Skip to content

GitLab

Unverified Commit c783e148 authored by elishowk's avatar elishowk Committed by GitHub
Browse files

upgrade sentencepiece version (#13564)

parent e86c02ea
Hide whitespace changes
Inline Side-by-side
Showing with 105 additions and 94 deletions
examples/research_projects/lxmert/demo.ipynb
View file @ c783e148
...@@ -46,10 +46,10 @@ ...@@ -46,10 +46,10 @@
"ATTR_URL = \"https://raw.githubusercontent.com/airsplay/py-bottom-up-attention/master/demo/data/genome/1600-400-20/attributes_vocab.txt\"\n", "ATTR_URL = \"https://raw.githubusercontent.com/airsplay/py-bottom-up-attention/master/demo/data/genome/1600-400-20/attributes_vocab.txt\"\n",
"GQA_URL = \"https://raw.githubusercontent.com/airsplay/lxmert/master/data/gqa/trainval_label2ans.json\"\n", "GQA_URL = \"https://raw.githubusercontent.com/airsplay/lxmert/master/data/gqa/trainval_label2ans.json\"\n",
"VQA_URL = \"https://raw.githubusercontent.com/airsplay/lxmert/master/data/vqa/trainval_label2ans.json\"\n", "VQA_URL = \"https://raw.githubusercontent.com/airsplay/lxmert/master/data/vqa/trainval_label2ans.json\"\n",
" \n", "\n",
"\n", "\n",
"# for visualizing output\n", "# for visualizing output\n",
"def showarray(a, fmt='jpeg'):\n", "def showarray(a, fmt=\"jpeg\"):\n",
" a = np.uint8(np.clip(a, 0, 255))\n", " a = np.uint8(np.clip(a, 0, 255))\n",
" f = io.BytesIO()\n", " f = io.BytesIO()\n",
" PIL.Image.fromarray(a).save(f, fmt)\n", " PIL.Image.fromarray(a).save(f, fmt)\n",
...@@ -118,17 +118,17 @@ ...@@ -118,17 +118,17 @@
} }
], ],
"source": [ "source": [
"#image viz\n", "# image viz\n",
"frcnn_visualizer = SingleImageViz(URL, id2obj=objids, id2attr=attrids)\n", "frcnn_visualizer = SingleImageViz(URL, id2obj=objids, id2attr=attrids)\n",
"# run frcnn\n", "# run frcnn\n",
"images, sizes, scales_yx = image_preprocess(URL)\n", "images, sizes, scales_yx = image_preprocess(URL)\n",
"output_dict = frcnn(\n", "output_dict = frcnn(\n",
" images, \n", " images,\n",
" sizes, \n", " sizes,\n",
" scales_yx=scales_yx, \n", " scales_yx=scales_yx,\n",
" padding=\"max_detections\",\n", " padding=\"max_detections\",\n",
" max_detections=frcnn_cfg.max_detections,\n", " max_detections=frcnn_cfg.max_detections,\n",
" return_tensors=\"pt\"\n", " return_tensors=\"pt\",\n",
")\n", ")\n",
"# add boxes and labels to the image\n", "# add boxes and labels to the image\n",
"\n", "\n",
...@@ -174,7 +174,7 @@ ...@@ -174,7 +174,7 @@
" \"Where is this scene?\",\n", " \"Where is this scene?\",\n",
" \"what is the man riding?\",\n", " \"what is the man riding?\",\n",
" \"What is the man wearing?\",\n", " \"What is the man wearing?\",\n",
" \"What is the color of the horse?\"\n", " \"What is the color of the horse?\",\n",
"]\n", "]\n",
"test_questions_for_url2 = [\n", "test_questions_for_url2 = [\n",
" \"Where is the cat?\",\n", " \"Where is the cat?\",\n",
...@@ -184,7 +184,7 @@ ...@@ -184,7 +184,7 @@
" \"What is the shape of the monitor?\",\n", " \"What is the shape of the monitor?\",\n",
"]\n", "]\n",
"\n", "\n",
"#Very important that the boxes are normalized\n", "# Very important that the boxes are normalized\n",
"normalized_boxes = output_dict.get(\"normalized_boxes\")\n", "normalized_boxes = output_dict.get(\"normalized_boxes\")\n",
"features = output_dict.get(\"roi_features\")\n", "features = output_dict.get(\"roi_features\")\n",
"\n", "\n",
...@@ -200,7 +200,7 @@ ...@@ -200,7 +200,7 @@
" return_token_type_ids=True,\n", " return_token_type_ids=True,\n",
" return_attention_mask=True,\n", " return_attention_mask=True,\n",
" add_special_tokens=True,\n", " add_special_tokens=True,\n",
" return_tensors=\"pt\"\n", " return_tensors=\"pt\",\n",
" )\n", " )\n",
"\n", "\n",
" # run lxmert(s)\n", " # run lxmert(s)\n",
......
examples/research_projects/movement-pruning/Saving_PruneBERT.ipynb
View file @ c783e148
...@@ -44,7 +44,7 @@ ...@@ -44,7 +44,7 @@
"\n", "\n",
"from transformers import *\n", "from transformers import *\n",
"\n", "\n",
"os.chdir('../../')" "os.chdir(\"../../\")"
] ]
}, },
{ {
...@@ -70,15 +70,15 @@ ...@@ -70,15 +70,15 @@
"# Load fine-pruned model and quantize the model\n", "# Load fine-pruned model and quantize the model\n",
"\n", "\n",
"model = BertForQuestionAnswering.from_pretrained(\"huggingface/prunebert-base-uncased-6-finepruned-w-distil-squad\")\n", "model = BertForQuestionAnswering.from_pretrained(\"huggingface/prunebert-base-uncased-6-finepruned-w-distil-squad\")\n",
"model.to('cpu')\n", "model.to(\"cpu\")\n",
"\n", "\n",
"quantized_model = torch.quantization.quantize_dynamic(\n", "quantized_model = torch.quantization.quantize_dynamic(\n",
" model=model,\n", " model=model,\n",
" qconfig_spec = {\n", " qconfig_spec={\n",
" nn.Linear : torch.quantization.default_dynamic_qconfig,\n", " nn.Linear: torch.quantization.default_dynamic_qconfig,\n",
" },\n", " },\n",
" dtype=torch.qint8,\n", " dtype=torch.qint8,\n",
" )\n", ")\n",
"# print(quantized_model)\n", "# print(quantized_model)\n",
"\n", "\n",
"qtz_st = quantized_model.state_dict()" "qtz_st = quantized_model.state_dict()"
...@@ -92,10 +92,14 @@ ...@@ -92,10 +92,14 @@
"source": [ "source": [
"# Saving the original (encoder + classifier) in the standard torch.save format\n", "# Saving the original (encoder + classifier) in the standard torch.save format\n",
"\n", "\n",
"dense_st = {name: param for name, param in model.state_dict().items() \n", "dense_st = {\n",
" if \"embedding\" not in name and \"pooler\" not in name}\n", " name: param for name, param in model.state_dict().items() if \"embedding\" not in name and \"pooler\" not in name\n",
"torch.save(dense_st, 'dbg/dense_squad.pt',)\n", "}\n",
"dense_mb_size = os.path.getsize(\"dbg/dense_squad.pt\")\n" "torch.save(\n",
" dense_st,\n",
" \"dbg/dense_squad.pt\",\n",
")\n",
"dense_mb_size = os.path.getsize(\"dbg/dense_squad.pt\")"
] ]
}, },
{ {
...@@ -198,23 +202,23 @@ ...@@ -198,23 +202,23 @@
" if \"dtype\" not in name and param.is_quantized:\n", " if \"dtype\" not in name and param.is_quantized:\n",
" print(\"Decompose quantization for\", name)\n", " print(\"Decompose quantization for\", name)\n",
" # We need to extract the scale, the zero_point and the int_repr for the quantized tensor and modules\n", " # We need to extract the scale, the zero_point and the int_repr for the quantized tensor and modules\n",
" scale = param.q_scale() # torch.tensor(1,) - float32\n", " scale = param.q_scale() # torch.tensor(1,) - float32\n",
" zero_point = param.q_zero_point() # torch.tensor(1,) - int32\n", " zero_point = param.q_zero_point() # torch.tensor(1,) - int32\n",
" elementary_qtz_st[f\"{name}.scale\"] = scale\n", " elementary_qtz_st[f\"{name}.scale\"] = scale\n",
" elementary_qtz_st[f\"{name}.zero_point\"] = zero_point\n", " elementary_qtz_st[f\"{name}.zero_point\"] = zero_point\n",
"\n", "\n",
" # We assume the int_repr is sparse and compute its CSR representation\n", " # We assume the int_repr is sparse and compute its CSR representation\n",
" # Only the FCs in the encoder are actually sparse\n", " # Only the FCs in the encoder are actually sparse\n",
" int_repr = param.int_repr() # torch.tensor(nb_rows, nb_columns) - int8\n", " int_repr = param.int_repr() # torch.tensor(nb_rows, nb_columns) - int8\n",
" int_repr_cs = sparse.csr_matrix(int_repr) # scipy.sparse.csr.csr_matrix\n", " int_repr_cs = sparse.csr_matrix(int_repr) # scipy.sparse.csr.csr_matrix\n",
"\n", "\n",
" elementary_qtz_st[f\"{name}.int_repr.data\"] = int_repr_cs.data # np.array int8\n", " elementary_qtz_st[f\"{name}.int_repr.data\"] = int_repr_cs.data # np.array int8\n",
" elementary_qtz_st[f\"{name}.int_repr.indptr\"] = int_repr_cs.indptr # np.array int32\n", " elementary_qtz_st[f\"{name}.int_repr.indptr\"] = int_repr_cs.indptr # np.array int32\n",
" assert max(int_repr_cs.indices) < 65535 # If not, we shall fall back to int32\n", " assert max(int_repr_cs.indices) < 65535 # If not, we shall fall back to int32\n",
" elementary_qtz_st[f\"{name}.int_repr.indices\"] = np.uint16(int_repr_cs.indices) # np.array uint16\n", " elementary_qtz_st[f\"{name}.int_repr.indices\"] = np.uint16(int_repr_cs.indices) # np.array uint16\n",
" elementary_qtz_st[f\"{name}.int_repr.shape\"] = int_repr_cs.shape # tuple(int, int)\n", " elementary_qtz_st[f\"{name}.int_repr.shape\"] = int_repr_cs.shape # tuple(int, int)\n",
" else:\n", " else:\n",
" elementary_qtz_st[name] = param\n" " elementary_qtz_st[name] = param"
] ]
}, },
{ {
...@@ -225,7 +229,7 @@ ...@@ -225,7 +229,7 @@
"source": [ "source": [
"# Create mapping from torch.dtype to string description (we could also used an int8 instead of string)\n", "# Create mapping from torch.dtype to string description (we could also used an int8 instead of string)\n",
"str_2_dtype = {\"qint8\": torch.qint8}\n", "str_2_dtype = {\"qint8\": torch.qint8}\n",
"dtype_2_str = {torch.qint8: \"qint8\"}\n" "dtype_2_str = {torch.qint8: \"qint8\"}"
] ]
}, },
{ {
...@@ -246,11 +250,17 @@ ...@@ -246,11 +250,17 @@
"source": [ "source": [
"# Saving the pruned (encoder + classifier) in the standard torch.save format\n", "# Saving the pruned (encoder + classifier) in the standard torch.save format\n",
"\n", "\n",
"dense_optimized_st = {name: param for name, param in elementary_qtz_st.items() \n", "dense_optimized_st = {\n",
" if \"embedding\" not in name and \"pooler\" not in name}\n", " name: param for name, param in elementary_qtz_st.items() if \"embedding\" not in name and \"pooler\" not in name\n",
"torch.save(dense_optimized_st, 'dbg/dense_squad_optimized.pt',)\n", "}\n",
"print(\"Encoder Size (MB) - Sparse & Quantized - `torch.save`:\",\n", "torch.save(\n",
" round(os.path.getsize(\"dbg/dense_squad_optimized.pt\")/1e6, 2))\n" " dense_optimized_st,\n",
" \"dbg/dense_squad_optimized.pt\",\n",
")\n",
"print(\n",
" \"Encoder Size (MB) - Sparse & Quantized - `torch.save`:\",\n",
" round(os.path.getsize(\"dbg/dense_squad_optimized.pt\") / 1e6, 2),\n",
")"
] ]
}, },
{ {
...@@ -287,7 +297,7 @@ ...@@ -287,7 +297,7 @@
"# Save the decomposed state_dict with an HDF5 file\n", "# Save the decomposed state_dict with an HDF5 file\n",
"# Saving only the encoder + QA Head\n", "# Saving only the encoder + QA Head\n",
"\n", "\n",
"with h5py.File('dbg/squad_sparse.h5','w') as hf:\n", "with h5py.File(\"dbg/squad_sparse.h5\", \"w\") as hf:\n",
" for name, param in elementary_qtz_st.items():\n", " for name, param in elementary_qtz_st.items():\n",
" if \"embedding\" in name:\n", " if \"embedding\" in name:\n",
" print(f\"Skip {name}\")\n", " print(f\"Skip {name}\")\n",
...@@ -318,18 +328,18 @@ ...@@ -318,18 +328,18 @@
" elif type(param) == torch.dtype:\n", " elif type(param) == torch.dtype:\n",
" # dtype - tensor _packed_params.dtype\n", " # dtype - tensor _packed_params.dtype\n",
" hf.attrs[name] = dtype_2_str[param]\n", " hf.attrs[name] = dtype_2_str[param]\n",
" \n", "\n",
" else:\n", " else:\n",
" hf.create_dataset(name, data=param, compression=\"gzip\", compression_opts=9)\n", " hf.create_dataset(name, data=param, compression=\"gzip\", compression_opts=9)\n",
"\n", "\n",
"\n", "\n",
"with open('dbg/metadata.json', 'w') as f:\n", "with open(\"dbg/metadata.json\", \"w\") as f:\n",
" f.write(json.dumps(qtz_st._metadata)) \n", " f.write(json.dumps(qtz_st._metadata))\n",
"\n", "\n",
"size = os.path.getsize(\"dbg/squad_sparse.h5\") + os.path.getsize(\"dbg/metadata.json\")\n", "size = os.path.getsize(\"dbg/squad_sparse.h5\") + os.path.getsize(\"dbg/metadata.json\")\n",
"print(\"\")\n", "print(\"\")\n",
"print(\"Encoder Size (MB) - Dense: \", round(dense_mb_size/1e6, 2))\n", "print(\"Encoder Size (MB) - Dense: \", round(dense_mb_size / 1e6, 2))\n",
"print(\"Encoder Size (MB) - Sparse & Quantized:\", round(size/1e6, 2))\n" "print(\"Encoder Size (MB) - Sparse & Quantized:\", round(size / 1e6, 2))"
] ]
}, },
{ {
...@@ -350,15 +360,15 @@ ...@@ -350,15 +360,15 @@
"# Save the decomposed state_dict to HDF5 storage\n", "# Save the decomposed state_dict to HDF5 storage\n",
"# Save everything in the architecutre (embedding + encoder + QA Head)\n", "# Save everything in the architecutre (embedding + encoder + QA Head)\n",
"\n", "\n",
"with h5py.File('dbg/squad_sparse_with_embs.h5','w') as hf:\n", "with h5py.File(\"dbg/squad_sparse_with_embs.h5\", \"w\") as hf:\n",
" for name, param in elementary_qtz_st.items():\n", " for name, param in elementary_qtz_st.items():\n",
"# if \"embedding\" in name:\n", " # if \"embedding\" in name:\n",
"# print(f\"Skip {name}\")\n", " # print(f\"Skip {name}\")\n",
"# continue\n", " # continue\n",
"\n", "\n",
"# if \"pooler\" in name:\n", " # if \"pooler\" in name:\n",
"# print(f\"Skip {name}\")\n", " # print(f\"Skip {name}\")\n",
"# continue\n", " # continue\n",
"\n", "\n",
" if type(param) == torch.Tensor:\n", " if type(param) == torch.Tensor:\n",
" if param.numel() == 1:\n", " if param.numel() == 1:\n",
...@@ -381,17 +391,16 @@ ...@@ -381,17 +391,16 @@
" elif type(param) == torch.dtype:\n", " elif type(param) == torch.dtype:\n",
" # dtype - tensor _packed_params.dtype\n", " # dtype - tensor _packed_params.dtype\n",
" hf.attrs[name] = dtype_2_str[param]\n", " hf.attrs[name] = dtype_2_str[param]\n",
" \n", "\n",
" else:\n", " else:\n",
" hf.create_dataset(name, data=param, compression=\"gzip\", compression_opts=9)\n", " hf.create_dataset(name, data=param, compression=\"gzip\", compression_opts=9)\n",
"\n", "\n",
"\n", "\n",
"\n", "with open(\"dbg/metadata.json\", \"w\") as f:\n",
"with open('dbg/metadata.json', 'w') as f:\n", " f.write(json.dumps(qtz_st._metadata))\n",
" f.write(json.dumps(qtz_st._metadata)) \n",
"\n", "\n",
"size = os.path.getsize(\"dbg/squad_sparse_with_embs.h5\") + os.path.getsize(\"dbg/metadata.json\")\n", "size = os.path.getsize(\"dbg/squad_sparse_with_embs.h5\") + os.path.getsize(\"dbg/metadata.json\")\n",
"print('\\nSize (MB):', round(size/1e6, 2))\n" "print(\"\\nSize (MB):\", round(size / 1e6, 2))"
] ]
}, },
{ {
...@@ -411,10 +420,10 @@ ...@@ -411,10 +420,10 @@
"\n", "\n",
"reconstructed_elementary_qtz_st = {}\n", "reconstructed_elementary_qtz_st = {}\n",
"\n", "\n",
"hf = h5py.File('dbg/squad_sparse_with_embs.h5','r')\n", "hf = h5py.File(\"dbg/squad_sparse_with_embs.h5\", \"r\")\n",
"\n", "\n",
"for attr_name, attr_param in hf.attrs.items():\n", "for attr_name, attr_param in hf.attrs.items():\n",
" if 'shape' in attr_name:\n", " if \"shape\" in attr_name:\n",
" attr_param = tuple(attr_param)\n", " attr_param = tuple(attr_param)\n",
" elif \".scale\" in attr_name:\n", " elif \".scale\" in attr_name:\n",
" if \"_packed_params\" in attr_name:\n", " if \"_packed_params\" in attr_name:\n",
...@@ -430,20 +439,20 @@ ...@@ -430,20 +439,20 @@
" attr_param = str_2_dtype[attr_param]\n", " attr_param = str_2_dtype[attr_param]\n",
" reconstructed_elementary_qtz_st[attr_name] = attr_param\n", " reconstructed_elementary_qtz_st[attr_name] = attr_param\n",
" # print(f\"Unpack {attr_name}\")\n", " # print(f\"Unpack {attr_name}\")\n",
" \n", "\n",
"# Get the tensors/arrays\n", "# Get the tensors/arrays\n",
"for data_name, data_param in hf.items():\n", "for data_name, data_param in hf.items():\n",
" if \"LayerNorm\" in data_name or \"_packed_params.bias\" in data_name:\n", " if \"LayerNorm\" in data_name or \"_packed_params.bias\" in data_name:\n",
" reconstructed_elementary_qtz_st[data_name] = torch.from_numpy(np.array(data_param))\n", " reconstructed_elementary_qtz_st[data_name] = torch.from_numpy(np.array(data_param))\n",
" elif \"embedding\" in data_name:\n", " elif \"embedding\" in data_name:\n",
" reconstructed_elementary_qtz_st[data_name] = torch.from_numpy(np.array(data_param))\n", " reconstructed_elementary_qtz_st[data_name] = torch.from_numpy(np.array(data_param))\n",
" else: # _packed_params.weight.int_repr.data, _packed_params.weight.int_repr.indices and _packed_params.weight.int_repr.indptr\n", " else: # _packed_params.weight.int_repr.data, _packed_params.weight.int_repr.indices and _packed_params.weight.int_repr.indptr\n",
" data_param = np.array(data_param)\n", " data_param = np.array(data_param)\n",
" if \"indices\" in data_name:\n", " if \"indices\" in data_name:\n",
" data_param = np.array(data_param, dtype=np.int32)\n", " data_param = np.array(data_param, dtype=np.int32)\n",
" reconstructed_elementary_qtz_st[data_name] = data_param\n", " reconstructed_elementary_qtz_st[data_name] = data_param\n",
" # print(f\"Unpack {data_name}\")\n", " # print(f\"Unpack {data_name}\")\n",
" \n", "\n",
"\n", "\n",
"hf.close()" "hf.close()"
] ]
...@@ -484,27 +493,29 @@ ...@@ -484,27 +493,29 @@
"for name, param in reconstructed_elementary_qtz_st.items():\n", "for name, param in reconstructed_elementary_qtz_st.items():\n",
" if \"weight.int_repr.indptr\" in name:\n", " if \"weight.int_repr.indptr\" in name:\n",
" prefix_ = name[:-16]\n", " prefix_ = name[:-16]\n",
" data = reconstructed_elementary_qtz_st[f\"{prefix_}.int_repr.data\"]\n", " data = reconstructed_elementary_qtz_st[f\"{prefix_}.int_repr.data\"]\n",
" indptr = reconstructed_elementary_qtz_st[f\"{prefix_}.int_repr.indptr\"]\n", " indptr = reconstructed_elementary_qtz_st[f\"{prefix_}.int_repr.indptr\"]\n",
" indices = reconstructed_elementary_qtz_st[f\"{prefix_}.int_repr.indices\"]\n", " indices = reconstructed_elementary_qtz_st[f\"{prefix_}.int_repr.indices\"]\n",
" shape = reconstructed_elementary_qtz_st[f\"{prefix_}.int_repr.shape\"]\n", " shape = reconstructed_elementary_qtz_st[f\"{prefix_}.int_repr.shape\"]\n",
"\n", "\n",
" int_repr = sparse.csr_matrix(arg1=(data, indices, indptr),\n", " int_repr = sparse.csr_matrix(arg1=(data, indices, indptr), shape=shape)\n",
" shape=shape)\n",
" int_repr = torch.tensor(int_repr.todense())\n", " int_repr = torch.tensor(int_repr.todense())\n",
"\n", "\n",
" scale = reconstructed_elementary_qtz_st[f\"{prefix_}.scale\"]\n", " scale = reconstructed_elementary_qtz_st[f\"{prefix_}.scale\"]\n",
" zero_point = reconstructed_elementary_qtz_st[f\"{prefix_}.zero_point\"]\n", " zero_point = reconstructed_elementary_qtz_st[f\"{prefix_}.zero_point\"]\n",
" weight = torch._make_per_tensor_quantized_tensor(int_repr,\n", " weight = torch._make_per_tensor_quantized_tensor(int_repr, scale, zero_point)\n",
" scale,\n",
" zero_point)\n",
"\n", "\n",
" reconstructed_qtz_st[f\"{prefix_}\"] = weight\n", " reconstructed_qtz_st[f\"{prefix_}\"] = weight\n",
" elif \"int_repr.data\" in name or \"int_repr.shape\" in name or \"int_repr.indices\" in name or \\\n", " elif (\n",
" \"weight.scale\" in name or \"weight.zero_point\" in name:\n", " \"int_repr.data\" in name\n",
" or \"int_repr.shape\" in name\n",
" or \"int_repr.indices\" in name\n",
" or \"weight.scale\" in name\n",
" or \"weight.zero_point\" in name\n",
" ):\n",
" continue\n", " continue\n",
" else:\n", " else:\n",
" reconstructed_qtz_st[name] = param\n" " reconstructed_qtz_st[name] = param"
] ]
}, },
{ {
...@@ -556,17 +567,17 @@ ...@@ -556,17 +567,17 @@
"source": [ "source": [
"# Load the re-constructed state dict into a model\n", "# Load the re-constructed state dict into a model\n",
"\n", "\n",
"dummy_model = BertForQuestionAnswering.from_pretrained('bert-base-uncased')\n", "dummy_model = BertForQuestionAnswering.from_pretrained(\"bert-base-uncased\")\n",
"dummy_model.to('cpu')\n", "dummy_model.to(\"cpu\")\n",
"\n", "\n",
"reconstructed_qtz_model = torch.quantization.quantize_dynamic(\n", "reconstructed_qtz_model = torch.quantization.quantize_dynamic(\n",
" model=dummy_model,\n", " model=dummy_model,\n",
" qconfig_spec = None,\n", " qconfig_spec=None,\n",
" dtype=torch.qint8,\n", " dtype=torch.qint8,\n",
" )\n", ")\n",
"\n", "\n",
"reconstructed_qtz_st = OrderedDict(reconstructed_qtz_st)\n", "reconstructed_qtz_st = OrderedDict(reconstructed_qtz_st)\n",
"with open('dbg/metadata.json', 'r') as read_file:\n", "with open(\"dbg/metadata.json\", \"r\") as read_file:\n",
" metadata = json.loads(read_file.read())\n", " metadata = json.loads(read_file.read())\n",
"reconstructed_qtz_st._metadata = metadata\n", "reconstructed_qtz_st._metadata = metadata\n",
"\n", "\n",
...@@ -596,8 +607,8 @@ ...@@ -596,8 +607,8 @@
" mask = torch.ones(size=(N, 128))\n", " mask = torch.ones(size=(N, 128))\n",
"\n", "\n",
" y_reconstructed = reconstructed_qtz_model(input_ids=inputs, attention_mask=mask)[0]\n", " y_reconstructed = reconstructed_qtz_model(input_ids=inputs, attention_mask=mask)[0]\n",
" y = quantized_model(input_ids=inputs, attention_mask=mask)[0]\n", " y = quantized_model(input_ids=inputs, attention_mask=mask)[0]\n",
" \n", "\n",
" assert torch.all(torch.eq(y, y_reconstructed))\n", " assert torch.all(torch.eq(y, y_reconstructed))\n",
"print(\"Sanity check passed\")" "print(\"Sanity check passed\")"
] ]
......
examples/research_projects/visual_bert/demo.ipynb
View file @ c783e148
...@@ -37,10 +37,10 @@ ...@@ -37,10 +37,10 @@
"OBJ_URL = \"https://raw.githubusercontent.com/airsplay/py-bottom-up-attention/master/demo/data/genome/1600-400-20/objects_vocab.txt\"\n", "OBJ_URL = \"https://raw.githubusercontent.com/airsplay/py-bottom-up-attention/master/demo/data/genome/1600-400-20/objects_vocab.txt\"\n",
"ATTR_URL = \"https://raw.githubusercontent.com/airsplay/py-bottom-up-attention/master/demo/data/genome/1600-400-20/attributes_vocab.txt\"\n", "ATTR_URL = \"https://raw.githubusercontent.com/airsplay/py-bottom-up-attention/master/demo/data/genome/1600-400-20/attributes_vocab.txt\"\n",
"VQA_URL = \"https://dl.fbaipublicfiles.com/pythia/data/answers_vqa.txt\"\n", "VQA_URL = \"https://dl.fbaipublicfiles.com/pythia/data/answers_vqa.txt\"\n",
" \n", "\n",
"\n", "\n",
"# for visualizing output\n", "# for visualizing output\n",
"def showarray(a, fmt='jpeg'):\n", "def showarray(a, fmt=\"jpeg\"):\n",
" a = np.uint8(np.clip(a, 0, 255))\n", " a = np.uint8(np.clip(a, 0, 255))\n",
" f = io.BytesIO()\n", " f = io.BytesIO()\n",
" PIL.Image.fromarray(a).save(f, fmt)\n", " PIL.Image.fromarray(a).save(f, fmt)\n",
...@@ -82,7 +82,7 @@ ...@@ -82,7 +82,7 @@
"image_preprocess = Preprocess(frcnn_cfg)\n", "image_preprocess = Preprocess(frcnn_cfg)\n",
"\n", "\n",
"bert_tokenizer = BertTokenizerFast.from_pretrained(\"bert-base-uncased\")\n", "bert_tokenizer = BertTokenizerFast.from_pretrained(\"bert-base-uncased\")\n",
"visualbert_vqa = VisualBertForQuestionAnswering.from_pretrained(\"uclanlp/visualbert-vqa\")\n" "visualbert_vqa = VisualBertForQuestionAnswering.from_pretrained(\"uclanlp/visualbert-vqa\")"
], ],
"outputs": [ "outputs": [
{ {
...@@ -104,17 +104,17 @@ ...@@ -104,17 +104,17 @@
"cell_type": "code", "cell_type": "code",
"execution_count": 5, "execution_count": 5,
"source": [ "source": [
"#image viz\n", "# image viz\n",
"frcnn_visualizer = SingleImageViz(URL, id2obj=objids, id2attr=attrids)\n", "frcnn_visualizer = SingleImageViz(URL, id2obj=objids, id2attr=attrids)\n",
"# run frcnn\n", "# run frcnn\n",
"images, sizes, scales_yx = image_preprocess(URL)\n", "images, sizes, scales_yx = image_preprocess(URL)\n",
"output_dict = frcnn(\n", "output_dict = frcnn(\n",
" images, \n", " images,\n",
" sizes, \n", " sizes,\n",
" scales_yx=scales_yx, \n", " scales_yx=scales_yx,\n",
" padding=\"max_detections\",\n", " padding=\"max_detections\",\n",
" max_detections=frcnn_cfg.max_detections,\n", " max_detections=frcnn_cfg.max_detections,\n",
" return_tensors=\"pt\"\n", " return_tensors=\"pt\",\n",
")\n", ")\n",
"# add boxes and labels to the image\n", "# add boxes and labels to the image\n",
"\n", "\n",
...@@ -167,7 +167,7 @@ ...@@ -167,7 +167,7 @@
" \"What is the shape of the monitor?\",\n", " \"What is the shape of the monitor?\",\n",
"]\n", "]\n",
"\n", "\n",
"#Very important that the boxes are normalized\n", "# Very important that the boxes are normalized\n",
"# normalized_boxes = output_dict.get(\"normalized_boxes\")\n", "# normalized_boxes = output_dict.get(\"normalized_boxes\")\n",
"features = output_dict.get(\"roi_features\")" "features = output_dict.get(\"roi_features\")"
], ],
...@@ -189,7 +189,7 @@ ...@@ -189,7 +189,7 @@
" return_token_type_ids=True,\n", " return_token_type_ids=True,\n",
" return_attention_mask=True,\n", " return_attention_mask=True,\n",
" add_special_tokens=True,\n", " add_special_tokens=True,\n",
" return_tensors=\"pt\"\n", " return_tensors=\"pt\",\n",
" )\n", " )\n",
"\n", "\n",
" output_vqa = visualbert_vqa(\n", " output_vqa = visualbert_vqa(\n",
......
setup.py
View file @ c783e148
...@@ -134,7 +134,7 @@ _deps = [ ...@@ -134,7 +134,7 @@ _deps = [
"sacremoses", "sacremoses",
"sagemaker>=2.31.0", "sagemaker>=2.31.0",
"scikit-learn", "scikit-learn",
"sentencepiece==0.1.91", "sentencepiece>=0.1.91,!=0.1.92",
"soundfile", "soundfile",
"sphinx-copybutton", "sphinx-copybutton",
"sphinx-markdown-tables", "sphinx-markdown-tables",
......
src/transformers/dependency_versions_table.py
View file @ c783e148
...@@ -52,7 +52,7 @@ deps = { ...@@ -52,7 +52,7 @@ deps = {
"sacremoses": "sacremoses", "sacremoses": "sacremoses",
"sagemaker": "sagemaker>=2.31.0", "sagemaker": "sagemaker>=2.31.0",
"scikit-learn": "scikit-learn", "scikit-learn": "scikit-learn",
"sentencepiece": "sentencepiece==0.1.91", "sentencepiece": "sentencepiece>=0.1.91,!=0.1.92",
"soundfile": "soundfile", "soundfile": "soundfile",
"sphinx-copybutton": "sphinx-copybutton", "sphinx-copybutton": "sphinx-copybutton",
"sphinx-markdown-tables": "sphinx-markdown-tables", "sphinx-markdown-tables": "sphinx-markdown-tables",
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment