def apply_regression_to_fmri(fmri_file, dt_key, model, frame_idx, subject=1):
# with h5py.File("/data1/home/guangjie/Data/vim-2-gallant/myOrig/imagenet128_embeds_from_vqvae.hdf5",'r') as ebdf:
# embeds = ebdf['embeds']
saveDir = "/data1/home/guangjie/Data/vim-2-gallant/regressed_k_of_vqvae/subject_{}/{}/".format(
subject, dt_key)
os.makedirs(saveDir, exist_ok=True)
model.load_state_dict(
torch.load(
"/data1/home/guangjie/Data/vim-2-gallant/regressionKModel/subject_{}/frame_{}/subject_{}_frame_{}_regression_k_model_i_4917_o_1024.pth"
.format(subject, frame_idx, subject, frame_idx)))
mean, std = None, None
dataset = fmri_dataset(fmri_file, mean, std, dt_key=dt_key)
dataloader = DataLoader(dataset, batch_size=128, shuffle=False)
with h5py.File(
os.path.join(saveDir,
"frame_{}_regressed_k.hdf5".format(frame_idx)),
'w') as sf:
k = sf.create_dataset('k',
shape=(len(dataset), 32, 32),
dtype=np.uint16)
with torch.no_grad():
model.eval()
begin_idx = 0
for step, fmri in enumerate(dataloader):
out = model(fmri.cuda())
end_idx = begin_idx + len(out)
out = torch.clamp(out.view(-1, 32, 32) * 512, 0,
511).cpu().numpy().astype(np.uint16)
k[begin_idx:end_idx] = out # todo 限制大小到0-511
begin_idx = end_idx
print(step)
def apply_regression_to_fmri_quick(dt_key, frame_idx, n_point, subject,
model_in_dim, model_out_dim):
model_dir = "/data1/home/guangjie/Data/vim-2-gallant/regressionFeaturePointModel/subject_{}/frame_{}".format(
subject, frame_idx)
save_dir = "/data1/home/guangjie/Data/vim-2-gallant/regressed_zq_of_vqvae_by_feature_point/subject_{}/{}/frame_{}".format(
subject, dt_key, frame_idx)
os.makedirs(save_dir, exist_ok=True)
# model = NonLinearRegressionModel(model_in_dim, model_out_dim).cuda()
model = LinearRegressionModel(model_in_dim, model_out_dim).cuda()
# mean, std = get_vim2_fmri_mean_std(
# "/data1/home/guangjie/Data/vim-2-gallant/myOrig/VoxelResponses_subject1_v1234_rt_rv_rva0.hdf5",
# 'rt') # todo 归一化方式
mean, std = None, None
dataset = fmri_dataset(
"/data1/home/guangjie/Data/vim-2-gallant/myOrig/VoxelResponses_subject1_v1234_rt_{}.hdf5"
.format('train' if dt_key == 'rt' else 'test'), mean, std,
'rt') # todo train test
# dataset = fmri_dataset(
# "/data1/home/guangjie/Data/vim-2-gallant/myOrig/VoxelResponses_subject1_v1234_rt_rv_rva0.hdf5",
# mean, std, dt_key) # todo
dataloader = DataLoader(dataset,
batch_size=6000,
shuffle=False,
num_workers=0)
with h5py.File(
os.path.join(
save_dir,
"subject_{}_frame_{}_ze_fpoint_all_quick_apply.hdf5".format(
subject, frame_idx)), 'w') as sf:
latent = sf.create_dataset('latent',
shape=(len(dataset), 32, 32, 128),
chunks=True)
for p_idx in range(n_point):
model.load_state_dict(
torch.load(
os.path.join(
model_dir,
"subject_{}_frame_{}_regression_model_i_{}_o_{}_fpoint_{}.pth"
.format(subject, frame_idx, i_dim, o_dim, p_idx))))
row_column = p_idx // 128
latent_idx = p_idx % 128
row = row_column // 32
column = row_column % 32
with torch.no_grad():
begin_idx = 0
model.eval()
for step, fmri in enumerate(dataloader):
out = model(fmri.cuda())
end_idx = begin_idx + len(out)
latent[begin_idx:end_idx, row, column,
latent_idx] = out.squeeze().cpu().numpy(
) # 需要cpu().numpy()?
begin_idx = end_idx
print(p_idx)
def apply_regression_to_fmri(dt_key,
frame_idx,
subject=1,
n_fmap=128,
model_in_dim=4917,
model_out_dim=1024,
fmap_size=1024):
model_dir = "/data1/home/guangjie/Data/purdue/exprimentData/regressionFeatureMapModel/subject_{}/frame_{}".format(
subject, frame_idx)
save_dir = "/data1/home/guangjie/Data/purdue/exprimentData/regressed_zq_of_vqvae_by_feature_map/subject_{}/{}/frame_{}".format(
subject, dt_key, frame_idx)
os.makedirs(save_dir, exist_ok=True)
# model = NonLinearRegressionModel(model_in_dim, model_out_dim).cuda()
model = LinearRegressionModel(model_in_dim, model_out_dim).cuda()
# mean, std = get_vim2_fmri_mean_std(
# "/data1/home/guangjie/Data/vim-2-gallant/myOrig/VoxelResponses_subject1_v1234_rt_rv_rva0.hdf5",
# 'rt') # todo 归一化方式
mean, std = None, None
dataset = fmri_dataset(
"/data1/home/guangjie/Data/purdue/exprimentData/fMRI/subject{}_{}_fmri.hdf5"
.format(subject, dt_key), mean, std, dt_key)
dataloader = DataLoader(dataset,
batch_size=128,
shuffle=False,
num_workers=0)
for fmap_idx in range(n_fmap):
model.load_state_dict(
torch.load(
os.path.join(
model_dir,
"subject_{}_frame_{}_regression_model_i_{}_o_{}_fmap_{}.pth"
.format(subject, frame_idx, i_dim, o_dim, fmap_idx))))
sf = h5py.File(
os.path.join(
save_dir, "subject_{}_frame_{}_zq_fmap_{}.hdf5".format(
subject, frame_idx, fmap_idx)), 'w') # todo zq
fmap = sf.create_dataset('fmap', shape=(len(dataset), fmap_size))
with torch.no_grad():
begin_idx = 0
model.eval()
for step, fmri in enumerate(dataloader):
out = model(fmri.cuda())
end_idx = begin_idx + len(out)
fmap[begin_idx:end_idx] = out.cpu().numpy() # 需要cpu().numpy()?
begin_idx = end_idx
sf.close()
print(fmap_idx)
def apply_regression_to_fmri(dt_key, frame_idx, n_fmap, n_slice, model_in_dim,
model_out_dim, subject):
save_dir = "/data1/home/guangjie/Data/vim-2-gallant/regressed_zq_of_vqvae_by_feature_map_slice/subject_{}/{}/frame_{}".format(
subject, dt_key, frame_idx)
os.makedirs(save_dir, exist_ok=True)
# model = NonLinearRegressionModel(model_in_dim, model_out_dim).cuda()
model = LinearRegressionModel(model_in_dim, model_out_dim).cuda()
# mean, std = get_vim2_fmri_mean_std(
# "/data1/home/guangjie/Data/vim-2-gallant/myOrig/VoxelResponses_subject1_v1234_rt_rv_rva0.hdf5",
# 'rt') # todo 归一化方式
mean, std = None, None
dataset = fmri_dataset(
"/data1/home/guangjie/Data/vim-2-gallant/myOrig/VoxelResponses_subject1_v1234_rt_{}.hdf5"
.format('train' if dt_key == 'rt' else 'test'), mean, std,
'rt') # todo train test
# dataset = fmri_dataset(
# "/data1/home/guangjie/Data/vim-2-gallant/myOrig/VoxelResponses_subject1_v1234_rt_rv_rva0.hdf5",
# mean, std, dt_key) # todo
dataloader = DataLoader(dataset,
batch_size=64,
shuffle=False,
num_workers=0)
with h5py.File(
os.path.join(
save_dir, "subject_{}_frame_{}_ze_fmap_all.hdf5".format(
subject, frame_idx)), 'w') as sf: # todo zq
fmap = sf.create_dataset('latent', shape=(len(dataset), 32, 32, 128))
for fmap_idx in range(n_fmap):
model_dir = "/data1/home/guangjie/Data/vim-2-gallant/regressionFeatureMapSliceModel/subject_{}/frame_{}/fmap_{}".format(
subject, frame_idx, fmap_idx)
for slice_idx in range(n_slice):
model.load_state_dict(
torch.load(
os.path.join(
model_dir,
"frame_{}_fmap_{}_slice_{}_regression_model.pth".
format(frame_idx, fmap_idx, slice_idx))))
with torch.no_grad():
begin_idx = 0
model.eval()
for step, fmri in enumerate(dataloader):
out = model(fmri.cuda()) # 32
end_idx = begin_idx + len(out)
fmap[begin_idx:end_idx, slice_idx, :,
fmap_idx] = out.cpu().numpy() # todo :,slice_idx
begin_idx = end_idx
print(fmap_idx)
def apply_regression_to_fmri(dt_key,
frame_idx,
subject=1,
n_lantent=1024,
model_in_dim=4917,
model_out_dim=128,
dim_lantent=128):
model_dir = "/data1/home/guangjie/Data/vim-2-gallant/regressionLatentModel/subject_{}/frame_{}".format(
subject, frame_idx)
save_dir = "/data1/home/guangjie/Data/vim-2-gallant/regressed_zq_of_vqvae_by_feature_latent/subject_{}/{}/frame_{}".format(
subject, dt_key, frame_idx)
os.makedirs(save_dir, exist_ok=True)
# model = NonLinearRegressionModel(model_in_dim, model_out_dim).cuda()
model = LinearRegressionModel(model_in_dim, model_out_dim).cuda()
# mean, std = get_vim2_fmri_mean_std(
# "/data1/home/guangjie/Data/vim-2-gallant/myOrig/VoxelResponses_subject1_v1234_rt_rv_rva0.hdf5",
# 'rt') # todo 归一化方式
mean, std = None, None
dataset = fmri_dataset(
"/data1/home/guangjie/Data/vim-2-gallant/myOrig/VoxelResponses_subject1_v1234_rt_{}.hdf5"
.format('train' if dt_key == 'rt' else 'test'), mean, std,
'rt') # todo train test
# dataset = fmri_dataset(
# "/data1/home/guangjie/Data/vim-2-gallant/myOrig/VoxelResponses_subject1_v1234_rt_rv_rva0.hdf5",
# mean, std, dt_key) # todo
dataloader = DataLoader(dataset,
batch_size=128,
shuffle=False,
num_workers=0)
for latent_idx in range(n_lantent):
# dataset = fmri_vector_dataset(
# fmri_file="/data1/home/guangjie/Data/vim-2-gallant/myOrig/VoxelResponses_subject{}_v1234_rt_rv_rva0.hdf5".format(
# subject), zq_file="/data1/home/guangjie/Data/vim-2-gallant/myOrig/zq_from_vqvae_{}.hdf5".format(
# 'st' if dt_key == 'rt' else 'sv'),
# mean=mean,
# std=std, fmri_key=dt_key, frame_idx=frame_idx, latent_idx=latent_idx)
# dataloader = DataLoader(dataset, batch_size=128, shuffle=False, num_workers=1)
# 逐次加载单个模型，抽取每张图片对应ze的第lantent_idx个隐含表征
model.load_state_dict(
torch.load(
os.path.join(
model_dir,
"subject_{}_regression_model_i_{}_o_{}_latent_{}.pth".
format(subject, model_in_dim, model_out_dim, latent_idx))))
sf = h5py.File(
os.path.join(
save_dir, "subject_{}_frame_{}_ze_latent_{}.hdf5".format(
subject, frame_idx, latent_idx)), 'w') # todo zq
latent = sf.create_dataset('latent', shape=(len(dataset), dim_lantent))
with torch.no_grad():
begin_idx = 0
model.eval()
for step, fmri in enumerate(dataloader):
out = model(fmri.cuda())
end_idx = begin_idx + len(out)
latent[begin_idx:end_idx] = out.cpu().numpy(
) # 需要cpu().numpy()?
begin_idx = end_idx
sf.close()
print(latent_idx)