'])
model.fit(train_x, train_y,
          validation_data=(valid_x, valid_y),
          epochs=5,
          batch_size=32)
Epoch 1/5 0.02s Step 25/25: loss: 1.719885, acc: 19.375000, val_loss: 1.667446, val_acc: 22.000000
Epoch 2/5 0.02s Step 25/25: loss: 1.705489, acc: 19.750000, val_loss: 1.660806, val_acc: 22.000000
Epoch 3/5 0.01s Step 25/25: loss: 1.692345, acc: 19.625000, val_loss: 1.655008, val_acc: 22.500000
...

Using PyTorch DataLoader:

import torch
from torch.utils.data import DataLoader, TensorDataset
from poutyne.framework import Model

num_features = 20
num_classes = 5

# Our training dataset with 800 samples.
num_train_samples = 800
train_x = torch.rand(num_train_samples, num_features)
train_y = torch.randint(num_classes, (num_train_samples,), dtype=torch.long)
train_dataset = TensorDataset(train_x, train_y)
train_generator = DataLoader(train_dataset, batch_size=32)

# Our validation dataset with 200 samples.
num_valid_samples = 200
valid_x = torch.rand(num_valid_samples, num_features)
valid_y = torch.randint(num_classes, (num_valid_samples,), dtype=torch.long)
valid_dataset = TensorDataset(valid_x, valid_y)
valid_generator = DataLoader(valid_dataset, batch_size=32)

pytorch_module = torch.nn.Linear(num_features, num_train_samples)

model = Model(pytorch_module, 'sgd', 'cross_entropy', batch_metrics=['accuracy'])
model.fit_generator(train_generator,
                    valid_generator,
                    epochs=5)
Epoch 1/5 0.05s Step 25/25: loss: 6.752676, acc: 0.000000, val_loss: 6.575071, val_acc: 0.000000
Epoch 2/5 0.03s Step 25/25: loss: 6.454859, acc: 0.125000, val_loss: 6.279577, val_acc: 0.000000
Epoch 3/5 0.03s Step 25/25: loss: 6.158523, acc: 2.125000, val_loss: 5.985811, val_acc: 9.500000
...
cpu(*args, **kwargs)[source]

Tranfers the network on the CPU. The arguments are passed to the torch.nn.Module.cpu() method. Notice that the device is saved so that the batches can send to the right device before passing it to the network.

Note

PyTorch optimizers assume that the parameters have been transfered to the right device before their creations. Furthermore, future versions of PyTorch will no longer modify the parameters of a PyTorch module in-place when transferring them to another device. See this issue and this pull request for details.

Since Poutyne supposes that the optimizer has been initialized before the Poutyne Model, necessarily the parameters are not guaranteed to be in sync with those contained in the optimizer once the PyTorch module is transferred to another device. Thus, this method takes care of this inconsistency by updating the parameters inside the optimizer.

Returns

self.

cuda(*args, **kwargs)[source]

Tranfers the network on the GPU. The arguments are passed to the torch.nn.Module.cuda() method. Notice that the device is saved so that the batches can send to the right device before passing it to the network.

Note

PyTorch optimizers assume that the parameters have been transfered to the right device before their creations. Furthermore, future versions of PyTorch will no longer modify the parameters of a PyTorch module in-place when transferring them to another device. See this issue and this pull request for details.

Since Poutyne supposes that the optimizer has been initialized before the Poutyne Model, necessarily the parameters are not guaranteed to be in sync with those contained in the optimizer once the PyTorch module is transferred to another device. Thus, this method takes care of this inconsistency by updating the parameters inside the optimizer.

Returns

self.

evaluate(x, y, batch_size=32, return_pred=False)[source]

Computes the loss and the metrics of the network on batches of samples and optionally returns the predictions.

Parameters
  • x (Union[Tensor, ndarray] or Union[tuple, list] of Union[Tensor, ndarray]) – Input to the model. Union[Tensor, ndarray] if the model has a single input. Union[tuple, list] of Union[Tensor, ndarray] if the model has multiple inputs.

  • y (Union[Tensor, ndarray] or Union[tuple, list] of Union[Tensor, ndarray]) – Target, corresponding ground truth. Union[Tensor, ndarray] if the model has a single output. Union[tuple, list] of Union[Tensor, ndarray] if the model has multiple outputs.

  • batch_size (int) – Number of samples given to the network at one time. (Default value = 32)

  • return_pred (bool, optional) – Whether to return the predictions. (Default value = False)

Returns

Tuple (loss, metrics, pred_y) where specific elements are omitted if not applicable. If only loss is applicable, then it is returned as a float.

metrics is a Numpy array of size n, where n is the number of batch metrics plus the number of epoch metrics if n > 1. If n == 1, then metrics is a float. If n == 0, the metrics is omitted. The first elements of metrics are the batch metrics and are followed by the epoch metrics. See the fit_generator() method for examples with batch metrics and epoch metrics.

If return_pred is True, pred_y is the list of the predictions of each batch with tensors converted into Numpy arrays. It is otherwise ommited.

evaluate_generator(generator, *, steps=None, return_pred=False, return_ground_truth=False)[source]

Computes the loss and the metrics of the network on batches of samples and optionaly returns the predictions.

Parameters
  • generator – Generator-like object for the dataset. See the fit_generator() method for details on the types of generators supported.

  • steps (int, optional) – Number of iterations done on generator. (Defaults the number of steps needed to see the entire dataset)

  • return_pred (bool, optional) – Whether to return the predictions. (Default value = False)

  • return_ground_truth (bool, optional) – Whether to return the ground truths. (Default value = False)

Returns

Tuple (loss, metrics, pred_y, true_y) where specific elements are omitted if not applicable. If only loss is applicable, then it is returned as a float.

metrics is a Numpy array of size n, where n is the number of batch metrics plus the number of epoch metrics if n > 1. If n == 1, then metrics is a float. If n == 0, the metrics is omitted. The first elements of metrics are the batch metrics and are followed by the epoch metrics.

If return_pred is True, pred_y is the list of the predictions of each batch with tensors converted into Numpy arrays. It is otherwise ommited.

If return_ground_truth is True, true_y is the list of the ground truths of each batch with tensors converted into Numpy arrays. It is otherwise ommited.

Example

With no metrics:

model = Model(pytorch_module, optimizer, loss_function,
              batch_metrics=None)
loss = model.evaluate_generator(test_generator)

With only one batch metric:

model = Model(pytorch_module, optimizer, loss_function,
              batch_metrics=[my_metric_fn])
loss, my_metric = model.evaluate_generator(test_generator)

With several batch metrics:

model = Model(pytorch_module, optimizer, loss_function,
              batch_metrics=[my_metric1_fn, my_metric2_fn])
loss, (my_metric1, my_metric2) = model.evaluate_generator(test_generator)

With one batch metric and one epoch metric:

model = Model(pytorch_module, optimizer, loss_function,
              batch_metrics=[my_metric_fn], epoch_metrics=[MyEpochMetricClass()])
loss, (my_batch_metric, my__epoch_metric) = model.evaluate_generator(test_generator)

With batch metrics and return_pred flag:

model = Model(pytorch_module, optimizer, loss_function,
              batch_metrics=[my_metric1_fn, my_metric2_fn])
loss, (my_metric1, my_metric2), pred_y = model.evaluate_generator(
    test_generator, return_pred=True
)

With batch metrics, return_pred and return_ground_truth flags:

model = Model(pytorch_module, optimizer, loss_function,
              batch_metrics=[my_metric1_fn, my_metric2_fn])
loss, (my_metric1, my_metric2), pred_y, true_y = model.evaluate_generator(
    test_generator, return_pred=True, return_ground_truth=True
)
evaluate_on_batch(x, y, *, return_pred=False)[source]

Computes the loss and the metrics of the network on a single batch of samples and optionally returns the predictions.

Parameters
  • x – Input data as a batch.

  • y – Target data as a batch.

  • return_pred (bool, optional) – Whether to return the predictions for batch. (Default value = False)

Returns

Tuple (loss, metrics, pred_y) where specific elements are omitted if not applicable. If only loss is applicable, then it is returned as a float.

metrics` is a Numpy array of size n, where n is the number of metrics if n > 1. If n == 1, then metrics is a float. If n == 0, the metrics is omitted.

If return_pred is True, pred_y is the list of the predictions of each batch with tensors converted into Numpy arrays. It is otherwise ommited.

fit(x, y, validation_data=None, *, batch_size=32, epochs=1000, steps_per_epoch=None, validation_steps=None, batches_per_step=1, initial_epoch=1, verbose=True, callbacks=None)[source]

Trains the model on a dataset. This method creates generators and calls the fit_generator() method.

Parameters
  • x (Union[Tensor, ndarray] or Union[tuple, list] of Union[Tensor, ndarray]) – Training dataset. Union[Tensor, ndarray] if the model has a single input. Union[tuple, list] of Union[Tensor, ndarray] if the model has multiple inputs.

  • y (Union[Tensor, ndarray] or Union[tuple, list] of Union[Tensor, ndarray]) – Target. Union[Tensor, ndarray] if the model has a single output. Union[tuple, list] of Union[Tensor, ndarray] if the model has multiple outputs.

  • validation_data (Tuple[x_val, y_val]) – Same format as x and y previously described. Validation dataset on which to evaluate the loss and any model metrics at the end of each epoch. The model will not be trained on this data. (Default value = None)

  • batch_size (int) – Number of samples given to the network at one time. (Default value = 32)

  • epochs (int) – Number of times the entire training dataset is seen. (Default value = 1000)

  • steps_per_epoch (int, optional) – Number of batch used during one epoch. Obviously, using this argument may cause one epoch not to see the entire training dataset or see it multiple times. (Defaults the number of steps needed to see the entire training dataset)

  • validation_steps (int, optional) – Same as for steps_per_epoch but for the validation dataset. (Defaults to steps_per_epoch if provided or the number of steps needed to see the entire validation dataset)

  • batches_per_step (int) – Number of batches on which to compute the running loss before backpropagating it through the network. Note that the total loss used for backpropagation is the mean of the batches_per_step batch losses. (Default value = 1)

  • initial_epoch (int, optional) – Epoch at which to start training (useful for resuming a previous training run). (Default value = 1)

  • verbose (bool) – Whether to display the progress of the training. (Default value = True)

  • callbacks (List[Callback]) – List of callbacks that will be called during training. (Default value = None)

Returns

List of dict containing the history of each epoch.

Example

model = Model(pytorch_module, optimizer, loss_function)
history = model.fit(train_x, train_y,
                    validation_data=(valid_x, valid_y)
                    epochs=num_epochs,
                    batch_size=batch_size,
                    verbose=False)
print(*history, sep="\n")
{'epoch': 1, 'loss': 1.7198852968215943, 'time': 0.019999928001197986, 'acc': 19.375, 'val_loss': 1.6674459838867188, 'val_acc': 22.0}
{'epoch': 2, 'loss': 1.7054892110824584, 'time': 0.015421080999658443, 'acc': 19.75, 'val_loss': 1.660806336402893, 'val_acc': 22.0}
{'epoch': 3, 'loss': 1.6923445892333984, 'time': 0.01363091799794347, 'acc': 19.625, 'val_loss': 1.6550078630447387, 'val_acc': 22.5}
...
fit_generator(train_generator, valid_generator=None, *, epochs=1000, steps_per_epoch=None, validation_steps=None, batches_per_step=1, initial_epoch=1, verbose=True, callbacks=None)[source]

Trains the model on a dataset using a generator.

Parameters
  • train_generator

    Generator-like object for the training dataset. The generator must yield a batch in the form of a tuple (x, y) where x is the input and y is the target len(x) is taken as the batch size (or the first element of x in case of multi inputs). The loss and the metrics are averaged using this batch size. If the batch size cannot be, inferred then a warning is raised and the “batch size” defaults to 1.

    If the generator does not have a method __len__(), either the steps_per_epoch argument must be provided, or the iterator returned raises a StopIteration exception at the end of the training dataset. PyTorch DataLoaders object do provide a __len__() method.

    Before each epoch, the method __iter__() on the generator is called and the method __next__() is called for each step on resulting object returned by __iter__(). Notice that a call to __iter__() on a generator made using the python keyword yield returns the generator itself.

  • valid_generator (optional) – Generator-like object for the validation dataset. This generator is optional. The generator is used the same way as the generator train_generator. If the generator does not have a method __len__(), either the validation_steps or the steps_per_epoch argument must be provided or the iterator returned raises a StopIteration exception at the end of the validation dataset. (Default value = None)

  • epochs (int) – Number of times the entire training dataset is seen. (Default value = 1000)

  • steps_per_epoch (int, optional) – Number of batch used during one epoch. Obviously, using this argument may cause one epoch not to see the entire training dataset or see it multiple times. (Defaults the number of steps needed to see the entire training dataset)

  • validation_steps (int, optional) – Same as for steps_per_epoch but for the validation dataset. (Defaults to steps_per_epoch if provided or the number of steps needed to see the entire validation dataset)

  • batches_per_step (int) – Number of batches on which to compute the running loss before backpropagating it through the network. Note that the total loss used for backpropagation is the mean of the batches_per_step batch losses. (Default value = 1)

  • initial_epoch (int, optional) – Epoch at which to start training (useful for resuming a previous training run). (Default value = 1)

  • verbose (bool) – Whether to display the progress of the training. (Default value = True)

  • callbacks (List[Callback]) – List of callbacks that will be called during training. (Default value = None)

Returns

List of dict containing the history of each epoch.

Example

model = Model(pytorch_module, optimizer, loss_function)
history = model.fit_generator(train_generator,
                              valid_generator,
                              epochs=num_epochs,
                              verbose=False)
print(*history, sep="\n")
{'epoch': 1, 'loss': 1.7198852968215943, 'time': 0.019999928001197986, 'acc': 19.375, 'val_loss': 1.6674459838867188, 'val_acc': 22.0}
{'epoch': 2, 'loss': 1.7054892110824584, 'time': 0.015421080999658443, 'acc': 19.75, 'val_loss': 1.660806336402893, 'val_acc': 22.0}
{'epoch': 3, 'loss': 1.6923445892333984, 'time': 0.01363091799794347, 'acc': 19.625, 'val_loss': 1.6550078630447387, 'val_acc': 22.5}
...
get_weight_copies()[source]

Returns a dictionary containing copies of the parameters of the network.

get_weights()[source]

Returns a dictionary containing the parameters of the network. The tensors are just references to the parameters. To get copies of the weights, see the get_weight_copies() method.

load_optimizer_state(f)[source]

Loads the optimizer state saved using the torch.save() method or the save_optimizer_state() method of this class.

Parameters

f – File-like object (has to implement fileno that returns a file descriptor) or string containing a file name.

load_weights(f)[source]

Loads the weights saved using the torch.save() method or the save_weights() method of this class. Contrary to torch.load(), the weights are not transfered to the device from which they were saved from. In other words, the PyTorch module will stay on the same device it already is on.

Parameters

f – File-like object (has to implement fileno that returns a file descriptor) or string containing a file name.

predict(x, batch_size=32)[source]

Returns the predictions of the network given a dataset x, where the tensors are converted into Numpy arrays.

Parameters
  • x (Union[Tensor, ndarray] or Union[tuple, list] of Union[Tensor, ndarray]) – Input to the model. Union[Tensor, ndarray] if the model has a single input. Union[tuple, list] of Union[Tensor, ndarray] if the model has multiple inputs.

  • batch_size (int) – Number of samples given to the network at one time. (Default value = 32)

Returns

Numpy arrays of the predictions.

predict_generator(generator, *, steps=None)[source]

Returns the predictions of the network given batches of samples x, where the tensors are converted into Numpy arrays.

generator: Generator-like object for the dataset. The generator must yield a batch of

samples. See the fit_generator() method for details on the types of generators supported. This should only yield input data x and not the target y.

steps (int, optional): Number of iterations done on generator.

(Defaults the number of steps needed to see the entire dataset)

Returns

List of the predictions of each batch with tensors converted into Numpy arrays.

predict_on_batch(x)[source]

Returns the predictions of the network given a batch x, where the tensors are converted into Numpy arrays.

Parameters

x – Input data as a batch.

Returns

The predictions with tensors converted into Numpy arrays.

save_optimizer_state(f)[source]

Saves the state of the current optimizer.

Parameters

f – File-like object (has to implement fileno that returns a file descriptor) or string containing a file name.

save_weights(f)[source]

Saves the weights of the current network.

Parameters

f – File-like object (has to implement fileno that returns a file descriptor) or string containing a file name.

set_weights(weights)[source]

Modifies the weights of the network with the given weights.

Parameters

weights (dict) – Weights returned by either get_weights() or get_weight_copies().

to(device)[source]

Tranfers the network on the specified device. The device is saved so that the batches can send to the right device before passing it to the network.

Note

PyTorch optimizers assume that the parameters have been transfered to the right device before their creations. Furthermore, future versions of PyTorch will no longer modify the parameters of a PyTorch module in-place when transferring them to another device. See this issue and this pull request for details.

Since Poutyne supposes that the optimizer has been initialized before the Poutyne Model, necessarily the parameters are not guaranteed to be in sync with those contained in the optimizer once the PyTorch module is transferred to another device. Thus, this method takes care of this inconsistency by updating the parameters inside the optimizer.

Parameters

device (torch.torch.device) – The device to which the network is sent.

Returns

self.

train_on_batch(x, y, return_pred=False)[source]

Trains the model for the batch (x, y) and computes the loss and the metrics, and optionaly returns the predictions.

Parameters
  • x – Input data as a batch.

  • y – Target data as a batch.

  • return_pred (bool, optional) – Whether to return the predictions. (Default value = False)

Returns

Float loss if no metrics were specified and return_pred is false.

Otherwise, tuple (loss, metrics) if return_pred is false. metrics is a Numpy array of size n, where n is the number of metrics if n > 1. If n == 1, then metrics is a float. If n == 0, the metrics is omitted.

Tuple (loss, metrics, pred_y) if return_pred is true where pred_y is the predictions with tensors converted into Numpy arrays.