action='store_true',
default=False,
help='Use Ingr embedding in encoder')
parser.add_argument('--shared-proj',
action='store_true',
default=False,
help='Share projection layers for name and steps')
parser.add_argument('--load-checkpoint',
type=str,
default=None,
help='Load from state dict checkpoint')
args = parser.parse_args()
start = datetime.now()
USE_CUDA, DEVICE = get_device()
# Filters
MAX_NAME = 15
MAX_INGR = 5
MAX_INGR_TOK = 20
MAX_STEP_TOK = 256
# Reproducibility
seed = 42
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
# Args
data_dir = args.data_dir
def main():
import os
import torch
import argparse
import torch.nn.init as init
from recipe_gen.utils import get_device, count_parameters
parser = argparse.ArgumentParser(
description='Baseline for recipe generation (dynamic attn)')
parser.add_argument('--gold-path',
type=str,
required=True,
help='location of the data corpus')
parser.add_argument('--pred-path',
type=str,
required=True,
help='location of the data corpus')
parser.add_argument('--model-path',
type=str,
required=True,
help='location of the data corpus')
args = parser.parse_args()
start = datetime.now()
USE_CUDA, DEVICE = get_device()
args = parser.parse_args()
gold_path = args.gold_path
pred_path = args.pred_path
model_path = args.model_path
model = AbsoluteOrderTeacher(embedding_dim=768,
hidden_size=256,
gru_layers=2)
gold_test_df = pd.read_msgpack(gold_path)
pred_test_df = pd.read_msgpack(pred_path)
gold_bert_feature_lists = []
for index, row in tqdm(gold_test_df.iterrows(), total=len(gold_test_df)):
row = row['bert_features_steps'].tolist()
if type(row[-1]) == tuple:
pads = [x for x in row[-1]]
new_row = np.array(row[:-1] + pads)
if new_row.shape[0] > 15:
new_row = new_row[:15]
gold_bert_feature_lists.append(new_row)
else:
new_row = np.array(row)
if new_row.shape[0] > 15:
new_row = new_row[:15]
gold_bert_feature_lists.append(np.array(new_row))
gold_test_df['bert_features_steps'] = gold_bert_feature_lists
print(gold_test_df.head(2))
pred_bert_feature_lists = []
for index, row in tqdm(pred_test_df.iterrows(), total=len(pred_test_df)):
row = row['bert_features_steps'].tolist()
if type(row[-1]) == tuple:
pads = [x for x in row[-1]]
new_row = np.array(row[:-1] + pads)
if new_row.shape[0] > 15:
new_row = new_row[:15]
pred_bert_feature_lists.append(new_row)
else:
new_row = np.array(row)
if new_row.shape[0] > 15:
new_row = new_row[:15]
pred_bert_feature_lists.append(np.array(new_row))
pred_test_df['bert_features_steps'] = pred_bert_feature_lists
eval_df = pd.merge(gold_test_df, pred_test_df, left_on='i', right_on='i')
eval_df.dropna()
eval_df.columns = [
'i', 'bert_features_step_lists', 'bert_features_pred_step_lists'
]
print(eval_df.head(2))
eval_df = eval_df.sample(frac=0.5)
print(len(eval_df))
# Load state dict
state_dict = torch.load(model_path)
model.load_state_dict(state_dict, strict=True)
model.to(DEVICE)
print('Model loaded')
model.eval()
with torch.no_grad():
eval_loss = test_run_epoch(device=DEVICE,
df=eval_df,
model=model,
batch_size=16)
print('The absolute oder score is: {}'.format(eval_loss))