def load_pretrain_weights(self):
"""Loading weights from trained MLP model & GMF model"""
config = self.config
config['latent_dim'] = config['latent_dim_mlp']
mlp_model = MLP(config)
if config['use_cuda'] is True:
mlp_model.cuda()
resume_checkpoint(mlp_model,
model_dir=config['pretrain_mlp'],
device_id=config['device_id'])
self.embedding_user_mlp.weight.data = mlp_model.embedding_user.weight.data
self.embedding_item_mlp.weight.data = mlp_model.embedding_item.weight.data
for idx in range(len(self.fc_layers)):
self.fc_layers[idx].weight.data = mlp_model.fc_layers[
idx].weight.data
config['latent_dim'] = config['latent_dim_mf']
gmf_model = GMF(config)
if config['use_cuda'] is True:
gmf_model.cuda()
resume_checkpoint(gmf_model,
model_dir=config['pretrain_mf'],
device_id=config['device_id'])
self.embedding_user_mf.weight.data = gmf_model.embedding_user.weight.data
self.embedding_item_mf.weight.data = gmf_model.embedding_item.weight.data
self.affine_output.weight.data = 0.5 * torch.cat([
mlp_model.affine_output.weight.data,
gmf_model.affine_output.weight.data
],
dim=-1)
self.affine_output.bias.data = 0.5 * (
mlp_model.affine_output.bias.data +
gmf_model.affine_output.bias.data)
def create_model(self):
print('-' * 15, "Creating model", '-' * 15)
latent_dim = 10
config = {
'num_virus': self.n_v,
'num_human': self.n_h,
'latent_dim': latent_dim,
'sparse':
False # set false for now because some optimizers dont work with sparse
}
self.model = GMF(config)
self.model.to(self.device)
print(self.model)
# print("params-------")
# print(list(self.model.parameters()))
# print("end-------")
# print('grad: ', list(self.model.parameters())[0].grad)
# print('grad: ', list(self.model.parameters())[3].grad)
# print('grad: ', list(self.model.parameters())[4].grad)
print('-' * 15, "Done with model", '-' * 15)
print()
def __init__(self, args, num_users, num_items):
BaseModel.__init__(self, args, num_users, num_items)
self.layers = eval(args.layers)
self.lambda_layers = eval(args.reg_layers)
self.num_factors = args.num_factors
self.model_GMF = GMF(args, num_users, num_items)
self.model_MLP = MLP(args, num_users, num_items)
def load_pretrain_weights(self):
"""Loading weights from trained MLP model & GMF model"""
config = self.config
mlp_model = MLP(config)
device_id = -1
if config['use_cuda'] is True:
mlp_model.cuda()
device_id = config['device_id']
resume_checkpoint(mlp_model,
model_dir=config['pretrain_mlp'],
device_id=device_id)
self.embedding_account_mlp.weight.data = mlp_model.embedding_account.weight.data
self.embedding_location_mlp.weight.data = mlp_model.embedding_location.weight.data
for idx in range(len(self.fc_layers)):
self.fc_layers[idx].weight.data = mlp_model.fc_layers[
idx].weight.data
config['latent_dim'] = config['latent_dim_mf']
gmf_model = GMF(config)
if config['use_cuda'] is True:
gmf_model.cuda()
resume_checkpoint(gmf_model,
model_dir=config['pretrain_mf'],
device_id=device_id)
self.embedding_account_mf.weight.data = gmf_model.embedding_account.weight.data
self.embedding_location_mf.weight.data = gmf_model.embedding_location.weight.data
self.embedding_account_mlp.require = False
self.embedding_location_mlp.require = False
self.embedding_account_mf.require = False
self.embedding_location_mf.require = False
def load_pretrain_weights(self):
"""Loading weights from trained GMF model"""
config = self.config
gmf_model = GMF(config)
if config['use_cuda'] is True:
gmf_model.cuda()
resume_checkpoint(gmf_model, model_dir=config['pretrain_mf'], device_id=config['device_id'])
self.embedding_user.weight.data = gmf_model.embedding_user.weight.data
self.embedding_item.weight.data = gmf_model.embedding_item.weight.data
def train_gmf():
model = GMF(gmf_config).to(device)
opt = optim.Adam(model.parameters(), lr=0.001, betas=(0.9, 0.999),weight_decay=1e-4)
criterion = torch.nn.MSELoss()
for epoch in range(gmf_config['num_epoch']):
print("running epoch ", epoch)
train_mse = epoch_run(model, train_generator, opt, criterion, "train")
val_mse = epoch_run(model, val_generator, opt, criterion,"val")
print("train mse loss => ", train_mse, "val mse loss => ", val_mse)
return model
def create_model(self):
print('-' * 15, "Creating model", '-' * 15)
latent_dim = 2799
config = {
'num_virus': self.n_v,
'num_human': self.n_h,
'latent_dim': latent_dim,
'sparse': False # set false for now because some optimizers dont work with sparse
}
self.model = GMF(config)
self.model.to(self.device)
print(self.model)
print('-' * 15, "Done with model", '-' * 15)
print()
def load_pretrain_weights(self):
"""Loading weights from trained GMF model"""
config = self.config
gmf_model = GMF(config)
if config['use_cuda'] is True:
gmf_model.cuda()
resume_checkpoint(gmf_model, model_dir=config['pretrain_mf'], device_id=config['device_id'])
self.embedding_user.weight.data = gmf_model.embedding_user.weight.data
self.embedding_item.weight.data = gmf_model.embedding_item.weight.data
# class MLPEngine(Engine):
# """Engine for training & evaluating GMF model"""
# def __init__(self, config):
# self.model = MLP(config)
# if config['use_cuda'] is True:
# use_cuda(True, config['device_id'])
# self.model.cuda()
# super(MLPEngine, self).__init__(config)
# print(self.model)
# if config['pretrain']:
# self.model.load_pretrain_weights()
class NeuMF(BaseModel):
def __init__(self, args, num_users, num_items):
BaseModel.__init__(self, args, num_users, num_items)
self.layers = eval(args.layers)
self.lambda_layers = eval(args.reg_layers)
self.num_factors = args.num_factors
self.model_GMF = GMF(args, num_users, num_items)
self.model_MLP = MLP(args, num_users, num_items)
def build_core_model(self, user_indices, item_indices):
vector_GMF, len_GMF, params_GMF = self.model_GMF.build_core_model(
user_indices, item_indices)
vector_MLP, len_MLP, params_MLP = self.model_MLP.build_core_model(
user_indices, item_indices)
model_vector = tf.concat([vector_GMF, vector_MLP], 1)
model_len = len_GMF + len_MLP
model_params = []
model_params.extend(params_GMF)
model_params.extend(params_MLP)
return model_vector, model_len, model_params
def build_model(self, user_indices=None, item_indices=None):
if not user_indices:
user_indices = tf.placeholder(tf.int32, [None],
name="user_indices")
self.user_indices = user_indices
if not item_indices:
item_indices = tf.placeholder(tf.int32, [None],
name="item_indices")
self.item_indices = item_indices
self.ratings = tf.placeholder(tf.float32, [None])
model_vector, model_len, model_params = self.build_core_model(
user_indices, item_indices)
self.output, self.loss, self.error, self.raw_error, self.train_step = \
self.build_train_model(model_vector, model_len,
self.ratings, model_params)
def train(args, train_data_path):
print("use_gpu:{}, NeuMF:{}, epochs:{}, batch_size:{}, num_factors:{}, num_neg:{}, lr:{}, model_dir:{}, layers:{}".format(
args.use_gpu, args.NeuMF, args.epochs, args.batch_size, args.num_factors, args.num_neg, args.lr, args.model_dir, args.layers))
dataset = Dataset(args.path + args.dataset)
testRatings, testNegatives = dataset.testRatings, dataset.testNegatives
train_data_generator = utils.Dataset()
train_reader = fluid.io.batch(train_data_generator.train(train_data_path, True), batch_size=args.batch_size)
inputs = utils.input_data(True)
if args.GMF:
model = GMF()
loss, pred = model.net(inputs, args.num_users, args.num_items, args.num_factors)
elif args.MLP:
model = MLP()
loss, pred = model.net(inputs, args.num_users, args.num_items, args.layers)
elif args.NeuMF:
model = NeuMF()
loss, pred = model.net(inputs, args.num_users, args.num_items, args.num_factors, args.layers)
optimizer = fluid.optimizer.AdamOptimizer(args.lr)
optimizer.minimize(loss)
place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
loader = fluid.io.DataLoader.from_generator(
feed_list=inputs, capacity=args.batch_size, iterable=True)
loader.set_sample_list_generator(train_reader, places=place)
for epoch in range(args.epochs):
for batch_id, data in enumerate(loader()):
begin = time.time()
loss_val = exe.run(program=fluid.default_main_program(),
feed=data,
fetch_list=[loss.name],
return_numpy=True)
end = time.time()
logger.info("epoch: {}, batch_id: {}, batch_time: {:.5f}s, loss: {:.5f}".format(epoch, batch_id, end - begin, np.array(loss_val)[0][0]))
save_dir = "%s/epoch_%d" % (args.model_dir, epoch)
feed_var_names = ["user_input", "item_input"]
fetch_vars = [pred]
fluid.io.save_inference_model(save_dir, feed_var_names, fetch_vars, exe)
def load_pretrain_weights(self):
"""Loading weights from trained MLP model & GMF model"""
config = self.config
config['latent_dim'] = config['latent_dim_mlp']
mlp_model = MLP(config)
# if config['use_cuda'] is True:
# mlp_model.cuda()
# resume_checkpoint(mlp_model, model_dir = config['pretrain_mlp'], device_id = config['device_id'])
resume_checkpoint(mlp_model, model_dir=config['pretrain_mlp'])
# Get the user and item weights from the trained MLP model
self.embedding_user_mlp.weight.data = mlp_model.embedding_user.weight.data
self.embedding_item_mlp.weight.data = mlp_model.embedding_item.weight.data
for idx in range(len(self.fc_layers)):
self.fc_layers[idx].weight.data = mlp_model.fc_layers[
idx].weight.data
config['latent_dim'] = config['latent_dim_mf']
gmf_model = GMF(config)
# if config['use_cuda'] is True:
# gmf_model.cuda()
# resume_checkpoint(gmf_model, model_dir = config['pretrain_mf'], device_id = config['device_id'])
resume_checkpoint(gmf_model, model_dir=config['pretrain_mf'])
# Get the user and item weights from the trained GMF model
self.embedding_user_mf.weight.data = gmf_model.embedding_user.weight.data
self.embedding_item_mf.weight.data = gmf_model.embedding_item.weight.data
# Perform linear transformation to get the final weight and bias values from both MLP and GMF weights
self.affine_output.weight.data = 0.5 * torch.cat([
mlp_model.affine_output.weight.data,
gmf_model.affine_output.weight.data
],
dim=-1)
self.affine_output.bias.data = 0.5 * (
mlp_model.affine_output.bias.data +
gmf_model.affine_output.bias.data)
modelpath = os.path.join(modeldir, modelfname)
resultsdfpath = os.path.join(modeldir, 'results_df.p')
dataset = np.load(os.path.join(datadir, dataname))
train_ratings = load_npz(os.path.join(datadir, train_matrix)).todok()
test_ratings, negatives = dataset['test_negative'], dataset['negatives']
n_users, n_items = dataset['n_users'].item(), dataset['n_items'].item()
test_loader = DataLoader(dataset=test_ratings,
batch_size=1000,
shuffle=False
)
model = NeuMF(n_users, n_items, n_emb, layers, dropouts)
if os.path.isfile(mf_pretrain) and os.path.isfile(mlp_pretrain):
gmf_model = GMF(n_users, n_items, n_emb)
gmf_model.load_state_dict(torch.load(mf_pretrain))
mlp_model = MLP(n_users, n_items, layers, dropouts)
mlp_model.load_state_dict(torch.load(mlp_pretrain))
model = load_pretrain_model(model, gmf_model, mlp_model)
print("Load pretrained GMF {} and MLP {} models done. ".format(mf_pretrain, mlp_pretrain))
use_cuda = torch.cuda.is_available()
if use_cuda:
model = model.cuda()
if freeze:
for name, layer in model.named_parameters():
if not ("out" in name):
layer.requires_grad = False
def parse(path):
g = gzip.open(path, 'rb')
for l in g:
yield eval(l)
def getDF(path):
i = 0
df = {}
for d in parse(path):
df[i] = d
i += 1
return pd.DataFrame.from_dict(df, orient='index')
DATA_PATH = Path(".")
MODEL_DIR = "models"
asin2id_map = pickle.load(open(DATA_PATH/'item_mappings.p', 'rb'))
id2asin_map = {k:v for v,k in asin2id_map.items()}
df_movies_meta_data = getDF(DATA_PATH/'meta_Movies_and_TV.json.gz')
keep_cols = ['asin', 'title']
df_movies_meta_data = df_movies_meta_data[keep_cols]
df_movies_meta_data = df_movies_meta_data[~df_movies_meta_data.title.isna()]
asin2title_map = dict(df_movies_meta_data.values)
print("number of items with missing title in the core dataset: {}".format(
np.setdiff1d(list(id2asin_map.values()), list(asin2title_map.keys())).shape[0]))
print("number of items with non missing titles in the core dataset: {}".format(
len(id2asin_map) \
- np.setdiff1d(list(id2asin_map.values()), list(asin2title_map.keys())).shape[0]))
id2title_map = {}
for k,v in id2asin_map.items():
try:
id2title_map[k] = asin2title_map[v]
except:
continue
df_results = pd.read_pickle(DATA_PATH/MODEL_DIR/'results_df.p')
best_gmf = (df_results[df_results.modelname.str.contains('GMF')]
.sort_values('best_hr', ascending=False)
.reset_index(drop=True)
).modelname[0]
n_emb_i = int(np.where([s == 'emb' for s in best_gmf.split("_")])[0])+1
n_emb = int(best_gmf.split("_")[n_emb_i])
dataset = np.load(DATA_PATH/'neuralcf_split.npz')
n_users, n_items = dataset['n_users'].item(), dataset['n_items'].item()
gmf_model = GMF(n_users, n_items, n_emb)
gmf_model.load_state_dict(torch.load(DATA_PATH/MODEL_DIR/best_gmf))
item_embeddings = gmf_model.embeddings_item.weight.data.numpy()
knn_model = NearestNeighbors(metric = 'cosine', algorithm = 'brute')
knn_model.fit(item_embeddings)
def get_movie_titles(input_id, n=20):
"""first movie will be the "query" movie and the remaining n-1 the similar
movies. Similar defined under the functioning of the algorithm, i.e.
leading to the same prediction"""
dist, nnidx = knn_model.kneighbors(
item_embeddings[input_id].reshape(1, -1),
n_neighbors = n)
titles = []
for idx in nnidx[0]:
try:
titles.append(id2title_map[idx])
except:
continue
return titles
similar_movies = get_movie_titles(1234)
class GMFConfig_dbg:
def __init__(self, device, n=150, m=150, prob=.5):
self.device = device
self.create_generator(n, m, prob)
self.create_model()
def create_model(self):
print('-' * 15, "Creating model", '-' * 15)
latent_dim = 10
config = {
'num_virus': self.n_v,
'num_human': self.n_h,
'latent_dim': latent_dim,
'sparse':
False # set false for now because some optimizers dont work with sparse
}
self.model = GMF(config)
self.model.to(self.device)
print(self.model)
# print("params-------")
# print(list(self.model.parameters()))
# print("end-------")
# print('grad: ', list(self.model.parameters())[0].grad)
# print('grad: ', list(self.model.parameters())[3].grad)
# print('grad: ', list(self.model.parameters())[4].grad)
print('-' * 15, "Done with model", '-' * 15)
print()
def create_generator(self, m, n, prob):
############################
## generate bipartite
###########################
print('-' * 15, "Generating graph", '-' * 15)
G = nx.bipartite.random_graph(n, m, prob)
observed = list(G.edges())
nodes = list(G.nodes())
virusUprot = []
humanUprot = []
edges = []
for i in tqdm(range(n)):
for j in tqdm(range(n, m + n)):
virusUprot.append(i)
humanUprot.append(j)
if (i, j) in observed:
edges.append(1.0)
else:
edges.append(0.0)
M = pd.DataFrame({
'virusUprot': virusUprot,
'humanUprot': humanUprot,
'edge': edges
})
htoi = {v: k for k, v in enumerate(M['humanUprot'].unique())}
vtoi = {v: k for k, v in enumerate(M['virusUprot'].unique())}
print('-' * 15, "Dataframe created", '-' * 15)
print()
############################
## Prepare data (dataloader)
############################
print('-' * 15, "Creating data loaders", '-' * 15)
data_config = {
'interactions': M,
'htoi': htoi,
'vtoi': vtoi,
'pct_test': .10,
'device': self.device
}
self.n_v = len(vtoi)
self.n_h = len(htoi)
self.gen = ProteinInteractionGenerator(data_config)
print('-' * 15, "Generator done", '-' * 15)
print()
def get_generator(self):
return self.gen
def get_model(self):
return self.model
class GMFConfig:
def __init__(self, path, debug, device):
self.device = device
self.create_generator(path, debug)
self.create_model()
def create_model(self):
print('-' * 15, "Creating model", '-' * 15)
latent_dim = 2799
config = {
'num_virus': self.n_v,
'num_human': self.n_h,
'latent_dim': latent_dim,
'sparse': False # set false for now because some optimizers dont work with sparse
}
self.model = GMF(config)
self.model.to(self.device)
print(self.model)
print('-' * 15, "Done with model", '-' * 15)
print()
def create_generator(self, path, debug):
############################
## paths
###########################
train_csv =f'{path}full_train.csv'
############################
## Load data
############################
print('-' * 15, "Loading data", '-' * 15)
print("loading traning matrix at: ", train_csv)
M = pd.read_csv(train_csv)
if debug:
print("Making debug dataset.....")
pos = M.loc[M['edge'] > 0].sample(frac=1)
negs = M.loc[M['edge'] == 0].sample(frac=1)
M = pd.concat([pos, negs[:len(pos)]], ignore_index=True).sample(frac=1)
htoi = {v:k for k,v in enumerate(M['humanUprot'].unique())}
vtoi = {v:k for k,v in enumerate(M['virusUprot'].unique())}
############################
## Prepare data (dataloader)
############################
print('-' * 15, "Creating data loaders", '-' * 15)
data_config = {
'interactions':M,
'htoi':htoi,
'vtoi':vtoi,
'pct_test':.10,
'device': self.device
}
self.n_v = len(vtoi)
self.n_h = len(htoi)
self.gen = ProteinInteractionGenerator(data_config)
print('-' * 15, "Generator done", '-' * 15)
print()
def get_generator(self):
return self.gen
def get_model(self):
return self.model