Revised importing QuantSim and updated front readme page

Signed-off-by: Bharath Ramaswamy <quic_bharathr@quicinc.com>

README.md

@@ -233,7 +233,7 @@ An original FP32 source model is quantized either using post-training quantizati
     <td>EfficientNet-lite0</td>
     <td><a href="https://github.com/rwightman/gen-efficientnet-pytorch">GitHub Repo</a></td>
     <td><a href="https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/efficientnet_lite0_ra-37913777.pth">Pretrained Model</a></td>
-    <td><a href="zoo_torch/examples/eval_efficientnetlite0.py">See Example</a></td>
+    <td><a href="/../../releases/download/pt-effnet-checkpoint/adaround_efficient_lite.pth">Quantized Model</a></td>
     <td>(ImageNet) Top-1 Accuracy <br> FP32: 75.42%<br> INT8: 74.44%</td>
     <td><a href="zoo_torch/Docs/EfficientNet-lite0.md">EfficientNet-lite0.md</a></td>
   </tr>

zoo_torch/Docs/MobileNetV2-SSD-lite.md

@@ -16,43 +16,7 @@ git apply ../aimet-model-zoo/zoo_torch/examples/torch_ssd_eval.patch
 mv vision ../aimet-model-zoo/zoo_torch/examples/
 mv eval_ssd.py ../aimet-model-zoo/zoo_torch/examples/
 ```
-3. Change __init__ function from line #27 in vision/ssd/ssd.py as follows:
-```
-self.config = None #############Change 1
-
-self.image_size = 300
-self.image_mean = np.array([127, 127, 127])  # RGB layout
-self.image_std = 128.0
-self.iou_threshold = 0.45
-self.center_variance = 0.1
-self.size_variance = 0.2
-
-self.specs = [box_utils.SSDSpec(19, 16, box_utils.SSDBoxSizes(60, 105), [2, 3]),
-              box_utils.SSDSpec(10, 32, box_utils.SSDBoxSizes(105, 150), [2, 3]),
-              box_utils.SSDSpec(5, 64, box_utils.SSDBoxSizes(150, 195), [2, 3]),
-              box_utils.SSDSpec(3, 100, box_utils.SSDBoxSizes(195, 240), [2, 3]),
-              box_utils.SSDSpec(2, 150, box_utils.SSDBoxSizes(240, 285), [2, 3]),
-              box_utils.SSDSpec(1, 300, box_utils.SSDBoxSizes(285, 330), [2, 3])]
-
-self.gen_priors = box_utils.generate_ssd_priors(self.specs, self.image_size)
 
-# register layers in source_layer_indexes by adding them to a module list
-self.source_layer_add_ons = nn.ModuleList([t[1] for t in source_layer_indexes
-                                           if isinstance(t, tuple) and not isinstance(t, GraphPath)])
-
-if device:
-    self.device = device
-else:
-    self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
-if is_test:
-    self.priors = self.gen_priors.to(self.device)
-```
-4. Change line #93 in vision/ssd/ssd.py as follows:
-```
-boxes = box_utils.convert_locations_to_boxes(
-                locations.cpu(), self.priors.cpu(), self.center_variance, self.size_variance
-)
-```
 
 ## Obtaining model checkpoint and dataset
 - The original MobileNetV2-SSD-lite checkpoint can be downloaded here:

zoo_torch/Docs/MobilenetV2.md

@@ -25,10 +25,6 @@ self.last_channel = int(last_channel * width_mult) if width_mult > 1.0 else last
   - Change line #91 as follows in MobileNetV2.py
 ```
 output_channel = int(c * width_mult)
-```
-  - Append line #100 as follows in MobileNetV2.py
-```
-self.features.append(nn.AvgPool2d(input_size // 32)
 ```
   - Change line #104 as follows in MobileNetV2.py
 ```
@@ -40,6 +36,10 @@ self.classifier = nn.Sequential(
   - Change line #110 as follows in MobileNetV2.py
 ```
 x = x.squeeze()
+```
+  - Append line #100 as follows in MobileNetV2.py
+```
+self.features.append(nn.AvgPool2d(input_size // 32))
 ```
 ## Obtaining model checkpoint and dataset
 
@@ -56,7 +56,6 @@ python eval_mobilenetv2.py \
 	--model-path <path to optimized mobilenetv2 checkpoint> \
 	--images-dir <path to imagenet root directory> \
 	--quant-scheme <quantization schme to run> \
-	--input-shape <input shape to model> \
 	--default-output-bw <bitwidth for activation quantization> \
 	--default-param-bw <bitwidth for weight quantization>     		
 ```

zoo_torch/examples/eval_deeplabv3.py

@@ -47,28 +47,26 @@ def func_wrapper(model, arguments):
     return func_wrapper
 
 
-
-
 def arguments():
     parser = argparse.ArgumentParser(description='Evaluation script for PyTorch ImageNet networks.')
 
     parser.add_argument('--checkpoint-path',            help='Path to optimized checkpoint directory to load from.', default = None, type=str)
-    parser.add_argument('--base-size',				    help='Base size for Random Crop', default=513)
-    parser.add_argument('--crop-size',                  help='Crop size for Random Crop', default=513)
-    parser.add_argument('--num-classes',                help='Number of classes in a dataset', default=21)
-    parser.add_argument('--dataset',                    help='dataset used for evaluation', default='pascal')
+    parser.add_argument('--base-size',				    help='Base size for Random Crop', type = int, default=513)
+    parser.add_argument('--crop-size',                  help='Crop size for Random Crop', type = int, default=513)
+    parser.add_argument('--num-classes',                help='Number of classes in a dataset', type = int, default=21)
+    parser.add_argument('--dataset',                    help='dataset used for evaluation', default='pascal', type = str)
 
     parser.add_argument('--seed',						help='Seed number for reproducibility', default=0)
     parser.add_argument('--use-sbd',                    help='Use SBD data for data augmentation during training', default=False)
 
     parser.add_argument('--quant-scheme',               help='Quant scheme to use for quantization (tf, tf_enhanced, range_learning_tf, range_learning_tf_enhanced).', default='tf', choices = ['tf', 'tf_enhanced', 'range_learning_tf', 'range_learning_tf_enhanced'])
     parser.add_argument('--round-mode',                 help='Round mode for quantization.', default='nearest')
-    parser.add_argument('--default-output-bw',          help='Default output bitwidth for quantization.', default=8)
-    parser.add_argument('--default-param-bw',           help='Default parameter bitwidth for quantization.', default=8)
+    parser.add_argument('--default-output-bw',          help='Default output bitwidth for quantization.', type = int, default=8)
+    parser.add_argument('--default-param-bw',           help='Default parameter bitwidth for quantization.', type = int, default=8)
     parser.add_argument('--config-file',       			help='Quantsim configuration file.', default=None, type=str)
     parser.add_argument('--cuda',						help='Enable cuda for a model', default=True)
 
-    parser.add_argument('--batch-size',					help='Data batch size for a model', default=16)
+    parser.add_argument('--batch-size',					help='Data batch size for a model', type = int, default=16)
     args = parser.parse_args()
     return args
 
@@ -87,13 +85,13 @@ def main():
     model = DeepLab(backbone='mobilenet', output_stride=16, num_classes=21,
                  sync_bn=False)
     model.eval()
-
+    
     from aimet_torch import batch_norm_fold
     from aimet_torch import utils
     args.input_shape = (1,3,513,513)
     batch_norm_fold.fold_all_batch_norms(model, args.input_shape)
     utils.replace_modules_of_type1_with_type2(model, torch.nn.ReLU6, torch.nn.ReLU)
-    # from IPython import embed; embed()
+
     if args.checkpoint_path:
         model.load_state_dict(torch.load(args.checkpoint_path))
     else:
@@ -105,7 +103,7 @@ def main():
     eval_func = model_eval(args, val_loader)
 
     from aimet_common.defs import QuantScheme
-    from aimet_torch.pro.quantsim import QuantizationSimModel
+    from aimet_torch.quantsim import QuantizationSimModel
     if hasattr(args, 'quant_scheme'):
         if args.quant_scheme == 'range_learning_tf':
             quant_scheme = QuantScheme.training_range_learning_with_tf_init

zoo_torch/examples/eval_efficientnetlite0.py

@@ -23,7 +23,7 @@
 from aimet_torch import cross_layer_equalization
 from aimet_torch import batch_norm_fold
 from aimet_common.defs import QuantScheme
-from aimet_torch.pro.quantsim import QuantizationSimModel
+from aimet_torch.quantsim import QuantizationSimModel
 from aimet_torch.onnx_utils import onnx_pytorch_conn_graph_type_pairs
 from aimet_common.utils import AimetLogger
 import logging
@@ -111,19 +111,18 @@ def arguments():
 
     parser.add_argument('--checkpoint',                 help='Path to optimized checkpoint', default=None, type=str)
     parser.add_argument('--images-dir',         		help='Imagenet eval image', default='./ILSVRC2012_PyTorch/', type=str)
-    parser.add_argument('--input-shape',				help='Model to an input image shape, (ex : [batch, channel, width, height]', default=(1,3,224,224))
-    parser.add_argument('--seed',						help='Seed number for reproducibility', default=0)
+    parser.add_argument('--seed',						help='Seed number for reproducibility', type = int, default=0)
 
     parser.add_argument('--quant-tricks', 				help='Preprocessing prior to Quantization', default=[], choices=['BNfold', 'CLE'], nargs = "+")
     parser.add_argument('--quant-scheme',               help='Quant scheme to use for quantization (tf, tf_enhanced, range_learning_tf, range_learning_tf_enhanced).', default='tf', choices = ['tf', 'tf_enhanced', 'range_learning_tf', 'range_learning_tf_enhanced'])
     parser.add_argument('--round-mode',                 help='Round mode for quantization.', default='nearest')
-    parser.add_argument('--default-output-bw',          help='Default output bitwidth for quantization.', default=8)
-    parser.add_argument('--default-param-bw',           help='Default parameter bitwidth for quantization.', default=8)
+    parser.add_argument('--default-output-bw',          help='Default output bitwidth for quantization.', type = int, default=8)
+    parser.add_argument('--default-param-bw',           help='Default parameter bitwidth for quantization.', type = int, default=8)
     parser.add_argument('--config-file',       			help='Quantsim configuration file.', default=None, type=str)
     parser.add_argument('--cuda',						help='Enable cuda for a model', default=True)
 
-    parser.add_argument('--batch-size',					help='Data batch size for a model', default=64)
-    parser.add_argument('--num-workers',                help='Number of workers to run data loader in parallel', default=16)
+    parser.add_argument('--batch-size',					help='Data batch size for a model', type = int, default=64)
+    parser.add_argument('--num-workers',                help='Number of workers to run data loader in parallel', type = int, default=16)
 
     args = parser.parse_args()
     return args
@@ -137,8 +136,9 @@ def main():
     else:
         model = load_model()
     model.eval()
-
-    image_size = args.input_shape[-1]
+    input_shape = (1,3,224,224)
+    args.input_shape = input_shape
+    image_size = input_shape[-1]
 
     data_loader_kwargs = { 'worker_init_fn':work_init, 'num_workers' : args.num_workers}
     normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
@@ -179,7 +179,7 @@ def main():
             'config_file': args.config_file
         }
     print(kwargs)
-    sim = QuantizationSimModel(model.cpu(), input_shapes=args.input_shape, **kwargs)
+    sim = QuantizationSimModel(model.cpu(), input_shapes=input_shape, **kwargs)
 
     # Manually Config Super group, AIMET currently does not support [Conv-ReLU6] in a supergroup
     from aimet_torch.qc_quantize_op import QcPostTrainingWrapper

zoo_torch/examples/eval_mobilenetv2.py

@@ -88,18 +88,17 @@ def arguments():
 
 	parser.add_argument('--model-path',             	help='Path to checkpoint directory to load from.', default = "./model/mv2qat_modeldef.pth", type=str)
 	parser.add_argument('--images-dir',         		help='Imagenet eval image', default='./ILSVRC2012/', type=str)
-	parser.add_argument('--input-shape',				help='Model to an input image shape, (ex : [batch, channel, width, height]', default=(1,3,224,224))
-	parser.add_argument('--seed',						help='Seed number for reproducibility', default=0)
+	parser.add_argument('--seed',						help='Seed number for reproducibility', type = int, default=0)
 
 	parser.add_argument('--quant-tricks', 				help='Preprocessing prior to Quantization', choices=['BNfold', 'CLS', 'HBF', 'CLE', 'BC', 'adaround'], nargs = "+")
 	parser.add_argument('--quant-scheme',               help='Quant scheme to use for quantization (tf, tf_enhanced, range_learning_tf, range_learning_tf_enhanced).', default='tf', choices = ['tf', 'tf_enhanced', 'range_learning_tf', 'range_learning_tf_enhanced'])
 	parser.add_argument('--round-mode',                 help='Round mode for quantization.', default='nearest')
-	parser.add_argument('--default-output-bw',          help='Default output bitwidth for quantization.', default=8)
-	parser.add_argument('--default-param-bw',           help='Default parameter bitwidth for quantization.', default=8)
+	parser.add_argument('--default-output-bw',          help='Default output bitwidth for quantization.', type = int, default=8)
+	parser.add_argument('--default-param-bw',           help='Default parameter bitwidth for quantization.', type = int, default=8)
 	parser.add_argument('--config-file',       			help='Quantsim configuration file.', default=None, type=str)
 	parser.add_argument('--cuda',						help='Enable cuda for a model', default=True)
 
-	parser.add_argument('--batch-size',					help='Data batch size for a model', default=64)
+	parser.add_argument('--batch-size',					help='Data batch size for a model', type = int, default=64)
 
 
 	args = parser.parse_args()
@@ -123,12 +122,13 @@ def main():
         raise ValueError('Model path {} must be specified'.format(args.model_path))
 
     model.eval()
-    image_size = args.input_shape[-1]
+    input_shape = (1,3,224,224)
+    image_size = input_shape[-1]
     eval_func_quant = model_eval(args.images_dir + '/val/', image_size, batch_size=args.batch_size, num_workers=0, quant = True)
     eval_func = model_eval(args.images_dir + '/val/', image_size, batch_size=args.batch_size, num_workers=16)
 
     from aimet_common.defs import QuantScheme
-    from aimet_torch.pro.quantsim import QuantizationSimModel
+    from aimet_torch.quantsim import QuantizationSimModel
     if hasattr(args, 'quant_scheme'):
         if args.quant_scheme == 'range_learning_tf':
             quant_scheme = QuantScheme.training_range_learning_with_tf_init
@@ -147,7 +147,7 @@ def main():
             'config_file': args.config_file
         }
     print(kwargs)
-    sim = QuantizationSimModel(model.cpu(), input_shapes=args.input_shape, **kwargs)
+    sim = QuantizationSimModel(model.cpu(), input_shapes=input_shape, **kwargs)
     sim.compute_encodings(eval_func_quant, (32, True))
     post_quant_top1 = eval_func(sim.model.cuda(), (0, True))
     print("Post Quant Top1 :", post_quant_top1)

zoo_torch/examples/ssd_utils.py

@@ -59,7 +59,7 @@ def func_quant(model, iterations, use_cuda = True):
 
 def get_simulations(model, args):
 	from aimet_common.defs import QuantScheme
-	from aimet_torch.pro.quantsim import QuantizationSimModel
+	from aimet_torch.quantsim import QuantizationSimModel
 	if hasattr(args, 'quant_scheme'):
 		if args.quant_scheme == 'range_learning_tf':
 		    quant_scheme = QuantScheme.training_range_learning_with_tf_init
@@ -79,4 +79,4 @@ def get_simulations(model, args):
 	}
 	print(kwargs)
 	sim = QuantizationSimModel(model.cpu(), input_shapes=args.input_shape, **kwargs)
-	return sim
+	return sim