Feature/plotting (#220)

Adding a flag to save metrics to a json file and adding a script to plot the metrics of the given file with matplotlib.

README.md

@@ -110,6 +110,8 @@ All the parameters in the examples and recipes below need to be further tuned to
 
 * Make sure to set the right path to the model in the [training config](src/llama_recipes/configs/training.py).
 
+* To save the loss and perplexity metrics for evaluation, enable this by passing `--save_metrics` to the finetuning script. The file can be plotted using the [plot_metrics.py](./examples/plot_metrics.py) script, `python examples/plot_metrics.py --file_path path/to/metrics.json`
+
 ### Single GPU:
 
 ```bash

examples/plot_metrics.py

@@ -0,0 +1,71 @@
+import json
+import matplotlib.pyplot as plt
+import argparse
+import os
+
+def plot_metric(data, metric_name, x_label, y_label, title, colors):
+    plt.figure(figsize=(7, 6))
+
+    plt.plot(data[f'train_epoch_{metric_name}'], label=f'Train Epoch {metric_name.capitalize()}', color=colors[0])
+    plt.plot(data[f'val_epoch_{metric_name}'], label=f'Validation Epoch {metric_name.capitalize()}', color=colors[1])
+    plt.xlabel(x_label)
+    plt.ylabel(y_label)
+    plt.title(f'Train and Validation Epoch {title}')
+    plt.legend()
+    plt.tight_layout()
+
+def plot_single_metric_by_step(data, metric_name, x_label, y_label, title, color):
+    plt.plot(data[f'{metric_name}'], label=f'{title}', color=color)
+    plt.xlabel(x_label)
+    plt.ylabel(y_label)
+    plt.title(title)
+    plt.legend()
+    plt.tight_layout()
+
+def plot_metrics_by_step(data, metric_name, x_label, y_label, colors):
+    plt.figure(figsize=(14, 6))
+
+    plt.subplot(1, 2, 1)
+    plot_single_metric_by_step(data, f'train_step_{metric_name}', x_label, y_label, f'Train Step {metric_name.capitalize()}', colors[0])
+    plt.subplot(1, 2, 2)
+    plot_single_metric_by_step(data, f'val_step_{metric_name}', x_label, y_label, f'Validation Step {metric_name.capitalize()}', colors[1])
+    plt.tight_layout()
+
+
+def plot_metrics(file_path):
+    if not os.path.exists(file_path):
+        print(f"File {file_path} does not exist.")
+        return
+
+    with open(file_path, 'r') as f:
+        try:
+            data = json.load(f)
+        except json.JSONDecodeError:
+            print("Invalid JSON file.")
+            return
+
+    directory = os.path.dirname(file_path)
+    filename_prefix = os.path.basename(file_path).split('.')[0]
+
+    plot_metric(data, 'loss', 'Epoch', 'Loss', 'Loss', ['b', 'r'])
+    plt.savefig(os.path.join(directory, f"{filename_prefix}_train_and_validation_loss.png"))
+    plt.close()
+
+    plot_metric(data, 'perplexity', 'Epoch', 'Perplexity', 'Perplexity', ['g', 'm'])
+    plt.savefig(os.path.join(directory, f"{filename_prefix}_train_and_validation_perplexity.png"))
+    plt.close()
+
+    plot_metrics_by_step(data, 'loss', 'Step', 'Loss', ['b', 'r'])
+    plt.savefig(os.path.join(directory, f"{filename_prefix}_train_and_validation_loss_by_step.png"))
+    plt.close()
+
+    plot_metrics_by_step(data, 'perplexity', 'Step', 'Loss', ['g', 'm'])
+    plt.savefig(os.path.join(directory, f"{filename_prefix}_train_and_validation_perplexity_by_step.png"))
+    plt.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Plot metrics from JSON file.')
+    parser.add_argument('--file_path', required=True, type=str, help='Path to the metrics JSON file.')
+    args = parser.parse_args()
+
+    plot_metrics(args.file_path)

requirements.txt

@@ -12,4 +12,5 @@ transformers>=4.34.1
 sentencepiece
 py7zr
 scipy
-optimum
+optimum
+matplotlib

src/llama_recipes/configs/training.py

@@ -38,3 +38,4 @@ class train_config:
     dist_checkpoint_folder: str="fine-tuned" # will be used if using FSDP
     save_optimizer: bool=False # will be used if using FSDP
     use_fast_kernels: bool = False # Enable using SDPA from PyTroch Accelerated Transformers, make use Flash Attention and Xformer memory-efficient kernels
+    save_metrics: bool = False # saves training metrics to a json file for later plotting

src/llama_recipes/utils/train_utils.py

@@ -7,6 +7,7 @@
 from contextlib import nullcontext
 from pathlib import Path
 from pkg_resources import packaging
+from datetime import datetime
 
 
 import torch
@@ -16,6 +17,7 @@
 from torch.distributed.fsdp.sharded_grad_scaler import ShardedGradScaler
 from tqdm import tqdm
 from transformers import LlamaTokenizer
+import json
 
 
 from llama_recipes.model_checkpointing import save_model_checkpoint, save_model_and_optimizer_sharded, save_optimizer_checkpoint
@@ -55,13 +57,24 @@ def train(model, train_dataloader,eval_dataloader, tokenizer, optimizer, lr_sche
     elif train_config.use_fp16 and not train_config.enable_fsdp:
         scaler = torch.cuda.amp.GradScaler()
     if train_config.enable_fsdp:
-        world_size = int(os.environ["WORLD_SIZE"])
+        world_size = int(os.environ["WORLD_SIZE"]) 
+
+
+
     autocast = torch.cuda.amp.autocast if train_config.use_fp16 else nullcontext
 
     train_prep = []
     train_loss = []
     val_prep = []
     val_loss =[]
+
+    if train_config.save_metrics:
+        metrics_filename = f"{train_config.output_dir}/metrics_data_{local_rank}-{datetime.now().strftime('%Y-%m-%d_%H-%M-%S')}.json"
+        train_step_perplexity = []
+        train_step_loss = []
+        val_step_loss = []
+        val_step_perplexity = []
+
     epoch_times = []
     checkpoint_times = []
     results = {}
@@ -82,6 +95,9 @@ def train(model, train_dataloader,eval_dataloader, tokenizer, optimizer, lr_sche
                 with autocast():
                     loss = model(**batch).loss
                 loss = loss / gradient_accumulation_steps
+                if train_config.save_metrics:
+                    train_step_loss.append(loss.detach().float().item())
+                    train_step_perplexity.append(float(torch.exp(loss.detach().float())))
                 total_loss += loss.detach().float()
                 if train_config.use_fp16:
                     # if fp16 is enabled, use gradient scaler to handle gradient update
@@ -111,6 +127,9 @@ def train(model, train_dataloader,eval_dataloader, tokenizer, optimizer, lr_sche
                         pbar.update(1)
 
                 pbar.set_description(f"Training Epoch: {epoch+1}/{train_config.num_epochs}, step {step}/{len(train_dataloader)} completed (loss: {loss.detach().float()})")
+
+                if train_config.save_metrics:
+                    save_to_json(metrics_filename, train_step_loss, train_loss, train_step_perplexity, train_prep, val_step_loss, val_loss, val_step_perplexity, val_prep)
             pbar.close()
 
         epoch_end_time = time.perf_counter()-epoch_start_time
@@ -122,10 +141,10 @@ def train(model, train_dataloader,eval_dataloader, tokenizer, optimizer, lr_sche
         if train_config.enable_fsdp:
             train_epoch_loss = train_epoch_loss/world_size
         train_perplexity = torch.exp(train_epoch_loss)
-
-        train_prep.append(train_perplexity)
-        train_loss.append(train_epoch_loss)
-
+        
+        train_prep.append(float(train_perplexity))
+        train_loss.append(float(train_epoch_loss))
+        
         if train_config.enable_fsdp:
             if rank==0:
                 print(f"Max CUDA memory allocated was {memtrace.peak} GB")
@@ -144,7 +163,11 @@ def train(model, train_dataloader,eval_dataloader, tokenizer, optimizer, lr_sche
         lr_scheduler.step()
 
         if train_config.run_validation:
-            eval_ppl, eval_epoch_loss = evaluation(model, train_config, eval_dataloader, local_rank, tokenizer)
+            eval_ppl, eval_epoch_loss, temp_val_loss, temp_step_perplexity = evaluation(model, train_config, eval_dataloader, local_rank, tokenizer)
+            if train_config.save_metrics:
+                val_step_loss.extend(temp_val_loss)
+                val_step_perplexity.extend(temp_step_perplexity)
+
             checkpoint_start_time = time.perf_counter()
             if train_config.save_model and eval_epoch_loss < best_val_loss:
                 if train_config.enable_fsdp:
@@ -195,13 +218,18 @@ def train(model, train_dataloader,eval_dataloader, tokenizer, optimizer, lr_sche
                         print(f"best eval loss on epoch {epoch+1} is {best_val_loss}")
                 else:
                     print(f"best eval loss on epoch {epoch+1} is {best_val_loss}")
-            val_loss.append(best_val_loss)
-            val_prep.append(eval_ppl)
+            val_loss.append(float(best_val_loss))
+            val_prep.append(float(eval_ppl))
         if train_config.enable_fsdp:
             if rank==0:
                 print(f"Epoch {epoch+1}: train_perplexity={train_perplexity:.4f}, train_epoch_loss={train_epoch_loss:.4f}, epoch time {epoch_end_time}s")
         else:
             print(f"Epoch {epoch+1}: train_perplexity={train_perplexity:.4f}, train_epoch_loss={train_epoch_loss:.4f}, epoch time {epoch_end_time}s")
+
+        # Saving the results every epoch to plot later
+        if train_config.save_metrics:
+            save_to_json(metrics_filename, train_step_loss, train_loss, train_step_perplexity, train_prep, val_step_loss, val_loss, val_step_perplexity, val_prep)
+
     avg_epoch_time = sum(epoch_times)/ len(epoch_times)
     avg_checkpoint_time = sum(checkpoint_times)/ len(checkpoint_times) if len(checkpoint_times) > 0 else 0
     avg_train_prep = sum(train_prep)/len(train_prep)
@@ -217,6 +245,8 @@ def train(model, train_dataloader,eval_dataloader, tokenizer, optimizer, lr_sche
         results['avg_eval_loss'] = avg_eval_loss
     results["avg_epoch_time"] = avg_epoch_time
     results["avg_checkpoint_time"] = avg_checkpoint_time
+    if train_config.save_metrics:
+        results["metrics_filename"] = metrics_filename
 
     #saving the training params including fsdp setting for reference.
     if train_config.enable_fsdp and not train_config.use_peft:
@@ -240,6 +270,8 @@ def evaluation(model,train_config, eval_dataloader, local_rank, tokenizer):
         world_size = int(os.environ["WORLD_SIZE"])
     model.eval()
     eval_preds = []
+    val_step_loss = []
+    val_step_perplexity = []
     eval_loss = 0.0  # Initialize evaluation loss
     with MemoryTrace() as memtrace:
         for step, batch in enumerate(tqdm(eval_dataloader,colour="green", desc="evaluating Epoch", dynamic_ncols=True)):
@@ -253,6 +285,10 @@ def evaluation(model,train_config, eval_dataloader, local_rank, tokenizer):
                 # Forward pass and compute loss
                 outputs = model(**batch)
                 loss = outputs.loss
+                if train_config.save_metrics:
+                    val_step_loss.append(loss.detach().float().item())
+                    val_step_perplexity.append(float(torch.exp(loss.detach().float())))  
+
                 eval_loss += loss.detach().float()
             # Decode predictions and add to evaluation predictions list
             preds = torch.argmax(outputs.logits, -1)
@@ -276,8 +312,8 @@ def evaluation(model,train_config, eval_dataloader, local_rank, tokenizer):
             print(f" {eval_ppl=} {eval_epoch_loss=}")
     else:
         print(f" {eval_ppl=} {eval_epoch_loss=}")
-
-    return eval_ppl, eval_epoch_loss
+        
+    return eval_ppl, eval_epoch_loss, val_step_loss, val_step_perplexity
 
 def freeze_transformer_layers(model, num_layer):
    for i, layer in enumerate(model.model.layers):
@@ -417,3 +453,17 @@ def save_train_params(train_config, fsdp_config, rank):
             f.write(config_yaml)
         if rank==0:
             print(f"training params are saved in {file_name}")
+
+def save_to_json(output_filename, train_step_loss, train_epoch_loss, train_step_ppl, train_epoch_ppl, val_step_loss, val_epoch_loss, val_step_ppl, val_epoch_ppl):
+    metrics_data = {
+        "train_step_loss": train_step_loss,
+        "train_epoch_loss": train_epoch_loss,
+        "train_step_perplexity": train_step_ppl,
+        "train_epoch_perplexity": train_epoch_ppl,
+        "val_step_loss": val_step_loss,
+        "val_epoch_loss": val_epoch_loss,
+        "val_step_perplexity": val_step_ppl,
+        "val_epoch_perplexity": val_epoch_ppl
+    }
+    with open(output_filename, "w") as f:
+        json.dump(metrics_data, f)

tests/test_train_utils.py

@@ -2,11 +2,27 @@
 # This software may be used and distributed according to the terms of the Llama 2 Community License Agreement.
 
 from unittest.mock import patch
+import pytest
 
 import torch
 
+import os
+import shutil
+
 from llama_recipes.utils.train_utils import train
 
+TEMP_OUTPUT_DIR = os.getcwd() + "/tmp"
+
+@pytest.fixture(scope="session")
+def temp_output_dir():
+    # Create the directory during the session-level setup
+    temp_output_dir = "tmp"
+    os.mkdir(os.path.join(os.getcwd(), temp_output_dir))
+    yield temp_output_dir
+    # Delete the directory during the session-level teardown
+    shutil.rmtree(temp_output_dir)
+
+
 @patch("llama_recipes.utils.train_utils.MemoryTrace")
 @patch("llama_recipes.utils.train_utils.nullcontext")
 @patch("llama_recipes.utils.train_utils.torch.cuda.amp.GradScaler")
@@ -28,6 +44,7 @@ def test_gradient_accumulation(autocast, scaler, nullcontext, mem_trace, mocker)
     train_config.use_fp16 = False
     train_config.run_validation = False
     train_config.gradient_clipping = False
+    train_config.save_metrics = False
 
     train(
         model,
@@ -63,3 +80,39 @@ def test_gradient_accumulation(autocast, scaler, nullcontext, mem_trace, mocker)
     assert optimizer.zero_grad.call_count == 3
     assert nullcontext.call_count == 0
     assert autocast.call_count == 5
+
+def test_save_to_json(temp_output_dir, mocker):
+    model = mocker.MagicMock(name="model")
+    model().loss.__truediv__().detach.return_value = torch.tensor(1)
+    mock_tensor = mocker.MagicMock(name="tensor")
+    batch = {"input": mock_tensor}
+    train_dataloader = [batch, batch, batch, batch, batch]
+    eval_dataloader = None
+    tokenizer = mocker.MagicMock()
+    optimizer = mocker.MagicMock()
+    lr_scheduler = mocker.MagicMock()
+    gradient_accumulation_steps = 1
+    train_config = mocker.MagicMock()
+    train_config.enable_fsdp = False
+    train_config.use_fp16 = False
+    train_config.run_validation = False
+    train_config.gradient_clipping = False
+    train_config.save_metrics = True
+    train_config.output_dir = temp_output_dir
+
+    results = train(
+        model,
+        train_dataloader,
+        eval_dataloader,
+        tokenizer,
+        optimizer,
+        lr_scheduler,
+        gradient_accumulation_steps,
+        train_config,
+        local_rank=0
+    )
+
+    assert results["metrics_filename"] not in ["", None]
+    assert os.path.isfile(results["metrics_filename"])
+
+