tensorlayer3/tensorlayer/models/core.py

439 lines
19 KiB
Python

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
from collections.abc import Iterable
from tensorlayer.files import utils
from tensorlayer import logging
import tensorlayer as tl
from tensorlayer.layers.core import Module
import numpy as np
import os
import time
if tl.BACKEND == 'tensorflow':
import tensorflow as tf
if tl.BACKEND == 'mindspore':
import mindspore as ms
from mindspore.ops import composite
from mindspore.ops import operations as P
from mindspore.ops import functional as F
# from mindspore.parallel._utils import (_get_device_num, _get_mirror_mean, _get_parallel_mode)
# from mindspore.train.parallel_utils import ParallelMode
from mindspore.nn.wrap import DistributedGradReducer
from mindspore.common import ParameterTuple
if tl.BACKEND == 'paddle':
import paddle as pd
class Model:
"""
High-Level API for Training or Testing.
`Model` groups layers into an object with training and inference features.
Args:
network : The training or testing network.
loss_fn : Objective function, if loss_fn is None, the
network should contain the logic of loss and grads calculation, and the logic
of parallel if needed. Default: None.
optimizer : Optimizer for updating the weights. Default: None.
metrics : Dict or set of metrics to be evaluated by the model during
Examples:
>>> import tensorlayer as tl
>>> class Net(Module):
>>> def __init__(self):
>>> super(Net, self).__init__()
>>> self.conv = tl.layers.Conv2d(n_filter=32, filter_size=(3, 3), strides=(2, 2), in_channels=5, name='conv2d')
>>> self.bn = tl.layers.BatchNorm2d(num_features=32, act=tl.ReLU)
>>> self.flatten = tl.layers.Flatten()
>>> self.fc = tl.layers.Dense(n_units=12, in_channels=32*224*224) # padding=0
>>>
>>> def construct(self, x):
>>> x = self.conv(x)
>>> x = self.bn(x)
>>> x = self.flatten(x)
>>> out = self.fc(x)
>>> return out
>>>
>>> net = Net()
>>> loss = tl.cost.SoftmaxCrossEntropyWithLogits(is_grad=False, sparse=True)
>>> optim = tl.layers.Momentum(params=net.trainable_weights, learning_rate=0.1, momentum=0.9)
>>> model = Model(net, loss_fn=loss, optimizer=optim, metrics=None)
>>> dataset = get_dataset()
>>> model.train(2, dataset)
"""
def __init__(self, network, loss_fn=None, optimizer=None, metrics=None, **kwargs):
self.network = network
self.loss_fn = loss_fn
self.optimizer = optimizer
self.metrics = metrics
self.all_weights = network.all_weights
self.train_weights = self.network.trainable_weights
def train(self, n_epoch, train_dataset=None, test_dataset=False, print_train_batch=False, print_freq=5):
if not isinstance(train_dataset, Iterable):
raise Exception("Expected type in (train_dataset, Iterable), but got {}.".format(type(train_dataset)))
if tl.BACKEND == 'tensorflow':
self.tf_train(
n_epoch=n_epoch, train_dataset=train_dataset, network=self.network, loss_fn=self.loss_fn,
train_weights=self.train_weights, optimizer=self.optimizer, metrics=self.metrics,
print_train_batch=print_train_batch, print_freq=print_freq, test_dataset=test_dataset
)
elif tl.BACKEND == 'mindspore':
self.ms_train(
n_epoch=n_epoch, train_dataset=train_dataset, network=self.network, loss_fn=self.loss_fn,
train_weights=self.train_weights, optimizer=self.optimizer, metrics=self.metrics,
print_train_batch=print_train_batch, print_freq=print_freq, test_dataset=test_dataset
)
elif tl.BACKEND == 'paddle':
self.pd_train(
n_epoch=n_epoch, train_dataset=train_dataset, network=self.network, loss_fn=self.loss_fn,
train_weights=self.train_weights, optimizer=self.optimizer, metrics=self.metrics,
print_train_batch=print_train_batch, print_freq=print_freq, test_dataset=test_dataset
)
def eval(self, test_dataset):
self.network.eval()
test_loss, test_acc, n_iter = 0, 0, 0
for X_batch, y_batch in test_dataset:
_logits = self.network(X_batch)
test_loss += self.loss_fn(_logits, y_batch)
if self.metrics:
test_acc += self.metrics(_logits, y_batch)
else:
test_acc += np.mean(np.equal(np.argmax(_logits, 1), y_batch))
n_iter += 1
print(" test loss: {}".format(test_loss / n_iter))
print(" test acc: {}".format(test_acc / n_iter))
def save_weights(self, file_path, format=None):
"""Input file_path, save model weights into a file of given format.
Use self.load_weights() to restore.
Parameters
----------
file_path : str
Filename to which the model weights will be saved.
format : str or None
Saved file format.
Value should be None, 'hdf5', 'npz', 'npz_dict' or 'ckpt'. Other format is not supported now.
1) If this is set to None, then the postfix of file_path will be used to decide saved format.
If the postfix is not in ['h5', 'hdf5', 'npz', 'ckpt'], then file will be saved in hdf5 format by default.
2) 'hdf5' will save model weights name in a list and each layer has its weights stored in a group of
the hdf5 file.
3) 'npz' will save model weights sequentially into a npz file.
4) 'npz_dict' will save model weights along with its name as a dict into a npz file.
5) 'ckpt' will save model weights into a tensorflow ckpt file.
Default None.
Examples
--------
1) Save model weights in hdf5 format by default.
>>> net = vgg16()
>>> net.save_weights('./model.h5')
...
>>> net.load_weights('./model.h5')
2) Save model weights in npz/npz_dict format
>>> net = vgg16()
>>> net.save_weights('./model.npz')
>>> net.save_weights('./model.npz', format='npz_dict')
"""
# self.all_weights = self.network.all_weights
if self.all_weights is None or len(self.all_weights) == 0:
logging.warning("Model contains no weights or layers haven't been built, nothing will be saved")
return
if format is None:
postfix = file_path.split('.')[-1]
if postfix in ['h5', 'hdf5', 'npz', 'ckpt']:
format = postfix
else:
format = 'hdf5'
if format == 'hdf5' or format == 'h5':
utils.save_weights_to_hdf5(file_path, self)
elif format == 'npz':
utils.save_npz(self.all_weights, file_path)
elif format == 'npz_dict':
utils.save_npz_dict(self.all_weights, file_path)
elif format == 'ckpt':
# TODO: enable this when tf save ckpt is enabled
raise NotImplementedError("ckpt load/save is not supported now.")
else:
raise ValueError(
"Save format must be 'hdf5', 'npz', 'npz_dict' or 'ckpt'."
"Other format is not supported now."
)
def load_weights(self, file_path, format=None, in_order=True, skip=False):
"""Load model weights from a given file, which should be previously saved by self.save_weights().
Parameters
----------
file_path : str
Filename from which the model weights will be loaded.
format : str or None
If not specified (None), the postfix of the file_path will be used to decide its format. If specified,
value should be 'hdf5', 'npz', 'npz_dict' or 'ckpt'. Other format is not supported now.
In addition, it should be the same format when you saved the file using self.save_weights().
Default is None.
in_order : bool
Allow loading weights into model in a sequential way or by name. Only useful when 'format' is 'hdf5'.
If 'in_order' is True, weights from the file will be loaded into model in a sequential way.
If 'in_order' is False, weights from the file will be loaded into model by matching the name
with the weights of the model, particularly useful when trying to restore model in eager(graph) mode from
a weights file which is saved in graph(eager) mode.
Default is True.
skip : bool
Allow skipping weights whose name is mismatched between the file and model. Only useful when 'format' is
'hdf5' or 'npz_dict'. If 'skip' is True, 'in_order' argument will be ignored and those loaded weights
whose name is not found in model weights (self.all_weights) will be skipped. If 'skip' is False, error will
occur when mismatch is found.
Default is False.
Examples
--------
1) load model from a hdf5 file.
>>> net = tl.models.vgg16()
>>> net.load_weights('./model_graph.h5', in_order=False, skip=True) # load weights by name, skipping mismatch
>>> net.load_weights('./model_eager.h5') # load sequentially
2) load model from a npz file
>>> net.load_weights('./model.npz')
2) load model from a npz file, which is saved as npz_dict previously
>>> net.load_weights('./model.npz', format='npz_dict')
Notes
-------
1) 'in_order' is only useful when 'format' is 'hdf5'. If you are trying to load a weights file which is
saved in a different mode, it is recommended to set 'in_order' be True.
2) 'skip' is useful when 'format' is 'hdf5' or 'npz_dict'. If 'skip' is True,
'in_order' argument will be ignored.
"""
if not os.path.exists(file_path):
raise FileNotFoundError("file {} doesn't exist.".format(file_path))
if format is None:
format = file_path.split('.')[-1]
if format == 'hdf5' or format == 'h5':
if skip ==True or in_order == False:
# load by weights name
utils.load_hdf5_to_weights(file_path, self, skip)
else:
# load in order
utils.load_hdf5_to_weights_in_order(file_path, self)
elif format == 'npz':
utils.load_and_assign_npz(file_path, self)
elif format == 'npz_dict':
utils.load_and_assign_npz_dict(file_path, self, skip)
elif format == 'ckpt':
# TODO: enable this when tf save ckpt is enabled
raise NotImplementedError("ckpt load/save is not supported now.")
else:
raise ValueError(
"File format must be 'hdf5', 'npz', 'npz_dict' or 'ckpt'. "
"Other format is not supported now."
)
def tf_train(
self, n_epoch, train_dataset, network, loss_fn, train_weights, optimizer, metrics, print_train_batch,
print_freq, test_dataset
):
for epoch in range(n_epoch):
start_time = time.time()
train_loss, train_acc, n_iter = 0, 0, 0
for X_batch, y_batch in train_dataset:
network.set_train()
with tf.GradientTape() as tape:
# compute outputs
_logits = network(X_batch)
# compute loss and update model
_loss_ce = loss_fn(_logits, y_batch)
grad = tape.gradient(_loss_ce, train_weights)
optimizer.apply_gradients(zip(grad, train_weights))
train_loss += _loss_ce
if metrics:
metrics.update(_logits, y_batch)
train_acc += metrics.result()
metrics.reset()
else:
train_acc += np.mean(np.equal(np.argmax(_logits, 1), y_batch))
n_iter += 1
if print_train_batch:
print("Epoch {} of {} took {}".format(epoch + 1, n_epoch, time.time() - start_time))
print(" train loss: {}".format(train_loss / n_iter))
print(" train acc: {}".format(train_acc / n_iter))
if epoch + 1 == 1 or (epoch + 1) % print_freq == 0:
print("Epoch {} of {} took {}".format(epoch + 1, n_epoch, time.time() - start_time))
print(" train loss: {}".format(train_loss / n_iter))
print(" train acc: {}".format(train_acc / n_iter))
if test_dataset:
# use training and evaluation sets to evaluate the model every print_freq epoch
if epoch + 1 == 1 or (epoch + 1) % print_freq == 0:
network.eval()
val_loss, val_acc, n_iter = 0, 0, 0
for X_batch, y_batch in test_dataset:
_logits = network(X_batch) # is_train=False, disable dropout
val_loss += loss_fn(_logits, y_batch, name='eval_loss')
if metrics:
metrics.update(_logits, y_batch)
val_acc += metrics.result()
metrics.reset()
else:
val_acc += np.mean(np.equal(np.argmax(_logits, 1), y_batch))
n_iter += 1
print(" val loss: {}".format(val_loss / n_iter))
print(" val acc: {}".format(val_acc / n_iter))
def ms_train(
self, n_epoch, train_dataset, network, loss_fn, train_weights, optimizer, metrics, print_train_batch,
print_freq, test_dataset
):
net_with_criterion = WithLoss(network, loss_fn)
train_network = GradWrap(net_with_criterion)
train_network.set_train()
for epoch in range(n_epoch):
start_time = time.time()
train_loss, train_acc, n_iter = 0, 0, 0
for X_batch, y_batch in train_dataset:
output = network(X_batch)
loss_output = loss_fn(output, y_batch)
grads = train_network(X_batch, y_batch)
success = optimizer.apply_gradients(zip(grads, train_weights))
loss = loss_output.asnumpy()
train_loss += loss
if metrics:
metrics.update(output, y_batch)
train_acc += metrics.result()
metrics.reset()
else:
train_acc += np.mean((P.Equal()(P.Argmax(axis=1)(output), y_batch).asnumpy()))
n_iter += 1
if print_train_batch:
print("Epoch {} of {} took {}".format(epoch + 1, n_epoch, time.time() - start_time))
print(" train loss: {}".format(train_loss / n_iter))
print(" train acc: {}".format(train_acc / n_iter))
if epoch + 1 == 1 or (epoch + 1) % print_freq == 0:
print("Epoch {} of {} took {}".format(epoch + 1, n_epoch, time.time() - start_time))
print(" train loss: {}".format(train_loss / n_iter))
print(" train acc: {}".format(train_acc / n_iter))
if test_dataset:
# use training and evaluation sets to evaluate the model every print_freq epoch
if epoch + 1 == 1 or (epoch + 1) % print_freq == 0:
network.eval()
val_loss, val_acc, n_iter = 0, 0, 0
for X_batch, y_batch in test_dataset:
_logits = network(X_batch)
val_loss += loss_fn(_logits, y_batch, name='eval_loss')
if metrics:
metrics.update(_logits, y_batch)
val_acc += metrics.result()
metrics.reset()
else:
val_acc += np.mean((P.Equal()(P.Argmax(axis=1)(_logits), y_batch).asnumpy()))
n_iter += 1
print(" val loss: {}".format(val_loss / n_iter))
print(" val acc: {}".format(val_acc / n_iter))
def pd_train(
self, n_epoch, train_dataset, network, loss_fn, train_weights, optimizer, metrics, print_train_batch,
print_freq, test_dataset
):
for epoch in range(n_epoch):
start_time = time.time()
train_loss, train_acc, n_iter = 0, 0, 0
for X_batch, y_batch in train_dataset:
network.set_train()
output = network(X_batch)
loss = loss_fn(output, y_batch)
loss_ce = loss.numpy()
params_grads = optimizer.gradient(loss, train_weights)
optimizer.apply_gradients(params_grads)
train_loss += loss_ce
if metrics:
metrics.update(output, y_batch)
train_acc += metrics.result()
metrics.reset()
else:
train_acc += pd.metric.accuracy(output, y_batch)
n_iter += 1
if print_train_batch:
print("Epoch {} of {} took {}".format(epoch + 1, n_epoch, time.time() - start_time))
print(" train loss: {}".format(train_loss / n_iter))
print(" train acc: {}".format(train_acc / n_iter))
if epoch + 1 == 1 or (epoch + 1) % print_freq == 0:
print("Epoch {} of {} took {}".format(epoch + 1, n_epoch, time.time() - start_time))
print(" train loss: {}".format(train_loss / n_iter))
print(" train acc: {}".format(train_acc / n_iter))
if test_dataset:
# use training and evaluation sets to evaluate the model every print_freq epoch
if epoch + 1 == 1 or (epoch + 1) % print_freq == 0:
network.eval()
val_loss, val_acc, n_iter = 0, 0, 0
for X_batch, y_batch in test_dataset:
_logits = network(X_batch) # is_train=False, disable dropout
val_loss += loss_fn(_logits, y_batch, name='eval_loss')
if metrics:
metrics.update(_logits, y_batch)
val_acc += metrics.result()
metrics.reset()
else:
val_acc += np.mean(np.equal(np.argmax(_logits, 1), y_batch))
n_iter += 1
print(" val loss: {}".format(val_loss / n_iter))
print(" val acc: {}".format(val_acc / n_iter))
class WithLoss(Module):
def __init__(self, backbone, loss_fn):
super(WithLoss, self).__init__()
self._backbone = backbone
self._loss_fn = loss_fn
def construct(self, data, label):
out = self._backbone(data)
return self._loss_fn(out, label)
@property
def backbone_network(self):
return self._backbone
class GradWrap(Module):
""" GradWrap definition """
def __init__(self, network):
super(GradWrap, self).__init__(auto_prefix=False)
self.network = network
self.weights = ParameterTuple(network.trainable_weights)
def forward(self, x, label):
return composite.GradOperation(get_by_list=True)(self.network, self.weights)(x, label)