mediasynthesismuseum/deepdream-googlenet

DeepDream GoogLeNet weights (PyTorch)

PyTorch-loadable weights for the original DeepDream GoogLeNet models, hosted for mediasynthesismuseum/deepdream. These are the classic Caffe models converted to PyTorch (via ProGamerGov), preserving the original caffe layer names (inception_4c_output, inception_3b_5x5_reduce, …) used in Alexander Mordvintsev's 2015 dream.ipynb.

File	Model	Trained on	Dream aesthetic
bvlc_googlenet.pth	BVLC GoogLeNet	ImageNet	the classic dogs / eyes / pagodas ("puppy-slug")
googlenet_places365.pth	GoogLeNet-Places	Places365	buildings, domes, temples, landscapes

Preprocessing (both models)

Caffe-style: BGR channel order, subtract mean [104, 116, 122], input range 0–255 (no /255 scaling). Feature maps are read from the inception_*_output modules.

Usage

import torch
from huggingface_hub import hf_hub_download
from dd_models import BVLC_GOOGLENET   # model definitions included in this repo

net = BVLC_GOOGLENET(); net.add_layers()
sd = torch.load(hf_hub_download("mediasynthesismuseum/deepdream-googlenet", "bvlc_googlenet.pth"),
                map_location="cpu", weights_only=False)
net.load_state_dict({k: v for k, v in sd.items() if k in net.state_dict()
                     and v.shape == net.state_dict()[k].shape}, strict=False)
net.eval()
# hook net.inception_4c_output, then gradient-ascend its L2 norm across octaves.

Model definitions (bvlc_googlenet.py, googlenetplaces.py, helper_layers.py) are included in this repo. Original credit: Google DeepDream, BVLC Caffe Model Zoo, MIT Places, and ProGamerGov's Caffe→PyTorch conversions.