('tech_mask_tensor', 'techniques_mask'),
]
tensor_names, tensor_cols = zip(*tensors_to_load)
# Load tensors into memory
memory_tensors = load_recipe_tensors(df_r,
DEVICE,
cols=tensor_cols,
types=[torch.LongTensor] *
len(tensors_to_load))
memory_tensor_map = dict(zip(tensor_names, memory_tensors))
print('{} - Tensors loaded in memory.'.format(datetime.now() - start))
# Samplers
train_data = DataFrameDataset(train_df, ['u', 'i'])
train_sampler = BatchSampler(train_data, batch_size, random=True)
valid_data = DataFrameDataset(valid_df, ['u', 'i'])
valid_sampler = BatchSampler(valid_data, batch_size)
test_data = DataFrameDataset(test_df, ['u', 'i'])
test_sampler = BatchSampler(test_data, batch_size)
'''
Create model
'''
model = create_model(
vocab_emb_dim=vocab_emb_dim,
calorie_emb_dim=calorie_emb_dim,
hidden_size=hidden_size,
n_layers=n_layers,
dropout=dropout,
max_ingr=MAX_INGR,
max_ingr_tok=MAX_INGR_TOK,
)
memory_tensor_map = dict(zip(tensor_names, memory_tensors))
print('{} - Tensors loaded in memory.'.format(datetime.now() - start))
# Name padding for item
memory_tensor_map['name_tensor'] = torch.cat(
[memory_tensor_map['name_tensor'],
torch.LongTensor([[PAD_INDEX] * MAX_NAME]).to(DEVICE)]
)
if n_samples < len(test_df) and not ppx_only:
sampled_test = test_df.sample(n=n_samples)
else:
sampled_test = test_df
test_data = DataFrameDataset(sampled_test, ['u', 'i'])
test_sampler = BatchSampler(test_data, batch_size)
'''
Create model
'''
model = create_model(
vocab_emb_dim=vocab_emb_dim, calorie_emb_dim=calorie_emb_dim,
item_emb_dim=item_emb_dim, n_items_w_pad=NUM_ITEM_EMBEDDING, hidden_size=hidden_size,
n_layers=n_layers, dropout=0.0, max_ingr=MAX_INGR, max_ingr_tok=MAX_INGR_TOK,
use_cuda=USE_CUDA, state_dict_path=model_path,
ingr_gru=ingr_gru, decode_name=decode_name, ingr_emb=ingr_emb,
num_ingr=N_INGREDIENTS, ingr_emb_dim=ingr_emb_dim, shared_projection=shared_proj,
item_emb=item_emb
)
model_id = os.path.basename(model_path)[:-3]