Module
sym, arg_params, aux_params = mx.model.load_checkpoint(model_prefix, 0)
mod = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
mod.bind(for_training=False, data_shapes=[('data', (1,3,224,224))], label_shapes=mod._label_shapes)
mod.set_params(arg_params, aux_params, allow_missing=True)Gluon
net = gluon.nn.SymbolBlock.imports('resnet18-symbol.json', ['data'], param_file='resnet18-0000.params', ctx=mx.gpu())https://discuss.mxnet.io/t/load-params-from-symbol-models-to-gluon/804
Module
sym, arg_params, aux_params = mx.model.load_checkpoint(model_prefix, 3)
mod.fit(train_data=train_dataiter, eval_data=val_dataiter,
optimizer='sgd',
optimizer_params={'learning_rate':0.01, 'momentum': 0.9},
arg_params=arg_params, aux_params=aux_params,
eval_metric='acc', num_epoch=10, begin_epoch=3)Gluon
Module
# construct a callback function to save checkpoints
model_prefix = 'mymodel'
checkpoint = mx.callback.do_checkpoint(model_prefix)
mod = mx.mod.Module(symbol=net)
mod.fit(train_iter, num_epoch=5, epoch_end_callback=checkpoint)Gluon
for epoch in epochs:
# Train here...
net.export(f"{model_name}", epoch) # Saves sym and params files
# or
net.save_parameters(f"{model_name}.params") # Saves params only, model defined in code.Module
sym, arg_params, aux_params = mx.model.load_checkpoint(model_path, epoch=epoch)
all_layers = sym.get_internals()
sym = all_layers['fc1_output']
model = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
model.bind(for_training=False, data_shapes=[('data', (1, 3, self.image_size, self.image_size))])
model.set_params(arg_params, aux_params)Gluon
net = gluoncv.model_zoo.get_model("ResNet50V2", pretrained=True)
all_params = net.collect_params()
print(net)
# Get one layer by using the index name/number from the output above
W1 = net.features[6][0].conv1.weight.data() # returns NDArray
W1params = all_params.get("resnetv20_stage2_conv1_weight") # Params from same layer as above.Combine data and labels in a Dataset object
mx.random.seed(42) # Fix the seed for reproducibility
X = mx.random.uniform(shape=(10, 3))
y = mx.random.uniform(shape=(10, 1))
dataset = mxnet.gluon.data.ArrayDataset(X, y)Randomly reading mini-batches
data_iter = mxnet.gluon.data.DataLoader(dataset, batch_size, shuffle=True)
# or with transformer
data_iter = mxnet.gluon.data.DataLoader(dataset.transform_first(transformer), batch_size, shuffle=True)Converts an image NDArray of shape
transformer = mxnet.gluon.data.vision.transform.ToTensor()Gluon
You can set grad_req attribute to 'null' (it is a string) to prevent changes of this parameter. Here is the example. I define a set of parameter names I want to freeze and freeze them after creating my model, but before the initialization.
layers_to_freeze = set(['dense0_weight', 'dense0_bias', 'dense1_weight', 'dense1_bias'])
for p in net.collect_params().values():
if p.name in layers_to_freeze:
p.grad_req = 'null'or
for param in net[24:].collect_params().values():
param.grad_req = 'write'
https://discuss.mxnet.io/t/gluon-access-layer-weights/1160/2
https://discuss.mxnet.io/t/layer-access-in-a-pre-trained-model/2248
https://mxnet.apache.org/api/python/docs/tutorials/packages/gluon/blocks/parameters.html
Gluon
Given VGG16:
>>> net = model_zoo.get_model("VGG16", pretrained=True)
>>> net
Out[50]:
VGG(
(features): HybridSequential(
(0): Conv2D(3 -> 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): Activation(relu)
(2): Conv2D(64 -> 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): Activation(relu)
(4): MaxPool2D(size=(2, 2), stride=(2, 2), padding=(0, 0), ceil_mode=False, global_pool=False, pool_type=max, layout=NCHW)
(5): Conv2D(64 -> 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(6): Activation(relu)
(7): Conv2D(128 -> 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(8): Activation(relu)
(9): MaxPool2D(size=(2, 2), stride=(2, 2), padding=(0, 0), ceil_mode=False, global_pool=False, pool_type=max, layout=NCHW)
(10): Conv2D(128 -> 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(11): Activation(relu)
(12): Conv2D(256 -> 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(13): Activation(relu)
(14): Conv2D(256 -> 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(15): Activation(relu)
(16): MaxPool2D(size=(2, 2), stride=(2, 2), padding=(0, 0), ceil_mode=False, global_pool=False, pool_type=max, layout=NCHW)
(17): Conv2D(256 -> 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(18): Activation(relu)
(19): Conv2D(512 -> 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(20): Activation(relu)
(21): Conv2D(512 -> 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(22): Activation(relu)
(23): MaxPool2D(size=(2, 2), stride=(2, 2), padding=(0, 0), ceil_mode=False, global_pool=False, pool_type=max, layout=NCHW)
(24): Conv2D(512 -> 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(25): Activation(relu)
(26): Conv2D(512 -> 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(27): Activation(relu)
(28): Conv2D(512 -> 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(29): Activation(relu)
(30): MaxPool2D(size=(2, 2), stride=(2, 2), padding=(0, 0), ceil_mode=False, global_pool=False, pool_type=max, layout=NCHW)
(31): Dense(25088 -> 4096, Activation(relu))
(32): Dropout(p = 0.5, axes=())
(33): Dense(4096 -> 4096, Activation(relu))
(34): Dropout(p = 0.5, axes=())
)
(output): Dense(4096 -> 1000, linear)
)Get the weight of the first convolutional layer:
W0 = net.features[0].weight.data()or
W0 = net.collect_params().get("conv0_weight").data()Note: it can be that the listed key contains a prefix. E.g. net.collect_params().keys() gives ['vgg0_conv0_weight', 'vgg0_conv0_bias', ...]. When quering the value by key with the get() function, you have to remove the prefix first.
or
for p in pretrained_net.collect_params().values():
print(p.name, p.data())