################################################################################
# Copyright (c) 2021 ContinualAI. #
# Copyrights licensed under the MIT License. #
# See the accompanying LICENSE file for terms. #
# #
# Date: 1-05-2020 #
# Author(s): Vincenzo Lomonaco #
# E-mail: contact@continualai.org #
# Website: avalanche.continualai.org #
################################################################################
"""
This is the definition od the Mid-caffenet high resolution in Pythorch
"""
import torch.nn as nn
import torch
from pytorchcv.models.mobilenet import mobilenet_w1
try:
from pytorchcv.models.mobilenet import DwsConvBlock
except Exception:
from pytorchcv.models.common import DwsConvBlock
def remove_sequential(network, all_layers):
for layer in network.children():
# if sequential layer, apply recursively to layers in sequential layer
if isinstance(layer, nn.Sequential):
# print(layer)
remove_sequential(layer, all_layers)
else: # if leaf node, add it to list
# print(layer)
all_layers.append(layer)
def remove_DwsConvBlock(cur_layers):
all_layers = []
for layer in cur_layers:
if isinstance(layer, DwsConvBlock):
# print("helloooo: ", layer)
for ch in layer.children():
all_layers.append(ch)
else:
all_layers.append(layer)
return all_layers
[docs]class MobilenetV1(nn.Module):
"""MobileNet v1 implementation. This model
can be instantiated from a pretrained network."""
[docs] def __init__(self, pretrained=True, latent_layer_num=20):
super().__init__()
model = mobilenet_w1(pretrained=pretrained)
model.features.final_pool = nn.AvgPool2d(4)
all_layers = []
remove_sequential(model, all_layers)
all_layers = remove_DwsConvBlock(all_layers)
lat_list = []
end_list = []
for i, layer in enumerate(all_layers[:-1]):
if i <= latent_layer_num:
lat_list.append(layer)
else:
end_list.append(layer)
self.lat_features = nn.Sequential(*lat_list)
self.end_features = nn.Sequential(*end_list)
self.output = nn.Linear(1024, 50, bias=False)
def forward(self, x, latent_input=None, return_lat_acts=False):
if latent_input is not None:
with torch.no_grad():
orig_acts = self.lat_features(x)
lat_acts = torch.cat((orig_acts, latent_input), 0)
else:
orig_acts = self.lat_features(x)
lat_acts = orig_acts
x = self.end_features(lat_acts)
x = x.view(x.size(0), -1)
logits = self.output(x)
if return_lat_acts:
return logits, orig_acts
else:
return logits
if __name__ == "__main__":
model = MobilenetV1(pretrained=True)
for name, param in model.named_parameters():
print(name)