def onnx2keras(onnx_model_path,
input_names,
output_dir,
swap_channel_ordering=True):
from onnx2keras import onnx_to_keras
onnx_model = onnx.load(onnx_model_path)
k_model = onnx_to_keras(onnx_model,
input_names,
change_ordering=swap_channel_ordering,
name_policy='renumerate')
weights = k_model.get_weights()
K.set_learning_phase(0)
with K.get_session() as sess:
init = tf.global_variables_initializer()
sess.run(init)
k_model.set_weights(weights)
tf.saved_model.simple_save(
sess,
output_dir,
inputs={
input.name: tensor
for input, tensor in zip(onnx_model.graph.input,
k_model.inputs)
},
outputs={
output.name: tensor
for output, tensor in zip(onnx_model.graph.output,
k_model.outputs)
})
def convertToKerasAndBack(fileName, outputFileName, inputs):
ret = None
onnxModel = onnx.load(fileName)
kwArgs = dict()
if "shuffle" in fileName:
kwArgs["input_shapes"] = [1, 3, 224, 224]
import shufflenet
import keras
kerasModel = shufflenet.ShuffleNet(groups=3)
kerasModel.load_weights(
"keras-shufflenet/weights/ShuffleNet_1X_g3_br_0.25_373.hdf5")
kerasModel.compile(optimizer=keras.optimizers.SGD(lr=.05,
decay=5e-4,
momentum=0.9),
metrics=['accuracy'],
loss='categorical_crossentropy')
ret = kerasModel.predict(inputs[0][0].transpose(0, 2, 3, 1))
else:
kerasModel = onnx_to_keras(onnxModel,
getOnnxInputNames(fileName),
verbose=False,
**kwArgs)
#tf.keras.utils.plot_model(kerasModel, show_shapes=True)
backconvOnnxModel = onnxmltools.convert_keras(kerasModel)
onnxmltools.utils.save_model(backconvOnnxModel, outputFileName)
return ret
def onnx2keras_pb(onnx_model_path,
input_names,
output_names,
output_path,
swap_channel_ordering=True):
from onnx2keras import onnx_to_keras
from convert.utils import freeze_session
output_dir, filename = os.path.split(output_path)
onnx_model = onnx.load(onnx_model_path)
k_model = onnx_to_keras(onnx_model,
input_names,
change_ordering=swap_channel_ordering,
name_policy='renumerate')
weights = k_model.get_weights()
K.set_learning_phase(0)
with K.get_session() as sess:
init = tf.global_variables_initializer()
sess.run(init)
k_model.set_weights(weights)
frozen_graph = freeze_session(sess,
keep_var_names=None,
output_names=output_names)
tf.train.write_graph(frozen_graph, output_dir, filename, as_text=False)
def pytorch2savedmodel(onnx_model_path, saved_model_dir):
onnx_model = onnx.load(onnx_model_path)
input_names = ['input_ids', "attention_mask", "token_type_ids", "task_id"]
k_model = onnx_to_keras(onnx_model=onnx_model,
input_names=input_names,
change_ordering=True,
verbose=False)
weights = k_model.get_weights()
K.set_learning_phase(0)
saved_model_dir = Path(saved_model_dir)
if saved_model_dir.exists():
shutil.rmtree(str(saved_model_dir))
saved_model_dir.mkdir()
with K.get_session() as sess:
init = tf.global_variables_initializer()
sess.run(init)
k_model.set_weights(weights)
tf.saved_model.simple_save(
sess,
str(saved_model_dir.joinpath('1')),
inputs={'image_array': k_model.input},
outputs=dict((output.name, tensor) for output, tensor in zip(
onnx_model.graph.output, k_model.outputs)))
def pytorch2savedmodel(onnx_model_path, saved_model_dir):
onnx_model = onnx.load(onnx_model_path)
input_names = ["image_array"]
k_model = onnx_to_keras(
onnx_model=onnx_model,
input_names=input_names,
change_ordering=True,
verbose=False,
)
weights = k_model.get_weights()
K.set_learning_phase(0)
saved_model_dir = Path(saved_model_dir)
if saved_model_dir.exists():
shutil.rmtree(str(saved_model_dir))
saved_model_dir.mkdir()
tf.saved_model.save(
k_model,
str(saved_model_dir.joinpath("1")),
)
def get_keras(dynamic_input=False):
onnx_model = onnx.load(
"/home/fico/DeepL/Gans/Mobile_Cyclegan/vangogh2photo_fixed.onnx")
# Call the converter (input - is the main model input name, can be different for your model)
k_model = onnx_to_keras(onnx_model, ['input'],
test_c_order=True,
dynamic_input=dynamic_input)
# if dynamic_input == True:
# model_config = k_model.get_config()
# print(model_config['layers'][0])
print(k_model.summary())
return k_model
def input_processing_onnx(file_path=None):
'''
Process input from onnx file.
Parameters:
-----------
file_path: path to onnx input file. (still in developed)
Return
-----------
number_of_layer: Number of layers in the neural network
number_of_neurons_each_layer: list of number of neurons in each layer
weight: list of weights of the neurons network
bias: list of bias at each layer
../sample_input/eranmnist_benchmark/onnx/tf/mnist_relu_3_50.onnx
'''
onnx_model = onnx.load(file_path)
onnx_model.graph.node[2].output = 'dense/Relu_0'
print(onnx_model.graph.node[2].output)
# print(onnx_model.graph.input)
exit(0)
kera_model = onnx_to_keras(onnx_model, input_names=['dense_input'])
number_of_layer = 0
number_of_neurons_each_layer = []
weight = []
bias = []
for layer in kera_model.layers:
layer_weight = layer.get_weights()
if not layer_weight:
continue
number_of_layer += 1
if number_of_layer == 1:
number_of_neurons_each_layer.append(
np.array(layer_weight[0]).shape[0])
number_of_neurons_each_layer.append(np.array(layer_weight[0]).shape[1])
weight.append(np.array(layer_weight[0]))
bias.append(np.array(layer_weight[1]))
number_of_layer += 1
print(number_of_layer)
print(number_of_neurons_each_layer)
print(weight)
print(bias)
return number_of_layer, number_of_neurons_each_layer, weight, bias
def ToKeras(self,
output_name,
path=None,
genotype=None,
init_channels=None,
layers=None):
model = self.load_model(path, genotype, init_channels, layers)
model.eval()
input_np = np.random.uniform(0, 1, (1, 3, 32, 32))
input_var = torch.FloatTensor(input_np)
output = model(input_var)
torch.onnx.export(model, (input_var),
output_name + ".onnx",
verbose=True,
input_names=['input'],
output_names=['output'])
import onnx
from onnx2keras import onnx_to_keras
onnx_model = onnx.load(output_name + ".onnx")
k_model = onnx_to_keras(onnx_model, ['input'])
k_model.save(output_name + ".h5")
def driver(inputdir,
outputdir,
datadir,
plotdir,
preddir,
trainflag,
validflag,
testflag,
normalize,
fmean,
fstdev,
scale,
fmin,
fmax,
scalelims,
fsize,
rmse_file,
r2_file,
inD,
outD,
ilog,
olog,
TFRfile,
batch_size,
ncores,
buffer_size,
gridsearch,
architectures,
layers,
lay_params,
activations,
act_params,
nodes,
lengthscale,
max_lr,
clr_mode,
clr_steps,
epochs,
patience,
weight_file,
resume,
plot_cases,
fxvals,
xlabel,
ylabel,
filters=None,
filtconv=1.):
"""
Driver function to handle model training and evaluation.
Inputs
------
inputdir : string. Path/to/directory of inputs.
outputdir : string. Path/to/directory of outputs.
datadir : string. Path/to/directory of data.
plotdir : string. Path/to/directory of plots.
preddir : string. Path/to/directory of predictions.
trainflag : bool. Determines whether to train the NN model.
validflag : bool. Determines whether to validate the NN model.
testflag : bool. Determines whether to test the NN model.
normalize : bool. Determines whether to normalize the data.
fmean : string. Path/to/file of mean of training data.
fstdev : string. Path/to/file of stdev of training data.
scale : bool. Determines whether to scale the data.
fmin : string. Path/to/file of minima of training data.
fmax : string. Path/to/file of maxima of training data.
scalelims : list, floats. [min, max] of range of scaled data.
rmse_file : string. Prefix for savefiles for RMSE calculations.
r2_file : string. Prefix for savefiles for R2 calculations.
inD : int. Dimensionality of the input data.
outD : int. Dimensionality of the output data.
ilog : bool. Determines whether to take the log10 of intput data.
olog : bool. Determines whether to take the log10 of output data.
TFRfile : string. Prefix for TFRecords files.
batch_size : int. Size of batches for training/validating/testing.
ncores : int. Number of cores to use to load data cases.
buffer_size: int. Number of data cases to pre-load in memory.
gridsearch : bool. Determines whether to perform a grid search over
`architectures`.
architectures: list. Model architectures to consider.
layers : list, str. Types of hidden layers.
lay_params : list, ints. Parameters for the layer type
E.g., kernel size
activations: list, str. Activation functions for each hidden layer.
act_params : list, floats. Parameters for the activation functions.
nodes : list, ints. For layers with nodes, number of nodes per layer.
lengthscale: float. Minimum learning rat.e
max_lr : float. Maximum learning rate.
clr_mode : string. Sets the cyclical learning rate function.
clr_steps : int. Number of steps per cycle of the learning rate.
epochs : int. Max number of iterations through dataset for training.
patience : int. If no model improvement after `patience` epochs,
halts training.
weight_file: string. Path/to/file where NN weights are saved.
resume : bool. Determines whether to resume training.
plot_cases : list, ints. Cases from test set to plot.
fxvals : string. Path/to/file of X-axis values to correspond to
predicted Y values.
xlabel : string. X-axis label for plotting.
ylabel : string. Y-axis label for plotting.
filters : list, strings. Paths/to/filter files. Default: None
If specified, will compute RMSE/R2 stats over the
integrated filter bandpasses.
filtconv : float. Conversion factor for filter file x-axis values to
desired unit. Default: 1.0
"""
# Get file names, calculate number of cases per file
print('Loading files & calculating total number of batches...')
try:
datsize = np.load(inputdir + fsize)
num_train = datsize[0]
num_valid = datsize[1]
num_test = datsize[2]
except:
ftrain = glob.glob(datadir + 'train' + os.sep + '*.npy')
fvalid = glob.glob(datadir + 'valid' + os.sep + '*.npy')
ftest = glob.glob(datadir + 'test' + os.sep + '*.npy')
num_train = U.data_set_size(ftrain, ncores)
num_valid = U.data_set_size(fvalid, ncores)
num_test = U.data_set_size(ftest, ncores)
np.save(inputdir + fsize,
np.array([num_train, num_valid, num_test], dtype=int))
del ftrain, fvalid, ftest
print("Data set sizes")
print("Training data:", num_train)
print("Validation data:", num_valid)
print("Testing data:", num_test)
print("Total data:", num_train + num_valid + num_test)
train_batches = num_train // batch_size
valid_batches = num_valid // batch_size
test_batches = num_test // batch_size
# Update `clr_steps`
if clr_steps == "range test":
clr_steps = train_batches * epochs
rng_test = True
else:
clr_steps = train_batches * int(clr_steps)
rng_test = False
# Get mean/stdev for normalizing
if normalize:
print('\nNormalizing the data...')
try:
mean = np.load(inputdir + fmean)
stdev = np.load(inputdir + fstdev)
except:
print("Calculating the mean and standard deviation of the data " +\
"using Welford's method.")
# Compute stats
ftrain = glob.glob(datadir + 'train' + os.sep + '*.npy')
mean, stdev, datmin, datmax = S.mean_stdev(ftrain, inD, ilog, olog)
np.save(inputdir + fmean, mean)
np.save(inputdir + fstdev, stdev)
np.save(inputdir + fmin, datmin)
np.save(inputdir + fmax, datmax)
del datmin, datmax, ftrain
print("mean :", mean)
print("stdev:", stdev)
# Slice desired indices
x_mean, y_mean = mean[:inD], mean[inD:]
x_std, y_std = stdev[:inD], stdev[inD:]
# Memory cleanup -- no longer need full mean/stdev arrays
del mean, stdev
else:
x_mean = 0.
x_std = 1.
y_mean = 0.
y_std = 1.
if olog:
# To properly calculate RMSE & R2 for log-scaled output
try:
y_mean_delog = np.load(inputdir +
fmean.replace(".npy", "_delog.npy"))
except:
ftrain = glob.glob(datadir + 'train' + os.sep + '*.npy')
mean_delog = S.mean_stdev(ftrain, inD, ilog, False)[0]
del ftrain
y_mean_delog = mean_delog[inD:]
np.save(inputdir + fmean.replace(".npy", "_delog.npy"),
y_mean_delog)
else:
y_mean_delog = y_mean
# Get min/max values for scaling
if scale:
print('\nScaling the data...')
try:
datmin = np.load(inputdir + fmin)
datmax = np.load(inputdir + fmax)
except:
ftrain = glob.glob(datadir + 'train' + os.sep + '*.npy')
mean, stdev, datmin, datmax = S.mean_stdev(ftrain, inD, ilog, olog)
np.save(inputdir + fmean, mean)
np.save(inputdir + fstdev, stdev)
np.save(inputdir + fmin, datmin)
np.save(inputdir + fmax, datmax)
del mean, stdev, ftrain
print("min :", datmin)
print("max :", datmax)
# Slice desired indices
x_min, y_min = datmin[:inD], datmin[inD:]
x_max, y_max = datmax[:inD], datmax[inD:]
# Memory cleanup -- no longer need min/max arrays
del datmin, datmax
# Normalize min/max values
if normalize:
x_min = U.normalize(x_min, x_mean, x_std)
x_max = U.normalize(x_max, x_mean, x_std)
y_min = U.normalize(y_min, y_mean, y_std)
y_max = U.normalize(y_max, y_mean, y_std)
else:
x_min = 0.
x_max = 1.
y_min = 0.
y_max = 1.
scalelims = [0., 1.]
# Get TFRecord file names
print('\nLoading TFRecords file names...')
TFRpath = inputdir + 'TFRecords' + os.sep + TFRfile
ftrain_TFR = glob.glob(TFRpath + 'train*.tfrecords')
fvalid_TFR = glob.glob(TFRpath + 'valid*.tfrecords')
ftest_TFR = glob.glob(TFRpath + 'test*.tfrecords')
if len(ftrain_TFR) == 0 or len(fvalid_TFR) == 0 or len(ftest_TFR) == 0:
# Doesn't exist -- make them
print("\nSome TFRecords files do not exist yet.")
ftrain = glob.glob(datadir + 'train' + os.sep + '*.npy')
fvalid = glob.glob(datadir + 'valid' + os.sep + '*.npy')
ftest = glob.glob(datadir + 'test' + os.sep + '*.npy')
if len(ftrain_TFR) == 0:
print("Making TFRecords for training data...")
U.make_TFRecord(
inputdir + 'TFRecords' + os.sep + TFRfile + 'train.tfrecords',
ftrain, inD, ilog, olog, batch_size, train_batches)
if len(fvalid_TFR) == 0:
print("\nMaking TFRecords for validation data...")
U.make_TFRecord(
inputdir + 'TFRecords' + os.sep + TFRfile + 'valid.tfrecords',
fvalid, inD, ilog, olog, batch_size, valid_batches)
if len(ftest_TFR) == 0:
print("\nMaking TFRecords for test data...")
U.make_TFRecord(
inputdir + 'TFRecords' + os.sep + TFRfile + 'test.tfrecords',
ftest, inD, ilog, olog, batch_size, test_batches)
print("\nTFRecords creation complete.")
# Free memory
del ftrain, fvalid, ftest
# Get TFR file names for real this time
ftrain_TFR = glob.glob(TFRpath + 'train*.tfrecords')
fvalid_TFR = glob.glob(TFRpath + 'valid*.tfrecords')
ftest_TFR = glob.glob(TFRpath + 'test*.tfrecords')
# Load the xvals
if fxvals is not None:
xvals = np.load(fxvals)
else:
xvals = None
# Perform grid search
if gridsearch:
# Train a model for each architecture, w/ unique directories
print("\nPerforming a grid search.\n")
maxlen = 0
for i, arch in enumerate(architectures):
if len(arch) > maxlen:
maxlen = len(arch)
archdir = os.path.join(outputdir, arch, '')
wsplit = weight_file.rsplit(os.sep, 1)[1].rsplit('.', 1)
wfile = ''.join([archdir, wsplit[0], '_', arch, '.', wsplit[1]])
U.make_dir(archdir)
nn = NNModel(ftrain_TFR,
fvalid_TFR,
ftest_TFR,
inD,
outD,
olog,
x_mean,
x_std,
y_mean,
y_std,
x_min,
x_max,
y_min,
y_max,
scalelims,
ncores,
buffer_size,
batch_size,
[train_batches, valid_batches, test_batches],
layers[i],
lay_params[i],
activations[i],
act_params[i],
nodes[i],
lengthscale,
max_lr,
clr_mode,
clr_steps,
wfile,
stop_file='./STOP',
resume=resume,
train_flag=True,
shuffle=True)
nn.train(train_batches, valid_batches, epochs, patience)
P.loss(nn, archdir)
# Print/save out the minmium validation loss for each architecture
minvl = np.ones(len(architectures)) * np.inf
print('Grid search summary')
print('-------------------')
with open(outputdir + 'gridsearch.txt', 'w') as foo:
foo.write('Grid search summary\n')
foo.write('-------------------\n')
for i, arch in enumerate(architectures):
archdir = os.path.join(outputdir, arch, '')
history = np.load(archdir + 'history.npz')
minvl[i] = np.amin(history['val_loss'])
print(arch.ljust(maxlen, ' ') + ': ' + str(minvl[i]))
with open(outputdir + 'gridsearch.txt', 'a') as foo:
foo.write(arch.ljust(maxlen, ' ') + ': ' \
+ str(minvl[i]) + '\n')
return
# Train a model
if trainflag:
print('\nBeginning model training.\n')
nn = NNModel(ftrain_TFR,
fvalid_TFR,
ftest_TFR,
inD,
outD,
olog,
x_mean,
x_std,
y_mean,
y_std,
x_min,
x_max,
y_min,
y_max,
scalelims,
ncores,
buffer_size,
batch_size, [train_batches, valid_batches, test_batches],
layers,
lay_params,
activations,
act_params,
nodes,
lengthscale,
max_lr,
clr_mode,
clr_steps,
weight_file,
stop_file='./STOP',
train_flag=True,
shuffle=True,
resume=resume)
nn.train(train_batches, valid_batches, epochs, patience)
# Plot the loss
P.loss(nn, plotdir)
# Call new model with shuffle=False
nn = NNModel(ftrain_TFR,
fvalid_TFR,
ftest_TFR,
inD,
outD,
olog,
x_mean,
x_std,
y_mean,
y_std,
x_min,
x_max,
y_min,
y_max,
scalelims,
ncores,
buffer_size,
batch_size, [train_batches, valid_batches, test_batches],
layers,
lay_params,
activations,
act_params,
nodes,
lengthscale,
max_lr,
clr_mode,
clr_steps,
weight_file,
stop_file='./STOP',
train_flag=False,
shuffle=False,
resume=False)
if '.h5' in weight_file or '.hdf5' in weight_file:
nn.model.load_weights(weight_file) # Load the model
# Save in ONNX format
try:
onnx_model = keras2onnx.convert_keras(nn.model)
onnx.save_model(onnx_model,
nn.weight_file.rsplit('.', 1)[0] + '.onnx')
except Exception as e:
print("Unable to convert the Keras model to ONNX:")
print(e)
else:
nn.model = onnx_to_keras(onnx.load_model(weight_file), ['input_1'])
# Validate model
if (validflag or trainflag) and not rng_test:
print('\nValidating the model...\n')
# Y values
print(' Predicting...')
fvalpred = nn.Yeval('pred',
'valid',
preddir,
denorm=(normalize == False and scale == False))
fvalpred = glob.glob(fvalpred + '*')
print(' Loading the true Y values...')
fvaltrue = nn.Yeval('true',
'valid',
preddir,
denorm=(normalize == False and scale == False))
fvaltrue = glob.glob(fvaltrue + '*')
### RMSE & R2
print('\n Calculating RMSE & R2...')
if not normalize and not scale:
val_stats = S.rmse_r2(fvalpred,
fvaltrue,
y_mean,
olog=olog,
y_mean_delog=y_mean_delog,
x_vals=xvals,
filters=filters,
filtconv=filtconv)
else:
val_stats = S.rmse_r2(fvalpred, fvaltrue, y_mean, y_std, y_min,
y_max, scalelims, olog, y_mean_delog, xvals,
filters, filtconv)
# RMSE
if np.any(val_stats[0] != -1) and np.any(val_stats[1] != -1):
print(' Normalized RMSE : ', val_stats[0])
print(' Mean normalized RMSE : ', np.mean(val_stats[0]))
print(' Denormalized RMSE : ', val_stats[1])
print(' Mean denormalized RMSE: ', np.mean(val_stats[1]))
np.savez(outputdir + rmse_file + '_val_norm.npz',
rmse=val_stats[0],
rmse_mean=np.mean(val_stats[0]))
saveRMSEnorm = True
saveRMSEdenorm = True
elif np.any(val_stats[0] != -1):
print(' RMSE : ', val_stats[0])
print(' Mean RMSE: ', np.mean(val_stats[0]))
saveRMSEnorm = True
saveRMSEdenorm = False
elif np.any(val_stats[1] != -1):
print(' RMSE : ', val_stats[1])
print(' Mean RMSE: ', np.mean(val_stats[1]))
saveRMSEnorm = False
saveRMSEdenorm = True
else:
print(" No files passed in to compute RMSE.")
saveRMSEnorm = False
saveRMSEdenorm = False
if saveRMSEnorm:
P.plot(''.join([plotdir, rmse_file, '_val_norm.png']), xvals,
val_stats[0], xlabel, 'RMSE')
np.savez(outputdir + rmse_file + '_val_norm.npz',
rmse=val_stats[0],
rmse_mean=np.mean(val_stats[0]))
if saveRMSEdenorm:
P.plot(''.join([plotdir, rmse_file, '_val_denorm.png']), xvals,
val_stats[1], xlabel, 'RMSE')
np.savez(outputdir + rmse_file + '_val_denorm.npz',
rmse=val_stats[1],
rmse_mean=np.mean(val_stats[1]))
# R2
if np.any(val_stats[2] != -1) and np.any(val_stats[3] != -1):
print(' Normalized R2 : ', val_stats[2])
print(' Mean normalized R2 : ', np.mean(val_stats[2]))
print(' Denormalized R2 : ', val_stats[3])
print(' Mean denormalized R2: ', np.mean(val_stats[3]))
saveR2norm = True
saveR2denorm = True
elif np.any(val_stats[2] != -1):
print(' R2 : ', val_stats[2])
print(' Mean R2: ', np.mean(val_stats[2]))
saveR2norm = True
saveR2denorm = False
elif np.any(val_stats[3] != -1):
print(' R2 : ', val_stats[3])
print(' Mean R2: ', np.mean(val_stats[3]))
saveR2norm = False
saveR2denorm = True
else:
print(" No files passed in to compute R2.")
saveR2norm = False
saveR2denorm = False
if saveR2norm:
P.plot(''.join([plotdir, r2_file, '_val_norm.png']), xvals,
val_stats[2], xlabel, '$R^2$')
np.savez(outputdir + r2_file + '_val_norm.npz',
r2=val_stats[2],
r2_mean=np.mean(val_stats[2]))
if saveR2denorm:
P.plot(''.join([plotdir, r2_file, '_val_denorm.png']), xvals,
val_stats[3], xlabel, '$R^2$')
np.savez(outputdir + r2_file + '_val_denorm.npz',
r2=val_stats[3],
r2_mean=np.mean(val_stats[3]))
# Evaluate model on test set
if testflag and not rng_test:
print('\nTesting the model...\n')
# Y values
print(' Predicting...')
ftestpred = nn.Yeval('pred',
'test',
preddir,
denorm=(normalize == False and scale == False))
ftestpred = glob.glob(ftestpred + '*')
print(' Loading the true Y values...')
ftesttrue = nn.Yeval('true',
'test',
preddir,
denorm=(normalize == False and scale == False))
ftesttrue = glob.glob(ftesttrue + '*')
### RMSE & R2
print('\n Calculating RMSE & R2...')
if not normalize and not scale:
test_stats = S.rmse_r2(ftestpred,
ftesttrue,
y_mean,
olog=olog,
y_mean_delog=y_mean_delog,
x_vals=xvals,
filters=filters,
filtconv=filtconv)
else:
test_stats = S.rmse_r2(ftestpred, ftesttrue, y_mean, y_std, y_min,
y_max, scalelims, olog, y_mean_delog, xvals,
filters, filtconv)
# RMSE
if np.any(test_stats[0] != -1) and np.any(test_stats[1] != -1):
print(' Normalized RMSE : ', test_stats[0])
print(' Mean normalized RMSE : ', np.mean(test_stats[0]))
print(' Denormalized RMSE : ', test_stats[1])
print(' Mean denormalized RMSE: ', np.mean(test_stats[1]))
np.savez(outputdir + rmse_file + '_val_norm.npz',
rmse=test_stats[0],
rmse_mean=np.mean(test_stats[0]))
saveRMSEnorm = True
saveRMSEdenorm = True
elif np.any(test_stats[0] != -1):
print(' RMSE : ', test_stats[0])
print(' Mean RMSE: ', np.mean(test_stats[0]))
saveRMSEnorm = True
saveRMSEdenorm = False
elif np.any(test_stats[1] != -1):
print(' RMSE : ', test_stats[1])
print(' Mean RMSE: ', np.mean(test_stats[1]))
saveRMSEnorm = False
saveRMSEdenorm = True
else:
print(" No files passed in to compute RMSE.")
saveRMSEnorm = False
saveRMSEdenorm = False
if saveRMSEnorm:
P.plot(''.join([plotdir, rmse_file, '_test_norm.png']), xvals,
test_stats[0], xlabel, 'RMSE')
np.savez(outputdir + rmse_file + '_test_norm.npz',
rmse=test_stats[0],
rmse_mean=np.mean(test_stats[0]))
if saveRMSEdenorm:
P.plot(''.join([plotdir, rmse_file, '_test_denorm.png']), xvals,
test_stats[1], xlabel, 'RMSE')
np.savez(outputdir + rmse_file + '_test_denorm.npz',
rmse=test_stats[1],
rmse_mean=np.mean(test_stats[1]))
# R2
if np.any(test_stats[2] != -1) and np.any(test_stats[3] != -1):
print(' Normalized R2 : ', test_stats[2])
print(' Mean normalized R2 : ', np.mean(test_stats[2]))
print(' Denormalized R2 : ', test_stats[3])
print(' Mean denormalized R2: ', np.mean(test_stats[3]))
saveR2norm = True
saveR2denorm = True
elif np.any(test_stats[2] != -1):
print(' R2 : ', test_stats[2])
print(' Mean R2: ', np.mean(test_stats[2]))
saveR2norm = True
saveR2denorm = False
elif np.any(test_stats[3] != -1):
print(' R2 : ', test_stats[3])
print(' Mean R2: ', np.mean(test_stats[3]))
saveR2norm = False
saveR2denorm = True
else:
print(" No files passed in to compute R2.")
saveR2norm = False
saveR2denorm = False
if saveR2norm:
P.plot(''.join([plotdir, r2_file, '_test_norm.png']), xvals,
test_stats[2], xlabel, '$R^2$')
np.savez(outputdir + r2_file + '_test_norm.npz',
r2=test_stats[2],
r2_mean=np.mean(test_stats[2]))
if saveR2denorm:
P.plot(''.join([plotdir, r2_file, '_test_denorm.png']), xvals,
test_stats[3], xlabel, '$R^2$')
np.savez(outputdir + r2_file + '_test_denorm.npz',
r2=test_stats[3],
r2_mean=np.mean(test_stats[3]))
# Plot requested cases
if not rng_test:
predfoo = sorted(glob.glob(preddir + 'test' + os.sep + 'pred*'))
truefoo = sorted(glob.glob(preddir + 'test' + os.sep + 'true*'))
if len(predfoo) > 0 and len(truefoo) > 0:
print("\nPlotting the requested cases...")
nplot = 0
for v in plot_cases:
fname = plotdir + 'spec' + str(v) + '_pred-vs-true.png'
predspec = np.load(predfoo[v // batch_size])[v % batch_size]
predspec = U.denormalize(
U.descale(predspec, y_min, y_max, scalelims), y_mean,
y_std)
truespec = np.load(truefoo[v // batch_size])[v % batch_size]
truespec = U.denormalize(
U.descale(truespec, y_min, y_max, scalelims), y_mean,
y_std)
if olog:
predspec[olog] = 10**predspec[olog]
truespec[olog] = 10**truespec[olog]
P.plot_spec(fname, predspec, truespec, xvals, xlabel, ylabel)
nplot += 1
print(" Plot " + str(nplot) + "/" + str(len(plot_cases)),
end='\r')
print("")
else:
raise Exception("No predictions found in " + preddir + "test.")
return
model = LayerTest(kernel_size=kernel_size,
padding=padding,
stride=stride)
model.eval()
input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
input_var = Variable(torch.FloatTensor(input_np))
torch.onnx.export(model,
input_var,
"_tmpnet.onnx",
verbose=True,
input_names=['test_in'],
output_names=['test_out'])
onnx_model = onnx.load('_tmpnet.onnx')
k_model = onnx_to_keras(onnx_model, ['test_in'],
change_ordering=change_ordering)
error = check_torch_keras_error(
model,
k_model,
input_np,
change_ordering=change_ordering)
print('Error:', error)
if max_error < error:
max_error = error
print('Max error: {0}'.format(max_error))
### sudo pip3 install onnx2keras tf-nightly
### tf-nightly-2.2.0-dev20200502
import onnx
from onnx2keras import onnx_to_keras
import tensorflow as tf
import shutil
onnx_model = onnx.load('human-pose-estimation-3d-0001.onnx')
k_model = onnx_to_keras(onnx_model=onnx_model,
input_names=['data'],
change_ordering=True)
shutil.rmtree('saved_model', ignore_errors=True)
tf.saved_model.save(k_model, 'saved_model')
"""
$ saved_model_cli show --dir saved_model --all
MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:
signature_def['__saved_model_init_op']:
The given SavedModel SignatureDef contains the following input(s):
The given SavedModel SignatureDef contains the following output(s):
outputs['__saved_model_init_op'] tensor_info:
dtype: DT_INVALID
shape: unknown_rank
name: NoOp
Method name is:
signature_def['serving_default']:
The given SavedModel SignatureDef contains the following input(s):
FHardtanhTest,
LayerReLUTest,
FReLUTest,
LayerELUTest,
FPELUTest,
]:
for i in range(10):
model = act_type()
model.eval()
input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
input_var = torch.FloatTensor(input_np)
torch.onnx.export(model,
input_var,
"_tmpnet.onnx",
verbose=True,
input_names=['test_in'],
output_names=['test_out'])
onnx_model = onnx.load('_tmpnet.onnx')
k_model = onnx_to_keras(onnx_model, ['test_in'])
os.unlink('_tmpnet.onnx')
error = check_torch_keras_error(model, k_model, input_np)
print('Error:', error)
if max_error < error:
max_error = error
print('Max error: {0}'.format(max_error))
import numpy as np
import cv2
import onnx
import onnxmltools
import onnx2keras
from onnx2keras import onnx_to_keras
# convert to keras
onnx_path = './base-model-new-320.onnx'
keras_path = onnx_path.split('/')[-1].split('.')[0]
onnx_model = onnxmltools.utils.load_model(onnx_path)
keras_model = onnx_to_keras(onnx_model, ['input'])
keras_model.save(keras_path)
keras_model.summary()
def load(path, filename):
"""Load network from file.
Parameters
----------
path: str
Path to directory where to load model from.
filename: str
Name of file to load model from.
Returns
-------
: dict[str, Union[keras.models.Sequential, function]]
A dictionary of objects that constitute the input model. It must
contain the following two keys:
- 'model': keras.models.Sequential
Keras model instance of the network.
- 'val_fn': function
Function that allows evaluating the original model.
"""
filepath = str(os.path.join(path, filename))
# Load the Pytorch model.
mod = import_script(path, filename)
kwargs = mod.kwargs if hasattr(mod, 'kwargs') else {}
model_pytorch = mod.Model(**kwargs)
map_location = 'cpu' if not torch.cuda.is_available() else None
for ext in ['.pth', '.pkl']:
model_path = filepath + ext
if os.path.exists(model_path):
break
assert model_path, "Pytorch state_dict not found at {}".format(model_path)
model_pytorch.load_state_dict(
torch.load(model_path, map_location=map_location))
# Switch from train to eval mode to ensure Dropout / BatchNorm is handled
# correctly.
model_pytorch.eval()
# Run on dummy input with correct shape to trace the Pytorch model.
input_shape = [1] + list(model_pytorch.input_shape)
input_numpy = np.random.random_sample(input_shape).astype(np.float32)
input_torch = torch.from_numpy(input_numpy).float()
output_torch = model_pytorch(input_torch)
output_numpy = to_numpy(output_torch)
# Export as onnx model, and then reload.
input_names = ['input_0']
output_names = ['output_{}'.format(i) for i in range(len(output_torch))]
dynamic_axes = {'input_0': {0: 'batch_size'}}
dynamic_axes.update({name: {0: 'batch_size'} for name in output_names})
torch.onnx.export(model_pytorch,
input_torch,
filepath + '.onnx',
input_names=input_names,
output_names=output_names,
dynamic_axes=dynamic_axes)
model_onnx = onnx.load(filepath + '.onnx')
# onnx.checker.check_model(model_onnx) # Crashes with segmentation fault.
# Compute ONNX Runtime output prediction.
ort_session = onnxruntime.InferenceSession(filepath + '.onnx')
input_onnx = {ort_session.get_inputs()[0].name: input_numpy}
output_onnx = ort_session.run(None, input_onnx)
# Compare ONNX Runtime and PyTorch results.
err_msg = "Pytorch model could not be ported to ONNX. Output difference: "
np.testing.assert_allclose(output_numpy,
output_onnx[0],
rtol=1e-03,
atol=1e-05,
err_msg=err_msg)
print("Pytorch model was successfully ported to ONNX.")
change_ordering = keras.backend.image_data_format() == 'channels_last'
if change_ordering:
input_numpy = np.moveaxis(input_numpy, 1, -1)
output_numpy = np.moveaxis(output_numpy, 1, -1)
# Import this here; import changes image_data_format to channels_first.
from onnx2keras import onnx_to_keras
# Port ONNX model to Keras.
model_keras = onnx_to_keras(model_onnx,
input_names, [input_shape[1:]],
change_ordering=change_ordering,
verbose=False)
if change_ordering:
keras.backend.set_image_data_format('channels_last')
# Save the keras model.
keras.models.save_model(model_keras, filepath + '.h5')
# Loading the model here is a workaround for version conflicts with
# TF > 2.0.1 and keras > 2.2.5. Should be able to remove this later.
model_keras = keras.models.load_model(filepath + '.h5')
model_keras.compile('sgd', 'categorical_crossentropy',
['accuracy', keras.metrics.top_k_categorical_accuracy])
# Compute Keras output and compare against ONNX.
output_keras = model_keras.predict(input_numpy)
err_msg = "ONNX model could not be ported to Keras. Output difference: "
np.testing.assert_allclose(output_numpy,
output_keras,
rtol=1e-03,
atol=1e-05,
err_msg=err_msg)
print("ONNX model was successfully ported to Keras.")
return {'model': model_keras, 'val_fn': model_keras.evaluate}
import warnings
from onnx_tf.backend import prepare
import onnx
import tensorflow as tf
import onnx2keras
import numpy as np
warnings.filterwarnings('ignore')
#sess = ort.InferenceSession('model.onnx')
model = onnx.load_model("model.onnx")
k_model = onnx2keras.onnx_to_keras(onnx_model=model,
input_names=["input_data"],
change_ordering=True,
verbose=False)
print(k_model.layers)
import torch
from torch.autograd import Variable
from onnx2keras import onnx_to_keras, check_torch_keras_error
import onnx
from torchvision.models.densenet import densenet121
if __name__ == '__main__':
model = densenet121()
model.eval()
input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
input_var = Variable(torch.FloatTensor(input_np), requires_grad=False)
output = model(input_var)
torch.onnx.export(model, (input_var),
"_tmpnet.onnx",
verbose=True,
input_names=['test_in1'],
output_names=['test_out'])
onnx_model = onnx.load('_tmpnet.onnx')
k_model = onnx_to_keras(onnx_model, ['test_in1', 'test_in2'],
change_ordering=True)
error = check_torch_keras_error(model,
k_model,
input_np,
change_ordering=True)
print('Max error: {0}'.format(error))
import onnx
from onnx2keras import onnx_to_keras
import tensorflow as tf
# Converts ONNX networks to keras networks
# Currently only for the actor network
# Load ONNX model
onnx_actor = onnx.load('actor.onnx')
# Replace incompatible placeholder node from Matlab
onnx_actor.graph.node[5].op_type = 'Tanh'
print('Loaded and replaced')
# Convert model to keras
actor_onnx = onnx_to_keras(onnx_actor, ['observation'])
# Rebuild model in tensorflow to get the right layers
actor_tf = tf.keras.Sequential()
actor_tf.add(tf.keras.Input(shape=(26, )))
actor_tf.add(tf.keras.layers.Dense(400))
actor_tf.add(tf.keras.layers.Dense(300, activation='relu'))
actor_tf.add(tf.keras.layers.Dense(8))
actor_tf.add(tf.keras.layers.Activation('tanh'))
# actor_tf.add(tf.keras.layers.Flatten())
# Apply weights and biases
weight_0 = actor_onnx.layers[1].get_weights()
actor_tf.layers[0].set_weights(
[tf.reshape(weight_0[0], (26, 400)), weight_0[1]])
weight_1 = actor_onnx.layers[3].get_weights()
actor_tf.layers[1].set_weights(
[tf.reshape(weight_1[0], (400, 300)), weight_1[1]])
import onnx2keras from onnx2keras import onnx_to_keras import keras import onnx path = "C:/Users/msa/Documents/datasets/pretrained architectures/PointPillars/pfe.onnx" onnx_model = onnx.load(path) k_model = onnx_to_keras(onnx_model) print(k_model)
# Versions of tensorflow used -------------------------
print("tf.__version__ is", tf.__version__)
print("tf.keras.__version__ is:", tf.keras.__version__)
print("Num GPUs Available: ",
len(tf.config.experimental.list_physical_devices('GPU')))
# Versions of tensorflow used -------------------------
# transform the given images --------------------------
PIL_transform = transforms.ToPILImage()
tensor_transform = transforms.ToTensor()
# transform the given images --------------------------
# Loading the onnx model ------------------------------
onnx_model = onnx.load("Material_Classifier.onnx")
k_model = onnx_to_keras(onnx_model, ['imageinput'])
keras.models.save_model(k_model,
"Material_Classifier.h5",
overwrite=True,
include_optimizer=True)
# Loading the onnx model ------------------------------
# Folder and image size initialization ----------------
IMG_SIZE = 200
CARBON = "CARBON/"
FIBERGLASS = "FIBERGLASS/"
LABELS = {CARBON: 0, FIBERGLASS: 1}
# Folder and image size initialization ----------------
# Counts and training data ----------------------------
training_data = []
return error
if __name__ == '__main__':
max_error = 0
for i in range(10):
model = FTest()
model.eval()
input_np1 = np.random.uniform(0, 1, (1, 3, 224, 224))
input_np2 = np.random.uniform(0, 1, (1, 3, 224, 224))
input_var1 = Variable(torch.FloatTensor(input_np1))
input_var2 = Variable(torch.FloatTensor(input_np2))
output = model(input_var1, input_var2)
torch.onnx.export(model, (input_var1, input_var2),
"_tmpnet.onnx",
verbose=True,
input_names=['test_in1', 'test_in2'],
output_names=['test_out'])
onnx_model = onnx.load('_tmpnet.onnx')
k_model = onnx_to_keras(onnx_model, ['test_in1', 'test_in2'])
error = check_error(output, k_model, [input_np1, input_np2])
if max_error < error:
max_error = error
print('Max error: {0}'.format(max_error))
input_names=["input"],
output_names=["output"],
dynamic_axes={
"input": {
0: "batch_size"
},
"output": {
0: "batch_size"
}
})
onnx_model = onnx.load("./temp.onnx")
onnx.checker.check_model(onnx_model)
inpt = ['input']
keras_model = onnx_to_keras(onnx_model=onnx_model,
input_names=inpt,
change_ordering=True,
verbose=False)
converter = tf.lite.TFLiteConverter.from_keras_model(keras_model)
converter.optimizations = [tf.lite.Optimize.OPTIMIZE_FOR_SIZE]
tflite_model = converter.convert()
open('modelzoo/model_' + str(i) + '.tflite', "wb").write(tflite_model)
data = ''
for itr in netEmbedding:
for itr2 in itr:
data = data + str(itr2) + ','
data = data[:-1]
data = data + '\n'
file.write(data)
file.close()
# print('==> Resuming from checkpoint..')
# assert os.path.isfile(resume), 'Error: no checkpoint directory found!'
# checkpoint = os.path.dirname(resume)
# checkpoint = torch.load(resume)
# model.load_state_dict(checkpoint['state_dict'])
# dummy_input = Variable(torch.randn(1, 3, 32, 32)) # nchw
# # dummy_output = model.module(dummy_input)
# # print(dummy_output)
# onnx_filename = "model.onnx"
# torch.onnx.export(model.module, dummy_input, onnx_filename, output_names=['test_output'])
model_onnx = onnx.load('./model.onnx')
k_model = onnx_to_keras(model_onnx, input_names=['0'])
# Export model as .pb file
# tf_rep.export_graph('model_tf.h5')
# def load_pb(path_to_pb):
# with tf.gfile.GFile(path_to_pb, 'rb') as f:
# graph_def = tf.GraphDef()
# graph_def.ParseFromString(f.read())
# with tf.Graph().as_default() as graph:
# tf.import_graph_def(graph_def, name='')
# return graph
# tf_graph = load_pb('model_tf.pb')
# sess = tf.Session(graph=tf_graph)
!pip install onnx
!git clone https://github.com/nerox8664/onnx2keras.git
'''
#%cd onnx2keras
import tensorflow as tf
import onnx
from onnx2keras import onnx_to_keras
from tensorflow.keras.models import load_model
# Load ONNX model
onnx_model = onnx.load('/content/pnet_video.onnx')
# Call the converter and save keras model
k_model = onnx_to_keras(onnx_model, ['input.1'],change_ordering=True)
k_model.save('/content/pnet_video.h5')
from tensorflow.keras.models import Model
from tensorflow.keras.models import load_model
from tensorflow.keras.layers import Activation, Lambda, Reshape
# Load keras model
k_model=load_model('/content/pnet_video.h5')
k_model.summary()
# Remove edge branch from output
edge_model=Model(inputs=k_model.input,outputs=k_model.layers[-2].output)
edge_model.summary()
# Add softmax on output
args = parser.parse_args()
print('Loading data...')
(x_train, y_train), (x_test,
y_test) = imdb.load_data(num_words=args.vocab_size,
maxlen=args.max_len)
print(len(x_train), 'train sequences')
print(len(x_test), 'test sequences')
print('Pad sequences (samples x time)')
x_train = sequence.pad_sequences(x_train, maxlen=args.max_len)
x_test = sequence.pad_sequences(x_test, maxlen=args.max_len)
print('x_train shape:', x_train.shape)
print('x_test shape:', x_test.shape)
print('Load ONNX model...')
onnx_model = onnx.load(args.model_path)
#onnx.checker.check_model(onnx_model)
print('Convert ONNX to Keras...')
k_model = onnx_to_keras(onnx_model, ['embedding_input'])
model.compile(loss='binary_crossentropy',
optimizer='adam',
metrics=['accuracy'])
loss, acc = model.evaluate(x_test, y_test, batch_size=args.batch_size)
print("Evaluation result: Loss:", loss, " Accuracy:", acc)
def pytorch_to_keras(model,
args,
input_shapes=None,
change_ordering=False,
verbose=False,
name_policy=None,
use_optimizer=False,
do_constant_folding=False):
"""
By given PyTorch model convert layers with ONNX.
Args:
model: pytorch model
args: pytorch model arguments
input_shapes: keras input shapes (using for each InputLayer)
change_ordering: change CHW to HWC
verbose: verbose output
name_policy: use short names, use random-suffix or keep original names for keras layers
Returns:
model: created keras model.
"""
logger = logging.getLogger('pytorch2keras')
if verbose:
logging.basicConfig(level=logging.DEBUG)
logger.info('Converter is called.')
if name_policy:
logger.warning('Name policy isn\'t supported now.')
if input_shapes:
logger.warning('Custom shapes isn\'t supported now.')
if input_shapes and not isinstance(input_shapes, list):
input_shapes = [input_shapes]
if not isinstance(args, list):
args = [args]
args = tuple(args)
dummy_output = model(*args)
if isinstance(dummy_output, torch.autograd.Variable):
dummy_output = [dummy_output]
input_names = ['input_{0}'.format(i) for i in range(len(args))]
output_names = ['output_{0}'.format(i) for i in range(len(dummy_output))]
logger.debug('Input_names:')
logger.debug(input_names)
logger.debug('Output_names:')
logger.debug(output_names)
stream = io.BytesIO()
torch.onnx.export(model,
args,
stream,
do_constant_folding=do_constant_folding,
verbose=verbose,
input_names=input_names,
output_names=output_names)
stream.seek(0)
onnx_model = onnx.load(stream)
if use_optimizer:
if use_optimizer is True:
optimizer2run = optimizer.get_available_passes()
else:
use_optimizer = set(use_optimizer)
optimizer2run = [
x for x in optimizer.get_available_passes()
if x in use_optimizer
]
logger.info("Running optimizer:\n%s", "\n".join(optimizer2run))
onnx_model = optimizer.optimize(onnx_model, optimizer2run)
k_model = onnx_to_keras(onnx_model=onnx_model,
input_names=input_names,
input_shapes=input_shapes,
name_policy=name_policy,
verbose=verbose,
change_ordering=change_ordering)
return k_model
### tensorflow-gpu==2.2.0
### onnx==1.7.0
### onnx2keras==0.0.21
### https://github.com/amir-abdi/keras_to_tensorflow.git
import onnx
from onnx2keras import onnx_to_keras
import tensorflow as tf
import shutil
from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2
onnx_model = onnx.load('dbface.onnx')
k_model = onnx_to_keras(onnx_model=onnx_model, input_names=['x'])
shutil.rmtree('saved_model', ignore_errors=True)
tf.saved_model.save(k_model, 'saved_model')
# Convert Keras model to ConcreteFunction
full_model = tf.function(lambda x: k_model(x))
full_model = full_model.get_concrete_function(
tf.TensorSpec(k_model.inputs[0].shape, k_model.inputs[0].dtype))
# Get frozen ConcreteFunction
frozen_func = convert_variables_to_constants_v2(full_model)
frozen_func.graph.as_graph_def()
tf.io.write_graph(graph_or_graph_def=frozen_func.graph,
logdir=".",
name="dbface.pb",
as_text=False)
