QM9 trainer up and running

applications/FLASK/Transformer/datasets/QM9.py

@@ -7,6 +7,7 @@
 
 import numpy as np
 from pretokenize.SMILES_tokenizer import MolTokenizer
+from pretokenize.data_utils import random_zero_array
 
 sequence_length = int(os.getenv("QM9_SEQUENCE_LENGTH", default="32"))
 
@@ -20,10 +21,12 @@
 tokenizer = MolTokenizer(os.path.join(data_dir, "QM9_vocab.json"))
 tokenizer.load_vocab_file()
 
-dataset_train = np.load(os.path.join(data_dir, "QM9_Pretokenized.npy"), allow_pickle=True)
+dataset_train = np.load(os.path.join(data_dir, "QM9_Pretokenized.npy"))
 
+# dataset_train = np.zeros((140000, 32), dtype=np.float32)
 _vocab_size = 46
-
+pad_index = tokenizer.token_to_id("<pad>")
+sep_index = tokenizer.token_to_id("<eos>")
 
 # ----------------------------------------------
 # Sample access functions
@@ -36,7 +39,16 @@ def num_train_samples():
 
 def get_train_sample(i):
     data = dataset_train[i]
-    return data
+
+    boundary = np.where(data == sep_index)[0][0]
+    masked_data = random_zero_array(
+        data[:boundary], 0.15, tokenizer.token_to_id(tokenizer.mask_token)
+    )
+    output = np.zeros((2 * sequence_length), dtype=np.int32)
+    output[0:boundary] = masked_data
+    output[boundary] = sep_index
+    output[sequence_length:] = data
+    return output
 
 
 def sample_dims():
@@ -50,4 +62,4 @@ def vocab_size():
 if __name__ == "__main__":
     print("Training samples:", num_train_samples())
     print("Training sample 101:")
-    print(get_train_sample(101))
+    print(get_train_sample(0))

applications/FLASK/Transformer/datasets/pretokenize/QM9_Pretokenize.py

@@ -1,26 +1,34 @@
 import numpy as np
 from SMILES_tokenizer import MolTokenizer
 from data_utils import random_zero_array
+import os
+import os.path
 
 
 def main():
-    tokenizer = MolTokenizer("SMILES_vocab.json")
+    data_dir = os.getenv("QM9_DATA_DIR", "/p/vast1/lbann/datasets/FLASK/QM9")
+
+    tokenizer = MolTokenizer(os.path.join(data_dir, "QM9_vocab.json"))
     tokenizer.load_vocab_file()
-    with open("QM9_smiles.txt", 'r') as smiles_data:
-      smiles_data = smiles_data.readlines()
-      num_samples = len(smiles_data)
-      max_length = 32
 
-      tokenized_data = np.ones((num_samples, max_length)) * tokenizer.encode(tokenizer.pad_token) 
-      tokenized_data[:, 0] = tokenizer.encode(tokenizer.sep_token)
+    data_file = os.path.join(data_dir, "QM9_smiles.txt")
+    with open(data_file, "r") as smiles_data:
+        smiles_data = smiles_data.readlines()
+        num_samples = len(smiles_data)
+        max_length = 32
+
+        tokenized_data = np.ones((num_samples, max_length)) * tokenizer.encode(
+            tokenizer.pad_token
+        )
+        tokenized_data[:, 0] = tokenizer.encode(tokenizer.sep_token)
 
-      for i, smiles in enumerate(smiles_data, start=1):
-        tokens = tokenizer.tokenize(smiles)
-        tokens = random_zero_array(tokens, 0.15, tokenizer.encode(tokenizer.mask_token))
-        tokenized_data[i, :len(tokens)] = tokens
-        tokenized_data[i, len(tokens)] = tokenizer.encode(tokenizer.cls_token)
+        for i, smiles in enumerate(smiles_data, start=0):
+            tokens = tokenizer.tokenize(smiles)
+            tokenized_data[i, : len(tokens)] = tokens
+            tokenized_data[i, len(tokens)] = tokenizer.encode(tokenizer.sep_token)
+    save_file_loc = os.path.join(data_dir, "QM9_Pretokenized.npy")
+    np.save(save_file_loc, tokenized_data)
 
-    np.save('QM9_Pretokenized.npy', tokenized_data)
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     main()

applications/FLASK/Transformer/network.py

@@ -144,11 +144,11 @@ def _add_input_encoding(
 
     # Apply encoder
     if encoder_input is not None:
-        encoder_input = positional_encoder(
+        encoder_input = positional_encoder.apply_input(
             encoder_input, encoder_sequence_length, **kwargs
         )
     if decoder_input is not None:
-        decoder_input = positional_encoder(
+        decoder_input = positional_encoder.apply_input(
             decoder_input, decoder_sequence_length, **kwargs
         )