I have created a python notebook for the molecular translation competition. I have used the visual attention model to generate the InChI name from the molecular structure image. This model has two parts, encoder, and decoder. For the encoder, I used a pre-trained Inceptionv3 model trained on the Imagenet dataset. For the decoder, I used the RNN model with GRU unit and attention mechanism.
\nThe same type of model is present on the Tensorflow website for Image Captioning: https://www.tensorflow.org/tutorials/text/image_captioning
\nWe can change the hyperparameters and the pre-trained model for encoding.
\nI used the following parameters -
\nBATCH_SIZE = 8
\nBUFFER_SIZE = 1000
\nembedding_dim = 256
\nunits = 512
\nvocab_size = 198
\nnum_steps = len(images) // BATCH_SIZE
\nfeatures_shape = 2048
\nattention_features_shape = 64
Tokenization of InChI names
\nThe list of tokens that I used -
\nTOKEN_LIST = [\"\", \"InChI=1S/\",\"\", \"\", \"/c\", \"/h\", \"/m\", \"/t\", \"/b\", \"/s\", \"/i\"] +\\
\n ['Si', 'Br', 'Cl', 'F', 'I', 'N', 'O', 'P', 'S', 'C', 'H', 'B', ] +\\
\n [str(i) for i in range(167,-1,-1)] +\\
\n [\"+\", \"(\", \")\", \"-\", \",\", \"D\", \"T\"]
For the tokenization, I used the re library.
\nFollowing is my function to convert into tokens
def convert_to_tensor(label):
\n token = [tok_2_int[\"\"]]
\n l = label.split('/')
\n token.append(tok_2_int[l[0]+'/'])
\n f = re.split('(\\d+)', l[1])
\n for c in f:
\n if c.isnumeric()==False:
\n st=0
\n for i in range(len(c)+1):
\n if c[st:i] in TOKEN_LIST:
\n token.append(tok_2_int[c[st:i]])
\n st=i
\n else:
\n token.append(tok_2_int[c])
for i in range(2,len(l)):
\n token.append(tok_2_int['/'+l[i][0]])
\n s = re.split(r'(\\W+)', l[i][1:])
\n for c in s:
\n if c.isnumeric()==False and len(c)>=2:
\n if c[0] == '-' or c[0] == '+' or c[0] == ',' or c[0] == ')' or c[0]== '(':
\n for sp in c:
\n token.append(tok_2_int[sp])
\n else:
\n cc = re.split('(\\d+)', c)
\n for b in cc:
\n if b.isnumeric()==False and len(b)>=2:
\n st=0
\n for i in range(len(b)+1):
\n if b[st:i] in TOKEN_LIST:
\n token.append(tok_2_int[b[st:i]])
\n st=i
\n else:
\n if len(b)>0:
\n token.append(tok_2_int[b])
\n else:
\n if len(c)>0:
\n token.append(tok_2_int[c])
\n token.append(tok_2_int[\"\"])
\n return token
Training
\nDuring training, it took almost 30 min for 1 epoch using the batch of size 8 on Colab GPU.
", "messages": [ { "id": 1326588, "postDate": "2021-05-28T14:55:50.183Z", "content": "I have created a python notebook for the molecular translation competition. I have used the visual attention model to generate the InChI name from the molecular structure image. This model has two parts, encoder, and decoder. For the encoder, I used a pre-trained Inceptionv3 model trained on the Imagenet dataset. For the decoder, I used the RNN model with GRU unit and attention mechanism.
\nThe same type of model is present on the Tensorflow website for Image Captioning: https://www.tensorflow.org/tutorials/text/image_captioning
\nWe can change the hyperparameters and the pre-trained model for encoding.
\nI used the following parameters -
\nBATCH_SIZE = 8
\nBUFFER_SIZE = 1000
\nembedding_dim = 256
\nunits = 512
\nvocab_size = 198
\nnum_steps = len(images) // BATCH_SIZE
\nfeatures_shape = 2048
\nattention_features_shape = 64
Tokenization of InChI names
\nThe list of tokens that I used -
\nTOKEN_LIST = [\"\", \"InChI=1S/\",\"\", \"\", \"/c\", \"/h\", \"/m\", \"/t\", \"/b\", \"/s\", \"/i\"] +\\
\n ['Si', 'Br', 'Cl', 'F', 'I', 'N', 'O', 'P', 'S', 'C', 'H', 'B', ] +\\
\n [str(i) for i in range(167,-1,-1)] +\\
\n [\"+\", \"(\", \")\", \"-\", \",\", \"D\", \"T\"]
For the tokenization, I used the re library.
\nFollowing is my function to convert into tokens
def convert_to_tensor(label):
\n token = [tok_2_int[\"\"]]
\n l = label.split('/')
\n token.append(tok_2_int[l[0]+'/'])
\n f = re.split('(\\d+)', l[1])
\n for c in f:
\n if c.isnumeric()==False:
\n st=0
\n for i in range(len(c)+1):
\n if c[st:i] in TOKEN_LIST:
\n token.append(tok_2_int[c[st:i]])
\n st=i
\n else:
\n token.append(tok_2_int[c])
for i in range(2,len(l)):
\n token.append(tok_2_int['/'+l[i][0]])
\n s = re.split(r'(\\W+)', l[i][1:])
\n for c in s:
\n if c.isnumeric()==False and len(c)>=2:
\n if c[0] == '-' or c[0] == '+' or c[0] == ',' or c[0] == ')' or c[0]== '(':
\n for sp in c:
\n token.append(tok_2_int[sp])
\n else:
\n cc = re.split('(\\d+)', c)
\n for b in cc:
\n if b.isnumeric()==False and len(b)>=2:
\n st=0
\n for i in range(len(b)+1):
\n if b[st:i] in TOKEN_LIST:
\n token.append(tok_2_int[b[st:i]])
\n st=i
\n else:
\n if len(b)>0:
\n token.append(tok_2_int[b])
\n else:
\n if len(c)>0:
\n token.append(tok_2_int[c])
\n token.append(tok_2_int[\"\"])
\n return token
Training
\nDuring training, it took almost 30 min for 1 epoch using the batch of size 8 on Colab GPU.
", "rawMarkdown": "I have created a python notebook for the molecular translation competition. I have used the visual attention model to generate the InChI name from the molecular structure image. This model has two parts, encoder, and decoder. For the encoder, I used a pre-trained Inceptionv3 model trained on the Imagenet dataset. For the decoder, I used the RNN model with GRU unit and attention mechanism.\n\nThe same type of model is present on the Tensorflow website for Image Captioning: https://www.tensorflow.org/tutorials/text/image_captioning\n\nWe can change the hyperparameters and the pre-trained model for encoding.\n\nI used the following parameters -\n\nBATCH_SIZE = 8\nBUFFER_SIZE = 1000\nembedding_dim = 256\nunits = 512\nvocab_size = 198\nnum_steps = len(images) // BATCH_SIZE\nfeatures_shape = 2048\nattention_features_shape = 64\n\n**Tokenization of InChI names**\n\nThe list of tokens that I used - \n\nTOKEN_LIST = [\"