fix typo and fp32 loading issue (#1294)

Author: xin3he

examples/pytorch/nlp/huggingface_models/question-answering/quantization/ptq_static/fx/run_qa.py

@@ -616,9 +616,9 @@ def take_eval_steps(model, trainer, metric_name, save_metrics=False):
         if save_metrics:
             trainer.save_metrics("eval", metrics)
         logger.info("metrics keys: {}".format(metrics.keys()))
-        print('Batch size = %d', batch_size)
+        print('Batch size = %d' % batch_size)
         print("Finally Eval {} Accuracy: {}".format(metric_name, metrics.get(metric_name)))
-        print("Latency: %.3f ms", (evalTime / samples * 1000))
+        print("Latency: %.3f ms" % (evalTime / samples * 1000))
         print("Throughput: {} samples/sec".format(samples / evalTime))
         return metrics.get(metric_name)
 

examples/pytorch/nlp/huggingface_models/question-answering/quantization/ptq_static/ipex/run_qa.py

@@ -620,9 +620,9 @@ def take_eval_steps(model, trainer, metric_name, save_metrics=False):
         if save_metrics:
          trainer.save_metrics("eval", metrics)
         logger.info("metrics keys: {}".format(metrics.keys()))
-        print('Batch size = %d', batch_size)
+        print('Batch size = %d' % batch_size)
         print("Finally Eval {} Accuracy: {}".format(metric_name, metrics.get(metric_name)))
-        print("Latency: %.3f ms", (evalTime / samples * 1000))
+        print("Latency: %.3f ms" % (evalTime / samples * 1000))
         print("Throughput: {} samples/sec".format(samples / evalTime))
         return metrics.get(metric_name)
 

examples/pytorch/nlp/huggingface_models/text-classification/quantization/ptq_dynamic/eager/run_benchmark.sh

@@ -103,13 +103,14 @@ function run_benchmark {
         model_name_or_path=$input_model
     fi
 
+    extra_cmd='--model_name_or_path '${input_model}
     if [[ ${int8} == "true" ]]; then
+        extra_cmd='--model_name_or_path '${tuned_checkpoint}
         extra_cmd=$extra_cmd" --int8"
     fi
     echo $extra_cmd
 
     python -u run_glue_tune.py \
-        --model_name_or_path ${tuned_checkpoint} \
         --task_name ${TASK_NAME} \
         --do_eval \
         --max_seq_length ${MAX_SEQ_LENGTH} \

examples/pytorch/nlp/huggingface_models/text-classification/quantization/ptq_dynamic/fx/run_benchmark.sh

@@ -71,13 +71,14 @@ function run_benchmark {
 
     fi
 
+    extra_cmd='--model_name_or_path '${input_model}
     if [[ ${int8} == "true" ]]; then
+        extra_cmd='--model_name_or_path '${tuned_checkpoint}
         extra_cmd=$extra_cmd" --int8"
     fi
     echo $extra_cmd
 
     python -u run_glue.py \
-        --model_name_or_path ${tuned_checkpoint} \
         --task_name ${TASK_NAME} \
         --do_eval \
         --max_seq_length ${MAX_SEQ_LENGTH} \

examples/pytorch/nlp/huggingface_models/text-classification/quantization/ptq_dynamic/fx/run_glue.py

@@ -486,9 +486,9 @@ def take_eval_steps(model, trainer, save_metrics=False):
         for key in bert_task_acc_keys:
             if key in metrics.keys():
                 throughput = metrics.get("eval_samples_per_second")
-                print('Batch size = %d', batch_size)
+                print('Batch size = %d' % batch_size)
                 print("Finally Eval {} Accuracy: {}".format(key, metrics[key]))
-                print("Latency: %.3f ms", (1000 / throughput))
+                print("Latency: %.3f ms" % (1000 / throughput))
                 print("Throughput: {} samples/sec".format(throughput))
                 return metrics[key]
         assert False, "No metric returned, Please check inference metric!"

examples/pytorch/nlp/huggingface_models/text-classification/quantization/ptq_static/fx/run_benchmark.sh

@@ -100,13 +100,14 @@ function run_benchmark {
         model_name_or_path=${input_model}
     fi
 
+    extra_cmd='--model_name_or_path '${input_model}
     if [[ ${int8} == "true" ]]; then
+        extra_cmd='--model_name_or_path '${tuned_checkpoint}
         extra_cmd=$extra_cmd" --int8"
     fi
     echo $extra_cmd
 
     python -u run_glue.py \
-        --model_name_or_path ${tuned_checkpoint} \
         --task_name ${TASK_NAME} \
         --do_eval \
         --max_seq_length ${MAX_SEQ_LENGTH} \

examples/pytorch/nlp/huggingface_models/text-classification/quantization/ptq_static/fx/run_glue.py

@@ -486,9 +486,9 @@ def take_eval_steps(model, trainer, save_metrics=False):
         for key in bert_task_acc_keys:
             if key in metrics.keys():
                 throughput = metrics.get("eval_samples_per_second")
-                print('Batch size = %d', batch_size)
+                print('Batch size = %d' % batch_size)
                 print("Finally Eval {} Accuracy: {}".format(key, metrics[key]))
-                print("Latency: %.3f ms", (1000 / throughput))
+                print("Latency: %.3f ms" % (1000 / throughput))
                 print("Throughput: {} samples/sec".format(throughput))
                 return metrics[key]
         assert False, "No metric returned, Please check inference metric!"

examples/pytorch/nlp/huggingface_models/text-classification/quantization/qat/fx/README.md

@@ -78,7 +78,7 @@ or
 or
 
     python run_glue_tune.py \
-        --model_name_or_path ${tuned_checkpoint} \
+        --model_name_or_path ${input_model}/${tuned_checkpoint} \
         --task_name ${task_name} \
         --do_train \
         --do_eval \

examples/pytorch/nlp/huggingface_models/text-classification/quantization/qat/fx/run_benchmark.sh

@@ -58,12 +58,13 @@ function run_benchmark {
         mode_cmd="--benchmark "
     fi
 
+    extra_cmd='--model_name_or_path '${input_model}
     if [[ ${int8} == "true" ]]; then
-        mode_cmd=$mode_cmd"--int8"
+        extra_cmd='--model_name_or_path '${tuned_checkpoint}
+        extra_cmd=$extra_cmd" --int8"
     fi
 
     python run_glue_tune.py \
-        --model_name_or_path ${tuned_checkpoint} \
         --task_name ${task_name} \
         --do_train \
         --do_eval \
@@ -74,6 +75,7 @@ function run_benchmark {
         --num_train_epochs 3 \
         --metric_for_best_model f1 \
         --output_dir ./output_log --overwrite_output_dir \
+        ${extra_cmd} \
         ${mode_cmd}
 }