Merge branch 'master' of https://gitlab.irit.fr/pnria/global-helper/deepgrail-rnn

# Conflicts:
#	.gitignore

Configuration/config.ini

@@ -5,17 +5,17 @@ dim_encoder = 768
 dim_decoder = 768
 num_rnn_layers=1
 dropout=0.1
-teacher_forcing=0.8
+teacher_forcing=0.05
 symbols_vocab_size=26
 max_symbols_in_sentence=1250
 max_len_sentence=112
 [MODEL_TRAINING]
 device=cpu
-batch_size=16
+batch_size=32
 epoch=20
 seed_val=42
 learning_rate=0.005
-use_checkpoint_SAVE=1
+use_checkpoint_SAVE=0
 output_path=Output
 use_checkpoint_LOAD=0
 input_path=Input

SuperTagger/Decoder/RNNDecoderLayer.py

@@ -2,7 +2,7 @@ import random
 
 import torch
 import torch.nn.functional as F
-from torch.nn import (Dropout, Module, Linear, LSTM)
+from torch.nn import (Dropout, Module, ModuleList, Linear, LSTM, GRU)
 
 from Configuration import Configuration
 from SuperTagger.Symbol.SymbolEmbedding import SymbolEmbedding
@@ -18,6 +18,8 @@ class RNNDecoderLayer(Module):
         self.symbols_vocab_size = int(Configuration.modelDecoderConfig['symbols_vocab_size'])
         dropout = float(Configuration.modelDecoderConfig['dropout'])
         self.num_rnn_layers = int(Configuration.modelDecoderConfig['num_rnn_layers'])
+        self.teacher_forcing = float(Configuration.modelDecoderConfig['teacher_forcing'])
+
         self.bidirectional = False
         self.symbols_map = symbols_map
         self.symbols_padding_id = self.symbols_map["[PAD]"]
@@ -25,8 +27,6 @@ class RNNDecoderLayer(Module):
         self.symbols_start_id = self.symbols_map["[START]"]
         self.symbols_sos_id = self.symbols_map["[SOS]"]
 
-        self.teacher_forcing = float(Configuration.modelDecoderConfig['teacher_forcing'])
-
         # Different layers
         # Symbols Embedding
         self.symbols_embedder = SymbolEmbedding(self.hidden_dim, self.symbols_vocab_size,
@@ -35,7 +35,14 @@ class RNNDecoderLayer(Module):
         self.dropout = Dropout(dropout)
         # rnn Layer
         self.rnn = LSTM(input_size=self.hidden_dim, hidden_size=self.hidden_dim, num_layers=self.num_rnn_layers,
+                        dropout=dropout,
                         bidirectional=self.bidirectional, batch_first=True)
+
+        self.intermediate = ModuleList()
+        for _ in range(3):
+            self.intermediate.append(Linear(self.hidden_dim, self.hidden_dim))
+        self.activation = F.gelu
+
         # Projection on vocab_size
         if self.bidirectional:
             self.proj = Linear(self.hidden_dim * 2, self.symbols_vocab_size)
@@ -57,22 +64,22 @@ class RNNDecoderLayer(Module):
 
         # y_hat[batch_size, max_len_sentence, vocab_size] init with probability pad =1
         y_hat = torch.zeros(batch_size, self.max_symbols_in_sentence, self.symbols_vocab_size,
-                            dtype=torch.float,device="cuda" if torch.cuda.is_available() else "cpu")
+                            dtype=torch.float, device="cuda" if torch.cuda.is_available() else "cpu")
         y_hat[:, :, self.symbols_padding_id] = 1
-        use_teacher_forcing = True if random.random() < 0.05 else False
-        if use_teacher_forcing:
-            print("\n FORCING TEACHING \n")
-            decoded_i = symbols_tokenized_batch[:, 0].unsqueeze(1)
-        else :
-            decoded_i = torch.ones(batch_size, 1, dtype=torch.long,device="cuda" if torch.cuda.is_available() else "cpu")* self.symbols_start_id
 
-        sos_mask = torch.zeros(batch_size, dtype=torch.bool,device="cuda" if torch.cuda.is_available() else "cpu")
+        decoded_i = torch.ones(batch_size, 1, dtype=torch.long,
+                               device="cuda" if torch.cuda.is_available() else "cpu") * self.symbols_start_id
+
+        sos_mask = torch.zeros(batch_size, dtype=torch.bool, device="cuda" if torch.cuda.is_available() else "cpu")
 
         # hidden_state goes through multiple linear layers
         hidden_state = pooler_output.unsqueeze(0).repeat(self.num_rnn_layers * (1 + self.bidirectional), 1, 1)
         c_state = torch.zeros(self.num_rnn_layers * (1 + self.bidirectional), batch_size, hidden_size,
-                              dtype=torch.float,device="cuda" if torch.cuda.is_available() else "cpu")
+                             dtype=torch.float, device="cuda" if torch.cuda.is_available() else "cpu")
+        for intermediate in self.intermediate:
+            hidden_state = self.dropout(self.activation(intermediate(hidden_state)))
 
+        use_teacher_forcing = True if random.random() < self.teacher_forcing else False
 
         # for each symbol
         for i in range(self.max_symbols_in_sentence):
@@ -97,8 +104,6 @@ class RNNDecoderLayer(Module):
             y_hat[~sos_mask, i, :-2] = proj[~sos_mask, -1, :]
             sos_mask = sos_mask_i | sos_mask
 
-            use_teacher_forcing = True if random.random() < 0.25 else False
-
             # Stop if every sentence says padding or if we are full
             if not torch.any(~sos_mask):
                 break
@@ -116,20 +121,23 @@ class RNNDecoderLayer(Module):
 
         # contains the predictions
         y_hat = torch.zeros(batch_size, self.max_symbols_in_sentence, self.symbols_vocab_size,
-                            dtype=torch.float,device="cuda" if torch.cuda.is_available() else "cpu")
+                            dtype=torch.float, device="cuda" if torch.cuda.is_available() else "cpu")
         y_hat[:, :, self.symbols_padding_id] = 1
         # input of the embedder, a created vector that replace the true value
-        decoded_i = torch.ones(batch_size, 1, dtype=torch.long,device="cuda" if torch.cuda.is_available() else "cpu") * self.symbols_start_id
+        decoded_i = torch.ones(batch_size, 1, dtype=torch.long,
+                               device="cuda" if torch.cuda.is_available() else "cpu") * self.symbols_start_id
 
-        sos_mask = torch.zeros(batch_size, dtype=torch.bool,device="cuda" if torch.cuda.is_available() else "cpu")
+        sos_mask = torch.zeros(batch_size, dtype=torch.bool, device="cuda" if torch.cuda.is_available() else "cpu")
 
         hidden_state = pooler_output.unsqueeze(0).repeat(self.num_rnn_layers * (1 + self.bidirectional), 1, 1)
         c_state = torch.zeros(self.num_rnn_layers * (1 + self.bidirectional), batch_size, hidden_size,
-                              dtype=torch.float,device="cuda" if torch.cuda.is_available() else "cpu")
+                             dtype=torch.float, device="cuda" if torch.cuda.is_available() else "cpu")
+        for intermediate in self.intermediate:
+            hidden_state = self.dropout(self.activation(intermediate(hidden_state)))
 
         for i in range(self.max_symbols_in_sentence):
             symbols_embedding = self.symbols_embedder(decoded_i)
-            #symbols_embedding = self.dropout(symbols_embedding)
+            symbols_embedding = self.dropout(symbols_embedding)
 
             output, (hidden_state, c_state) = self.rnn(symbols_embedding, (hidden_state, c_state))
 

SuperTagger/EncoderDecoder.py

@@ -41,7 +41,7 @@ class EncoderDecoder(Module):
             symbols_tokenized_batch: [batch_size, max_symbols_in_sentence] the true symbols for each sentence.
         """
         last_hidden_state, pooler_output = self.encoder([sents_tokenized_batch, sents_mask_batch])
-        #last_hidden_state = self.dropout(last_hidden_state)
+        last_hidden_state = self.dropout(last_hidden_state)
         return self.decoder(symbols_tokenized_batch, last_hidden_state, pooler_output)
 
     def decode_greedy_rnn(self, sents_tokenized_batch, sents_mask_batch):
@@ -52,7 +52,7 @@ class EncoderDecoder(Module):
             sents_mask_batch : mask output from the encoder tokenizer
         """
         last_hidden_state, pooler_output = self.encoder([sents_tokenized_batch, sents_mask_batch])
-        #last_hidden_state = self.dropout(last_hidden_state)
+        last_hidden_state = self.dropout(last_hidden_state)
 
         predictions = self.decoder.predict_rnn(last_hidden_state, pooler_output)
 
@@ -78,10 +78,11 @@ class EncoderDecoder(Module):
         print("\nsub true (", l, ") : ",
               [token for token in true_trad if token != '[PAD]'])
         print("\nsub predict (", len([i for i in predict_trad if i != '[PAD]']), ") : ",
-              [token for token in predict_trad[:l] if token != '[PAD]'])
+              [token for token in predict_trad if token != '[PAD]'])
 
         return measure_supertagging_accuracy(pred, b_symbols_tokenized,
-                                             ignore_idx=self.symbols_map["[PAD]"]), float(cross_entropy_loss(type_predictions, b_symbols_tokenized))
+                                             ignore_idx=self.symbols_map["[PAD]"]), float(
+            cross_entropy_loss(type_predictions, b_symbols_tokenized))
 
     def eval_epoch(self, dataloader, cross_entropy_loss):
         r"""Average the evaluation of all the batch.
@@ -93,7 +94,7 @@ class EncoderDecoder(Module):
 
         for step, batch in enumerate(dataloader):
             batch = batch
-            batch_output , loss = self.eval_batch(batch, cross_entropy_loss)
+            batch_output, loss = self.eval_batch(batch, cross_entropy_loss)
             ((bs_correct, bs_total), (bw_correct, bw_total)) = batch_output
             s_total += bs_total
             s_correct += bs_correct

SuperTagger/utils.py

@@ -5,7 +5,6 @@ import torch
 from tqdm import tqdm
 
 
-
 def read_csv_pgbar(csv_path, nrows=float('inf'), chunksize=500):
     print("\n" + "#" * 20)
     print("Loading csv...")
@@ -38,8 +37,8 @@ def format_time(elapsed):
     # Format as hh:mm:ss
     return str(datetime.timedelta(seconds=elapsed_rounded))
 
-def checkpoint_save(model, opt, epoch, dir, loss):
 
+def checkpoint_save(model, opt, epoch, dir, loss):
     torch.save({
         'epoch': epoch,
         'model_state_dict': model.state_dict(),
@@ -51,7 +50,7 @@ def checkpoint_save(model, opt, epoch, dir, loss):
 def checkpoint_load(model, opt, path):
     epoch = 0
     loss = 0
-    print("#" *15)
+    print("#" * 15)
 
     try:
         checkpoint = torch.load(path)
@@ -62,8 +61,6 @@ def checkpoint_load(model, opt, path):
         print("\n The loading checkpoint was successful  ! \n")
         print("#" * 10)
     except Exception as e:
-        print("\nCan't load checkpoint model because : "+ str(e) +"\n\nUse default model \n")
+        print("\nCan't load checkpoint model because : " + str(e) + "\n\nUse default model \n")
     print("#" * 15)
     return model, opt, epoch, loss
-
-

bash_GPU.sh

@@ -4,10 +4,10 @@
 #SBATCH --gres=gpu:1
 #SBATCH --mem=32000
 #SBATCH --gres-flags=enforce-binding
-#SBATCH --error="/users/celdev/jrabault/PNRIA - DeepGrail/OUT/error_rtx1.err"
-#SBATCH --output="/users/celdev/jrabault/PNRIA - DeepGrail/OUT/out_rtx1.out"
+#SBATCH --error="error_rtx1.err"
+#SBATCH --output="out_rtx1.out"
 
 module purge
 module load singularity/3.0.3
 
-srun singularity exec /logiciels/containerCollections/CUDA11/pytorch-NGC-21-03-py3.sif python "/users/celdev/jrabault/PNRIA - DeepGrail/train.py" 
+srun singularity exec /logiciels/containerCollections/CUDA11/pytorch-NGC-21-03-py3.sif python "train.py"
\ No newline at end of file

main.py

-# This is a sample Python script.
-
-# Press Maj+F10 to execute it or replace it with your code.
-# Press Double Shift to search everywhere for classes, files, tool windows, actions, and settings.
-
-
-def print_hi(name):
-    # Use a breakpoint in the code line below to debug your script.
-    print(f'Hi, {name}')  # Press Ctrl+F8 to toggle the breakpoint.
-
-
-# Press the green button in the gutter to run the script.
-if __name__ == '__main__':
-    print_hi('PyCharm')
-
-# See PyCharm help at https://www.jetbrains.com/help/pycharm/

push pull texte

+ https://gitlab-ci-token:glpat-AZdpzmAPDFCSK8nPZxCw@gitlab.irit.fr/pnria/global-helper/deepgrail-rnn.git
\ No newline at end of file

train.py

-import math
+import os
 import os
 import time
+from datetime import datetime
 
 import numpy as np
 import torch
-from torch.optim import AdamW
-import transformers
 import torch.nn.functional as F
+import transformers
+from torch.optim import SGD
 from torch.utils.data import Dataset, TensorDataset, random_split
 from transformers import (AutoTokenizer, get_cosine_schedule_with_warmup)
 from transformers import (CamembertModel)
-from torch.utils.tensorboard import SummaryWriter
 
 from Configuration import Configuration
 from SuperTagger.Encoder.EncoderInput import EncoderInput
@@ -20,7 +20,7 @@ from SuperTagger.Symbol.symbol_map import symbol_map
 from SuperTagger.eval import NormCrossEntropy
 from SuperTagger.utils import format_time, read_csv_pgbar, checkpoint_save, checkpoint_load
 
-from datetime import datetime
+from torch.utils.tensorboard import SummaryWriter
 
 transformers.TOKENIZERS_PARALLELISM = True
 torch.cuda.empty_cache()
@@ -38,8 +38,7 @@ num_gru_layers = int(Configuration.modelDecoderConfig['num_rnn_layers'])
 
 file_path = 'Datasets/m2_dataset.csv'
 batch_size = int(Configuration.modelTrainingConfig['batch_size'])
-nb_sentences = batch_size * 10
-# Number of training epochs. The BERT authors recommend between 2 and 4.
+nb_sentences = batch_size * 50
 epochs = int(Configuration.modelTrainingConfig['epoch'])
 seed_val = int(Configuration.modelTrainingConfig['seed_val'])
 learning_rate = float(Configuration.modelTrainingConfig['learning_rate'])
@@ -57,7 +56,7 @@ logs_dir = os.path.join(training_dir, 'logs')
 checkpoint_dir = training_dir
 writer = SummaryWriter(log_dir=logs_dir)
 
-use_checkpoint_SAVE = bool(Configuration.modelTrainingConfig['use_checkpoint_SAVE'])
+use_checkpoint_SAVE = bool(Configuration.modelTrainingConfig.getboolean('use_checkpoint_SAVE'))
 
 # endregion OutputTraining
 
@@ -66,8 +65,7 @@ use_checkpoint_SAVE = bool(Configuration.modelTrainingConfig['use_checkpoint_SAV
 input_path = str(Configuration.modelTrainingConfig['input_path'])
 model_to_load = str(Configuration.modelTrainingConfig['model_to_load'])
 model_to_load_path = os.path.join(input_path, model_to_load)
-use_checkpoint_LOAD = bool(Configuration.modelTrainingConfig['use_checkpoint_LOAD'])
-print(use_checkpoint_LOAD)
+use_checkpoint_LOAD = bool(Configuration.modelTrainingConfig.getboolean('use_checkpoint_LOAD'))
 
 # endregion InputTraining
 
@@ -136,7 +134,7 @@ sents_tokenized, sents_mask = sents_tokenizer.fit_transform_tensors(df['Sentence
 dataset = TensorDataset(sents_tokenized, sents_mask, symbols_tokenized)
 
 # Calculate the number of samples to include in each set.
-train_size = int(0.95 * len(dataset))
+train_size = int(0.9 * len(dataset))
 val_size = len(dataset) - train_size
 
 # Divide the dataset by randomly selecting samples.
@@ -153,23 +151,20 @@ validation_dataloader = torch.utils.data.DataLoader(val_dataset, batch_size=batc
 # region Fit tunning
 
 # Optimizer
-# optimizer_encoder = AdamW(model.encoder.parameters(),
-#                           lr=5e-5,
-#                           eps=1e-8)
-# optimizer_decoder = AdamW(model.decoder.parameters(),
-#                           lr=learning_rate,
-#                           eps=1e-8)
-
-optimizer = AdamW(model.parameters(),
-                  lr=learning_rate,
-                  eps=1e-8)
+optimizer_encoder = SGD(model.encoder.parameters(),
+                        lr=5e-5)
+optimizer_decoder = SGD(model.decoder.parameters(),
+                        lr=learning_rate)
 
 # Total number of training steps is [number of batches] x [number of epochs].
 # (Note that this is not the same as the number of training samples).
 total_steps = len(training_dataloader) * epochs
 
 # Create the learning rate scheduler.
-scheduler = get_cosine_schedule_with_warmup(optimizer,
+scheduler_encoder = get_cosine_schedule_with_warmup(optimizer_encoder,
+                                                    num_warmup_steps=0,
+                                                    num_training_steps=5)
+scheduler_decoder = get_cosine_schedule_with_warmup(optimizer_decoder,
                                             num_warmup_steps=0,
                                             num_training_steps=total_steps)
 
@@ -199,7 +194,7 @@ total_t0 = time.time()
 validate = True
 
 if use_checkpoint_LOAD:
-    model, optimizer, last_epoch, loss = checkpoint_load(model, optimizer, model_to_load_path)
+    model, optimizer_decoder, last_epoch, loss = checkpoint_load(model, optimizer_decoder, model_to_load_path)
     epochs = epochs - last_epoch
 
 
@@ -242,46 +237,40 @@ def run_epochs(epochs):
             b_sents_mask = batch[1].to("cuda" if torch.cuda.is_available() else "cpu")
             b_symbols_tokenized = batch[2].to("cuda" if torch.cuda.is_available() else "cpu")
 
-            # optimizer_encoder.zero_grad()
-            # optimizer_decoder.zero_grad()
-
-            optimizer.zero_grad()
+            optimizer_encoder.zero_grad()
+            optimizer_decoder.zero_grad()
 
             logits_predictions = model(b_sents_tokenized, b_sents_mask, b_symbols_tokenized)
 
             predict_trad = [{v: k for k, v in symbol_map.items()}[int(i)] for i in
                             torch.argmax(F.softmax(logits_predictions, dim=2), dim=2)[0]]
             true_trad = [{v: k for k, v in symbol_map.items()}[int(i)] for i in b_symbols_tokenized[0]]
-
             l = len([i for i in true_trad if i != '[PAD]'])
-
             if step % 40 == 0 and not step == 0:
-                writer.add_text("Sample", "\ntrain true (" + str(l) + ") : " + str([token for token in true_trad if token != '[PAD]']) + "\ntrain predict (" + str(len([i for i in predict_trad if i != '[PAD]'])) + ") : " + str([token for token in predict_trad[:l] if token != '[PAD]']))
+                writer.add_text("Sample", "\ntrain true (" + str(l) + ") : " + str(
+                    [token for token in true_trad if token != '[PAD]']) + "\ntrain predict (" + str(
+                    len([i for i in predict_trad if i != '[PAD]'])) + ") : " + str(
+                    [token for token in predict_trad[:l] if token != '[PAD]']))
 
             loss = cross_entropy_loss(logits_predictions, b_symbols_tokenized)
             # Perform a backward pass to calculate the gradients.
             total_train_loss += float(loss)
             loss.backward()
 
-            # Clip the norm of the gradients to 1.0.
             # This is to help prevent the "exploding gradients" problem.
-            # torch.nn.utils.clip_grad_norm_(model.encoder.parameters(), 1.0)
-            # torch.nn.utils.clip_grad_norm_(model.decoder.parameters(), 1.0)
+            torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=5.0, norm_type=2)
 
             # Update parameters and take a step using the computed gradient.
-            # The optimizer dictates the "update rule"--how the parameters are
-            # modified based on their gradients, the learning rate, etc.
-            # optimizer_encoder.step()
-            # optimizer_decoder.step()
+            optimizer_encoder.step()
+            optimizer_decoder.step()
 
-            optimizer.step()
-
-            scheduler.step()
+            scheduler_encoder.step()
+            scheduler_decoder.step()
 
         # checkpoint
 
         if use_checkpoint_SAVE:
-            checkpoint_save(model, optimizer, epoch_i, checkpoint_dir, loss)
+            checkpoint_save(model, optimizer_decoder, epoch_i, checkpoint_dir, loss)
 
         avg_train_loss = total_train_loss / len(training_dataloader)
 
@@ -313,5 +302,3 @@ def run_epochs(epochs):
 
 run_epochs(epochs)
 # endregion Train
-
-